diff --git a/.gitattributes b/.gitattributes index a6344aac8c09253b3b630fb776ae94478aa0275b..916f3058dccede36c04904a0c60226b598ecbd5e 100644 --- a/.gitattributes +++ b/.gitattributes @@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text *.zip filter=lfs diff=lfs merge=lfs -text *.zst filter=lfs diff=lfs merge=lfs -text *tfevents* filter=lfs diff=lfs merge=lfs -text +libbitsandbytes_cuda116.dll filter=lfs diff=lfs merge=lfs -text +libbitsandbytes_cuda118.dll filter=lfs diff=lfs merge=lfs -text diff --git a/FUNDING.yml b/FUNDING.yml new file mode 100644 index 0000000000000000000000000000000000000000..3b8943c35cd7e3a68de183cdbf2eeca35c5f62f4 --- /dev/null +++ b/FUNDING.yml @@ -0,0 +1,3 @@ +# These are supported funding model platforms + +github: kohya-ss diff --git a/__init__.py b/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/adafactor_fused.py b/adafactor_fused.py new file mode 100644 index 0000000000000000000000000000000000000000..bdfc32ced30786489c8e535dc0c6f0351cd949b3 --- /dev/null +++ b/adafactor_fused.py @@ -0,0 +1,106 @@ +import math +import torch +from transformers import Adafactor + +@torch.no_grad() +def adafactor_step_param(self, p, group): + if p.grad is None: + return + grad = p.grad + if grad.dtype in {torch.float16, torch.bfloat16}: + grad = grad.float() + if grad.is_sparse: + raise RuntimeError("Adafactor does not support sparse gradients.") + + state = self.state[p] + grad_shape = grad.shape + + factored, use_first_moment = Adafactor._get_options(group, grad_shape) + # State Initialization + if len(state) == 0: + state["step"] = 0 + + if use_first_moment: + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like(grad) + if factored: + state["exp_avg_sq_row"] = torch.zeros(grad_shape[:-1]).to(grad) + state["exp_avg_sq_col"] = torch.zeros(grad_shape[:-2] + grad_shape[-1:]).to(grad) + else: + state["exp_avg_sq"] = torch.zeros_like(grad) + + state["RMS"] = 0 + else: + if use_first_moment: + state["exp_avg"] = state["exp_avg"].to(grad) + if factored: + state["exp_avg_sq_row"] = state["exp_avg_sq_row"].to(grad) + state["exp_avg_sq_col"] = state["exp_avg_sq_col"].to(grad) + else: + state["exp_avg_sq"] = state["exp_avg_sq"].to(grad) + + p_data_fp32 = p + if p.dtype in {torch.float16, torch.bfloat16}: + p_data_fp32 = p_data_fp32.float() + + state["step"] += 1 + state["RMS"] = Adafactor._rms(p_data_fp32) + lr = Adafactor._get_lr(group, state) + + beta2t = 1.0 - math.pow(state["step"], group["decay_rate"]) + update = (grad ** 2) + group["eps"][0] + if factored: + exp_avg_sq_row = state["exp_avg_sq_row"] + exp_avg_sq_col = state["exp_avg_sq_col"] + + exp_avg_sq_row.mul_(beta2t).add_(update.mean(dim=-1), alpha=(1.0 - beta2t)) + exp_avg_sq_col.mul_(beta2t).add_(update.mean(dim=-2), alpha=(1.0 - beta2t)) + + # Approximation of exponential moving average of square of gradient + update = Adafactor._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col) + update.mul_(grad) + else: + exp_avg_sq = state["exp_avg_sq"] + + exp_avg_sq.mul_(beta2t).add_(update, alpha=(1.0 - beta2t)) + update = exp_avg_sq.rsqrt().mul_(grad) + + update.div_((Adafactor._rms(update) / group["clip_threshold"]).clamp_(min=1.0)) + update.mul_(lr) + + if use_first_moment: + exp_avg = state["exp_avg"] + exp_avg.mul_(group["beta1"]).add_(update, alpha=(1 - group["beta1"])) + update = exp_avg + + if group["weight_decay"] != 0: + p_data_fp32.add_(p_data_fp32, alpha=(-group["weight_decay"] * lr)) + + p_data_fp32.add_(-update) + + if p.dtype in {torch.float16, torch.bfloat16}: + p.copy_(p_data_fp32) + + +@torch.no_grad() +def adafactor_step(self, closure=None): + """ + Performs a single optimization step + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + adafactor_step_param(self, p, group) + + return loss + +def patch_adafactor_fused(optimizer: Adafactor): + optimizer.step_param = adafactor_step_param.__get__(optimizer) + optimizer.step = adafactor_step.__get__(optimizer) diff --git a/attention.py b/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..400b59b66822d4e1d455adf43e19d9e3628786db --- /dev/null +++ b/attention.py @@ -0,0 +1,119 @@ +import os +import torch +from functools import cache, wraps + +# pylint: disable=protected-access, missing-function-docstring, line-too-long + +# ARC GPUs can't allocate more than 4GB to a single block so we slice the attention layers + +sdpa_slice_trigger_rate = float(os.environ.get('IPEX_SDPA_SLICE_TRIGGER_RATE', 1)) +attention_slice_rate = float(os.environ.get('IPEX_ATTENTION_SLICE_RATE', 0.5)) + +# Find something divisible with the input_tokens +@cache +def find_split_size(original_size, slice_block_size, slice_rate=2): + split_size = original_size + while True: + if (split_size * slice_block_size) <= slice_rate and original_size % split_size == 0: + return split_size + split_size = split_size - 1 + if split_size <= 1: + return 1 + return split_size + + +# Find slice sizes for SDPA +@cache +def find_sdpa_slice_sizes(query_shape, key_shape, query_element_size, slice_rate=2, trigger_rate=3): + batch_size, attn_heads, query_len, _ = query_shape + _, _, key_len, _ = key_shape + + slice_batch_size = attn_heads * (query_len * key_len) * query_element_size / 1024 / 1024 / 1024 + + split_batch_size = batch_size + split_head_size = attn_heads + split_query_size = query_len + + do_batch_split = False + do_head_split = False + do_query_split = False + + if batch_size * slice_batch_size >= trigger_rate: + do_batch_split = True + split_batch_size = find_split_size(batch_size, slice_batch_size, slice_rate=slice_rate) + + if split_batch_size * slice_batch_size > slice_rate: + slice_head_size = split_batch_size * (query_len * key_len) * query_element_size / 1024 / 1024 / 1024 + do_head_split = True + split_head_size = find_split_size(attn_heads, slice_head_size, slice_rate=slice_rate) + + if split_head_size * slice_head_size > slice_rate: + slice_query_size = split_batch_size * split_head_size * (key_len) * query_element_size / 1024 / 1024 / 1024 + do_query_split = True + split_query_size = find_split_size(query_len, slice_query_size, slice_rate=slice_rate) + + return do_batch_split, do_head_split, do_query_split, split_batch_size, split_head_size, split_query_size + + +original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention +@wraps(torch.nn.functional.scaled_dot_product_attention) +def dynamic_scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, **kwargs): + if query.device.type != "xpu": + return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs) + is_unsqueezed = False + if len(query.shape) == 3: + query = query.unsqueeze(0) + is_unsqueezed = True + if len(key.shape) == 3: + key = key.unsqueeze(0) + if len(value.shape) == 3: + value = value.unsqueeze(0) + do_batch_split, do_head_split, do_query_split, split_batch_size, split_head_size, split_query_size = find_sdpa_slice_sizes(query.shape, key.shape, query.element_size(), slice_rate=attention_slice_rate, trigger_rate=sdpa_slice_trigger_rate) + + # Slice SDPA + if do_batch_split: + batch_size, attn_heads, query_len, _ = query.shape + _, _, _, head_dim = value.shape + hidden_states = torch.zeros((batch_size, attn_heads, query_len, head_dim), device=query.device, dtype=query.dtype) + if attn_mask is not None: + attn_mask = attn_mask.expand((query.shape[0], query.shape[1], query.shape[2], key.shape[-2])) + for ib in range(batch_size // split_batch_size): + start_idx = ib * split_batch_size + end_idx = (ib + 1) * split_batch_size + if do_head_split: + for ih in range(attn_heads // split_head_size): # pylint: disable=invalid-name + start_idx_h = ih * split_head_size + end_idx_h = (ih + 1) * split_head_size + if do_query_split: + for iq in range(query_len // split_query_size): # pylint: disable=invalid-name + start_idx_q = iq * split_query_size + end_idx_q = (iq + 1) * split_query_size + hidden_states[start_idx:end_idx, start_idx_h:end_idx_h, start_idx_q:end_idx_q, :] = original_scaled_dot_product_attention( + query[start_idx:end_idx, start_idx_h:end_idx_h, start_idx_q:end_idx_q, :], + key[start_idx:end_idx, start_idx_h:end_idx_h, :, :], + value[start_idx:end_idx, start_idx_h:end_idx_h, :, :], + attn_mask=attn_mask[start_idx:end_idx, start_idx_h:end_idx_h, start_idx_q:end_idx_q, :] if attn_mask is not None else attn_mask, + dropout_p=dropout_p, is_causal=is_causal, **kwargs + ) + else: + hidden_states[start_idx:end_idx, start_idx_h:end_idx_h, :, :] = original_scaled_dot_product_attention( + query[start_idx:end_idx, start_idx_h:end_idx_h, :, :], + key[start_idx:end_idx, start_idx_h:end_idx_h, :, :], + value[start_idx:end_idx, start_idx_h:end_idx_h, :, :], + attn_mask=attn_mask[start_idx:end_idx, start_idx_h:end_idx_h, :, :] if attn_mask is not None else attn_mask, + dropout_p=dropout_p, is_causal=is_causal, **kwargs + ) + else: + hidden_states[start_idx:end_idx, :, :, :] = original_scaled_dot_product_attention( + query[start_idx:end_idx, :, :, :], + key[start_idx:end_idx, :, :, :], + value[start_idx:end_idx, :, :, :], + attn_mask=attn_mask[start_idx:end_idx, :, :, :] if attn_mask is not None else attn_mask, + dropout_p=dropout_p, is_causal=is_causal, **kwargs + ) + torch.xpu.synchronize(query.device) + else: + hidden_states = original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs) + if is_unsqueezed: + hidden_states.squeeze(0) + return hidden_states diff --git a/attention_processors.py b/attention_processors.py new file mode 100644 index 0000000000000000000000000000000000000000..310c2cb1c63955f8f03296c54fd47c21f1a981c9 --- /dev/null +++ b/attention_processors.py @@ -0,0 +1,227 @@ +import math +from typing import Any +from einops import rearrange +import torch +from diffusers.models.attention_processor import Attention + + +# flash attention forwards and backwards + +# https://arxiv.org/abs/2205.14135 + +EPSILON = 1e-6 + + +class FlashAttentionFunction(torch.autograd.function.Function): + @staticmethod + @torch.no_grad() + def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size): + """Algorithm 2 in the paper""" + + device = q.device + dtype = q.dtype + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + o = torch.zeros_like(q) + all_row_sums = torch.zeros((*q.shape[:-1], 1), dtype=dtype, device=device) + all_row_maxes = torch.full( + (*q.shape[:-1], 1), max_neg_value, dtype=dtype, device=device + ) + + scale = q.shape[-1] ** -0.5 + + if mask is None: + mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size) + else: + mask = rearrange(mask, "b n -> b 1 1 n") + mask = mask.split(q_bucket_size, dim=-1) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + mask, + all_row_sums.split(q_bucket_size, dim=-2), + all_row_maxes.split(q_bucket_size, dim=-2), + ) + + for ind, (qc, oc, row_mask, row_sums, row_maxes) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = ( + torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale + ) + + if row_mask is not None: + attn_weights.masked_fill_(~row_mask, max_neg_value) + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones( + (qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device + ).triu(q_start_index - k_start_index + 1) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + block_row_maxes = attn_weights.amax(dim=-1, keepdims=True) + attn_weights -= block_row_maxes + exp_weights = torch.exp(attn_weights) + + if row_mask is not None: + exp_weights.masked_fill_(~row_mask, 0.0) + + block_row_sums = exp_weights.sum(dim=-1, keepdims=True).clamp( + min=EPSILON + ) + + new_row_maxes = torch.maximum(block_row_maxes, row_maxes) + + exp_values = torch.einsum( + "... i j, ... j d -> ... i d", exp_weights, vc + ) + + exp_row_max_diff = torch.exp(row_maxes - new_row_maxes) + exp_block_row_max_diff = torch.exp(block_row_maxes - new_row_maxes) + + new_row_sums = ( + exp_row_max_diff * row_sums + + exp_block_row_max_diff * block_row_sums + ) + + oc.mul_((row_sums / new_row_sums) * exp_row_max_diff).add_( + (exp_block_row_max_diff / new_row_sums) * exp_values + ) + + row_maxes.copy_(new_row_maxes) + row_sums.copy_(new_row_sums) + + ctx.args = (causal, scale, mask, q_bucket_size, k_bucket_size) + ctx.save_for_backward(q, k, v, o, all_row_sums, all_row_maxes) + + return o + + @staticmethod + @torch.no_grad() + def backward(ctx, do): + """Algorithm 4 in the paper""" + + causal, scale, mask, q_bucket_size, k_bucket_size = ctx.args + q, k, v, o, l, m = ctx.saved_tensors + + device = q.device + + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + dq = torch.zeros_like(q) + dk = torch.zeros_like(k) + dv = torch.zeros_like(v) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + do.split(q_bucket_size, dim=-2), + mask, + l.split(q_bucket_size, dim=-2), + m.split(q_bucket_size, dim=-2), + dq.split(q_bucket_size, dim=-2), + ) + + for ind, (qc, oc, doc, row_mask, lc, mc, dqc) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + dk.split(k_bucket_size, dim=-2), + dv.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc, dkc, dvc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = ( + torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale + ) + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones( + (qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device + ).triu(q_start_index - k_start_index + 1) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + exp_attn_weights = torch.exp(attn_weights - mc) + + if row_mask is not None: + exp_attn_weights.masked_fill_(~row_mask, 0.0) + + p = exp_attn_weights / lc + + dv_chunk = torch.einsum("... i j, ... i d -> ... j d", p, doc) + dp = torch.einsum("... i d, ... j d -> ... i j", doc, vc) + + D = (doc * oc).sum(dim=-1, keepdims=True) + ds = p * scale * (dp - D) + + dq_chunk = torch.einsum("... i j, ... j d -> ... i d", ds, kc) + dk_chunk = torch.einsum("... i j, ... i d -> ... j d", ds, qc) + + dqc.add_(dq_chunk) + dkc.add_(dk_chunk) + dvc.add_(dv_chunk) + + return dq, dk, dv, None, None, None, None + + +class FlashAttnProcessor: + def __call__( + self, + attn: Attention, + hidden_states, + encoder_hidden_states=None, + attention_mask=None, + ) -> Any: + q_bucket_size = 512 + k_bucket_size = 1024 + + h = attn.heads + q = attn.to_q(hidden_states) + + encoder_hidden_states = ( + encoder_hidden_states + if encoder_hidden_states is not None + else hidden_states + ) + encoder_hidden_states = encoder_hidden_states.to(hidden_states.dtype) + + if hasattr(attn, "hypernetwork") and attn.hypernetwork is not None: + context_k, context_v = attn.hypernetwork.forward( + hidden_states, encoder_hidden_states + ) + context_k = context_k.to(hidden_states.dtype) + context_v = context_v.to(hidden_states.dtype) + else: + context_k = encoder_hidden_states + context_v = encoder_hidden_states + + k = attn.to_k(context_k) + v = attn.to_v(context_v) + del encoder_hidden_states, hidden_states + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v)) + + out = FlashAttentionFunction.apply( + q, k, v, attention_mask, False, q_bucket_size, k_bucket_size + ) + + out = rearrange(out, "b h n d -> b n (h d)") + + out = attn.to_out[0](out) + out = attn.to_out[1](out) + return out diff --git a/blip.py b/blip.py new file mode 100644 index 0000000000000000000000000000000000000000..13b69ffd33e3239e666e598dc0f1b6e55427d3a3 --- /dev/null +++ b/blip.py @@ -0,0 +1,245 @@ +''' + * Copyright (c) 2022, salesforce.com, inc. + * All rights reserved. + * SPDX-License-Identifier: BSD-3-Clause + * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause + * By Junnan Li +''' +import warnings +warnings.filterwarnings("ignore") + +# from models.vit import VisionTransformer, interpolate_pos_embed +# from models.med import BertConfig, BertModel, BertLMHeadModel +from blip.vit import VisionTransformer, interpolate_pos_embed +from blip.med import BertConfig, BertModel, BertLMHeadModel +from transformers import BertTokenizer + +import torch +from torch import nn +import torch.nn.functional as F + +import os +from urllib.parse import urlparse +from timm.models.hub import download_cached_file +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +class BLIP_Base(nn.Module): + def __init__(self, + med_config = 'configs/med_config.json', + image_size = 224, + vit = 'base', + vit_grad_ckpt = False, + vit_ckpt_layer = 0, + ): + """ + Args: + med_config (str): path for the mixture of encoder-decoder model's configuration file + image_size (int): input image size + vit (str): model size of vision transformer + """ + super().__init__() + + self.visual_encoder, vision_width = create_vit(vit,image_size, vit_grad_ckpt, vit_ckpt_layer) + self.tokenizer = init_tokenizer() + med_config = BertConfig.from_json_file(med_config) + med_config.encoder_width = vision_width + self.text_encoder = BertModel(config=med_config, add_pooling_layer=False) + + + def forward(self, image, caption, mode): + + assert mode in ['image', 'text', 'multimodal'], "mode parameter must be image, text, or multimodal" + text = self.tokenizer(caption, return_tensors="pt").to(image.device) + + if mode=='image': + # return image features + image_embeds = self.visual_encoder(image) + return image_embeds + + elif mode=='text': + # return text features + text_output = self.text_encoder(text.input_ids, attention_mask = text.attention_mask, + return_dict = True, mode = 'text') + return text_output.last_hidden_state + + elif mode=='multimodal': + # return multimodel features + image_embeds = self.visual_encoder(image) + image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(image.device) + + text.input_ids[:,0] = self.tokenizer.enc_token_id + output = self.text_encoder(text.input_ids, + attention_mask = text.attention_mask, + encoder_hidden_states = image_embeds, + encoder_attention_mask = image_atts, + return_dict = True, + ) + return output.last_hidden_state + + + +class BLIP_Decoder(nn.Module): + def __init__(self, + med_config = 'configs/med_config.json', + image_size = 384, + vit = 'base', + vit_grad_ckpt = False, + vit_ckpt_layer = 0, + prompt = 'a picture of ', + ): + """ + Args: + med_config (str): path for the mixture of encoder-decoder model's configuration file + image_size (int): input image size + vit (str): model size of vision transformer + """ + super().__init__() + + self.visual_encoder, vision_width = create_vit(vit,image_size, vit_grad_ckpt, vit_ckpt_layer) + self.tokenizer = init_tokenizer() + med_config = BertConfig.from_json_file(med_config) + med_config.encoder_width = vision_width + self.text_decoder = BertLMHeadModel(config=med_config) + + self.prompt = prompt + self.prompt_length = len(self.tokenizer(self.prompt).input_ids)-1 + + + def forward(self, image, caption): + + image_embeds = self.visual_encoder(image) + image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(image.device) + + text = self.tokenizer(caption, padding='longest', truncation=True, max_length=40, return_tensors="pt").to(image.device) + + text.input_ids[:,0] = self.tokenizer.bos_token_id + + decoder_targets = text.input_ids.masked_fill(text.input_ids == self.tokenizer.pad_token_id, -100) + decoder_targets[:,:self.prompt_length] = -100 + + decoder_output = self.text_decoder(text.input_ids, + attention_mask = text.attention_mask, + encoder_hidden_states = image_embeds, + encoder_attention_mask = image_atts, + labels = decoder_targets, + return_dict = True, + ) + loss_lm = decoder_output.loss + + return loss_lm + + def generate(self, image, sample=False, num_beams=3, max_length=30, min_length=10, top_p=0.9, repetition_penalty=1.0): + image_embeds = self.visual_encoder(image) + + # recent version of transformers seems to do repeat_interleave automatically + # if not sample: + # image_embeds = image_embeds.repeat_interleave(num_beams,dim=0) + + image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(image.device) + model_kwargs = {"encoder_hidden_states": image_embeds, "encoder_attention_mask":image_atts} + + prompt = [self.prompt] * image.size(0) + input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids.to(image.device) + input_ids[:,0] = self.tokenizer.bos_token_id + input_ids = input_ids[:, :-1] + + if sample: + #nucleus sampling + outputs = self.text_decoder.generate(input_ids=input_ids, + max_length=max_length, + min_length=min_length, + do_sample=True, + top_p=top_p, + num_return_sequences=1, + eos_token_id=self.tokenizer.sep_token_id, + pad_token_id=self.tokenizer.pad_token_id, + repetition_penalty=1.1, + **model_kwargs) + else: + #beam search + outputs = self.text_decoder.generate(input_ids=input_ids, + max_length=max_length, + min_length=min_length, + num_beams=num_beams, + eos_token_id=self.tokenizer.sep_token_id, + pad_token_id=self.tokenizer.pad_token_id, + repetition_penalty=repetition_penalty, + **model_kwargs) + + captions = [] + for output in outputs: + caption = self.tokenizer.decode(output, skip_special_tokens=True) + captions.append(caption[len(self.prompt):]) + return captions + + +def blip_decoder(pretrained='',**kwargs): + model = BLIP_Decoder(**kwargs) + if pretrained: + model,msg = load_checkpoint(model,pretrained) + assert(len(msg.missing_keys)==0) + return model + +def blip_feature_extractor(pretrained='',**kwargs): + model = BLIP_Base(**kwargs) + if pretrained: + model,msg = load_checkpoint(model,pretrained) + assert(len(msg.missing_keys)==0) + return model + +def init_tokenizer(): + tokenizer = BertTokenizer.from_pretrained('bert-base-uncased') + tokenizer.add_special_tokens({'bos_token':'[DEC]'}) + tokenizer.add_special_tokens({'additional_special_tokens':['[ENC]']}) + tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0] + return tokenizer + + +def create_vit(vit, image_size, use_grad_checkpointing=False, ckpt_layer=0, drop_path_rate=0): + + assert vit in ['base', 'large'], "vit parameter must be base or large" + if vit=='base': + vision_width = 768 + visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=12, + num_heads=12, use_grad_checkpointing=use_grad_checkpointing, ckpt_layer=ckpt_layer, + drop_path_rate=0 or drop_path_rate + ) + elif vit=='large': + vision_width = 1024 + visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=24, + num_heads=16, use_grad_checkpointing=use_grad_checkpointing, ckpt_layer=ckpt_layer, + drop_path_rate=0.1 or drop_path_rate + ) + return visual_encoder, vision_width + +def is_url(url_or_filename): + parsed = urlparse(url_or_filename) + return parsed.scheme in ("http", "https") + +def load_checkpoint(model,url_or_filename): + if is_url(url_or_filename): + cached_file = download_cached_file(url_or_filename, check_hash=False, progress=True) + checkpoint = torch.load(cached_file, map_location='cpu') + elif os.path.isfile(url_or_filename): + checkpoint = torch.load(url_or_filename, map_location='cpu') + else: + raise RuntimeError('checkpoint url or path is invalid') + + state_dict = checkpoint['model'] + + state_dict['visual_encoder.pos_embed'] = interpolate_pos_embed(state_dict['visual_encoder.pos_embed'],model.visual_encoder) + if 'visual_encoder_m.pos_embed' in model.state_dict().keys(): + state_dict['visual_encoder_m.pos_embed'] = interpolate_pos_embed(state_dict['visual_encoder_m.pos_embed'], + model.visual_encoder_m) + for key in model.state_dict().keys(): + if key in state_dict.keys(): + if state_dict[key].shape!=model.state_dict()[key].shape: + del state_dict[key] + + msg = model.load_state_dict(state_dict,strict=False) + logger.info('load checkpoint from %s'%url_or_filename) + return model,msg + diff --git a/cache_latents.py b/cache_latents.py new file mode 100644 index 0000000000000000000000000000000000000000..2f0098b42d700ad446019a4e2e5f011870816e58 --- /dev/null +++ b/cache_latents.py @@ -0,0 +1,205 @@ +# latentsのdiskへの事前キャッシュを行う / cache latents to disk + +import argparse +import math +from multiprocessing import Value +import os + +from accelerate.utils import set_seed +import torch +from tqdm import tqdm + +from library import config_util +from library import train_util +from library import sdxl_train_util +from library.config_util import ( + ConfigSanitizer, + BlueprintGenerator, +) +from library.utils import setup_logging, add_logging_arguments +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +def cache_to_disk(args: argparse.Namespace) -> None: + setup_logging(args, reset=True) + train_util.prepare_dataset_args(args, True) + + # check cache latents arg + assert args.cache_latents_to_disk, "cache_latents_to_disk must be True / cache_latents_to_diskはTrueである必要があります" + + use_dreambooth_method = args.in_json is None + + if args.seed is not None: + set_seed(args.seed) # 乱数系列を初期化する + + # tokenizerを準備する：datasetを動かすために必要 + if args.sdxl: + tokenizer1, tokenizer2 = sdxl_train_util.load_tokenizers(args) + tokenizers = [tokenizer1, tokenizer2] + else: + tokenizer = train_util.load_tokenizer(args) + tokenizers = [tokenizer] + + # データセットを準備する + if args.dataset_class is None: + blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, True)) + if args.dataset_config is not None: + logger.info(f"Load dataset config from {args.dataset_config}") + user_config = config_util.load_user_config(args.dataset_config) + ignored = ["train_data_dir", "in_json"] + if any(getattr(args, attr) is not None for attr in ignored): + logger.warning( + "ignore following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format( + ", ".join(ignored) + ) + ) + else: + if use_dreambooth_method: + logger.info("Using DreamBooth method.") + user_config = { + "datasets": [ + { + "subsets": config_util.generate_dreambooth_subsets_config_by_subdirs( + args.train_data_dir, args.reg_data_dir + ) + } + ] + } + else: + logger.info("Training with captions.") + user_config = { + "datasets": [ + { + "subsets": [ + { + "image_dir": args.train_data_dir, + "metadata_file": args.in_json, + } + ] + } + ] + } + + blueprint = blueprint_generator.generate(user_config, args, tokenizer=tokenizers) + train_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group) + else: + train_dataset_group = train_util.load_arbitrary_dataset(args, tokenizers) + + # datasetのcache_latentsを呼ばなければ、生の画像が返る + + current_epoch = Value("i", 0) + current_step = Value("i", 0) + ds_for_collator = train_dataset_group if args.max_data_loader_n_workers == 0 else None + collator = train_util.collator_class(current_epoch, current_step, ds_for_collator) + + # acceleratorを準備する + logger.info("prepare accelerator") + args.deepspeed = False + accelerator = train_util.prepare_accelerator(args) + + # mixed precisionに対応した型を用意しておき適宜castする + weight_dtype, _ = train_util.prepare_dtype(args) + vae_dtype = torch.float32 if args.no_half_vae else weight_dtype + + # モデルを読み込む + logger.info("load model") + if args.sdxl: + (_, _, _, vae, _, _, _) = sdxl_train_util.load_target_model(args, accelerator, "sdxl", weight_dtype) + else: + _, vae, _, _ = train_util.load_target_model(args, weight_dtype, accelerator) + + if torch.__version__ >= "2.0.0": # PyTorch 2.0.0 以上対応のxformersなら以下が使える + vae.set_use_memory_efficient_attention_xformers(args.xformers) + vae.to(accelerator.device, dtype=vae_dtype) + vae.requires_grad_(False) + vae.eval() + + # dataloaderを準備する + train_dataset_group.set_caching_mode("latents") + + # DataLoaderのプロセス数：0 は persistent_workers が使えないので注意 + n_workers = min(args.max_data_loader_n_workers, os.cpu_count()) # cpu_count or max_data_loader_n_workers + + train_dataloader = torch.utils.data.DataLoader( + train_dataset_group, + batch_size=1, + shuffle=True, + collate_fn=collator, + num_workers=n_workers, + persistent_workers=args.persistent_data_loader_workers, + ) + + # acceleratorを使ってモデルを準備する：マルチGPUで使えるようになるはず + train_dataloader = accelerator.prepare(train_dataloader) + + # データ取得のためのループ + for batch in tqdm(train_dataloader): + b_size = len(batch["images"]) + vae_batch_size = b_size if args.vae_batch_size is None else args.vae_batch_size + flip_aug = batch["flip_aug"] + alpha_mask = batch["alpha_mask"] + random_crop = batch["random_crop"] + bucket_reso = batch["bucket_reso"] + + # バッチを分割して処理する + for i in range(0, b_size, vae_batch_size): + images = batch["images"][i : i + vae_batch_size] + absolute_paths = batch["absolute_paths"][i : i + vae_batch_size] + resized_sizes = batch["resized_sizes"][i : i + vae_batch_size] + + image_infos = [] + for i, (image, absolute_path, resized_size) in enumerate(zip(images, absolute_paths, resized_sizes)): + image_info = train_util.ImageInfo(absolute_path, 1, "dummy", False, absolute_path) + image_info.image = image + image_info.bucket_reso = bucket_reso + image_info.resized_size = resized_size + image_info.latents_npz = os.path.splitext(absolute_path)[0] + ".npz" + + if args.skip_existing: + if train_util.is_disk_cached_latents_is_expected( + image_info.bucket_reso, image_info.latents_npz, flip_aug, alpha_mask + ): + logger.warning(f"Skipping {image_info.latents_npz} because it already exists.") + continue + + image_infos.append(image_info) + + if len(image_infos) > 0: + train_util.cache_batch_latents(vae, True, image_infos, flip_aug, alpha_mask, random_crop) + + accelerator.wait_for_everyone() + accelerator.print(f"Finished caching latents for {len(train_dataset_group)} batches.") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + + add_logging_arguments(parser) + train_util.add_sd_models_arguments(parser) + train_util.add_training_arguments(parser, True) + train_util.add_dataset_arguments(parser, True, True, True) + config_util.add_config_arguments(parser) + parser.add_argument("--sdxl", action="store_true", help="Use SDXL model / SDXLモデルを使用する") + parser.add_argument( + "--no_half_vae", + action="store_true", + help="do not use fp16/bf16 VAE in mixed precision (use float VAE) / mixed precisionでも fp16/bf16 VAEを使わずfloat VAEを使う", + ) + parser.add_argument( + "--skip_existing", + action="store_true", + help="skip images if npz already exists (both normal and flipped exists if flip_aug is enabled) / npzが既に存在する画像をスキップする（flip_aug有効時は通常、反転の両方が存在する画像をスキップ）", + ) + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + args = train_util.read_config_from_file(args, parser) + + cache_to_disk(args) diff --git a/cache_text_encoder_outputs.py b/cache_text_encoder_outputs.py new file mode 100644 index 0000000000000000000000000000000000000000..a75d9da74a2a6ef4cc1bbd412edb04a0b1565e2e --- /dev/null +++ b/cache_text_encoder_outputs.py @@ -0,0 +1,197 @@ +# text encoder出力のdiskへの事前キャッシュを行う / cache text encoder outputs to disk in advance + +import argparse +import math +from multiprocessing import Value +import os + +from accelerate.utils import set_seed +import torch +from tqdm import tqdm + +from library import config_util +from library import train_util +from library import sdxl_train_util +from library.config_util import ( + ConfigSanitizer, + BlueprintGenerator, +) +from library.utils import setup_logging, add_logging_arguments +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def cache_to_disk(args: argparse.Namespace) -> None: + setup_logging(args, reset=True) + train_util.prepare_dataset_args(args, True) + + # check cache arg + assert ( + args.cache_text_encoder_outputs_to_disk + ), "cache_text_encoder_outputs_to_disk must be True / cache_text_encoder_outputs_to_diskはTrueである必要があります" + + # できるだけ準備はしておくが今のところSDXLのみしか動かない + assert ( + args.sdxl + ), "cache_text_encoder_outputs_to_disk is only available for SDXL / cache_text_encoder_outputs_to_diskはSDXLのみ利用可能です" + + use_dreambooth_method = args.in_json is None + + if args.seed is not None: + set_seed(args.seed) # 乱数系列を初期化する + + # tokenizerを準備する：datasetを動かすために必要 + if args.sdxl: + tokenizer1, tokenizer2 = sdxl_train_util.load_tokenizers(args) + tokenizers = [tokenizer1, tokenizer2] + else: + tokenizer = train_util.load_tokenizer(args) + tokenizers = [tokenizer] + + # データセットを準備する + if args.dataset_class is None: + blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, True)) + if args.dataset_config is not None: + logger.info(f"Load dataset config from {args.dataset_config}") + user_config = config_util.load_user_config(args.dataset_config) + ignored = ["train_data_dir", "in_json"] + if any(getattr(args, attr) is not None for attr in ignored): + logger.warning( + "ignore following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format( + ", ".join(ignored) + ) + ) + else: + if use_dreambooth_method: + logger.info("Using DreamBooth method.") + user_config = { + "datasets": [ + { + "subsets": config_util.generate_dreambooth_subsets_config_by_subdirs( + args.train_data_dir, args.reg_data_dir + ) + } + ] + } + else: + logger.info("Training with captions.") + user_config = { + "datasets": [ + { + "subsets": [ + { + "image_dir": args.train_data_dir, + "metadata_file": args.in_json, + } + ] + } + ] + } + + blueprint = blueprint_generator.generate(user_config, args, tokenizer=tokenizers) + train_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group) + else: + train_dataset_group = train_util.load_arbitrary_dataset(args, tokenizers) + + current_epoch = Value("i", 0) + current_step = Value("i", 0) + ds_for_collator = train_dataset_group if args.max_data_loader_n_workers == 0 else None + collator = train_util.collator_class(current_epoch, current_step, ds_for_collator) + + # acceleratorを準備する + logger.info("prepare accelerator") + args.deepspeed = False + accelerator = train_util.prepare_accelerator(args) + + # mixed precisionに対応した型を用意しておき適宜castする + weight_dtype, _ = train_util.prepare_dtype(args) + + # モデルを読み込む + logger.info("load model") + if args.sdxl: + (_, text_encoder1, text_encoder2, _, _, _, _) = sdxl_train_util.load_target_model(args, accelerator, "sdxl", weight_dtype) + text_encoders = [text_encoder1, text_encoder2] + else: + text_encoder1, _, _, _ = train_util.load_target_model(args, weight_dtype, accelerator) + text_encoders = [text_encoder1] + + for text_encoder in text_encoders: + text_encoder.to(accelerator.device, dtype=weight_dtype) + text_encoder.requires_grad_(False) + text_encoder.eval() + + # dataloaderを準備する + train_dataset_group.set_caching_mode("text") + + # DataLoaderのプロセス数：0 は persistent_workers が使えないので注意 + n_workers = min(args.max_data_loader_n_workers, os.cpu_count()) # cpu_count or max_data_loader_n_workers + + train_dataloader = torch.utils.data.DataLoader( + train_dataset_group, + batch_size=1, + shuffle=True, + collate_fn=collator, + num_workers=n_workers, + persistent_workers=args.persistent_data_loader_workers, + ) + + # acceleratorを使ってモデルを準備する：マルチGPUで使えるようになるはず + train_dataloader = accelerator.prepare(train_dataloader) + + # データ取得のためのループ + for batch in tqdm(train_dataloader): + absolute_paths = batch["absolute_paths"] + input_ids1_list = batch["input_ids1_list"] + input_ids2_list = batch["input_ids2_list"] + + image_infos = [] + for absolute_path, input_ids1, input_ids2 in zip(absolute_paths, input_ids1_list, input_ids2_list): + image_info = train_util.ImageInfo(absolute_path, 1, "dummy", False, absolute_path) + image_info.text_encoder_outputs_npz = os.path.splitext(absolute_path)[0] + train_util.TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX + image_info + + if args.skip_existing: + if os.path.exists(image_info.text_encoder_outputs_npz): + logger.warning(f"Skipping {image_info.text_encoder_outputs_npz} because it already exists.") + continue + + image_info.input_ids1 = input_ids1 + image_info.input_ids2 = input_ids2 + image_infos.append(image_info) + + if len(image_infos) > 0: + b_input_ids1 = torch.stack([image_info.input_ids1 for image_info in image_infos]) + b_input_ids2 = torch.stack([image_info.input_ids2 for image_info in image_infos]) + train_util.cache_batch_text_encoder_outputs( + image_infos, tokenizers, text_encoders, args.max_token_length, True, b_input_ids1, b_input_ids2, weight_dtype + ) + + accelerator.wait_for_everyone() + accelerator.print(f"Finished caching latents for {len(train_dataset_group)} batches.") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + + add_logging_arguments(parser) + train_util.add_sd_models_arguments(parser) + train_util.add_training_arguments(parser, True) + train_util.add_dataset_arguments(parser, True, True, True) + config_util.add_config_arguments(parser) + sdxl_train_util.add_sdxl_training_arguments(parser) + parser.add_argument("--sdxl", action="store_true", help="Use SDXL model / SDXLモデルを使用する") + parser.add_argument( + "--skip_existing", + action="store_true", + help="skip images if npz already exists (both normal and flipped exists if flip_aug is enabled) / npzが既に存在する画像をスキップする（flip_aug有効時は通常、反転の両方が存在する画像をスキップ）", + ) + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + args = train_util.read_config_from_file(args, parser) + + cache_to_disk(args) diff --git a/canny.py b/canny.py new file mode 100644 index 0000000000000000000000000000000000000000..f2190975c1fe88a02bfe668c4750dda67c24bb5e --- /dev/null +++ b/canny.py @@ -0,0 +1,34 @@ +import argparse +import cv2 + +import logging +from library.utils import setup_logging +setup_logging() +logger = logging.getLogger(__name__) + +def canny(args): + img = cv2.imread(args.input) + img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) + + canny_img = cv2.Canny(img, args.thres1, args.thres2) + # canny_img = 255 - canny_img + + cv2.imwrite(args.output, canny_img) + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("--input", type=str, default=None, help="input path") + parser.add_argument("--output", type=str, default=None, help="output path") + parser.add_argument("--thres1", type=int, default=32, help="thres1") + parser.add_argument("--thres2", type=int, default=224, help="thres2") + + return parser + + +if __name__ == '__main__': + parser = setup_parser() + + args = parser.parse_args() + canny(args) diff --git a/cextension.py b/cextension.py new file mode 100644 index 0000000000000000000000000000000000000000..d38684a2038bc598d2a8f6f0791217598891de82 --- /dev/null +++ b/cextension.py @@ -0,0 +1,54 @@ +import ctypes as ct +from pathlib import Path +from warnings import warn + +from .cuda_setup.main import evaluate_cuda_setup + + +class CUDALibrary_Singleton(object): + _instance = None + + def __init__(self): + raise RuntimeError("Call get_instance() instead") + + def initialize(self): + binary_name = evaluate_cuda_setup() + package_dir = Path(__file__).parent + binary_path = package_dir / binary_name + + if not binary_path.exists(): + print(f"CUDA SETUP: TODO: compile library for specific version: {binary_name}") + legacy_binary_name = "libbitsandbytes.so" + print(f"CUDA SETUP: Defaulting to {legacy_binary_name}...") + binary_path = package_dir / legacy_binary_name + if not binary_path.exists(): + print('CUDA SETUP: CUDA detection failed. Either CUDA driver not installed, CUDA not installed, or you have multiple conflicting CUDA libraries!') + print('CUDA SETUP: If you compiled from source, try again with `make CUDA_VERSION=DETECTED_CUDA_VERSION` for example, `make CUDA_VERSION=113`.') + raise Exception('CUDA SETUP: Setup Failed!') + # self.lib = ct.cdll.LoadLibrary(binary_path) + self.lib = ct.cdll.LoadLibrary(str(binary_path)) # $$$ + else: + print(f"CUDA SETUP: Loading binary {binary_path}...") + # self.lib = ct.cdll.LoadLibrary(binary_path) + self.lib = ct.cdll.LoadLibrary(str(binary_path)) # $$$ + + @classmethod + def get_instance(cls): + if cls._instance is None: + cls._instance = cls.__new__(cls) + cls._instance.initialize() + return cls._instance + + +lib = CUDALibrary_Singleton.get_instance().lib +try: + lib.cadam32bit_g32 + lib.get_context.restype = ct.c_void_p + lib.get_cusparse.restype = ct.c_void_p + COMPILED_WITH_CUDA = True +except AttributeError: + warn( + "The installed version of bitsandbytes was compiled without GPU support. " + "8-bit optimizers and GPU quantization are unavailable." + ) + COMPILED_WITH_CUDA = False diff --git a/check_lora_weights.py b/check_lora_weights.py new file mode 100644 index 0000000000000000000000000000000000000000..f8eab53ba528c842e5209d69d4a7e4abdbae5ab8 --- /dev/null +++ b/check_lora_weights.py @@ -0,0 +1,48 @@ +import argparse +import os +import torch +from safetensors.torch import load_file +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def main(file): + logger.info(f"loading: {file}") + if os.path.splitext(file)[1] == ".safetensors": + sd = load_file(file) + else: + sd = torch.load(file, map_location="cpu") + + values = [] + + keys = list(sd.keys()) + for key in keys: + if "lora_up" in key or "lora_down" in key or "lora_A" in key or "lora_B" in key or "oft_" in key: + values.append((key, sd[key])) + print(f"number of LoRA modules: {len(values)}") + + if args.show_all_keys: + for key in [k for k in keys if k not in values]: + values.append((key, sd[key])) + print(f"number of all modules: {len(values)}") + + for key, value in values: + value = value.to(torch.float32) + print(f"{key},{str(tuple(value.size())).replace(', ', '-')},{torch.mean(torch.abs(value))},{torch.min(torch.abs(value))}") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("file", type=str, help="model file to check / 重みを確認するモデルファイル") + parser.add_argument("-s", "--show_all_keys", action="store_true", help="show all keys / 全てのキーを表示する") + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + + main(args.file) diff --git a/clean_captions_and_tags.py b/clean_captions_and_tags.py new file mode 100644 index 0000000000000000000000000000000000000000..5aeb174259d8f9b55b42b3bbc3c910548b8a224d --- /dev/null +++ b/clean_captions_and_tags.py @@ -0,0 +1,194 @@ +# このスクリプトのライセンスは、Apache License 2.0とします +# (c) 2022 Kohya S. @kohya_ss + +import argparse +import glob +import os +import json +import re + +from tqdm import tqdm +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +PATTERN_HAIR_LENGTH = re.compile(r', (long|short|medium) hair, ') +PATTERN_HAIR_CUT = re.compile(r', (bob|hime) cut, ') +PATTERN_HAIR = re.compile(r', ([\w\-]+) hair, ') +PATTERN_WORD = re.compile(r', ([\w\-]+|hair ornament), ') + +# 複数人がいるとき、複数の髪色や目の色が定義されていれば削除する +PATTERNS_REMOVE_IN_MULTI = [ + PATTERN_HAIR_LENGTH, + PATTERN_HAIR_CUT, + re.compile(r', [\w\-]+ eyes, '), + re.compile(r', ([\w\-]+ sleeves|sleeveless), '), + # 複数の髪型定義がある場合は削除する + re.compile( + r', (ponytail|braid|ahoge|twintails|[\w\-]+ bun|single hair bun|single side bun|two side up|two tails|[\w\-]+ braid|sidelocks), '), +] + + +def clean_tags(image_key, tags): + # replace '_' to ' ' + tags = tags.replace('^_^', '^@@@^') + tags = tags.replace('_', ' ') + tags = tags.replace('^@@@^', '^_^') + + # remove rating: deepdanbooruのみ + tokens = tags.split(", rating") + if len(tokens) == 1: + # WD14 taggerのときはこちらになるのでメッセージは出さない + # logger.info("no rating:") + # logger.info(f"{image_key} {tags}") + pass + else: + if len(tokens) > 2: + logger.info("multiple ratings:") + logger.info(f"{image_key} {tags}") + tags = tokens[0] + + tags = ", " + tags.replace(", ", ", , ") + ", " # カンマ付きで検索をするための身も蓋もない対策 + + # 複数の人物がいる場合は髪色等のタグを削除する + if 'girls' in tags or 'boys' in tags: + for pat in PATTERNS_REMOVE_IN_MULTI: + found = pat.findall(tags) + if len(found) > 1: # 二つ以上、タグがある + tags = pat.sub("", tags) + + # 髪の特殊対応 + srch_hair_len = PATTERN_HAIR_LENGTH.search(tags) # 髪の長さタグは例外なので避けておく（全員が同じ髪の長さの場合） + if srch_hair_len: + org = srch_hair_len.group() + tags = PATTERN_HAIR_LENGTH.sub(", @@@, ", tags) + + found = PATTERN_HAIR.findall(tags) + if len(found) > 1: + tags = PATTERN_HAIR.sub("", tags) + + if srch_hair_len: + tags = tags.replace(", @@@, ", org) # 戻す + + # white shirtとshirtみたいな重複タグの削除 + found = PATTERN_WORD.findall(tags) + for word in found: + if re.search(f", ((\w+) )+{word}, ", tags): + tags = tags.replace(f", {word}, ", "") + + tags = tags.replace(", , ", ", ") + assert tags.startswith(", ") and tags.endswith(", ") + tags = tags[2:-2] + return tags + + +# 上から順に検索、置換される +# ('置換元文字列', '置換後文字列') +CAPTION_REPLACEMENTS = [ + ('anime anime', 'anime'), + ('young ', ''), + ('anime girl', 'girl'), + ('cartoon female', 'girl'), + ('cartoon lady', 'girl'), + ('cartoon character', 'girl'), # a or ~s + ('cartoon woman', 'girl'), + ('cartoon women', 'girls'), + ('cartoon girl', 'girl'), + ('anime female', 'girl'), + ('anime lady', 'girl'), + ('anime character', 'girl'), # a or ~s + ('anime woman', 'girl'), + ('anime women', 'girls'), + ('lady', 'girl'), + ('female', 'girl'), + ('woman', 'girl'), + ('women', 'girls'), + ('people', 'girls'), + ('person', 'girl'), + ('a cartoon figure', 'a figure'), + ('a cartoon image', 'an image'), + ('a cartoon picture', 'a picture'), + ('an anime cartoon image', 'an image'), + ('a cartoon anime drawing', 'a drawing'), + ('a cartoon drawing', 'a drawing'), + ('girl girl', 'girl'), +] + + +def clean_caption(caption): + for rf, rt in CAPTION_REPLACEMENTS: + replaced = True + while replaced: + bef = caption + caption = caption.replace(rf, rt) + replaced = bef != caption + return caption + + +def main(args): + if os.path.exists(args.in_json): + logger.info(f"loading existing metadata: {args.in_json}") + with open(args.in_json, "rt", encoding='utf-8') as f: + metadata = json.load(f) + else: + logger.error("no metadata / メタデータファイルがありません") + return + + logger.info("cleaning captions and tags.") + image_keys = list(metadata.keys()) + for image_key in tqdm(image_keys): + tags = metadata[image_key].get('tags') + if tags is None: + logger.error(f"image does not have tags / メタデータにタグがありません: {image_key}") + else: + org = tags + tags = clean_tags(image_key, tags) + metadata[image_key]['tags'] = tags + if args.debug and org != tags: + logger.info("FROM: " + org) + logger.info("TO: " + tags) + + caption = metadata[image_key].get('caption') + if caption is None: + logger.error(f"image does not have caption / メタデータにキャプションがありません: {image_key}") + else: + org = caption + caption = clean_caption(caption) + metadata[image_key]['caption'] = caption + if args.debug and org != caption: + logger.info("FROM: " + org) + logger.info("TO: " + caption) + + # metadataを書き出して終わり + logger.info(f"writing metadata: {args.out_json}") + with open(args.out_json, "wt", encoding='utf-8') as f: + json.dump(metadata, f, indent=2) + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + # parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ") + parser.add_argument("in_json", type=str, help="metadata file to input / 読み込むメタデータファイル") + parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先") + parser.add_argument("--debug", action="store_true", help="debug mode") + + return parser + + +if __name__ == '__main__': + parser = setup_parser() + + args, unknown = parser.parse_known_args() + if len(unknown) == 1: + logger.warning("WARNING: train_data_dir argument is removed. This script will not work with three arguments in future. Please specify two arguments: in_json and out_json.") + logger.warning("All captions and tags in the metadata are processed.") + logger.warning("警告: train_data_dir引数は不要になりました。将来的には三つの引数を指定すると動かなくなる予定です。読み込み元のメタデータと書き出し先の二つの引数だけ指定してください。") + logger.warning("メタデータ内のすべてのキャプションとタグが処理されます。") + args.in_json = args.out_json + args.out_json = unknown[0] + elif len(unknown) > 0: + raise ValueError(f"error: unrecognized arguments: {unknown}") + + main(args) diff --git a/config_README-en.md b/config_README-en.md new file mode 100644 index 0000000000000000000000000000000000000000..66a50dc09244608192fa0f5e93e9f3f35147dcea --- /dev/null +++ b/config_README-en.md @@ -0,0 +1,386 @@ +Original Source by kohya-ss + +First version: +A.I Translation by Model: NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO, editing by Darkstorm2150 + +Some parts are manually added. + +# Config Readme + +This README is about the configuration files that can be passed with the `--dataset_config` option. + +## Overview + +By passing a configuration file, users can make detailed settings. + +* Multiple datasets can be configured + * For example, by setting `resolution` for each dataset, they can be mixed and trained. + * In training methods that support both the DreamBooth approach and the fine-tuning approach, datasets of the DreamBooth method and the fine-tuning method can be mixed. +* Settings can be changed for each subset + * A subset is a partition of the dataset by image directory or metadata. Several subsets make up a dataset. + * Options such as `keep_tokens` and `flip_aug` can be set for each subset. On the other hand, options such as `resolution` and `batch_size` can be set for each dataset, and their values are common among subsets belonging to the same dataset. More details will be provided later. + +The configuration file format can be JSON or TOML. Considering the ease of writing, it is recommended to use [TOML](https://toml.io/ja/v1.0.0-rc.2). The following explanation assumes the use of TOML. + + +Here is an example of a configuration file written in TOML. + +```toml +[general] +shuffle_caption = true +caption_extension = '.txt' +keep_tokens = 1 + +# This is a DreamBooth-style dataset +[[datasets]] +resolution = 512 +batch_size = 4 +keep_tokens = 2 + + [[datasets.subsets]] + image_dir = 'C:\hoge' + class_tokens = 'hoge girl' + # This subset uses keep_tokens = 2 (the value of the parent datasets) + + [[datasets.subsets]] + image_dir = 'C:\fuga' + class_tokens = 'fuga boy' + keep_tokens = 3 + + [[datasets.subsets]] + is_reg = true + image_dir = 'C:\reg' + class_tokens = 'human' + keep_tokens = 1 + +# This is a fine-tuning dataset +[[datasets]] +resolution = [768, 768] +batch_size = 2 + + [[datasets.subsets]] + image_dir = 'C:\piyo' + metadata_file = 'C:\piyo\piyo_md.json' + # This subset uses keep_tokens = 1 (the value of [general]) +``` + +In this example, three directories are trained as a DreamBooth-style dataset at 512x512 (batch size 4), and one directory is trained as a fine-tuning dataset at 768x768 (batch size 2). + +## Settings for datasets and subsets + +Settings for datasets and subsets are divided into several registration locations. + +* `[general]` + * This is where options that apply to all datasets or all subsets are specified. + * If there are options with the same name in the dataset-specific or subset-specific settings, the dataset-specific or subset-specific settings take precedence. +* `[[datasets]]` + * `datasets` is where settings for datasets are registered. This is where options that apply individually to each dataset are specified. + * If there are subset-specific settings, the subset-specific settings take precedence. +* `[[datasets.subsets]]` + * `datasets.subsets` is where settings for subsets are registered. This is where options that apply individually to each subset are specified. + +Here is an image showing the correspondence between image directories and registration locations in the previous example. + +``` +C:\ +├─ hoge -> [[datasets.subsets]] No.1 ┐ ┐ +├─ fuga -> [[datasets.subsets]] No.2 |-> [[datasets]] No.1 |-> [general] +├─ reg -> [[datasets.subsets]] No.3 ┘ | +└─ piyo -> [[datasets.subsets]] No.4 --> [[datasets]] No.2 ┘ +``` + +The image directory corresponds to each `[[datasets.subsets]]`. Then, multiple `[[datasets.subsets]]` are combined to form one `[[datasets]]`. All `[[datasets]]` and `[[datasets.subsets]]` belong to `[general]`. + +The available options for each registration location may differ, but if the same option is specified, the value in the lower registration location will take precedence. You can check how the `keep_tokens` option is handled in the previous example for better understanding. + +Additionally, the available options may vary depending on the method that the learning approach supports. + +* Options specific to the DreamBooth method +* Options specific to the fine-tuning method +* Options available when using the caption dropout technique + +When using both the DreamBooth method and the fine-tuning method, they can be used together with a learning approach that supports both. +When using them together, a point to note is that the method is determined based on the dataset, so it is not possible to mix DreamBooth method subsets and fine-tuning method subsets within the same dataset. +In other words, if you want to use both methods together, you need to set up subsets of different methods belonging to different datasets. + +In terms of program behavior, if the `metadata_file` option exists, it is determined to be a subset of fine-tuning. Therefore, for subsets belonging to the same dataset, as long as they are either "all have the `metadata_file` option" or "all have no `metadata_file` option," there is no problem. + +Below, the available options will be explained. For options with the same name as the command-line argument, the explanation will be omitted in principle. Please refer to other READMEs. + +### Common options for all learning methods + +These are options that can be specified regardless of the learning method. + +#### Data set specific options + +These are options related to the configuration of the data set. They cannot be described in `datasets.subsets`. + + +| Option Name | Example Setting | `[general]` | `[[datasets]]` | +| ---- | ---- | ---- | ---- | +| `batch_size` | `1` | o | o | +| `bucket_no_upscale` | `true` | o | o | +| `bucket_reso_steps` | `64` | o | o | +| `enable_bucket` | `true` | o | o | +| `max_bucket_reso` | `1024` | o | o | +| `min_bucket_reso` | `128` | o | o | +| `resolution` | `256`, `[512, 512]` | o | o | + +* `batch_size` + * This corresponds to the command-line argument `--train_batch_size`. +* `max_bucket_reso`, `min_bucket_reso` + * Specify the maximum and minimum resolutions of the bucket. It must be divisible by `bucket_reso_steps`. + +These settings are fixed per dataset. That means that subsets belonging to the same dataset will share these settings. For example, if you want to prepare datasets with different resolutions, you can define them as separate datasets as shown in the example above, and set different resolutions for each. + +#### Options for Subsets + +These options are related to subset configuration. + +| Option Name | Example | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | ---- | +| `color_aug` | `false` | o | o | o | +| `face_crop_aug_range` | `[1.0, 3.0]` | o | o | o | +| `flip_aug` | `true` | o | o | o | +| `keep_tokens` | `2` | o | o | o | +| `num_repeats` | `10` | o | o | o | +| `random_crop` | `false` | o | o | o | +| `shuffle_caption` | `true` | o | o | o | +| `caption_prefix` | `"masterpiece, best quality, "` | o | o | o | +| `caption_suffix` | `", from side"` | o | o | o | +| `caption_separator` | (not specified) | o | o | o | +| `keep_tokens_separator` | `“|||”` | o | o | o | +| `secondary_separator` | `“;;;”` | o | o | o | +| `enable_wildcard` | `true` | o | o | o | + +* `num_repeats` + * Specifies the number of repeats for images in a subset. This is equivalent to `--dataset_repeats` in fine-tuning but can be specified for any training method. +* `caption_prefix`, `caption_suffix` + * Specifies the prefix and suffix strings to be appended to the captions. Shuffling is performed with these strings included. Be cautious when using `keep_tokens`. +* `caption_separator` + * Specifies the string to separate the tags. The default is `,`. This option is usually not necessary to set. +* `keep_tokens_separator` + * Specifies the string to separate the parts to be fixed in the caption. For example, if you specify `aaa, bbb ||| ccc, ddd, eee, fff ||| ggg, hhh`, the parts `aaa, bbb` and `ggg, hhh` will remain, and the rest will be shuffled and dropped. The comma in between is not necessary. As a result, the prompt will be `aaa, bbb, eee, ccc, fff, ggg, hhh` or `aaa, bbb, fff, ccc, eee, ggg, hhh`, etc. +* `secondary_separator` + * Specifies an additional separator. The part separated by this separator is treated as one tag and is shuffled and dropped. It is then replaced by `caption_separator`. For example, if you specify `aaa;;;bbb;;;ccc`, it will be replaced by `aaa,bbb,ccc` or dropped together. +* `enable_wildcard` + * Enables wildcard notation. This will be explained later. + +### DreamBooth-specific options + +DreamBooth-specific options only exist as subsets-specific options. + +#### Subset-specific options + +Options related to the configuration of DreamBooth subsets. + +| Option Name | Example Setting | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | ---- | +| `image_dir` | `'C:\hoge'` | - | - | o (required) | +| `caption_extension` | `".txt"` | o | o | o | +| `class_tokens` | `"sks girl"` | - | - | o | +| `cache_info` | `false` | o | o | o | +| `is_reg` | `false` | - | - | o | + +Firstly, note that for `image_dir`, the path to the image files must be specified as being directly in the directory. Unlike the previous DreamBooth method, where images had to be placed in subdirectories, this is not compatible with that specification. Also, even if you name the folder something like "5_cat", the number of repeats of the image and the class name will not be reflected. If you want to set these individually, you will need to explicitly specify them using `num_repeats` and `class_tokens`. + +* `image_dir` + * Specifies the path to the image directory. This is a required option. + * Images must be placed directly under the directory. +* `class_tokens` + * Sets the class tokens. + * Only used during training when a corresponding caption file does not exist. The determination of whether or not to use it is made on a per-image basis. If `class_tokens` is not specified and a caption file is not found, an error will occur. +* `cache_info` + * Specifies whether to cache the image size and caption. If not specified, it is set to `false`. The cache is saved in `metadata_cache.json` in `image_dir`. + * Caching speeds up the loading of the dataset after the first time. It is effective when dealing with thousands of images or more. +* `is_reg` + * Specifies whether the subset images are for normalization. If not specified, it is set to `false`, meaning that the images are not for normalization. + +### Fine-tuning method specific options + +The options for the fine-tuning method only exist for subset-specific options. + +#### Subset-specific options + +These options are related to the configuration of the fine-tuning method's subsets. + +| Option name | Example setting | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | ---- | +| `image_dir` | `'C:\hoge'` | - | - | o | +| `metadata_file` | `'C:\piyo\piyo_md.json'` | - | - | o (required) | + +* `image_dir` + * Specify the path to the image directory. Unlike the DreamBooth method, specifying it is not mandatory, but it is recommended to do so. + * The case where it is not necessary to specify is when the `--full_path` is added to the command line when generating the metadata file. + * The images must be placed directly under the directory. +* `metadata_file` + * Specify the path to the metadata file used for the subset. This is a required option. + * It is equivalent to the command-line argument `--in_json`. + * Due to the specification that a metadata file must be specified for each subset, it is recommended to avoid creating a metadata file with images from different directories as a single metadata file. It is strongly recommended to prepare a separate metadata file for each image directory and register them as separate subsets. + +### Options available when caption dropout method can be used + +The options available when the caption dropout method can be used exist only for subsets. Regardless of whether it's the DreamBooth method or fine-tuning method, if it supports caption dropout, it can be specified. + +#### Subset-specific options + +Options related to the setting of subsets that caption dropout can be used for. + +| Option Name | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | +| `caption_dropout_every_n_epochs` | o | o | o | +| `caption_dropout_rate` | o | o | o | +| `caption_tag_dropout_rate` | o | o | o | + +## Behavior when there are duplicate subsets + +In the case of the DreamBooth dataset, if there are multiple `image_dir` directories with the same content, they are considered to be duplicate subsets. For the fine-tuning dataset, if there are multiple `metadata_file` files with the same content, they are considered to be duplicate subsets. If duplicate subsets exist in the dataset, subsequent subsets will be ignored. + +However, if they belong to different datasets, they are not considered duplicates. For example, if you have subsets with the same `image_dir` in different datasets, they will not be considered duplicates. This is useful when you want to train with the same image but with different resolutions. + +```toml +# If data sets exist separately, they are not considered duplicates and are both used for training. + +[[datasets]] +resolution = 512 + + [[datasets.subsets]] + image_dir = 'C:\hoge' + +[[datasets]] +resolution = 768 + + [[datasets.subsets]] + image_dir = 'C:\hoge' +``` + +## Command Line Argument and Configuration File + +There are options in the configuration file that have overlapping roles with command line argument options. + +The following command line argument options are ignored if a configuration file is passed: + +* `--train_data_dir` +* `--reg_data_dir` +* `--in_json` + +The following command line argument options are given priority over the configuration file options if both are specified simultaneously. In most cases, they have the same names as the corresponding options in the configuration file. + +| Command Line Argument Option | Prioritized Configuration File Option | +| ------------------------------- | ------------------------------------- | +| `--bucket_no_upscale` | | +| `--bucket_reso_steps` | | +| `--caption_dropout_every_n_epochs` | | +| `--caption_dropout_rate` | | +| `--caption_extension` | | +| `--caption_tag_dropout_rate` | | +| `--color_aug` | | +| `--dataset_repeats` | `num_repeats` | +| `--enable_bucket` | | +| `--face_crop_aug_range` | | +| `--flip_aug` | | +| `--keep_tokens` | | +| `--min_bucket_reso` | | +| `--random_crop` | | +| `--resolution` | | +| `--shuffle_caption` | | +| `--train_batch_size` | `batch_size` | + +## Error Guide + +Currently, we are using an external library to check if the configuration file is written correctly, but the development has not been completed, and there is a problem that the error message is not clear. In the future, we plan to improve this problem. + +As a temporary measure, we will list common errors and their solutions. If you encounter an error even though it should be correct or if the error content is not understandable, please contact us as it may be a bug. + +* `voluptuous.error.MultipleInvalid: required key not provided @ ...`: This error occurs when a required option is not provided. It is highly likely that you forgot to specify the option or misspelled the option name. + * The error location is indicated by `...` in the error message. For example, if you encounter an error like `voluptuous.error.MultipleInvalid: required key not provided @ data['datasets'][0]['subsets'][0]['image_dir']`, it means that the `image_dir` option does not exist in the 0th `subsets` of the 0th `datasets` setting. +* `voluptuous.error.MultipleInvalid: expected int for dictionary value @ ...`: This error occurs when the specified value format is incorrect. It is highly likely that the value format is incorrect. The `int` part changes depending on the target option. The example configurations in this README may be helpful. +* `voluptuous.error.MultipleInvalid: extra keys not allowed @ ...`: This error occurs when there is an option name that is not supported. It is highly likely that you misspelled the option name or mistakenly included it. + +## Miscellaneous + +### Multi-line captions + +By setting `enable_wildcard = true`, multiple-line captions are also enabled. If the caption file consists of multiple lines, one line is randomly selected as the caption. + +```txt +1girl, hatsune miku, vocaloid, upper body, looking at viewer, microphone, stage +a girl with a microphone standing on a stage +detailed digital art of a girl with a microphone on a stage +``` + +It can be combined with wildcard notation. + +In metadata files, you can also specify multiple-line captions. In the `.json` metadata file, use `\n` to represent a line break. If the caption file consists of multiple lines, `merge_captions_to_metadata.py` will create a metadata file in this format. + +The tags in the metadata (`tags`) are added to each line of the caption. + +```json +{ + "/path/to/image.png": { + "caption": "a cartoon of a frog with the word frog on it\ntest multiline caption1\ntest multiline caption2", + "tags": "open mouth, simple background, standing, no humans, animal, black background, frog, animal costume, animal focus" + }, + ... +} +``` + +In this case, the actual caption will be `a cartoon of a frog with the word frog on it, open mouth, simple background ...`, `test multiline caption1, open mouth, simple background ...`, `test multiline caption2, open mouth, simple background ...`, etc. + +### Example of configuration file : `secondary_separator`, wildcard notation, `keep_tokens_separator`, etc. + +```toml +[general] +flip_aug = true +color_aug = false +resolution = [1024, 1024] + +[[datasets]] +batch_size = 6 +enable_bucket = true +bucket_no_upscale = true +caption_extension = ".txt" +keep_tokens_separator= "|||" +shuffle_caption = true +caption_tag_dropout_rate = 0.1 +secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side +enable_wildcard = true # 同上 / same as above + + [[datasets.subsets]] + image_dir = "/path/to/image_dir" + num_repeats = 1 + + # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically) + caption_prefix = "1girl, hatsune miku, vocaloid |||" + + # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains + # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself + caption_suffix = ", anime screencap ||| masterpiece, rating: general" +``` + +### Example of caption, secondary_separator notation: `secondary_separator = ";;;"` + +```txt +1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors +``` +The part `sky;;;cloud;;;day` is replaced with `sky,cloud,day` without shuffling or dropping. When shuffling and dropping are enabled, it is processed as a whole (as one tag). For example, it becomes `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` (shuffled) or `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` (dropped). + +### Example of caption, enable_wildcard notation: `enable_wildcard = true` + +```txt +1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background +``` +`simple` or `white` is randomly selected, and it becomes `simple background` or `white background`. + +```txt +1girl, hatsune miku, vocaloid, {{retro style}} +``` +If you want to include `{` or `}` in the tag string, double them like `{{` or `}}` (in this example, the actual caption used for training is `{retro style}`). + +### Example of caption, `keep_tokens_separator` notation: `keep_tokens_separator = "|||"` + +```txt +1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general +``` +It becomes `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` or `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` etc. + diff --git a/config_README-ja.md b/config_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..0ed95e0eb794f0d18cf2661da0b1c5fbb74ebbb6 --- /dev/null +++ b/config_README-ja.md @@ -0,0 +1,388 @@ +`--dataset_config` で渡すことができる設定ファイルに関する説明です。 + +## 概要 + +設定ファイルを渡すことにより、ユーザが細かい設定を行えるようにします。 + +* 複数のデータセットが設定可能になります + * 例えば `resolution` をデータセットごとに設定して、それらを混合して学習できます。 + * DreamBooth の手法と fine tuning の手法の両方に対応している学習方法では、DreamBooth 方式と fine tuning 方式のデータセットを混合することが可能です。 +* サブセットごとに設定を変更することが可能になります + * データセットを画像ディレクトリ別またはメタデータ別に分割したものがサブセットです。いくつかのサブセットが集まってデータセットを構成します。 + * `keep_tokens` や `flip_aug` 等のオプションはサブセットごとに設定可能です。一方、`resolution` や `batch_size` といったオプションはデータセットごとに設定可能で、同じデータセットに属するサブセットでは値が共通になります。詳しくは後述します。 + +設定ファイルの形式は JSON か TOML を利用できます。記述のしやすさを考えると [TOML](https://toml.io/ja/v1.0.0-rc.2) を利用するのがオススメです。以下、TOML の利用を前提に説明します。 + +TOML で記述した設定ファイルの例です。 + +```toml +[general] +shuffle_caption = true +caption_extension = '.txt' +keep_tokens = 1 + +# これは DreamBooth 方式のデータセット +[[datasets]] +resolution = 512 +batch_size = 4 +keep_tokens = 2 + + [[datasets.subsets]] + image_dir = 'C:\hoge' + class_tokens = 'hoge girl' + # このサブセットは keep_tokens = 2 （所属する datasets の値が使われる） + + [[datasets.subsets]] + image_dir = 'C:\fuga' + class_tokens = 'fuga boy' + keep_tokens = 3 + + [[datasets.subsets]] + is_reg = true + image_dir = 'C:\reg' + class_tokens = 'human' + keep_tokens = 1 + +# これは fine tuning 方式のデータセット +[[datasets]] +resolution = [768, 768] +batch_size = 2 + + [[datasets.subsets]] + image_dir = 'C:\piyo' + metadata_file = 'C:\piyo\piyo_md.json' + # このサブセットは keep_tokens = 1 （general の値が使われる） +``` + +この例では、3 つのディレクトリを DreamBooth 方式のデータセットとして 512x512 (batch size 4) で学習させ、1 つのディレクトリを fine tuning 方式のデータセットとして 768x768 (batch size 2) で学習させることになります。 + +## データセット・サブセットに関する設定 + +データセット・サブセットに関する設定は、登録可能な箇所がいくつかに分かれています。 + +* `[general]` + * 全データセットまたは全サブセットに適用されるオプションを指定する箇所です。 + * データセットごとの設定及びサブセットごとの設定に同名のオプションが存在していた場合には、データセット・サブセットごとの設定が優先されます。 +* `[[datasets]]` + * `datasets` はデータセットに関する設定の登録箇所になります。各データセットに個別に適用されるオプションを指定する箇所です。 + * サブセットごとの設定が存在していた場合には、サブセットごとの設定が優先されます。 +* `[[datasets.subsets]]` + * `datasets.subsets` はサブセットに関する設定の登録箇所になります。各サブセットに個別に適用されるオプションを指定する箇所です。 + +先程の例における、画像ディレクトリと登録箇所の対応に関するイメージ図です。 + +``` +C:\ +├─ hoge -> [[datasets.subsets]] No.1 ┐ ┐ +├─ fuga -> [[datasets.subsets]] No.2 |-> [[datasets]] No.1 |-> [general] +├─ reg -> [[datasets.subsets]] No.3 ┘ | +└─ piyo -> [[datasets.subsets]] No.4 --> [[datasets]] No.2 ┘ +``` + +画像ディレクトリがそれぞれ1つの `[[datasets.subsets]]` に対応しています。そして `[[datasets.subsets]]` が1つ以上組み合わさって1つの `[[datasets]]` を構成します。`[general]` には全ての `[[datasets]]`, `[[datasets.subsets]]` が属します。 + +登録箇所ごとに指定可能なオプションは異なりますが、同名のオプションが指定された場合は下位の登録箇所にある値が優先されます。先程の例の `keep_tokens` オプションの扱われ方を確認してもらうと理解しやすいかと思います。 + +加えて、学習方法が対応している手法によっても指定可能なオプションが変化します。 + +* DreamBooth 方式専用のオプション +* fine tuning 方式専用のオプション +* caption dropout の手法が使える場合のオプション + +DreamBooth の手法と fine tuning の手法の両方とも利用可能な学習方法では、両者を併用することができます。 +併用する際の注意点として、DreamBooth 方式なのか fine tuning 方式なのかはデータセット単位で判別を行っているため、同じデータセット中に DreamBooth 方式のサブセットと fine tuning 方式のサブセットを混在させることはできません。 +つまり、これらを併用したい場合には異なる方式のサブセットが異なるデータセットに所属するように設定する必要があります。 + +プログラムの挙動としては、後述する `metadata_file` オプションが存在していたら fine tuning 方式のサブセットだと判断します。 +そのため、同一のデータセットに所属するサブセットについて言うと、「全てが `metadata_file` オプションを持つ」か「全てが `metadata_file` オプションを持たない」かのどちらかになっていれば問題ありません。 + +以下、利用可能なオプションを説明します。コマンドライン引数と名称が同一のオプションについては、基本的に説明を割愛します。他の README を参照してください。 + +### 全学習方法で共通のオプション + +学習方法によらずに指定可能なオプションです。 + +#### データセット向けオプション + +データセットの設定に関わるオプションです。`datasets.subsets` には記述できません。 + +| オプション名 | 設定例 | `[general]` | `[[datasets]]` | +| ---- | ---- | ---- | ---- | +| `batch_size` | `1` | o | o | +| `bucket_no_upscale` | `true` | o | o | +| `bucket_reso_steps` | `64` | o | o | +| `enable_bucket` | `true` | o | o | +| `max_bucket_reso` | `1024` | o | o | +| `min_bucket_reso` | `128` | o | o | +| `resolution` | `256`, `[512, 512]` | o | o | + +* `batch_size` + * コマンドライン引数の `--train_batch_size` と同等です。 +* `max_bucket_reso`, `min_bucket_reso` + * bucketの最大、最小解像度を指定します。`bucket_reso_steps` で割り切れる必要があります。 + +これらの設定はデータセットごとに固定です。 +つまり、データセットに所属するサブセットはこれらの設定を共有することになります。 +例えば解像度が異なるデータセットを用意したい場合は、上に挙げた例のように別々のデータセットとして定義すれば別々の解像度を設定可能です。 + +#### サブセット向けオプション + +サブセットの設定に関わるオプションです。 + +| オプション名 | 設定例 | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | ---- | +| `color_aug` | `false` | o | o | o | +| `face_crop_aug_range` | `[1.0, 3.0]` | o | o | o | +| `flip_aug` | `true` | o | o | o | +| `keep_tokens` | `2` | o | o | o | +| `num_repeats` | `10` | o | o | o | +| `random_crop` | `false` | o | o | o | +| `shuffle_caption` | `true` | o | o | o | +| `caption_prefix` | `“masterpiece, best quality, ”` | o | o | o | +| `caption_suffix` | `“, from side”` | o | o | o | +| `caption_separator` | （通常は設定しません） | o | o | o | +| `keep_tokens_separator` | `“|||”` | o | o | o | +| `secondary_separator` | `“;;;”` | o | o | o | +| `enable_wildcard` | `true` | o | o | o | + +* `num_repeats` + * サブセットの画像の繰り返し回数を指定します。fine tuning における `--dataset_repeats` に相当しますが、`num_repeats` はどの学習方法でも指定可能です。 +* `caption_prefix`, `caption_suffix` + * キャプションの前、後に付与する文字列を指定します。シャッフルはこれらの文字列を含めた状態で行われます。`keep_tokens` を指定する場合には注意してください。 + +* `caption_separator` + * タグを区切る文字列を指定します。デフォルトは `,` です。このオプションは通常は設定する必要はありません。 + +* `keep_tokens_separator` + * キャプションで固定したい部分を区切る文字列を指定します。たとえば `aaa, bbb ||| ccc, ddd, eee, fff ||| ggg, hhh` のように指定すると、`aaa, bbb` と `ggg, hhh` の部分はシャッフル、drop されず残ります。間のカンマは不要です。結果としてプロンプトは `aaa, bbb, eee, ccc, fff, ggg, hhh` や `aaa, bbb, fff, ccc, eee, ggg, hhh` などになります。 + +* `secondary_separator` + * 追加の区切り文字を指定します。この区切り文字で区切られた部分は一つのタグとして扱われ、シャッフル、drop されます。その後、`caption_separator` に置き換えられます。たとえば `aaa;;;bbb;;;ccc` のように指定すると、`aaa,bbb,ccc` に置き換えられるか、まとめて drop されます。 + +* `enable_wildcard` + * ワイルドカード記法および複数行キャプションを有効にします。ワイルドカード記法、複数行キャプションについては後述します。 + +### DreamBooth 方式専用のオプション + +DreamBooth 方式のオプションは、サブセット向けオプションのみ存在します。 + +#### サブセット向けオプション + +DreamBooth 方式のサブセットの設定に関わるオプションです。 + +| オプション名 | 設定例 | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | ---- | +| `image_dir` | `‘C:\hoge’` | - | - | o（必須） | +| `caption_extension` | `".txt"` | o | o | o | +| `class_tokens` | `“sks girl”` | - | - | o | +| `cache_info` | `false` | o | o | o | +| `is_reg` | `false` | - | - | o | + +まず注意点として、 `image_dir` には画像ファイルが直下に置かれているパスを指定する必要があります。従来の DreamBooth の手法ではサブディレクトリに画像を置く必要がありましたが、そちらとは仕様に互換性がありません。また、`5_cat` のようなフォルダ名にしても、画像の繰り返し回数とクラス名は反映されません。これらを個別に設定したい場合、`num_repeats` と `class_tokens` で明示的に指定する必要があることに注意してください。 + +* `image_dir` + * 画像ディレクトリのパスを指定します。指定必須オプションです。 + * 画像はディレクトリ直下に置かれている必要があります。 +* `class_tokens` + * クラストークンを設定します。 + * 画像に対応する caption ファイルが存在しない場合にのみ学習時に利用されます。利用するかどうかの判定は画像ごとに行います。`class_tokens` を指定しなかった場合に caption ファイルも見つからなかった場合にはエラーになります。 +* `cache_info` + * 画像サイズ、キャプションをキャッシュするかどうかを指定します。指定しなかった場合は `false` になります。キャッシュは `image_dir` に `metadata_cache.json` というファイル名で保存されます。 + * キャッシュを行うと、二回目以降のデータセット読み込みが高速化されます。数千枚以上の画像を扱う場合には有効です。 +* `is_reg` + * サブセットの画像が正規化用かどうかを指定します。指定しなかった場合は `false` として、つまり正規化画像ではないとして扱います。 + +### fine tuning 方式専用のオプション + +fine tuning 方式のオプションは、サブセット向けオプションのみ存在します。 + +#### サブセット向けオプション + +fine tuning 方式のサブセットの設定に関わるオプションです。 + +| オプション名 | 設定例 | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | ---- | +| `image_dir` | `‘C:\hoge’` | - | - | o | +| `metadata_file` | `'C:\piyo\piyo_md.json'` | - | - | o（必須） | + +* `image_dir` + * 画像ディレクトリのパスを指定します。DreamBooth の手法の方とは異なり指定は必須ではありませんが、設定することを推奨します。 + * 指定する必要がない状況としては、メタデータファイルの生成時に `--full_path` を付与して実行していた場合です。 + * 画像はディレクトリ直下に置かれている必要があります。 +* `metadata_file` + * サブセットで利用されるメタデータファイルのパスを指定します。指定必須オプションです。 + * コマンドライン引数の `--in_json` と同等です。 + * サブセットごとにメタデータファイルを指定する必要がある仕様上、ディレクトリを跨いだメタデータを1つのメタデータファイルとして作成することは避けた方が良いでしょう。画像ディレクトリごとにメタデータファイルを用意し、それらを別々のサブセットとして登録することを強く推奨します。 + +### caption dropout の手法が使える場合に指定可能なオプション + +caption dropout の手法が使える場合のオプションは、サブセット向けオプションのみ存在します。 +DreamBooth 方式か fine tuning 方式かに関わらず、caption dropout に対応している学習方法であれば指定可能です。 + +#### サブセット向けオプション + +caption dropout が使えるサブセットの設定に関わるオプションです。 + +| オプション名 | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` | +| ---- | ---- | ---- | ---- | +| `caption_dropout_every_n_epochs` | o | o | o | +| `caption_dropout_rate` | o | o | o | +| `caption_tag_dropout_rate` | o | o | o | + +## 重複したサブセットが存在する時の挙動 + +DreamBooth 方式のデータセットの場合、その中にある `image_dir` が同一のサブセットは重複していると見なされます。 +fine tuning 方式のデータセットの場合は、その中にある `metadata_file` が同一のサブセットは重複していると見なされます。 +データセット中に重複したサブセットが存在する場合、2個目以降は無視されます。 + +一方、異なるデータセットに所属している場合は、重複しているとは見なされません。 +例えば、以下のように同一の `image_dir` を持つサブセットを別々のデータセットに入れた場合には、重複していないと見なします。 +これは、同じ画像でも異なる解像度で学習したい場合に役立ちます。 + +```toml +# 別々のデータセットに存在している場合は重複とは見なされず、両方とも学習に使われる + +[[datasets]] +resolution = 512 + + [[datasets.subsets]] + image_dir = 'C:\hoge' + +[[datasets]] +resolution = 768 + + [[datasets.subsets]] + image_dir = 'C:\hoge' +``` + +## コマンドライン引数との併用 + +設定ファイルのオプションの中には、コマンドライン引数のオプションと役割が重複しているものがあります。 + +以下に挙げるコマンドライン引数のオプションは、設定ファイルを渡した場合には無視されます。 + +* `--train_data_dir` +* `--reg_data_dir` +* `--in_json` + +以下に挙げるコマンドライン引数のオプションは、コマンドライン引数と設定ファイルで同時に指定された場合、コマンドライン引数の値よりも設定ファイルの値が優先されます。特に断りがなければ同名のオプションとなります。 + +| コマンドライン引数のオプション | 優先される設定ファイルのオプション | +| ---------------------------------- | ---------------------------------- | +| `--bucket_no_upscale` | | +| `--bucket_reso_steps` | | +| `--caption_dropout_every_n_epochs` | | +| `--caption_dropout_rate` | | +| `--caption_extension` | | +| `--caption_tag_dropout_rate` | | +| `--color_aug` | | +| `--dataset_repeats` | `num_repeats` | +| `--enable_bucket` | | +| `--face_crop_aug_range` | | +| `--flip_aug` | | +| `--keep_tokens` | | +| `--min_bucket_reso` | | +| `--random_crop` | | +| `--resolution` | | +| `--shuffle_caption` | | +| `--train_batch_size` | `batch_size` | + +## エラーの手引き + +現在、外部ライブラリを利用して設定ファイルの記述が正しいかどうかをチェックしているのですが、整備が行き届いておらずエラーメッセージがわかりづらいという問題があります。 +将来的にはこの問題の改善に取り組む予定です。 + +次善策として、頻出のエラーとその対処法について載せておきます。 +正しいはずなのにエラーが出る場合、エラー内容がどうしても分からない場合は、バグかもしれないのでご連絡ください。 + +* `voluptuous.error.MultipleInvalid: required key not provided @ ...`: 指定必須のオプションが指定されていないというエラーです。指定を忘れているか、オプション名を間違って記述している可能性が高いです。 + * `...` の箇所にはエラーが発生した場所が載っています。例えば `voluptuous.error.MultipleInvalid: required key not provided @ data['datasets'][0]['subsets'][0]['image_dir']` のようなエラーが出たら、0 番目の `datasets` 中の 0 番目の `subsets` の設定に `image_dir` が存在しないということになります。 +* `voluptuous.error.MultipleInvalid: expected int for dictionary value @ ...`: 指定する値の形式が不正というエラーです。値の形式が間違っている可能性が高いです。`int` の部分は対象となるオプションによって変わります。この README に載っているオプションの「設定例」が役立つかもしれません。 +* `voluptuous.error.MultipleInvalid: extra keys not allowed @ ...`: 対応していないオプション名が存在している場合に発生するエラーです。オプション名を間違って記述しているか、誤って紛れ込んでいる可能性が高いです。 + +## その他 + +### 複数行キャプション + +`enable_wildcard = true` を設定することで、複数行キャプションも同時に有効になります。キャプションファイルが複数の行からなる場合、ランダムに一つの行が選ばれてキャプションとして利用されます。 + +```txt +1girl, hatsune miku, vocaloid, upper body, looking at viewer, microphone, stage +a girl with a microphone standing on a stage +detailed digital art of a girl with a microphone on a stage +``` + +ワイルドカード記法と組み合わせることも可能です。 + +メタデータファイルでも同様に複数行キャプションを指定することができます。メタデータの .json 内には、`\n` を使って改行を表現してください。キャプションファイルが複数行からなる場合、`merge_captions_to_metadata.py` を使うと、この形式でメタデータファイルが作成されます。 + +メタデータのタグ (`tags`) は、キャプションの各行に追加されます。 + +```json +{ + "/path/to/image.png": { + "caption": "a cartoon of a frog with the word frog on it\ntest multiline caption1\ntest multiline caption2", + "tags": "open mouth, simple background, standing, no humans, animal, black background, frog, animal costume, animal focus" + }, + ... +} +``` + +この場合、実際のキャプションは `a cartoon of a frog with the word frog on it, open mouth, simple background ...` または `test multiline caption1, open mouth, simple background ...`、 `test multiline caption2, open mouth, simple background ...` 等になります。 + +### 設定ファイルの記述例：追加の区切り文字、ワイルドカード記法、`keep_tokens_separator` 等 + +```toml +[general] +flip_aug = true +color_aug = false +resolution = [1024, 1024] + +[[datasets]] +batch_size = 6 +enable_bucket = true +bucket_no_upscale = true +caption_extension = ".txt" +keep_tokens_separator= "|||" +shuffle_caption = true +caption_tag_dropout_rate = 0.1 +secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side +enable_wildcard = true # 同上 / same as above + + [[datasets.subsets]] + image_dir = "/path/to/image_dir" + num_repeats = 1 + + # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically) + caption_prefix = "1girl, hatsune miku, vocaloid |||" + + # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains + # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself + caption_suffix = ", anime screencap ||| masterpiece, rating: general" +``` + +### キャプション記述例、secondary_separator 記法：`secondary_separator = ";;;"` の場合 + +```txt +1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors +``` +`sky;;;cloud;;;day` の部分はシャッフル、drop されず `sky,cloud,day` に置換されます。シャッフル、drop が有効な場合、まとめて（一つのタグとして）処理されます。つまり `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` （シャッフル）や `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` （drop されたケース）などになります。 + +### キャプション記述例、ワイルドカード記法： `enable_wildcard = true` の場合 + +```txt +1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background +``` +ランダムに `simple` または `white` が選ばれ、`simple background` または `white background` になります。 + +```txt +1girl, hatsune miku, vocaloid, {{retro style}} +``` +タグ文字列に `{` や `}` そのものを含めたい場合は `{{` や `}}` のように二つ重ねてください（この例では実際に学習に用いられるキャプションは `{retro style}` になります）。 + +### キャプション記述例、`keep_tokens_separator` 記法： `keep_tokens_separator = "|||"` の場合 + +```txt +1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general +``` +`1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` や `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` などになります。 diff --git a/config_util.py b/config_util.py new file mode 100644 index 0000000000000000000000000000000000000000..10b2457f3930dbf26ed5f34179c17f38011b985b --- /dev/null +++ b/config_util.py @@ -0,0 +1,721 @@ +import argparse +from dataclasses import ( + asdict, + dataclass, +) +import functools +import random +from textwrap import dedent, indent +import json +from pathlib import Path + +# from toolz import curry +from typing import ( + List, + Optional, + Sequence, + Tuple, + Union, +) + +import toml +import voluptuous +from voluptuous import ( + Any, + ExactSequence, + MultipleInvalid, + Object, + Required, + Schema, +) +from transformers import CLIPTokenizer + +from . import train_util +from .train_util import ( + DreamBoothSubset, + FineTuningSubset, + ControlNetSubset, + DreamBoothDataset, + FineTuningDataset, + ControlNetDataset, + DatasetGroup, +) +from .utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +def add_config_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--dataset_config", type=Path, default=None, help="config file for detail settings / 詳細な設定用の設定ファイル" + ) + + +# TODO: inherit Params class in Subset, Dataset + + +@dataclass +class BaseSubsetParams: + image_dir: Optional[str] = None + num_repeats: int = 1 + shuffle_caption: bool = False + caption_separator: str = (",",) + keep_tokens: int = 0 + keep_tokens_separator: str = (None,) + secondary_separator: Optional[str] = None + enable_wildcard: bool = False + color_aug: bool = False + flip_aug: bool = False + face_crop_aug_range: Optional[Tuple[float, float]] = None + random_crop: bool = False + caption_prefix: Optional[str] = None + caption_suffix: Optional[str] = None + caption_dropout_rate: float = 0.0 + caption_dropout_every_n_epochs: int = 0 + caption_tag_dropout_rate: float = 0.0 + token_warmup_min: int = 1 + token_warmup_step: float = 0 + + +@dataclass +class DreamBoothSubsetParams(BaseSubsetParams): + is_reg: bool = False + class_tokens: Optional[str] = None + caption_extension: str = ".caption" + cache_info: bool = False + alpha_mask: bool = False + + +@dataclass +class FineTuningSubsetParams(BaseSubsetParams): + metadata_file: Optional[str] = None + alpha_mask: bool = False + + +@dataclass +class ControlNetSubsetParams(BaseSubsetParams): + conditioning_data_dir: str = None + caption_extension: str = ".caption" + cache_info: bool = False + + +@dataclass +class BaseDatasetParams: + tokenizer: Union[CLIPTokenizer, List[CLIPTokenizer]] = None + max_token_length: int = None + resolution: Optional[Tuple[int, int]] = None + network_multiplier: float = 1.0 + debug_dataset: bool = False + + +@dataclass +class DreamBoothDatasetParams(BaseDatasetParams): + batch_size: int = 1 + enable_bucket: bool = False + min_bucket_reso: int = 256 + max_bucket_reso: int = 1024 + bucket_reso_steps: int = 64 + bucket_no_upscale: bool = False + prior_loss_weight: float = 1.0 + + +@dataclass +class FineTuningDatasetParams(BaseDatasetParams): + batch_size: int = 1 + enable_bucket: bool = False + min_bucket_reso: int = 256 + max_bucket_reso: int = 1024 + bucket_reso_steps: int = 64 + bucket_no_upscale: bool = False + + +@dataclass +class ControlNetDatasetParams(BaseDatasetParams): + batch_size: int = 1 + enable_bucket: bool = False + min_bucket_reso: int = 256 + max_bucket_reso: int = 1024 + bucket_reso_steps: int = 64 + bucket_no_upscale: bool = False + + +@dataclass +class SubsetBlueprint: + params: Union[DreamBoothSubsetParams, FineTuningSubsetParams] + + +@dataclass +class DatasetBlueprint: + is_dreambooth: bool + is_controlnet: bool + params: Union[DreamBoothDatasetParams, FineTuningDatasetParams] + subsets: Sequence[SubsetBlueprint] + + +@dataclass +class DatasetGroupBlueprint: + datasets: Sequence[DatasetBlueprint] + + +@dataclass +class Blueprint: + dataset_group: DatasetGroupBlueprint + + +class ConfigSanitizer: + # @curry + @staticmethod + def __validate_and_convert_twodim(klass, value: Sequence) -> Tuple: + Schema(ExactSequence([klass, klass]))(value) + return tuple(value) + + # @curry + @staticmethod + def __validate_and_convert_scalar_or_twodim(klass, value: Union[float, Sequence]) -> Tuple: + Schema(Any(klass, ExactSequence([klass, klass])))(value) + try: + Schema(klass)(value) + return (value, value) + except: + return ConfigSanitizer.__validate_and_convert_twodim(klass, value) + + # subset schema + SUBSET_ASCENDABLE_SCHEMA = { + "color_aug": bool, + "face_crop_aug_range": functools.partial(__validate_and_convert_twodim.__func__, float), + "flip_aug": bool, + "num_repeats": int, + "random_crop": bool, + "shuffle_caption": bool, + "keep_tokens": int, + "keep_tokens_separator": str, + "secondary_separator": str, + "caption_separator": str, + "enable_wildcard": bool, + "token_warmup_min": int, + "token_warmup_step": Any(float, int), + "caption_prefix": str, + "caption_suffix": str, + } + # DO means DropOut + DO_SUBSET_ASCENDABLE_SCHEMA = { + "caption_dropout_every_n_epochs": int, + "caption_dropout_rate": Any(float, int), + "caption_tag_dropout_rate": Any(float, int), + } + # DB means DreamBooth + DB_SUBSET_ASCENDABLE_SCHEMA = { + "caption_extension": str, + "class_tokens": str, + "cache_info": bool, + } + DB_SUBSET_DISTINCT_SCHEMA = { + Required("image_dir"): str, + "is_reg": bool, + "alpha_mask": bool, + } + # FT means FineTuning + FT_SUBSET_DISTINCT_SCHEMA = { + Required("metadata_file"): str, + "image_dir": str, + "alpha_mask": bool, + } + CN_SUBSET_ASCENDABLE_SCHEMA = { + "caption_extension": str, + "cache_info": bool, + } + CN_SUBSET_DISTINCT_SCHEMA = { + Required("image_dir"): str, + Required("conditioning_data_dir"): str, + } + + # datasets schema + DATASET_ASCENDABLE_SCHEMA = { + "batch_size": int, + "bucket_no_upscale": bool, + "bucket_reso_steps": int, + "enable_bucket": bool, + "max_bucket_reso": int, + "min_bucket_reso": int, + "resolution": functools.partial(__validate_and_convert_scalar_or_twodim.__func__, int), + "network_multiplier": float, + } + + # options handled by argparse but not handled by user config + ARGPARSE_SPECIFIC_SCHEMA = { + "debug_dataset": bool, + "max_token_length": Any(None, int), + "prior_loss_weight": Any(float, int), + } + # for handling default None value of argparse + ARGPARSE_NULLABLE_OPTNAMES = [ + "face_crop_aug_range", + "resolution", + ] + # prepare map because option name may differ among argparse and user config + ARGPARSE_OPTNAME_TO_CONFIG_OPTNAME = { + "train_batch_size": "batch_size", + "dataset_repeats": "num_repeats", + } + + def __init__(self, support_dreambooth: bool, support_finetuning: bool, support_controlnet: bool, support_dropout: bool) -> None: + assert support_dreambooth or support_finetuning or support_controlnet, ( + "Neither DreamBooth mode nor fine tuning mode nor controlnet mode specified. Please specify one mode or more." + + " / DreamBooth モードか fine tuning モードか controlnet モードのどれも指定されていません。1つ以上指定してください。" + ) + + self.db_subset_schema = self.__merge_dict( + self.SUBSET_ASCENDABLE_SCHEMA, + self.DB_SUBSET_DISTINCT_SCHEMA, + self.DB_SUBSET_ASCENDABLE_SCHEMA, + self.DO_SUBSET_ASCENDABLE_SCHEMA if support_dropout else {}, + ) + + self.ft_subset_schema = self.__merge_dict( + self.SUBSET_ASCENDABLE_SCHEMA, + self.FT_SUBSET_DISTINCT_SCHEMA, + self.DO_SUBSET_ASCENDABLE_SCHEMA if support_dropout else {}, + ) + + self.cn_subset_schema = self.__merge_dict( + self.SUBSET_ASCENDABLE_SCHEMA, + self.CN_SUBSET_DISTINCT_SCHEMA, + self.CN_SUBSET_ASCENDABLE_SCHEMA, + self.DO_SUBSET_ASCENDABLE_SCHEMA if support_dropout else {}, + ) + + self.db_dataset_schema = self.__merge_dict( + self.DATASET_ASCENDABLE_SCHEMA, + self.SUBSET_ASCENDABLE_SCHEMA, + self.DB_SUBSET_ASCENDABLE_SCHEMA, + self.DO_SUBSET_ASCENDABLE_SCHEMA if support_dropout else {}, + {"subsets": [self.db_subset_schema]}, + ) + + self.ft_dataset_schema = self.__merge_dict( + self.DATASET_ASCENDABLE_SCHEMA, + self.SUBSET_ASCENDABLE_SCHEMA, + self.DO_SUBSET_ASCENDABLE_SCHEMA if support_dropout else {}, + {"subsets": [self.ft_subset_schema]}, + ) + + self.cn_dataset_schema = self.__merge_dict( + self.DATASET_ASCENDABLE_SCHEMA, + self.SUBSET_ASCENDABLE_SCHEMA, + self.CN_SUBSET_ASCENDABLE_SCHEMA, + self.DO_SUBSET_ASCENDABLE_SCHEMA if support_dropout else {}, + {"subsets": [self.cn_subset_schema]}, + ) + + if support_dreambooth and support_finetuning: + + def validate_flex_dataset(dataset_config: dict): + subsets_config = dataset_config.get("subsets", []) + + if support_controlnet and all(["conditioning_data_dir" in subset for subset in subsets_config]): + return Schema(self.cn_dataset_schema)(dataset_config) + # check dataset meets FT style + # NOTE: all FT subsets should have "metadata_file" + elif all(["metadata_file" in subset for subset in subsets_config]): + return Schema(self.ft_dataset_schema)(dataset_config) + # check dataset meets DB style + # NOTE: all DB subsets should have no "metadata_file" + elif all(["metadata_file" not in subset for subset in subsets_config]): + return Schema(self.db_dataset_schema)(dataset_config) + else: + raise voluptuous.Invalid( + "DreamBooth subset and fine tuning subset cannot be mixed in the same dataset. Please split them into separate datasets. / DreamBoothのサブセットとfine tuninのサブセットを同一のデータセットに混在させることはできません。別々のデータセットに分割してください。" + ) + + self.dataset_schema = validate_flex_dataset + elif support_dreambooth: + if support_controlnet: + self.dataset_schema = self.cn_dataset_schema + else: + self.dataset_schema = self.db_dataset_schema + elif support_finetuning: + self.dataset_schema = self.ft_dataset_schema + elif support_controlnet: + self.dataset_schema = self.cn_dataset_schema + + self.general_schema = self.__merge_dict( + self.DATASET_ASCENDABLE_SCHEMA, + self.SUBSET_ASCENDABLE_SCHEMA, + self.DB_SUBSET_ASCENDABLE_SCHEMA if support_dreambooth else {}, + self.CN_SUBSET_ASCENDABLE_SCHEMA if support_controlnet else {}, + self.DO_SUBSET_ASCENDABLE_SCHEMA if support_dropout else {}, + ) + + self.user_config_validator = Schema( + { + "general": self.general_schema, + "datasets": [self.dataset_schema], + } + ) + + self.argparse_schema = self.__merge_dict( + self.general_schema, + self.ARGPARSE_SPECIFIC_SCHEMA, + {optname: Any(None, self.general_schema[optname]) for optname in self.ARGPARSE_NULLABLE_OPTNAMES}, + {a_name: self.general_schema[c_name] for a_name, c_name in self.ARGPARSE_OPTNAME_TO_CONFIG_OPTNAME.items()}, + ) + + self.argparse_config_validator = Schema(Object(self.argparse_schema), extra=voluptuous.ALLOW_EXTRA) + + def sanitize_user_config(self, user_config: dict) -> dict: + try: + return self.user_config_validator(user_config) + except MultipleInvalid: + # TODO: エラー発生時のメッセージをわかりやすくする + logger.error("Invalid user config / ユーザ設定の形式が正しくないようです") + raise + + # NOTE: In nature, argument parser result is not needed to be sanitize + # However this will help us to detect program bug + def sanitize_argparse_namespace(self, argparse_namespace: argparse.Namespace) -> argparse.Namespace: + try: + return self.argparse_config_validator(argparse_namespace) + except MultipleInvalid: + # XXX: this should be a bug + logger.error( + "Invalid cmdline parsed arguments. This should be a bug. / コマンドラインのパース結果が正しくないようです。プログラムのバグの可能性が高いです。" + ) + raise + + # NOTE: value would be overwritten by latter dict if there is already the same key + @staticmethod + def __merge_dict(*dict_list: dict) -> dict: + merged = {} + for schema in dict_list: + # merged |= schema + for k, v in schema.items(): + merged[k] = v + return merged + + +class BlueprintGenerator: + BLUEPRINT_PARAM_NAME_TO_CONFIG_OPTNAME = {} + + def __init__(self, sanitizer: ConfigSanitizer): + self.sanitizer = sanitizer + + # runtime_params is for parameters which is only configurable on runtime, such as tokenizer + def generate(self, user_config: dict, argparse_namespace: argparse.Namespace, **runtime_params) -> Blueprint: + sanitized_user_config = self.sanitizer.sanitize_user_config(user_config) + sanitized_argparse_namespace = self.sanitizer.sanitize_argparse_namespace(argparse_namespace) + + # convert argparse namespace to dict like config + # NOTE: it is ok to have extra entries in dict + optname_map = self.sanitizer.ARGPARSE_OPTNAME_TO_CONFIG_OPTNAME + argparse_config = { + optname_map.get(optname, optname): value for optname, value in vars(sanitized_argparse_namespace).items() + } + + general_config = sanitized_user_config.get("general", {}) + + dataset_blueprints = [] + for dataset_config in sanitized_user_config.get("datasets", []): + # NOTE: if subsets have no "metadata_file", these are DreamBooth datasets/subsets + subsets = dataset_config.get("subsets", []) + is_dreambooth = all(["metadata_file" not in subset for subset in subsets]) + is_controlnet = all(["conditioning_data_dir" in subset for subset in subsets]) + if is_controlnet: + subset_params_klass = ControlNetSubsetParams + dataset_params_klass = ControlNetDatasetParams + elif is_dreambooth: + subset_params_klass = DreamBoothSubsetParams + dataset_params_klass = DreamBoothDatasetParams + else: + subset_params_klass = FineTuningSubsetParams + dataset_params_klass = FineTuningDatasetParams + + subset_blueprints = [] + for subset_config in subsets: + params = self.generate_params_by_fallbacks( + subset_params_klass, [subset_config, dataset_config, general_config, argparse_config, runtime_params] + ) + subset_blueprints.append(SubsetBlueprint(params)) + + params = self.generate_params_by_fallbacks( + dataset_params_klass, [dataset_config, general_config, argparse_config, runtime_params] + ) + dataset_blueprints.append(DatasetBlueprint(is_dreambooth, is_controlnet, params, subset_blueprints)) + + dataset_group_blueprint = DatasetGroupBlueprint(dataset_blueprints) + + return Blueprint(dataset_group_blueprint) + + @staticmethod + def generate_params_by_fallbacks(param_klass, fallbacks: Sequence[dict]): + name_map = BlueprintGenerator.BLUEPRINT_PARAM_NAME_TO_CONFIG_OPTNAME + search_value = BlueprintGenerator.search_value + default_params = asdict(param_klass()) + param_names = default_params.keys() + + params = {name: search_value(name_map.get(name, name), fallbacks, default_params.get(name)) for name in param_names} + + return param_klass(**params) + + @staticmethod + def search_value(key: str, fallbacks: Sequence[dict], default_value=None): + for cand in fallbacks: + value = cand.get(key) + if value is not None: + return value + + return default_value + + +def generate_dataset_group_by_blueprint(dataset_group_blueprint: DatasetGroupBlueprint): + datasets: List[Union[DreamBoothDataset, FineTuningDataset, ControlNetDataset]] = [] + + for dataset_blueprint in dataset_group_blueprint.datasets: + if dataset_blueprint.is_controlnet: + subset_klass = ControlNetSubset + dataset_klass = ControlNetDataset + elif dataset_blueprint.is_dreambooth: + subset_klass = DreamBoothSubset + dataset_klass = DreamBoothDataset + else: + subset_klass = FineTuningSubset + dataset_klass = FineTuningDataset + + subsets = [subset_klass(**asdict(subset_blueprint.params)) for subset_blueprint in dataset_blueprint.subsets] + dataset = dataset_klass(subsets=subsets, **asdict(dataset_blueprint.params)) + datasets.append(dataset) + + # print info + info = "" + for i, dataset in enumerate(datasets): + is_dreambooth = isinstance(dataset, DreamBoothDataset) + is_controlnet = isinstance(dataset, ControlNetDataset) + info += dedent( + f"""\ + [Dataset {i}] + batch_size: {dataset.batch_size} + resolution: {(dataset.width, dataset.height)} + enable_bucket: {dataset.enable_bucket} + network_multiplier: {dataset.network_multiplier} + """ + ) + + if dataset.enable_bucket: + info += indent( + dedent( + f"""\ + min_bucket_reso: {dataset.min_bucket_reso} + max_bucket_reso: {dataset.max_bucket_reso} + bucket_reso_steps: {dataset.bucket_reso_steps} + bucket_no_upscale: {dataset.bucket_no_upscale} + \n""" + ), + " ", + ) + else: + info += "\n" + + for j, subset in enumerate(dataset.subsets): + info += indent( + dedent( + f"""\ + [Subset {j} of Dataset {i}] + image_dir: "{subset.image_dir}" + image_count: {subset.img_count} + num_repeats: {subset.num_repeats} + shuffle_caption: {subset.shuffle_caption} + keep_tokens: {subset.keep_tokens} + keep_tokens_separator: {subset.keep_tokens_separator} + caption_separator: {subset.caption_separator} + secondary_separator: {subset.secondary_separator} + enable_wildcard: {subset.enable_wildcard} + caption_dropout_rate: {subset.caption_dropout_rate} + caption_dropout_every_n_epoches: {subset.caption_dropout_every_n_epochs} + caption_tag_dropout_rate: {subset.caption_tag_dropout_rate} + caption_prefix: {subset.caption_prefix} + caption_suffix: {subset.caption_suffix} + color_aug: {subset.color_aug} + flip_aug: {subset.flip_aug} + face_crop_aug_range: {subset.face_crop_aug_range} + random_crop: {subset.random_crop} + token_warmup_min: {subset.token_warmup_min}, + token_warmup_step: {subset.token_warmup_step}, + alpha_mask: {subset.alpha_mask}, + """ + ), + " ", + ) + + if is_dreambooth: + info += indent( + dedent( + f"""\ + is_reg: {subset.is_reg} + class_tokens: {subset.class_tokens} + caption_extension: {subset.caption_extension} + \n""" + ), + " ", + ) + elif not is_controlnet: + info += indent( + dedent( + f"""\ + metadata_file: {subset.metadata_file} + \n""" + ), + " ", + ) + + logger.info(f"{info}") + + # make buckets first because it determines the length of dataset + # and set the same seed for all datasets + seed = random.randint(0, 2**31) # actual seed is seed + epoch_no + for i, dataset in enumerate(datasets): + logger.info(f"[Dataset {i}]") + dataset.make_buckets() + dataset.set_seed(seed) + + return DatasetGroup(datasets) + + +def generate_dreambooth_subsets_config_by_subdirs(train_data_dir: Optional[str] = None, reg_data_dir: Optional[str] = None): + def extract_dreambooth_params(name: str) -> Tuple[int, str]: + tokens = name.split("_") + try: + n_repeats = int(tokens[0]) + except ValueError as e: + logger.warning(f"ignore directory without repeats / 繰り返し回数のないディレクトリを無視します: {name}") + return 0, "" + caption_by_folder = "_".join(tokens[1:]) + return n_repeats, caption_by_folder + + def generate(base_dir: Optional[str], is_reg: bool): + if base_dir is None: + return [] + + base_dir: Path = Path(base_dir) + if not base_dir.is_dir(): + return [] + + subsets_config = [] + for subdir in base_dir.iterdir(): + if not subdir.is_dir(): + continue + + num_repeats, class_tokens = extract_dreambooth_params(subdir.name) + if num_repeats < 1: + continue + + subset_config = {"image_dir": str(subdir), "num_repeats": num_repeats, "is_reg": is_reg, "class_tokens": class_tokens} + subsets_config.append(subset_config) + + return subsets_config + + subsets_config = [] + subsets_config += generate(train_data_dir, False) + subsets_config += generate(reg_data_dir, True) + + return subsets_config + + +def generate_controlnet_subsets_config_by_subdirs( + train_data_dir: Optional[str] = None, conditioning_data_dir: Optional[str] = None, caption_extension: str = ".txt" +): + def generate(base_dir: Optional[str]): + if base_dir is None: + return [] + + base_dir: Path = Path(base_dir) + if not base_dir.is_dir(): + return [] + + subsets_config = [] + subset_config = { + "image_dir": train_data_dir, + "conditioning_data_dir": conditioning_data_dir, + "caption_extension": caption_extension, + "num_repeats": 1, + } + subsets_config.append(subset_config) + + return subsets_config + + subsets_config = [] + subsets_config += generate(train_data_dir) + + return subsets_config + + +def load_user_config(file: str) -> dict: + file: Path = Path(file) + if not file.is_file(): + raise ValueError(f"file not found / ファイルが見つかりません: {file}") + + if file.name.lower().endswith(".json"): + try: + with open(file, "r") as f: + config = json.load(f) + except Exception: + logger.error( + f"Error on parsing JSON config file. Please check the format. / JSON 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}" + ) + raise + elif file.name.lower().endswith(".toml"): + try: + config = toml.load(file) + except Exception: + logger.error( + f"Error on parsing TOML config file. Please check the format. / TOML 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}" + ) + raise + else: + raise ValueError(f"not supported config file format / 対応していない設定ファイルの形式です: {file}") + + return config + + +# for config test +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--support_dreambooth", action="store_true") + parser.add_argument("--support_finetuning", action="store_true") + parser.add_argument("--support_controlnet", action="store_true") + parser.add_argument("--support_dropout", action="store_true") + parser.add_argument("dataset_config") + config_args, remain = parser.parse_known_args() + + parser = argparse.ArgumentParser() + train_util.add_dataset_arguments( + parser, config_args.support_dreambooth, config_args.support_finetuning, config_args.support_dropout + ) + train_util.add_training_arguments(parser, config_args.support_dreambooth) + argparse_namespace = parser.parse_args(remain) + train_util.prepare_dataset_args(argparse_namespace, config_args.support_finetuning) + + logger.info("[argparse_namespace]") + logger.info(f"{vars(argparse_namespace)}") + + user_config = load_user_config(config_args.dataset_config) + + logger.info("") + logger.info("[user_config]") + logger.info(f"{user_config}") + + sanitizer = ConfigSanitizer( + config_args.support_dreambooth, config_args.support_finetuning, config_args.support_controlnet, config_args.support_dropout + ) + sanitized_user_config = sanitizer.sanitize_user_config(user_config) + + logger.info("") + logger.info("[sanitized_user_config]") + logger.info(f"{sanitized_user_config}") + + blueprint = BlueprintGenerator(sanitizer).generate(user_config, argparse_namespace) + + logger.info("") + logger.info("[blueprint]") + logger.info(f"{blueprint}") diff --git a/control_net_lllite.py b/control_net_lllite.py new file mode 100644 index 0000000000000000000000000000000000000000..c9377bee89c5ead73d5f4334503f609123953839 --- /dev/null +++ b/control_net_lllite.py @@ -0,0 +1,449 @@ +import os +from typing import Optional, List, Type +import torch +from library import sdxl_original_unet +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +# input_blocksに適用するかどうか / if True, input_blocks are not applied +SKIP_INPUT_BLOCKS = False + +# output_blocksに適用するかどうか / if True, output_blocks are not applied +SKIP_OUTPUT_BLOCKS = True + +# conv2dに適用するかどうか / if True, conv2d are not applied +SKIP_CONV2D = False + +# transformer_blocksのみに適用するかどうか。Trueの場合、ResBlockには適用されない +# if True, only transformer_blocks are applied, and ResBlocks are not applied +TRANSFORMER_ONLY = True # if True, SKIP_CONV2D is ignored because conv2d is not used in transformer_blocks + +# Trueならattn1とattn2にのみ適用し、ffなどには適用しない / if True, apply only to attn1 and attn2, not to ff etc. +ATTN1_2_ONLY = True + +# Trueならattn1のQKV、attn2のQにのみ適用する、ATTN1_2_ONLY指定時のみ有効 / if True, apply only to attn1 QKV and attn2 Q, only valid when ATTN1_2_ONLY is specified +ATTN_QKV_ONLY = True + +# Trueならattn1やffなどにのみ適用し、attn2などには適用しない / if True, apply only to attn1 and ff, not to attn2 +# ATTN1_2_ONLYと同時にTrueにできない / cannot be True at the same time as ATTN1_2_ONLY +ATTN1_ETC_ONLY = False # True + +# transformer_blocksの最大インデックス。Noneなら全てのtransformer_blocksに適用 +# max index of transformer_blocks. if None, apply to all transformer_blocks +TRANSFORMER_MAX_BLOCK_INDEX = None + + +class LLLiteModule(torch.nn.Module): + def __init__(self, depth, cond_emb_dim, name, org_module, mlp_dim, dropout=None, multiplier=1.0): + super().__init__() + + self.is_conv2d = org_module.__class__.__name__ == "Conv2d" + self.lllite_name = name + self.cond_emb_dim = cond_emb_dim + self.org_module = [org_module] + self.dropout = dropout + self.multiplier = multiplier + + if self.is_conv2d: + in_dim = org_module.in_channels + else: + in_dim = org_module.in_features + + # conditioning1はconditioning imageを embedding する。timestepごとに呼ばれない + # conditioning1 embeds conditioning image. it is not called for each timestep + modules = [] + modules.append(torch.nn.Conv2d(3, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) # to latent (from VAE) size + if depth == 1: + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0)) + elif depth == 2: + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=4, stride=4, padding=0)) + elif depth == 3: + # kernel size 8は大きすぎるので、4にする / kernel size 8 is too large, so set it to 4 + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0)) + + self.conditioning1 = torch.nn.Sequential(*modules) + + # downで入力の次元数を削減する。LoRAにヒントを得ていることにする + # midでconditioning image embeddingと入力を結合する + # upで元の次元数に戻す + # これらはtimestepごとに呼ばれる + # reduce the number of input dimensions with down. inspired by LoRA + # combine conditioning image embedding and input with mid + # restore to the original dimension with up + # these are called for each timestep + + if self.is_conv2d: + self.down = torch.nn.Sequential( + torch.nn.Conv2d(in_dim, mlp_dim, kernel_size=1, stride=1, padding=0), + torch.nn.ReLU(inplace=True), + ) + self.mid = torch.nn.Sequential( + torch.nn.Conv2d(mlp_dim + cond_emb_dim, mlp_dim, kernel_size=1, stride=1, padding=0), + torch.nn.ReLU(inplace=True), + ) + self.up = torch.nn.Sequential( + torch.nn.Conv2d(mlp_dim, in_dim, kernel_size=1, stride=1, padding=0), + ) + else: + # midの前にconditioningをreshapeすること / reshape conditioning before mid + self.down = torch.nn.Sequential( + torch.nn.Linear(in_dim, mlp_dim), + torch.nn.ReLU(inplace=True), + ) + self.mid = torch.nn.Sequential( + torch.nn.Linear(mlp_dim + cond_emb_dim, mlp_dim), + torch.nn.ReLU(inplace=True), + ) + self.up = torch.nn.Sequential( + torch.nn.Linear(mlp_dim, in_dim), + ) + + # Zero-Convにする / set to Zero-Conv + torch.nn.init.zeros_(self.up[0].weight) # zero conv + + self.depth = depth # 1~3 + self.cond_emb = None + self.batch_cond_only = False # Trueなら推論時のcondにのみ適用する / if True, apply only to cond at inference + self.use_zeros_for_batch_uncond = False # Trueならuncondのconditioningを0にする / if True, set uncond conditioning to 0 + + # batch_cond_onlyとuse_zeros_for_batch_uncondはどちらも適用すると生成画像の色味がおかしくなるので実際には使えそうにない + # Controlの種類によっては使えるかも + # both batch_cond_only and use_zeros_for_batch_uncond make the color of the generated image strange, so it doesn't seem to be usable in practice + # it may be available depending on the type of Control + + def set_cond_image(self, cond_image): + r""" + 中でモデルを呼び出すので必要ならwith torch.no_grad()で囲む + / call the model inside, so if necessary, surround it with torch.no_grad() + """ + if cond_image is None: + self.cond_emb = None + return + + # timestepごとに呼ばれないので、あらかじめ計算しておく / it is not called for each timestep, so calculate it in advance + # logger.info(f"C {self.lllite_name}, cond_image.shape={cond_image.shape}") + cx = self.conditioning1(cond_image) + if not self.is_conv2d: + # reshape / b,c,h,w -> b,h*w,c + n, c, h, w = cx.shape + cx = cx.view(n, c, h * w).permute(0, 2, 1) + self.cond_emb = cx + + def set_batch_cond_only(self, cond_only, zeros): + self.batch_cond_only = cond_only + self.use_zeros_for_batch_uncond = zeros + + def apply_to(self): + self.org_forward = self.org_module[0].forward + self.org_module[0].forward = self.forward + + def forward(self, x): + r""" + 学習用の便利forward。元のモジュールのforwardを呼び出す + / convenient forward for training. call the forward of the original module + """ + if self.multiplier == 0.0 or self.cond_emb is None: + return self.org_forward(x) + + cx = self.cond_emb + + if not self.batch_cond_only and x.shape[0] // 2 == cx.shape[0]: # inference only + cx = cx.repeat(2, 1, 1, 1) if self.is_conv2d else cx.repeat(2, 1, 1) + if self.use_zeros_for_batch_uncond: + cx[0::2] = 0.0 # uncond is zero + # logger.info(f"C {self.lllite_name}, x.shape={x.shape}, cx.shape={cx.shape}") + + # downで入力の次元数を削減し、conditioning image embeddingと結合する + # 加算ではなくchannel方向に結合することで、うまいこと混ぜてくれることを期待している + # down reduces the number of input dimensions and combines it with conditioning image embedding + # we expect that it will mix well by combining in the channel direction instead of adding + + cx = torch.cat([cx, self.down(x if not self.batch_cond_only else x[1::2])], dim=1 if self.is_conv2d else 2) + cx = self.mid(cx) + + if self.dropout is not None and self.training: + cx = torch.nn.functional.dropout(cx, p=self.dropout) + + cx = self.up(cx) * self.multiplier + + # residual (x) を加算して元のforwardを呼び出す / add residual (x) and call the original forward + if self.batch_cond_only: + zx = torch.zeros_like(x) + zx[1::2] += cx + cx = zx + + x = self.org_forward(x + cx) # ここで元のモジュールを呼び出す / call the original module here + return x + + +class ControlNetLLLite(torch.nn.Module): + UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"] + UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + + def __init__( + self, + unet: sdxl_original_unet.SdxlUNet2DConditionModel, + cond_emb_dim: int = 16, + mlp_dim: int = 16, + dropout: Optional[float] = None, + varbose: Optional[bool] = False, + multiplier: Optional[float] = 1.0, + ) -> None: + super().__init__() + # self.unets = [unet] + + def create_modules( + root_module: torch.nn.Module, + target_replace_modules: List[torch.nn.Module], + module_class: Type[object], + ) -> List[torch.nn.Module]: + prefix = "lllite_unet" + + modules = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + is_linear = child_module.__class__.__name__ == "Linear" + is_conv2d = child_module.__class__.__name__ == "Conv2d" + + if is_linear or (is_conv2d and not SKIP_CONV2D): + # block indexからdepthを計算: depthはconditioningのサイズやチャネルを計算するのに使う + # block index to depth: depth is using to calculate conditioning size and channels + block_name, index1, index2 = (name + "." + child_name).split(".")[:3] + index1 = int(index1) + if block_name == "input_blocks": + if SKIP_INPUT_BLOCKS: + continue + depth = 1 if index1 <= 2 else (2 if index1 <= 5 else 3) + elif block_name == "middle_block": + depth = 3 + elif block_name == "output_blocks": + if SKIP_OUTPUT_BLOCKS: + continue + depth = 3 if index1 <= 2 else (2 if index1 <= 5 else 1) + if int(index2) >= 2: + depth -= 1 + else: + raise NotImplementedError() + + lllite_name = prefix + "." + name + "." + child_name + lllite_name = lllite_name.replace(".", "_") + + if TRANSFORMER_MAX_BLOCK_INDEX is not None: + p = lllite_name.find("transformer_blocks") + if p >= 0: + tf_index = int(lllite_name[p:].split("_")[2]) + if tf_index > TRANSFORMER_MAX_BLOCK_INDEX: + continue + + # time embは適用外とする + # attn2のconditioning (CLIPからの入力) はshapeが違うので適用できない + # time emb is not applied + # attn2 conditioning (input from CLIP) cannot be applied because the shape is different + if "emb_layers" in lllite_name or ( + "attn2" in lllite_name and ("to_k" in lllite_name or "to_v" in lllite_name) + ): + continue + + if ATTN1_2_ONLY: + if not ("attn1" in lllite_name or "attn2" in lllite_name): + continue + if ATTN_QKV_ONLY: + if "to_out" in lllite_name: + continue + + if ATTN1_ETC_ONLY: + if "proj_out" in lllite_name: + pass + elif "attn1" in lllite_name and ( + "to_k" in lllite_name or "to_v" in lllite_name or "to_out" in lllite_name + ): + pass + elif "ff_net_2" in lllite_name: + pass + else: + continue + + module = module_class( + depth, + cond_emb_dim, + lllite_name, + child_module, + mlp_dim, + dropout=dropout, + multiplier=multiplier, + ) + modules.append(module) + return modules + + target_modules = ControlNetLLLite.UNET_TARGET_REPLACE_MODULE + if not TRANSFORMER_ONLY: + target_modules = target_modules + ControlNetLLLite.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + + # create module instances + self.unet_modules: List[LLLiteModule] = create_modules(unet, target_modules, LLLiteModule) + logger.info(f"create ControlNet LLLite for U-Net: {len(self.unet_modules)} modules.") + + def forward(self, x): + return x # dummy + + def set_cond_image(self, cond_image): + r""" + 中でモデルを呼び出すので必要ならwith torch.no_grad()で囲む + / call the model inside, so if necessary, surround it with torch.no_grad() + """ + for module in self.unet_modules: + module.set_cond_image(cond_image) + + def set_batch_cond_only(self, cond_only, zeros): + for module in self.unet_modules: + module.set_batch_cond_only(cond_only, zeros) + + def set_multiplier(self, multiplier): + for module in self.unet_modules: + module.multiplier = multiplier + + def load_weights(self, file): + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + info = self.load_state_dict(weights_sd, False) + return info + + def apply_to(self): + logger.info("applying LLLite for U-Net...") + for module in self.unet_modules: + module.apply_to() + self.add_module(module.lllite_name, module) + + # マージできるかどうかを返す + def is_mergeable(self): + return False + + def merge_to(self, text_encoder, unet, weights_sd, dtype, device): + raise NotImplementedError() + + def enable_gradient_checkpointing(self): + # not supported + pass + + def prepare_optimizer_params(self): + self.requires_grad_(True) + return self.parameters() + + def prepare_grad_etc(self): + self.requires_grad_(True) + + def on_epoch_start(self): + self.train() + + def get_trainable_params(self): + return self.parameters() + + def save_weights(self, file, dtype, metadata): + if metadata is not None and len(metadata) == 0: + metadata = None + + state_dict = self.state_dict() + + if dtype is not None: + for key in list(state_dict.keys()): + v = state_dict[key] + v = v.detach().clone().to("cpu").to(dtype) + state_dict[key] = v + + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import save_file + + save_file(state_dict, file, metadata) + else: + torch.save(state_dict, file) + + +if __name__ == "__main__": + # デバッグ用 / for debug + + # sdxl_original_unet.USE_REENTRANT = False + + # test shape etc + logger.info("create unet") + unet = sdxl_original_unet.SdxlUNet2DConditionModel() + unet.to("cuda").to(torch.float16) + + logger.info("create ControlNet-LLLite") + control_net = ControlNetLLLite(unet, 32, 64) + control_net.apply_to() + control_net.to("cuda") + + logger.info(control_net) + + # logger.info number of parameters + logger.info(f"number of parameters {sum(p.numel() for p in control_net.parameters() if p.requires_grad)}") + + input() + + unet.set_use_memory_efficient_attention(True, False) + unet.set_gradient_checkpointing(True) + unet.train() # for gradient checkpointing + + control_net.train() + + # # visualize + # import torchviz + # logger.info("run visualize") + # controlnet.set_control(conditioning_image) + # output = unet(x, t, ctx, y) + # logger.info("make_dot") + # image = torchviz.make_dot(output, params=dict(controlnet.named_parameters())) + # logger.info("render") + # image.format = "svg" # "png" + # image.render("NeuralNet") # すごく時間がかかるので注意 / be careful because it takes a long time + # input() + + import bitsandbytes + + optimizer = bitsandbytes.adam.Adam8bit(control_net.prepare_optimizer_params(), 1e-3) + + scaler = torch.cuda.amp.GradScaler(enabled=True) + + logger.info("start training") + steps = 10 + + sample_param = [p for p in control_net.named_parameters() if "up" in p[0]][0] + for step in range(steps): + logger.info(f"step {step}") + + batch_size = 1 + conditioning_image = torch.rand(batch_size, 3, 1024, 1024).cuda() * 2.0 - 1.0 + x = torch.randn(batch_size, 4, 128, 128).cuda() + t = torch.randint(low=0, high=10, size=(batch_size,)).cuda() + ctx = torch.randn(batch_size, 77, 2048).cuda() + y = torch.randn(batch_size, sdxl_original_unet.ADM_IN_CHANNELS).cuda() + + with torch.cuda.amp.autocast(enabled=True): + control_net.set_cond_image(conditioning_image) + + output = unet(x, t, ctx, y) + target = torch.randn_like(output) + loss = torch.nn.functional.mse_loss(output, target) + + scaler.scale(loss).backward() + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad(set_to_none=True) + logger.info(f"{sample_param}") + + # from safetensors.torch import save_file + + # save_file(control_net.state_dict(), "logs/control_net.safetensors") diff --git a/control_net_lllite_for_train.py b/control_net_lllite_for_train.py new file mode 100644 index 0000000000000000000000000000000000000000..366451b7f2e8f2cca2ee252c93ee598f335447cd --- /dev/null +++ b/control_net_lllite_for_train.py @@ -0,0 +1,501 @@ +# cond_imageをU-Netのforwardで渡すバージョンのControlNet-LLLite検証用実装 +# ControlNet-LLLite implementation for verification with cond_image passed in U-Net's forward + +import os +import re +from typing import Optional, List, Type +import torch +from library import sdxl_original_unet +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +# input_blocksに適用するかどうか / if True, input_blocks are not applied +SKIP_INPUT_BLOCKS = False + +# output_blocksに適用するかどうか / if True, output_blocks are not applied +SKIP_OUTPUT_BLOCKS = True + +# conv2dに適用するかどうか / if True, conv2d are not applied +SKIP_CONV2D = False + +# transformer_blocksのみに適用するかどうか。Trueの場合、ResBlockには適用されない +# if True, only transformer_blocks are applied, and ResBlocks are not applied +TRANSFORMER_ONLY = True # if True, SKIP_CONV2D is ignored because conv2d is not used in transformer_blocks + +# Trueならattn1とattn2にのみ適用し、ffなどには適用しない / if True, apply only to attn1 and attn2, not to ff etc. +ATTN1_2_ONLY = True + +# Trueならattn1のQKV、attn2のQにのみ適用する、ATTN1_2_ONLY指定時のみ有効 / if True, apply only to attn1 QKV and attn2 Q, only valid when ATTN1_2_ONLY is specified +ATTN_QKV_ONLY = True + +# Trueならattn1やffなどにのみ適用し、attn2などには適用しない / if True, apply only to attn1 and ff, not to attn2 +# ATTN1_2_ONLYと同時にTrueにできない / cannot be True at the same time as ATTN1_2_ONLY +ATTN1_ETC_ONLY = False # True + +# transformer_blocksの最大インデックス。Noneなら全てのtransformer_blocksに適用 +# max index of transformer_blocks. if None, apply to all transformer_blocks +TRANSFORMER_MAX_BLOCK_INDEX = None + +ORIGINAL_LINEAR = torch.nn.Linear +ORIGINAL_CONV2D = torch.nn.Conv2d + + +def add_lllite_modules(module: torch.nn.Module, in_dim: int, depth, cond_emb_dim, mlp_dim) -> None: + # conditioning1はconditioning imageを embedding する。timestepごとに呼ばれない + # conditioning1 embeds conditioning image. it is not called for each timestep + modules = [] + modules.append(ORIGINAL_CONV2D(3, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) # to latent (from VAE) size + if depth == 1: + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0)) + elif depth == 2: + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim, kernel_size=4, stride=4, padding=0)) + elif depth == 3: + # kernel size 8は大きすぎるので、4にする / kernel size 8 is too large, so set it to 4 + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) + modules.append(torch.nn.ReLU(inplace=True)) + modules.append(ORIGINAL_CONV2D(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0)) + + module.lllite_conditioning1 = torch.nn.Sequential(*modules) + + # downで入力の次元数を削減する。LoRAにヒントを得ていることにする + # midでconditioning image embeddingと入力を結合する + # upで元の次元数に戻す + # これらはtimestepごとに呼ばれる + # reduce the number of input dimensions with down. inspired by LoRA + # combine conditioning image embedding and input with mid + # restore to the original dimension with up + # these are called for each timestep + + module.lllite_down = torch.nn.Sequential( + ORIGINAL_LINEAR(in_dim, mlp_dim), + torch.nn.ReLU(inplace=True), + ) + module.lllite_mid = torch.nn.Sequential( + ORIGINAL_LINEAR(mlp_dim + cond_emb_dim, mlp_dim), + torch.nn.ReLU(inplace=True), + ) + module.lllite_up = torch.nn.Sequential( + ORIGINAL_LINEAR(mlp_dim, in_dim), + ) + + # Zero-Convにする / set to Zero-Conv + torch.nn.init.zeros_(module.lllite_up[0].weight) # zero conv + + +class LLLiteLinear(ORIGINAL_LINEAR): + def __init__(self, in_features: int, out_features: int, **kwargs): + super().__init__(in_features, out_features, **kwargs) + self.enabled = False + + def set_lllite(self, depth, cond_emb_dim, name, mlp_dim, dropout=None, multiplier=1.0): + self.enabled = True + self.lllite_name = name + self.cond_emb_dim = cond_emb_dim + self.dropout = dropout + self.multiplier = multiplier # ignored + + in_dim = self.in_features + add_lllite_modules(self, in_dim, depth, cond_emb_dim, mlp_dim) + + self.cond_image = None + + def set_cond_image(self, cond_image): + self.cond_image = cond_image + + def forward(self, x): + if not self.enabled: + return super().forward(x) + + cx = self.lllite_conditioning1(self.cond_image) # make forward and backward compatible + + # reshape / b,c,h,w -> b,h*w,c + n, c, h, w = cx.shape + cx = cx.view(n, c, h * w).permute(0, 2, 1) + + cx = torch.cat([cx, self.lllite_down(x)], dim=2) + cx = self.lllite_mid(cx) + + if self.dropout is not None and self.training: + cx = torch.nn.functional.dropout(cx, p=self.dropout) + + cx = self.lllite_up(cx) * self.multiplier + + x = super().forward(x + cx) # ここで元のモジュールを呼び出す / call the original module here + return x + + +class LLLiteConv2d(ORIGINAL_CONV2D): + def __init__(self, in_channels: int, out_channels: int, kernel_size, **kwargs): + super().__init__(in_channels, out_channels, kernel_size, **kwargs) + self.enabled = False + + def set_lllite(self, depth, cond_emb_dim, name, mlp_dim, dropout=None, multiplier=1.0): + self.enabled = True + self.lllite_name = name + self.cond_emb_dim = cond_emb_dim + self.dropout = dropout + self.multiplier = multiplier # ignored + + in_dim = self.in_channels + add_lllite_modules(self, in_dim, depth, cond_emb_dim, mlp_dim) + + self.cond_image = None + self.cond_emb = None + + def set_cond_image(self, cond_image): + self.cond_image = cond_image + self.cond_emb = None + + def forward(self, x): # , cond_image=None): + if not self.enabled: + return super().forward(x) + + cx = self.lllite_conditioning1(self.cond_image) + + cx = torch.cat([cx, self.down(x)], dim=1) + cx = self.mid(cx) + + if self.dropout is not None and self.training: + cx = torch.nn.functional.dropout(cx, p=self.dropout) + + cx = self.up(cx) * self.multiplier + + x = super().forward(x + cx) # ここで元のモジュールを呼び出す / call the original module here + return x + + +class SdxlUNet2DConditionModelControlNetLLLite(sdxl_original_unet.SdxlUNet2DConditionModel): + UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"] + UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + LLLITE_PREFIX = "lllite_unet" + + def __init__(self, **kwargs): + super().__init__(**kwargs) + + def apply_lllite( + self, + cond_emb_dim: int = 16, + mlp_dim: int = 16, + dropout: Optional[float] = None, + varbose: Optional[bool] = False, + multiplier: Optional[float] = 1.0, + ) -> None: + def apply_to_modules( + root_module: torch.nn.Module, + target_replace_modules: List[torch.nn.Module], + ) -> List[torch.nn.Module]: + prefix = "lllite_unet" + + modules = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + is_linear = child_module.__class__.__name__ == "LLLiteLinear" + is_conv2d = child_module.__class__.__name__ == "LLLiteConv2d" + + if is_linear or (is_conv2d and not SKIP_CONV2D): + # block indexからdepthを計算: depthはconditioningのサイズやチャネルを計算するのに使う + # block index to depth: depth is using to calculate conditioning size and channels + block_name, index1, index2 = (name + "." + child_name).split(".")[:3] + index1 = int(index1) + if block_name == "input_blocks": + if SKIP_INPUT_BLOCKS: + continue + depth = 1 if index1 <= 2 else (2 if index1 <= 5 else 3) + elif block_name == "middle_block": + depth = 3 + elif block_name == "output_blocks": + if SKIP_OUTPUT_BLOCKS: + continue + depth = 3 if index1 <= 2 else (2 if index1 <= 5 else 1) + if int(index2) >= 2: + depth -= 1 + else: + raise NotImplementedError() + + lllite_name = prefix + "." + name + "." + child_name + lllite_name = lllite_name.replace(".", "_") + + if TRANSFORMER_MAX_BLOCK_INDEX is not None: + p = lllite_name.find("transformer_blocks") + if p >= 0: + tf_index = int(lllite_name[p:].split("_")[2]) + if tf_index > TRANSFORMER_MAX_BLOCK_INDEX: + continue + + # time embは適用外とする + # attn2のconditioning (CLIPからの入力) はshapeが違うので適用できない + # time emb is not applied + # attn2 conditioning (input from CLIP) cannot be applied because the shape is different + if "emb_layers" in lllite_name or ( + "attn2" in lllite_name and ("to_k" in lllite_name or "to_v" in lllite_name) + ): + continue + + if ATTN1_2_ONLY: + if not ("attn1" in lllite_name or "attn2" in lllite_name): + continue + if ATTN_QKV_ONLY: + if "to_out" in lllite_name: + continue + + if ATTN1_ETC_ONLY: + if "proj_out" in lllite_name: + pass + elif "attn1" in lllite_name and ( + "to_k" in lllite_name or "to_v" in lllite_name or "to_out" in lllite_name + ): + pass + elif "ff_net_2" in lllite_name: + pass + else: + continue + + child_module.set_lllite(depth, cond_emb_dim, lllite_name, mlp_dim, dropout, multiplier) + modules.append(child_module) + + return modules + + target_modules = SdxlUNet2DConditionModelControlNetLLLite.UNET_TARGET_REPLACE_MODULE + if not TRANSFORMER_ONLY: + target_modules = target_modules + SdxlUNet2DConditionModelControlNetLLLite.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + + # create module instances + self.lllite_modules = apply_to_modules(self, target_modules) + logger.info(f"enable ControlNet LLLite for U-Net: {len(self.lllite_modules)} modules.") + + # def prepare_optimizer_params(self): + def prepare_params(self): + train_params = [] + non_train_params = [] + for name, p in self.named_parameters(): + if "lllite" in name: + train_params.append(p) + else: + non_train_params.append(p) + logger.info(f"count of trainable parameters: {len(train_params)}") + logger.info(f"count of non-trainable parameters: {len(non_train_params)}") + + for p in non_train_params: + p.requires_grad_(False) + + # without this, an error occurs in the optimizer + # RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn + non_train_params[0].requires_grad_(True) + + for p in train_params: + p.requires_grad_(True) + + return train_params + + # def prepare_grad_etc(self): + # self.requires_grad_(True) + + # def on_epoch_start(self): + # self.train() + + def get_trainable_params(self): + return [p[1] for p in self.named_parameters() if "lllite" in p[0]] + + def save_lllite_weights(self, file, dtype, metadata): + if metadata is not None and len(metadata) == 0: + metadata = None + + org_state_dict = self.state_dict() + + # copy LLLite keys from org_state_dict to state_dict with key conversion + state_dict = {} + for key in org_state_dict.keys(): + # split with ".lllite" + pos = key.find(".lllite") + if pos < 0: + continue + lllite_key = SdxlUNet2DConditionModelControlNetLLLite.LLLITE_PREFIX + "." + key[:pos] + lllite_key = lllite_key.replace(".", "_") + key[pos:] + lllite_key = lllite_key.replace(".lllite_", ".") + state_dict[lllite_key] = org_state_dict[key] + + if dtype is not None: + for key in list(state_dict.keys()): + v = state_dict[key] + v = v.detach().clone().to("cpu").to(dtype) + state_dict[key] = v + + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import save_file + + save_file(state_dict, file, metadata) + else: + torch.save(state_dict, file) + + def load_lllite_weights(self, file, non_lllite_unet_sd=None): + r""" + LLLiteの重みを読み込まない（initされた値を使う）場合はfileにNoneを指定する。 + この場合、non_lllite_unet_sdにはU-Netのstate_dictを指定する。 + + If you do not want to load LLLite weights (use initialized values), specify None for file. + In this case, specify the state_dict of U-Net for non_lllite_unet_sd. + """ + if not file: + state_dict = self.state_dict() + for key in non_lllite_unet_sd: + if key in state_dict: + state_dict[key] = non_lllite_unet_sd[key] + info = self.load_state_dict(state_dict, False) + return info + + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + # module_name = module_name.replace("_block", "@blocks") + # module_name = module_name.replace("_layer", "@layer") + # module_name = module_name.replace("to_", "to@") + # module_name = module_name.replace("time_embed", "time@embed") + # module_name = module_name.replace("label_emb", "label@emb") + # module_name = module_name.replace("skip_connection", "skip@connection") + # module_name = module_name.replace("proj_in", "proj@in") + # module_name = module_name.replace("proj_out", "proj@out") + pattern = re.compile(r"(_block|_layer|to_|time_embed|label_emb|skip_connection|proj_in|proj_out)") + + # convert to lllite with U-Net state dict + state_dict = non_lllite_unet_sd.copy() if non_lllite_unet_sd is not None else {} + for key in weights_sd.keys(): + # split with "." + pos = key.find(".") + if pos < 0: + continue + + module_name = key[:pos] + weight_name = key[pos + 1 :] # exclude "." + module_name = module_name.replace(SdxlUNet2DConditionModelControlNetLLLite.LLLITE_PREFIX + "_", "") + + # これはうまくいかない。逆変換を考えなかった設計が悪い / this does not work well. bad design because I didn't think about inverse conversion + # module_name = module_name.replace("_", ".") + + # ださいけどSDXLのU-Netの "_" を "@" に変換する / ugly but convert "_" of SDXL U-Net to "@" + matches = pattern.findall(module_name) + if matches is not None: + for m in matches: + logger.info(f"{module_name} {m}") + module_name = module_name.replace(m, m.replace("_", "@")) + module_name = module_name.replace("_", ".") + module_name = module_name.replace("@", "_") + + lllite_key = module_name + ".lllite_" + weight_name + + state_dict[lllite_key] = weights_sd[key] + + info = self.load_state_dict(state_dict, False) + return info + + def forward(self, x, timesteps=None, context=None, y=None, cond_image=None, **kwargs): + for m in self.lllite_modules: + m.set_cond_image(cond_image) + return super().forward(x, timesteps, context, y, **kwargs) + + +def replace_unet_linear_and_conv2d(): + logger.info("replace torch.nn.Linear and torch.nn.Conv2d to LLLiteLinear and LLLiteConv2d in U-Net") + sdxl_original_unet.torch.nn.Linear = LLLiteLinear + sdxl_original_unet.torch.nn.Conv2d = LLLiteConv2d + + +if __name__ == "__main__": + # デバッグ用 / for debug + + # sdxl_original_unet.USE_REENTRANT = False + replace_unet_linear_and_conv2d() + + # test shape etc + logger.info("create unet") + unet = SdxlUNet2DConditionModelControlNetLLLite() + + logger.info("enable ControlNet-LLLite") + unet.apply_lllite(32, 64, None, False, 1.0) + unet.to("cuda") # .to(torch.float16) + + # from safetensors.torch import load_file + + # model_sd = load_file(r"E:\Work\SD\Models\sdxl\sd_xl_base_1.0_0.9vae.safetensors") + # unet_sd = {} + + # # copy U-Net keys from unet_state_dict to state_dict + # prefix = "model.diffusion_model." + # for key in model_sd.keys(): + # if key.startswith(prefix): + # converted_key = key[len(prefix) :] + # unet_sd[converted_key] = model_sd[key] + + # info = unet.load_lllite_weights("r:/lllite_from_unet.safetensors", unet_sd) + # logger.info(info) + + # logger.info(unet) + + # logger.info number of parameters + params = unet.prepare_params() + logger.info(f"number of parameters {sum(p.numel() for p in params)}") + # logger.info("type any key to continue") + # input() + + unet.set_use_memory_efficient_attention(True, False) + unet.set_gradient_checkpointing(True) + unet.train() # for gradient checkpointing + + # # visualize + # import torchviz + # logger.info("run visualize") + # controlnet.set_control(conditioning_image) + # output = unet(x, t, ctx, y) + # logger.info("make_dot") + # image = torchviz.make_dot(output, params=dict(controlnet.named_parameters())) + # logger.info("render") + # image.format = "svg" # "png" + # image.render("NeuralNet") # すごく時間がかかるので注意 / be careful because it takes a long time + # input() + + import bitsandbytes + + optimizer = bitsandbytes.adam.Adam8bit(params, 1e-3) + + scaler = torch.cuda.amp.GradScaler(enabled=True) + + logger.info("start training") + steps = 10 + batch_size = 1 + + sample_param = [p for p in unet.named_parameters() if ".lllite_up." in p[0]][0] + for step in range(steps): + logger.info(f"step {step}") + + conditioning_image = torch.rand(batch_size, 3, 1024, 1024).cuda() * 2.0 - 1.0 + x = torch.randn(batch_size, 4, 128, 128).cuda() + t = torch.randint(low=0, high=10, size=(batch_size,)).cuda() + ctx = torch.randn(batch_size, 77, 2048).cuda() + y = torch.randn(batch_size, sdxl_original_unet.ADM_IN_CHANNELS).cuda() + + with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16): + output = unet(x, t, ctx, y, conditioning_image) + target = torch.randn_like(output) + loss = torch.nn.functional.mse_loss(output, target) + + scaler.scale(loss).backward() + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad(set_to_none=True) + logger.info(sample_param) + + # from safetensors.torch import save_file + + # logger.info("save weights") + # unet.save_lllite_weights("r:/lllite_from_unet.safetensors", torch.float16, None) diff --git a/convert_diffusers20_original_sd.py b/convert_diffusers20_original_sd.py new file mode 100644 index 0000000000000000000000000000000000000000..572ee2f0c0c1744fda73f640c83785015e5b5b54 --- /dev/null +++ b/convert_diffusers20_original_sd.py @@ -0,0 +1,163 @@ +# convert Diffusers v1.x/v2.0 model to original Stable Diffusion + +import argparse +import os +import torch +from diffusers import StableDiffusionPipeline + +import library.model_util as model_util +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def convert(args): + # 引数を確認する + load_dtype = torch.float16 if args.fp16 else None + + save_dtype = None + if args.fp16 or args.save_precision_as == "fp16": + save_dtype = torch.float16 + elif args.bf16 or args.save_precision_as == "bf16": + save_dtype = torch.bfloat16 + elif args.float or args.save_precision_as == "float": + save_dtype = torch.float + + is_load_ckpt = os.path.isfile(args.model_to_load) + is_save_ckpt = len(os.path.splitext(args.model_to_save)[1]) > 0 + + assert not is_load_ckpt or args.v1 != args.v2, "v1 or v2 is required to load checkpoint / checkpointの読み込みにはv1/v2指定が必要です" + # assert ( + # is_save_ckpt or args.reference_model is not None + # ), f"reference model is required to save as Diffusers / Diffusers形式での保存には参照モデルが必要です" + + # モデルを読み込む + msg = "checkpoint" if is_load_ckpt else ("Diffusers" + (" as fp16" if args.fp16 else "")) + logger.info(f"loading {msg}: {args.model_to_load}") + + if is_load_ckpt: + v2_model = args.v2 + text_encoder, vae, unet = model_util.load_models_from_stable_diffusion_checkpoint( + v2_model, args.model_to_load, unet_use_linear_projection_in_v2=args.unet_use_linear_projection + ) + else: + pipe = StableDiffusionPipeline.from_pretrained( + args.model_to_load, torch_dtype=load_dtype, tokenizer=None, safety_checker=None, variant=args.variant + ) + text_encoder = pipe.text_encoder + vae = pipe.vae + unet = pipe.unet + + if args.v1 == args.v2: + # 自動判定する + v2_model = unet.config.cross_attention_dim == 1024 + logger.info("checking model version: model is " + ("v2" if v2_model else "v1")) + else: + v2_model = not args.v1 + + # 変換して保存する + msg = ("checkpoint" + ("" if save_dtype is None else f" in {save_dtype}")) if is_save_ckpt else "Diffusers" + logger.info(f"converting and saving as {msg}: {args.model_to_save}") + + if is_save_ckpt: + original_model = args.model_to_load if is_load_ckpt else None + key_count = model_util.save_stable_diffusion_checkpoint( + v2_model, + args.model_to_save, + text_encoder, + unet, + original_model, + args.epoch, + args.global_step, + None if args.metadata is None else eval(args.metadata), + save_dtype=save_dtype, + vae=vae, + ) + logger.info(f"model saved. total converted state_dict keys: {key_count}") + else: + logger.info( + f"copy scheduler/tokenizer config from: {args.reference_model if args.reference_model is not None else 'default model'}" + ) + model_util.save_diffusers_checkpoint( + v2_model, args.model_to_save, text_encoder, unet, args.reference_model, vae, args.use_safetensors + ) + logger.info("model saved.") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument( + "--v1", action="store_true", help="load v1.x model (v1 or v2 is required to load checkpoint) / 1.xのモデルを読み込む" + ) + parser.add_argument( + "--v2", action="store_true", help="load v2.0 model (v1 or v2 is required to load checkpoint) / 2.0のモデルを読み込む" + ) + parser.add_argument( + "--unet_use_linear_projection", + action="store_true", + help="When saving v2 model as Diffusers, set U-Net config to `use_linear_projection=true` (to match stabilityai's model) / Diffusers形式でv2モデルを保存するときにU-Netの設定を`use_linear_projection=true`にする（stabilityaiのモデルと合わせる）", + ) + parser.add_argument( + "--fp16", + action="store_true", + help="load as fp16 (Diffusers only) and save as fp16 (checkpoint only) / fp16形式で読み込み（Diffusers形式のみ対応）、保存する（checkpointのみ対応）", + ) + parser.add_argument("--bf16", action="store_true", help="save as bf16 (checkpoint only) / bf16形式で保存する（checkpointのみ対応）") + parser.add_argument( + "--float", action="store_true", help="save as float (checkpoint only) / float(float32)形式で保存する（checkpointのみ対応）" + ) + parser.add_argument( + "--save_precision_as", + type=str, + default="no", + choices=["fp16", "bf16", "float"], + help="save precision, do not specify with --fp16/--bf16/--float / 保存する精度、--fp16/--bf16/--floatと併用しないでください", + ) + parser.add_argument("--epoch", type=int, default=0, help="epoch to write to checkpoint / checkpointに記録するepoch数の値") + parser.add_argument( + "--global_step", type=int, default=0, help="global_step to write to checkpoint / checkpointに記録するglobal_stepの値" + ) + parser.add_argument( + "--metadata", + type=str, + default=None, + help='モデルに保存されるメタデータ、Pythonの辞書形式で指定 / metadata: metadata written in to the model in Python Dictionary. Example metadata: \'{"name": "model_name", "resolution": "512x512"}\'', + ) + parser.add_argument( + "--variant", + type=str, + default=None, + help="読む込むDiffusersのvariantを指定する、例: fp16 / variant: Diffusers variant to load. Example: fp16", + ) + parser.add_argument( + "--reference_model", + type=str, + default=None, + help="scheduler/tokenizerのコピー元Diffusersモデル、Diffusers形式で保存するときに使用される、省略時は`runwayml/stable-diffusion-v1-5` または `stabilityai/stable-diffusion-2-1` / reference Diffusers model to copy scheduler/tokenizer config from, used when saving as Diffusers format, default is `runwayml/stable-diffusion-v1-5` or `stabilityai/stable-diffusion-2-1`", + ) + parser.add_argument( + "--use_safetensors", + action="store_true", + help="use safetensors format to save Diffusers model (checkpoint depends on the file extension) / Duffusersモデルをsafetensors形式で保存する（checkpointは拡張子で自動判定）", + ) + + parser.add_argument( + "model_to_load", + type=str, + default=None, + help="model to load: checkpoint file or Diffusers model's directory / 読み込むモデル、checkpointかDiffusers形式モデルのディレクトリ", + ) + parser.add_argument( + "model_to_save", + type=str, + default=None, + help="model to save: checkpoint (with extension) or Diffusers model's directory (without extension) / 変換後のモデル、拡張子がある場合はcheckpoint、ない場合はDiffusesモデルとして保存", + ) + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + convert(args) diff --git a/custom_train_functions.py b/custom_train_functions.py new file mode 100644 index 0000000000000000000000000000000000000000..faf443048370b90d196045640280b4e12b4957fa --- /dev/null +++ b/custom_train_functions.py @@ -0,0 +1,559 @@ +import torch +import argparse +import random +import re +from typing import List, Optional, Union +from .utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +def prepare_scheduler_for_custom_training(noise_scheduler, device): + if hasattr(noise_scheduler, "all_snr"): + return + + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod) + sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod) + alpha = sqrt_alphas_cumprod + sigma = sqrt_one_minus_alphas_cumprod + all_snr = (alpha / sigma) ** 2 + + noise_scheduler.all_snr = all_snr.to(device) + + +def fix_noise_scheduler_betas_for_zero_terminal_snr(noise_scheduler): + # fix beta: zero terminal SNR + logger.info(f"fix noise scheduler betas: https://arxiv.org/abs/2305.08891") + + def enforce_zero_terminal_snr(betas): + # Convert betas to alphas_bar_sqrt + alphas = 1 - betas + alphas_bar = alphas.cumprod(0) + alphas_bar_sqrt = alphas_bar.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + # Shift so last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + # Scale so first timestep is back to old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 + alphas = alphas_bar[1:] / alphas_bar[:-1] + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + return betas + + betas = noise_scheduler.betas + betas = enforce_zero_terminal_snr(betas) + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + + # logger.info(f"original: {noise_scheduler.betas}") + # logger.info(f"fixed: {betas}") + + noise_scheduler.betas = betas + noise_scheduler.alphas = alphas + noise_scheduler.alphas_cumprod = alphas_cumprod + + +def apply_snr_weight(loss, timesteps, noise_scheduler, gamma, v_prediction=False): + snr = torch.stack([noise_scheduler.all_snr[t] for t in timesteps]) + min_snr_gamma = torch.minimum(snr, torch.full_like(snr, gamma)) + if v_prediction: + snr_weight = torch.div(min_snr_gamma, snr + 1).float().to(loss.device) + else: + snr_weight = torch.div(min_snr_gamma, snr).float().to(loss.device) + loss = loss * snr_weight + return loss + + +def scale_v_prediction_loss_like_noise_prediction(loss, timesteps, noise_scheduler): + scale = get_snr_scale(timesteps, noise_scheduler) + loss = loss * scale + return loss + + +def get_snr_scale(timesteps, noise_scheduler): + snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps]) # batch_size + snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000) # if timestep is 0, snr_t is inf, so limit it to 1000 + scale = snr_t / (snr_t + 1) + # # show debug info + # logger.info(f"timesteps: {timesteps}, snr_t: {snr_t}, scale: {scale}") + return scale + + +def add_v_prediction_like_loss(loss, timesteps, noise_scheduler, v_pred_like_loss): + scale = get_snr_scale(timesteps, noise_scheduler) + # logger.info(f"add v-prediction like loss: {v_pred_like_loss}, scale: {scale}, loss: {loss}, time: {timesteps}") + loss = loss + loss / scale * v_pred_like_loss + return loss + + +def apply_debiased_estimation(loss, timesteps, noise_scheduler, v_prediction=False): + snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps]) # batch_size + snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000) # if timestep is 0, snr_t is inf, so limit it to 1000 + if v_prediction: + weight = 1 / (snr_t + 1) + else: + weight = 1 / torch.sqrt(snr_t) + loss = weight * loss + return loss + + +# TODO train_utilと分散しているのでどちらかに寄せる + + +def add_custom_train_arguments(parser: argparse.ArgumentParser, support_weighted_captions: bool = True): + parser.add_argument( + "--min_snr_gamma", + type=float, + default=None, + help="gamma for reducing the weight of high loss timesteps. Lower numbers have stronger effect. 5 is recommended by paper. / 低いタイムステップでの高いlossに対して重みを減らすためのgamma値、低いほど効果が強く、論文では5が推奨", + ) + parser.add_argument( + "--scale_v_pred_loss_like_noise_pred", + action="store_true", + help="scale v-prediction loss like noise prediction loss / v-prediction lossをnoise prediction lossと同じようにスケーリングする", + ) + parser.add_argument( + "--v_pred_like_loss", + type=float, + default=None, + help="add v-prediction like loss multiplied by this value / v-prediction lossをこの値をかけたものをlossに加算する", + ) + parser.add_argument( + "--debiased_estimation_loss", + action="store_true", + help="debiased estimation loss / debiased estimation loss", + ) + if support_weighted_captions: + parser.add_argument( + "--weighted_captions", + action="store_true", + default=False, + help="Enable weighted captions in the standard style (token:1.3). No commas inside parens, or shuffle/dropout may break the decoder. / 「[token]」、「(token)」「(token:1.3)」のような重み付きキャプションを有効にする。カンマを括弧内に入れるとシャッフルやdropoutで重みづけがおかしくなるので注意", + ) + + +re_attention = re.compile( + r""" +\\\(| +\\\)| +\\\[| +\\]| +\\\\| +\\| +\(| +\[| +:([+-]?[.\d]+)\)| +\)| +]| +[^\\()\[\]:]+| +: +""", + re.X, +) + + +def parse_prompt_attention(text): + """ + Parses a string with attention tokens and returns a list of pairs: text and its associated weight. + Accepted tokens are: + (abc) - increases attention to abc by a multiplier of 1.1 + (abc:3.12) - increases attention to abc by a multiplier of 3.12 + [abc] - decreases attention to abc by a multiplier of 1.1 + \( - literal character '(' + \[ - literal character '[' + \) - literal character ')' + \] - literal character ']' + \\ - literal character '\' + anything else - just text + >>> parse_prompt_attention('normal text') + [['normal text', 1.0]] + >>> parse_prompt_attention('an (important) word') + [['an ', 1.0], ['important', 1.1], [' word', 1.0]] + >>> parse_prompt_attention('(unbalanced') + [['unbalanced', 1.1]] + >>> parse_prompt_attention('\(literal\]') + [['(literal]', 1.0]] + >>> parse_prompt_attention('(unnecessary)(parens)') + [['unnecessaryparens', 1.1]] + >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).') + [['a ', 1.0], + ['house', 1.5730000000000004], + [' ', 1.1], + ['on', 1.0], + [' a ', 1.1], + ['hill', 0.55], + [', sun, ', 1.1], + ['sky', 1.4641000000000006], + ['.', 1.1]] + """ + + res = [] + round_brackets = [] + square_brackets = [] + + round_bracket_multiplier = 1.1 + square_bracket_multiplier = 1 / 1.1 + + def multiply_range(start_position, multiplier): + for p in range(start_position, len(res)): + res[p][1] *= multiplier + + for m in re_attention.finditer(text): + text = m.group(0) + weight = m.group(1) + + if text.startswith("\\"): + res.append([text[1:], 1.0]) + elif text == "(": + round_brackets.append(len(res)) + elif text == "[": + square_brackets.append(len(res)) + elif weight is not None and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), float(weight)) + elif text == ")" and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), round_bracket_multiplier) + elif text == "]" and len(square_brackets) > 0: + multiply_range(square_brackets.pop(), square_bracket_multiplier) + else: + res.append([text, 1.0]) + + for pos in round_brackets: + multiply_range(pos, round_bracket_multiplier) + + for pos in square_brackets: + multiply_range(pos, square_bracket_multiplier) + + if len(res) == 0: + res = [["", 1.0]] + + # merge runs of identical weights + i = 0 + while i + 1 < len(res): + if res[i][1] == res[i + 1][1]: + res[i][0] += res[i + 1][0] + res.pop(i + 1) + else: + i += 1 + + return res + + +def get_prompts_with_weights(tokenizer, prompt: List[str], max_length: int): + r""" + Tokenize a list of prompts and return its tokens with weights of each token. + + No padding, starting or ending token is included. + """ + tokens = [] + weights = [] + truncated = False + for text in prompt: + texts_and_weights = parse_prompt_attention(text) + text_token = [] + text_weight = [] + for word, weight in texts_and_weights: + # tokenize and discard the starting and the ending token + token = tokenizer(word).input_ids[1:-1] + text_token += token + # copy the weight by length of token + text_weight += [weight] * len(token) + # stop if the text is too long (longer than truncation limit) + if len(text_token) > max_length: + truncated = True + break + # truncate + if len(text_token) > max_length: + truncated = True + text_token = text_token[:max_length] + text_weight = text_weight[:max_length] + tokens.append(text_token) + weights.append(text_weight) + if truncated: + logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples") + return tokens, weights + + +def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, no_boseos_middle=True, chunk_length=77): + r""" + Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length. + """ + max_embeddings_multiples = (max_length - 2) // (chunk_length - 2) + weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length + for i in range(len(tokens)): + tokens[i] = [bos] + tokens[i] + [eos] * (max_length - 1 - len(tokens[i])) + if no_boseos_middle: + weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i])) + else: + w = [] + if len(weights[i]) == 0: + w = [1.0] * weights_length + else: + for j in range(max_embeddings_multiples): + w.append(1.0) # weight for starting token in this chunk + w += weights[i][j * (chunk_length - 2) : min(len(weights[i]), (j + 1) * (chunk_length - 2))] + w.append(1.0) # weight for ending token in this chunk + w += [1.0] * (weights_length - len(w)) + weights[i] = w[:] + + return tokens, weights + + +def get_unweighted_text_embeddings( + tokenizer, + text_encoder, + text_input: torch.Tensor, + chunk_length: int, + clip_skip: int, + eos: int, + pad: int, + no_boseos_middle: Optional[bool] = True, +): + """ + When the length of tokens is a multiple of the capacity of the text encoder, + it should be split into chunks and sent to the text encoder individually. + """ + max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2) + if max_embeddings_multiples > 1: + text_embeddings = [] + for i in range(max_embeddings_multiples): + # extract the i-th chunk + text_input_chunk = text_input[:, i * (chunk_length - 2) : (i + 1) * (chunk_length - 2) + 2].clone() + + # cover the head and the tail by the starting and the ending tokens + text_input_chunk[:, 0] = text_input[0, 0] + if pad == eos: # v1 + text_input_chunk[:, -1] = text_input[0, -1] + else: # v2 + for j in range(len(text_input_chunk)): + if text_input_chunk[j, -1] != eos and text_input_chunk[j, -1] != pad: # 最後に普通の文字がある + text_input_chunk[j, -1] = eos + if text_input_chunk[j, 1] == pad: # BOSだけであとはPAD + text_input_chunk[j, 1] = eos + + if clip_skip is None or clip_skip == 1: + text_embedding = text_encoder(text_input_chunk)[0] + else: + enc_out = text_encoder(text_input_chunk, output_hidden_states=True, return_dict=True) + text_embedding = enc_out["hidden_states"][-clip_skip] + text_embedding = text_encoder.text_model.final_layer_norm(text_embedding) + + if no_boseos_middle: + if i == 0: + # discard the ending token + text_embedding = text_embedding[:, :-1] + elif i == max_embeddings_multiples - 1: + # discard the starting token + text_embedding = text_embedding[:, 1:] + else: + # discard both starting and ending tokens + text_embedding = text_embedding[:, 1:-1] + + text_embeddings.append(text_embedding) + text_embeddings = torch.concat(text_embeddings, axis=1) + else: + if clip_skip is None or clip_skip == 1: + text_embeddings = text_encoder(text_input)[0] + else: + enc_out = text_encoder(text_input, output_hidden_states=True, return_dict=True) + text_embeddings = enc_out["hidden_states"][-clip_skip] + text_embeddings = text_encoder.text_model.final_layer_norm(text_embeddings) + return text_embeddings + + +def get_weighted_text_embeddings( + tokenizer, + text_encoder, + prompt: Union[str, List[str]], + device, + max_embeddings_multiples: Optional[int] = 3, + no_boseos_middle: Optional[bool] = False, + clip_skip=None, +): + r""" + Prompts can be assigned with local weights using brackets. For example, + prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful', + and the embedding tokens corresponding to the words get multiplied by a constant, 1.1. + + Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + no_boseos_middle (`bool`, *optional*, defaults to `False`): + If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and + ending token in each of the chunk in the middle. + skip_parsing (`bool`, *optional*, defaults to `False`): + Skip the parsing of brackets. + skip_weighting (`bool`, *optional*, defaults to `False`): + Skip the weighting. When the parsing is skipped, it is forced True. + """ + max_length = (tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + if isinstance(prompt, str): + prompt = [prompt] + + prompt_tokens, prompt_weights = get_prompts_with_weights(tokenizer, prompt, max_length - 2) + + # round up the longest length of tokens to a multiple of (model_max_length - 2) + max_length = max([len(token) for token in prompt_tokens]) + + max_embeddings_multiples = min( + max_embeddings_multiples, + (max_length - 1) // (tokenizer.model_max_length - 2) + 1, + ) + max_embeddings_multiples = max(1, max_embeddings_multiples) + max_length = (tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + + # pad the length of tokens and weights + bos = tokenizer.bos_token_id + eos = tokenizer.eos_token_id + pad = tokenizer.pad_token_id + prompt_tokens, prompt_weights = pad_tokens_and_weights( + prompt_tokens, + prompt_weights, + max_length, + bos, + eos, + no_boseos_middle=no_boseos_middle, + chunk_length=tokenizer.model_max_length, + ) + prompt_tokens = torch.tensor(prompt_tokens, dtype=torch.long, device=device) + + # get the embeddings + text_embeddings = get_unweighted_text_embeddings( + tokenizer, + text_encoder, + prompt_tokens, + tokenizer.model_max_length, + clip_skip, + eos, + pad, + no_boseos_middle=no_boseos_middle, + ) + prompt_weights = torch.tensor(prompt_weights, dtype=text_embeddings.dtype, device=device) + + # assign weights to the prompts and normalize in the sense of mean + previous_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings = text_embeddings * prompt_weights.unsqueeze(-1) + current_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings = text_embeddings * (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + + return text_embeddings + + +# https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2 +def pyramid_noise_like(noise, device, iterations=6, discount=0.4): + b, c, w, h = noise.shape # EDIT: w and h get over-written, rename for a different variant! + u = torch.nn.Upsample(size=(w, h), mode="bilinear").to(device) + for i in range(iterations): + r = random.random() * 2 + 2 # Rather than always going 2x, + wn, hn = max(1, int(w / (r**i))), max(1, int(h / (r**i))) + noise += u(torch.randn(b, c, wn, hn).to(device)) * discount**i + if wn == 1 or hn == 1: + break # Lowest resolution is 1x1 + return noise / noise.std() # Scaled back to roughly unit variance + + +# https://www.crosslabs.org//blog/diffusion-with-offset-noise +def apply_noise_offset(latents, noise, noise_offset, adaptive_noise_scale): + if noise_offset is None: + return noise + if adaptive_noise_scale is not None: + # latent shape: (batch_size, channels, height, width) + # abs mean value for each channel + latent_mean = torch.abs(latents.mean(dim=(2, 3), keepdim=True)) + + # multiply adaptive noise scale to the mean value and add it to the noise offset + noise_offset = noise_offset + adaptive_noise_scale * latent_mean + noise_offset = torch.clamp(noise_offset, 0.0, None) # in case of adaptive noise scale is negative + + noise = noise + noise_offset * torch.randn((latents.shape[0], latents.shape[1], 1, 1), device=latents.device) + return noise + + +def apply_masked_loss(loss, batch): + if "conditioning_images" in batch: + # conditioning image is -1 to 1. we need to convert it to 0 to 1 + mask_image = batch["conditioning_images"].to(dtype=loss.dtype)[:, 0].unsqueeze(1) # use R channel + mask_image = mask_image / 2 + 0.5 + # print(f"conditioning_image: {mask_image.shape}") + elif "alpha_masks" in batch and batch["alpha_masks"] is not None: + # alpha mask is 0 to 1 + mask_image = batch["alpha_masks"].to(dtype=loss.dtype).unsqueeze(1) # add channel dimension + # print(f"mask_image: {mask_image.shape}, {mask_image.mean()}") + else: + return loss + + # resize to the same size as the loss + mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area") + loss = loss * mask_image + return loss + + +""" +########################################## +# Perlin Noise +def rand_perlin_2d(device, shape, res, fade=lambda t: 6 * t**5 - 15 * t**4 + 10 * t**3): + delta = (res[0] / shape[0], res[1] / shape[1]) + d = (shape[0] // res[0], shape[1] // res[1]) + + grid = ( + torch.stack( + torch.meshgrid(torch.arange(0, res[0], delta[0], device=device), torch.arange(0, res[1], delta[1], device=device)), + dim=-1, + ) + % 1 + ) + angles = 2 * torch.pi * torch.rand(res[0] + 1, res[1] + 1, device=device) + gradients = torch.stack((torch.cos(angles), torch.sin(angles)), dim=-1) + + tile_grads = ( + lambda slice1, slice2: gradients[slice1[0] : slice1[1], slice2[0] : slice2[1]] + .repeat_interleave(d[0], 0) + .repeat_interleave(d[1], 1) + ) + dot = lambda grad, shift: ( + torch.stack((grid[: shape[0], : shape[1], 0] + shift[0], grid[: shape[0], : shape[1], 1] + shift[1]), dim=-1) + * grad[: shape[0], : shape[1]] + ).sum(dim=-1) + + n00 = dot(tile_grads([0, -1], [0, -1]), [0, 0]) + n10 = dot(tile_grads([1, None], [0, -1]), [-1, 0]) + n01 = dot(tile_grads([0, -1], [1, None]), [0, -1]) + n11 = dot(tile_grads([1, None], [1, None]), [-1, -1]) + t = fade(grid[: shape[0], : shape[1]]) + return 1.414 * torch.lerp(torch.lerp(n00, n10, t[..., 0]), torch.lerp(n01, n11, t[..., 0]), t[..., 1]) + + +def rand_perlin_2d_octaves(device, shape, res, octaves=1, persistence=0.5): + noise = torch.zeros(shape, device=device) + frequency = 1 + amplitude = 1 + for _ in range(octaves): + noise += amplitude * rand_perlin_2d(device, shape, (frequency * res[0], frequency * res[1])) + frequency *= 2 + amplitude *= persistence + return noise + + +def perlin_noise(noise, device, octaves): + _, c, w, h = noise.shape + perlin = lambda: rand_perlin_2d_octaves(device, (w, h), (4, 4), octaves) + noise_perlin = [] + for _ in range(c): + noise_perlin.append(perlin()) + noise_perlin = torch.stack(noise_perlin).unsqueeze(0) # (1, c, w, h) + noise += noise_perlin # broadcast for each batch + return noise / noise.std() # Scaled back to roughly unit variance +""" diff --git a/deepspeed_utils.py b/deepspeed_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..99a7b2b3bee1925739e54a9c7840212fdc4c98ba --- /dev/null +++ b/deepspeed_utils.py @@ -0,0 +1,139 @@ +import os +import argparse +import torch +from accelerate import DeepSpeedPlugin, Accelerator + +from .utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +def add_deepspeed_arguments(parser: argparse.ArgumentParser): + # DeepSpeed Arguments. https://huggingface.co/docs/accelerate/usage_guides/deepspeed + parser.add_argument("--deepspeed", action="store_true", help="enable deepspeed training") + parser.add_argument("--zero_stage", type=int, default=2, choices=[0, 1, 2, 3], help="Possible options are 0,1,2,3.") + parser.add_argument( + "--offload_optimizer_device", + type=str, + default=None, + choices=[None, "cpu", "nvme"], + help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3.", + ) + parser.add_argument( + "--offload_optimizer_nvme_path", + type=str, + default=None, + help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.", + ) + parser.add_argument( + "--offload_param_device", + type=str, + default=None, + choices=[None, "cpu", "nvme"], + help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stage 3.", + ) + parser.add_argument( + "--offload_param_nvme_path", + type=str, + default=None, + help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.", + ) + parser.add_argument( + "--zero3_init_flag", + action="store_true", + help="Flag to indicate whether to enable `deepspeed.zero.Init` for constructing massive models." + "Only applicable with ZeRO Stage-3.", + ) + parser.add_argument( + "--zero3_save_16bit_model", + action="store_true", + help="Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3.", + ) + parser.add_argument( + "--fp16_master_weights_and_gradients", + action="store_true", + help="fp16_master_and_gradients requires optimizer to support keeping fp16 master and gradients while keeping the optimizer states in fp32.", + ) + + +def prepare_deepspeed_args(args: argparse.Namespace): + if not args.deepspeed: + return + + # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1. + args.max_data_loader_n_workers = 1 + + +def prepare_deepspeed_plugin(args: argparse.Namespace): + if not args.deepspeed: + return None + + try: + import deepspeed + except ImportError as e: + logger.error( + "deepspeed is not installed. please install deepspeed in your environment with following command. DS_BUILD_OPS=0 pip install deepspeed" + ) + exit(1) + + deepspeed_plugin = DeepSpeedPlugin( + zero_stage=args.zero_stage, + gradient_accumulation_steps=args.gradient_accumulation_steps, + gradient_clipping=args.max_grad_norm, + offload_optimizer_device=args.offload_optimizer_device, + offload_optimizer_nvme_path=args.offload_optimizer_nvme_path, + offload_param_device=args.offload_param_device, + offload_param_nvme_path=args.offload_param_nvme_path, + zero3_init_flag=args.zero3_init_flag, + zero3_save_16bit_model=args.zero3_save_16bit_model, + ) + deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] = args.train_batch_size + deepspeed_plugin.deepspeed_config["train_batch_size"] = ( + args.train_batch_size * args.gradient_accumulation_steps * int(os.environ["WORLD_SIZE"]) + ) + deepspeed_plugin.set_mixed_precision(args.mixed_precision) + if args.mixed_precision.lower() == "fp16": + deepspeed_plugin.deepspeed_config["fp16"]["initial_scale_power"] = 0 # preventing overflow. + if args.full_fp16 or args.fp16_master_weights_and_gradients: + if args.offload_optimizer_device == "cpu" and args.zero_stage == 2: + deepspeed_plugin.deepspeed_config["fp16"]["fp16_master_weights_and_grads"] = True + logger.info("[DeepSpeed] full fp16 enable.") + else: + logger.info( + "[DeepSpeed]full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam on ZeRO-2 stage." + ) + + if args.offload_optimizer_device is not None: + logger.info("[DeepSpeed] start to manually build cpu_adam.") + deepspeed.ops.op_builder.CPUAdamBuilder().load() + logger.info("[DeepSpeed] building cpu_adam done.") + + return deepspeed_plugin + + +# Accelerate library does not support multiple models for deepspeed. So, we need to wrap multiple models into a single model. +def prepare_deepspeed_model(args: argparse.Namespace, **models): + # remove None from models + models = {k: v for k, v in models.items() if v is not None} + + class DeepSpeedWrapper(torch.nn.Module): + def __init__(self, **kw_models) -> None: + super().__init__() + self.models = torch.nn.ModuleDict() + + for key, model in kw_models.items(): + if isinstance(model, list): + model = torch.nn.ModuleList(model) + assert isinstance( + model, torch.nn.Module + ), f"model must be an instance of torch.nn.Module, but got {key} is {type(model)}" + self.models.update(torch.nn.ModuleDict({key: model})) + + def get_models(self): + return self.models + + ds_model = DeepSpeedWrapper(**models) + return ds_model diff --git a/dependabot.yml b/dependabot.yml new file mode 100644 index 0000000000000000000000000000000000000000..64284b90748cb10e03e909e7c0530fc1d12a934e --- /dev/null +++ b/dependabot.yml @@ -0,0 +1,7 @@ +--- +version: 2 +updates: + - package-ecosystem: "github-actions" + directory: "/" + schedule: + interval: "monthly" diff --git a/detect_face_rotate.py b/detect_face_rotate.py new file mode 100644 index 0000000000000000000000000000000000000000..d2a4d9cfb878df6b1f624fed7410fa3c716ab7cc --- /dev/null +++ b/detect_face_rotate.py @@ -0,0 +1,253 @@ +# このスクリプトのライセンスは、train_dreambooth.pyと同じくApache License 2.0とします +# (c) 2022 Kohya S. @kohya_ss + +# 横長の画像から顔検出して正立するように回転し、そこを中心に正方形に切り出す + +# v2: extract max face if multiple faces are found +# v3: add crop_ratio option +# v4: add multiple faces extraction and min/max size + +import argparse +import math +import cv2 +import glob +import os +from anime_face_detector import create_detector +from tqdm import tqdm +import numpy as np +from library.utils import setup_logging, pil_resize +setup_logging() +import logging +logger = logging.getLogger(__name__) + +KP_REYE = 11 +KP_LEYE = 19 + +SCORE_THRES = 0.90 + + +def detect_faces(detector, image, min_size): + preds = detector(image) # bgr + # logger.info(len(preds)) + + faces = [] + for pred in preds: + bb = pred['bbox'] + score = bb[-1] + if score < SCORE_THRES: + continue + + left, top, right, bottom = bb[:4] + cx = int((left + right) / 2) + cy = int((top + bottom) / 2) + fw = int(right - left) + fh = int(bottom - top) + + lex, ley = pred['keypoints'][KP_LEYE, 0:2] + rex, rey = pred['keypoints'][KP_REYE, 0:2] + angle = math.atan2(ley - rey, lex - rex) + angle = angle / math.pi * 180 + + faces.append((cx, cy, fw, fh, angle)) + + faces.sort(key=lambda x: max(x[2], x[3]), reverse=True) # 大きい順 + return faces + + +def rotate_image(image, angle, cx, cy): + h, w = image.shape[0:2] + rot_mat = cv2.getRotationMatrix2D((cx, cy), angle, 1.0) + + # # 回転する分、すこし画像サイズを大きくする→とりあえず無効化 + # nh = max(h, int(w * math.sin(angle))) + # nw = max(w, int(h * math.sin(angle))) + # if nh > h or nw > w: + # pad_y = nh - h + # pad_t = pad_y // 2 + # pad_x = nw - w + # pad_l = pad_x // 2 + # m = np.array([[0, 0, pad_l], + # [0, 0, pad_t]]) + # rot_mat = rot_mat + m + # h, w = nh, nw + # cx += pad_l + # cy += pad_t + + result = cv2.warpAffine(image, rot_mat, (w, h), flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REFLECT) + return result, cx, cy + + +def process(args): + assert (not args.resize_fit) or args.resize_face_size is None, f"resize_fit and resize_face_size can't be specified both / resize_fitとresize_face_sizeはどちらか片方しか指定できません" + assert args.crop_ratio is None or args.resize_face_size is None, f"crop_ratio指定時はresize_face_sizeは指定できません" + + # アニメ顔検出モデルを読み込む + logger.info("loading face detector.") + detector = create_detector('yolov3') + + # cropの引数を解析する + if args.crop_size is None: + crop_width = crop_height = None + else: + tokens = args.crop_size.split(',') + assert len(tokens) == 2, f"crop_size must be 'width,height' / crop_sizeは'幅,高さ'で指定してください" + crop_width, crop_height = [int(t) for t in tokens] + + if args.crop_ratio is None: + crop_h_ratio = crop_v_ratio = None + else: + tokens = args.crop_ratio.split(',') + assert len(tokens) == 2, f"crop_ratio must be 'horizontal,vertical' / crop_ratioは'幅,高さ'の倍率で指定してください" + crop_h_ratio, crop_v_ratio = [float(t) for t in tokens] + + # 画像を処理する + logger.info("processing.") + output_extension = ".png" + + os.makedirs(args.dst_dir, exist_ok=True) + paths = glob.glob(os.path.join(args.src_dir, "*.png")) + glob.glob(os.path.join(args.src_dir, "*.jpg")) + \ + glob.glob(os.path.join(args.src_dir, "*.webp")) + for path in tqdm(paths): + basename = os.path.splitext(os.path.basename(path))[0] + + # image = cv2.imread(path) # 日本語ファイル名でエラーになる + image = cv2.imdecode(np.fromfile(path, np.uint8), cv2.IMREAD_UNCHANGED) + if len(image.shape) == 2: + image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR) + if image.shape[2] == 4: + logger.warning(f"image has alpha. ignore / 画像の透明度が設定されているため無視します: {path}") + image = image[:, :, :3].copy() # copyをしないと内部的に透明度情報が付いたままになるらしい + + h, w = image.shape[:2] + + faces = detect_faces(detector, image, args.multiple_faces) + for i, face in enumerate(faces): + cx, cy, fw, fh, angle = face + face_size = max(fw, fh) + if args.min_size is not None and face_size < args.min_size: + continue + if args.max_size is not None and face_size >= args.max_size: + continue + face_suffix = f"_{i+1:02d}" if args.multiple_faces else "" + + # オプション指定があれば回転する + face_img = image + if args.rotate: + face_img, cx, cy = rotate_image(face_img, angle, cx, cy) + + # オプション指定があれば顔を中心に切り出す + if crop_width is not None or crop_h_ratio is not None: + cur_crop_width, cur_crop_height = crop_width, crop_height + if crop_h_ratio is not None: + cur_crop_width = int(face_size * crop_h_ratio + .5) + cur_crop_height = int(face_size * crop_v_ratio + .5) + + # リサイズを必要なら行う + scale = 1.0 + if args.resize_face_size is not None: + # 顔サイズを基準にリサイズする + scale = args.resize_face_size / face_size + if scale < cur_crop_width / w: + logger.warning( + f"image width too small in face size based resizing / 顔を基準にリサイズすると画像の幅がcrop sizeより小さい（顔が相対的に大きすぎる）ので顔サイズが変わります: {path}") + scale = cur_crop_width / w + if scale < cur_crop_height / h: + logger.warning( + f"image height too small in face size based resizing / 顔を基準にリサイズすると画像の高さがcrop sizeより小さい（顔が相対的に大きすぎる）ので顔サイズが変わります: {path}") + scale = cur_crop_height / h + elif crop_h_ratio is not None: + # 倍率指定の時にはリサイズしない + pass + else: + # 切り出しサイズ指定あり + if w < cur_crop_width: + logger.warning(f"image width too small/ 画像の幅がcrop sizeより小さいので画質が劣化します: {path}") + scale = cur_crop_width / w + if h < cur_crop_height: + logger.warning(f"image height too small/ 画像の高さがcrop sizeより小さいので画質が劣化します: {path}") + scale = cur_crop_height / h + if args.resize_fit: + scale = max(cur_crop_width / w, cur_crop_height / h) + + if scale != 1.0: + w = int(w * scale + .5) + h = int(h * scale + .5) + if scale < 1.0: + face_img = cv2.resize(face_img, (w, h), interpolation=cv2.INTER_AREA) + else: + face_img = pil_resize(face_img, (w, h)) + cx = int(cx * scale + .5) + cy = int(cy * scale + .5) + fw = int(fw * scale + .5) + fh = int(fh * scale + .5) + + cur_crop_width = min(cur_crop_width, face_img.shape[1]) + cur_crop_height = min(cur_crop_height, face_img.shape[0]) + + x = cx - cur_crop_width // 2 + cx = cur_crop_width // 2 + if x < 0: + cx = cx + x + x = 0 + elif x + cur_crop_width > w: + cx = cx + (x + cur_crop_width - w) + x = w - cur_crop_width + face_img = face_img[:, x:x+cur_crop_width] + + y = cy - cur_crop_height // 2 + cy = cur_crop_height // 2 + if y < 0: + cy = cy + y + y = 0 + elif y + cur_crop_height > h: + cy = cy + (y + cur_crop_height - h) + y = h - cur_crop_height + face_img = face_img[y:y + cur_crop_height] + + # # debug + # logger.info(path, cx, cy, angle) + # crp = cv2.resize(image, (image.shape[1]//8, image.shape[0]//8)) + # cv2.imshow("image", crp) + # if cv2.waitKey() == 27: + # break + # cv2.destroyAllWindows() + + # debug + if args.debug: + cv2.rectangle(face_img, (cx-fw//2, cy-fh//2), (cx+fw//2, cy+fh//2), (255, 0, 255), fw//20) + + _, buf = cv2.imencode(output_extension, face_img) + with open(os.path.join(args.dst_dir, f"{basename}{face_suffix}_{cx:04d}_{cy:04d}_{fw:04d}_{fh:04d}{output_extension}"), "wb") as f: + buf.tofile(f) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("--src_dir", type=str, help="directory to load images / 画像を読み込むディレクトリ") + parser.add_argument("--dst_dir", type=str, help="directory to save images / 画像を保存するディレクトリ") + parser.add_argument("--rotate", action="store_true", help="rotate images to align faces / 顔が正立するように画像を回転する") + parser.add_argument("--resize_fit", action="store_true", + help="resize to fit smaller side after cropping / 切り出し後の画像の短辺がcrop_sizeにあうようにリサイズする") + parser.add_argument("--resize_face_size", type=int, default=None, + help="resize image before cropping by face size / 切り出し前に顔がこのサイズになるようにリサイズする") + parser.add_argument("--crop_size", type=str, default=None, + help="crop images with 'width,height' pixels, face centered / 顔を中心として'幅,高さ'のサイズで切り出す") + parser.add_argument("--crop_ratio", type=str, default=None, + help="crop images with 'horizontal,vertical' ratio to face, face centered / 顔を中心として顔サイズの'幅倍率,高さ倍率'のサイズで切り出す") + parser.add_argument("--min_size", type=int, default=None, + help="minimum face size to output (included) / 処理対象とする顔の最小サイズ（この値以上）") + parser.add_argument("--max_size", type=int, default=None, + help="maximum face size to output (excluded) / 処理対象とする顔の最大サイズ（この値未満）") + parser.add_argument("--multiple_faces", action="store_true", + help="output each faces / 複数の顔が見つかった場合、それぞれを切り出す") + parser.add_argument("--debug", action="store_true", help="render rect for face / 処理後画像の顔位置に矩形を描画します") + + return parser + + +if __name__ == '__main__': + parser = setup_parser() + + args = parser.parse_args() + + process(args) diff --git a/device_utils.py b/device_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..d2e1974504fcc6010100b1c55222144fb596a54b --- /dev/null +++ b/device_utils.py @@ -0,0 +1,89 @@ +import functools +import gc + +import torch +try: + # intel gpu support for pytorch older than 2.5 + # ipex is not needed after pytorch 2.5 + import intel_extension_for_pytorch as ipex # noqa +except Exception: + pass + + +try: + HAS_CUDA = torch.cuda.is_available() +except Exception: + HAS_CUDA = False + +try: + HAS_MPS = torch.backends.mps.is_available() +except Exception: + HAS_MPS = False + +try: + HAS_XPU = torch.xpu.is_available() +except Exception: + HAS_XPU = False + + +def clean_memory(): + gc.collect() + if HAS_CUDA: + torch.cuda.empty_cache() + if HAS_XPU: + torch.xpu.empty_cache() + if HAS_MPS: + torch.mps.empty_cache() + + +def clean_memory_on_device(device: torch.device): + r""" + Clean memory on the specified device, will be called from training scripts. + """ + gc.collect() + + # device may "cuda" or "cuda:0", so we need to check the type of device + if device.type == "cuda": + torch.cuda.empty_cache() + if device.type == "xpu": + torch.xpu.empty_cache() + if device.type == "mps": + torch.mps.empty_cache() + + +@functools.lru_cache(maxsize=None) +def get_preferred_device() -> torch.device: + r""" + Do not call this function from training scripts. Use accelerator.device instead. + """ + if HAS_CUDA: + device = torch.device("cuda") + elif HAS_XPU: + device = torch.device("xpu") + elif HAS_MPS: + device = torch.device("mps") + else: + device = torch.device("cpu") + print(f"get_preferred_device() -> {device}") + return device + + +def init_ipex(): + """ + Apply IPEX to CUDA hijacks using `library.ipex.ipex_init`. + + This function should run right after importing torch and before doing anything else. + + If xpu is not available, this function does nothing. + """ + try: + if HAS_XPU: + from library.ipex import ipex_init + + is_initialized, error_message = ipex_init() + if not is_initialized: + print("failed to initialize ipex:", error_message) + else: + return + except Exception as e: + print("failed to initialize ipex:", e) diff --git a/diffusers.py b/diffusers.py new file mode 100644 index 0000000000000000000000000000000000000000..75715d161ceff964455f3492f1c68e66e4f08e7f --- /dev/null +++ b/diffusers.py @@ -0,0 +1,47 @@ +from functools import wraps +import torch +import diffusers # pylint: disable=import-error + +# pylint: disable=protected-access, missing-function-docstring, line-too-long + + +# Diffusers FreeU +original_fourier_filter = diffusers.utils.torch_utils.fourier_filter +@wraps(diffusers.utils.torch_utils.fourier_filter) +def fourier_filter(x_in, threshold, scale): + return_dtype = x_in.dtype + return original_fourier_filter(x_in.to(dtype=torch.float32), threshold, scale).to(dtype=return_dtype) + + +# fp64 error +class FluxPosEmbed(torch.nn.Module): + def __init__(self, theta: int, axes_dim): + super().__init__() + self.theta = theta + self.axes_dim = axes_dim + + def forward(self, ids: torch.Tensor) -> torch.Tensor: + n_axes = ids.shape[-1] + cos_out = [] + sin_out = [] + pos = ids.float() + for i in range(n_axes): + cos, sin = diffusers.models.embeddings.get_1d_rotary_pos_embed( + self.axes_dim[i], + pos[:, i], + theta=self.theta, + repeat_interleave_real=True, + use_real=True, + freqs_dtype=torch.float32, + ) + cos_out.append(cos) + sin_out.append(sin) + freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device) + freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device) + return freqs_cos, freqs_sin + + +def ipex_diffusers(device_supports_fp64=False, can_allocate_plus_4gb=False): + diffusers.utils.torch_utils.fourier_filter = fourier_filter + if not device_supports_fp64: + diffusers.models.embeddings.FluxPosEmbed = FluxPosEmbed diff --git a/dylora.py b/dylora.py new file mode 100644 index 0000000000000000000000000000000000000000..82d96f59b68f5e9e775e8c157e4fce7fd76892b2 --- /dev/null +++ b/dylora.py @@ -0,0 +1,529 @@ +# some codes are copied from: +# https://github.com/huawei-noah/KD-NLP/blob/main/DyLoRA/ + +# Copyright (C) 2022. Huawei Technologies Co., Ltd. All rights reserved. +# Changes made to the original code: +# 2022.08.20 - Integrate the DyLoRA layer for the LoRA Linear layer +# ------------------------------------------------------------------------------------------ +# Copyright (c) Microsoft Corporation. All rights reserved. +# Licensed under the MIT License (MIT). See LICENSE in the repo root for license information. +# ------------------------------------------------------------------------------------------ + +import math +import os +import random +from typing import Dict, List, Optional, Tuple, Type, Union +from diffusers import AutoencoderKL +from transformers import CLIPTextModel +import torch +from torch import nn +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +class DyLoRAModule(torch.nn.Module): + """ + replaces forward method of the original Linear, instead of replacing the original Linear module. + """ + + # NOTE: support dropout in future + def __init__(self, lora_name, org_module: torch.nn.Module, multiplier=1.0, lora_dim=4, alpha=1, unit=1): + super().__init__() + self.lora_name = lora_name + self.lora_dim = lora_dim + self.unit = unit + assert self.lora_dim % self.unit == 0, "rank must be a multiple of unit" + + if org_module.__class__.__name__ == "Conv2d": + in_dim = org_module.in_channels + out_dim = org_module.out_channels + else: + in_dim = org_module.in_features + out_dim = org_module.out_features + + if type(alpha) == torch.Tensor: + alpha = alpha.detach().float().numpy() # without casting, bf16 causes error + alpha = self.lora_dim if alpha is None or alpha == 0 else alpha + self.scale = alpha / self.lora_dim + self.register_buffer("alpha", torch.tensor(alpha)) # 定数として扱える + + self.is_conv2d = org_module.__class__.__name__ == "Conv2d" + self.is_conv2d_3x3 = self.is_conv2d and org_module.kernel_size == (3, 3) + + if self.is_conv2d and self.is_conv2d_3x3: + kernel_size = org_module.kernel_size + self.stride = org_module.stride + self.padding = org_module.padding + self.lora_A = nn.ParameterList([org_module.weight.new_zeros((1, in_dim, *kernel_size)) for _ in range(self.lora_dim)]) + self.lora_B = nn.ParameterList([org_module.weight.new_zeros((out_dim, 1, 1, 1)) for _ in range(self.lora_dim)]) + else: + self.lora_A = nn.ParameterList([org_module.weight.new_zeros((1, in_dim)) for _ in range(self.lora_dim)]) + self.lora_B = nn.ParameterList([org_module.weight.new_zeros((out_dim, 1)) for _ in range(self.lora_dim)]) + + # same as microsoft's + for lora in self.lora_A: + torch.nn.init.kaiming_uniform_(lora, a=math.sqrt(5)) + for lora in self.lora_B: + torch.nn.init.zeros_(lora) + + self.multiplier = multiplier + self.org_module = org_module # remove in applying + + def apply_to(self): + self.org_forward = self.org_module.forward + self.org_module.forward = self.forward + del self.org_module + + def forward(self, x): + result = self.org_forward(x) + + # specify the dynamic rank + trainable_rank = random.randint(0, self.lora_dim - 1) + trainable_rank = trainable_rank - trainable_rank % self.unit # make sure the rank is a multiple of unit + + # 一部のパラメータを固定して、残りのパラメータを学習する + for i in range(0, trainable_rank): + self.lora_A[i].requires_grad = False + self.lora_B[i].requires_grad = False + for i in range(trainable_rank, trainable_rank + self.unit): + self.lora_A[i].requires_grad = True + self.lora_B[i].requires_grad = True + for i in range(trainable_rank + self.unit, self.lora_dim): + self.lora_A[i].requires_grad = False + self.lora_B[i].requires_grad = False + + lora_A = torch.cat(tuple(self.lora_A), dim=0) + lora_B = torch.cat(tuple(self.lora_B), dim=1) + + # calculate with lora_A and lora_B + if self.is_conv2d_3x3: + ab = torch.nn.functional.conv2d(x, lora_A, stride=self.stride, padding=self.padding) + ab = torch.nn.functional.conv2d(ab, lora_B) + else: + ab = x + if self.is_conv2d: + ab = ab.reshape(ab.size(0), ab.size(1), -1).transpose(1, 2) # (N, C, H, W) -> (N, H*W, C) + + ab = torch.nn.functional.linear(ab, lora_A) + ab = torch.nn.functional.linear(ab, lora_B) + + if self.is_conv2d: + ab = ab.transpose(1, 2).reshape(ab.size(0), -1, *x.size()[2:]) # (N, H*W, C) -> (N, C, H, W) + + # 最後の項は、低rankをより大きくするためのスケーリング（じゃないかな） + result = result + ab * self.scale * math.sqrt(self.lora_dim / (trainable_rank + self.unit)) + + # NOTE weightに加算してからlinear/conv2dを呼んだほうが速いかも + return result + + def state_dict(self, destination=None, prefix="", keep_vars=False): + # state dictを通常のLoRAと同じにする: + # nn.ParameterListは `.lora_A.0` みたいな名前になるので、forwardと同様にcatして入れ替える + sd = super().state_dict(destination=destination, prefix=prefix, keep_vars=keep_vars) + + lora_A_weight = torch.cat(tuple(self.lora_A), dim=0) + if self.is_conv2d and not self.is_conv2d_3x3: + lora_A_weight = lora_A_weight.unsqueeze(-1).unsqueeze(-1) + + lora_B_weight = torch.cat(tuple(self.lora_B), dim=1) + if self.is_conv2d and not self.is_conv2d_3x3: + lora_B_weight = lora_B_weight.unsqueeze(-1).unsqueeze(-1) + + sd[self.lora_name + ".lora_down.weight"] = lora_A_weight if keep_vars else lora_A_weight.detach() + sd[self.lora_name + ".lora_up.weight"] = lora_B_weight if keep_vars else lora_B_weight.detach() + + i = 0 + while True: + key_a = f"{self.lora_name}.lora_A.{i}" + key_b = f"{self.lora_name}.lora_B.{i}" + if key_a in sd: + sd.pop(key_a) + sd.pop(key_b) + else: + break + i += 1 + return sd + + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): + # 通常のLoRAと同じstate dictを読み込めるようにする：この方法はchatGPTに聞いた + lora_A_weight = state_dict.pop(self.lora_name + ".lora_down.weight", None) + lora_B_weight = state_dict.pop(self.lora_name + ".lora_up.weight", None) + + if lora_A_weight is None or lora_B_weight is None: + if strict: + raise KeyError(f"{self.lora_name}.lora_down/up.weight is not found") + else: + return + + if self.is_conv2d and not self.is_conv2d_3x3: + lora_A_weight = lora_A_weight.squeeze(-1).squeeze(-1) + lora_B_weight = lora_B_weight.squeeze(-1).squeeze(-1) + + state_dict.update( + {f"{self.lora_name}.lora_A.{i}": nn.Parameter(lora_A_weight[i].unsqueeze(0)) for i in range(lora_A_weight.size(0))} + ) + state_dict.update( + {f"{self.lora_name}.lora_B.{i}": nn.Parameter(lora_B_weight[:, i].unsqueeze(1)) for i in range(lora_B_weight.size(1))} + ) + + super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + + +def create_network( + multiplier: float, + network_dim: Optional[int], + network_alpha: Optional[float], + vae: AutoencoderKL, + text_encoder: Union[CLIPTextModel, List[CLIPTextModel]], + unet, + **kwargs, +): + if network_dim is None: + network_dim = 4 # default + if network_alpha is None: + network_alpha = 1.0 + + # extract dim/alpha for conv2d, and block dim + conv_dim = kwargs.get("conv_dim", None) + conv_alpha = kwargs.get("conv_alpha", None) + unit = kwargs.get("unit", None) + if conv_dim is not None: + conv_dim = int(conv_dim) + assert conv_dim == network_dim, "conv_dim must be same as network_dim" + if conv_alpha is None: + conv_alpha = 1.0 + else: + conv_alpha = float(conv_alpha) + + if unit is not None: + unit = int(unit) + else: + unit = 1 + + network = DyLoRANetwork( + text_encoder, + unet, + multiplier=multiplier, + lora_dim=network_dim, + alpha=network_alpha, + apply_to_conv=conv_dim is not None, + unit=unit, + varbose=True, + ) + + loraplus_lr_ratio = kwargs.get("loraplus_lr_ratio", None) + loraplus_unet_lr_ratio = kwargs.get("loraplus_unet_lr_ratio", None) + loraplus_text_encoder_lr_ratio = kwargs.get("loraplus_text_encoder_lr_ratio", None) + loraplus_lr_ratio = float(loraplus_lr_ratio) if loraplus_lr_ratio is not None else None + loraplus_unet_lr_ratio = float(loraplus_unet_lr_ratio) if loraplus_unet_lr_ratio is not None else None + loraplus_text_encoder_lr_ratio = float(loraplus_text_encoder_lr_ratio) if loraplus_text_encoder_lr_ratio is not None else None + if loraplus_lr_ratio is not None or loraplus_unet_lr_ratio is not None or loraplus_text_encoder_lr_ratio is not None: + network.set_loraplus_lr_ratio(loraplus_lr_ratio, loraplus_unet_lr_ratio, loraplus_text_encoder_lr_ratio) + + return network + + +# Create network from weights for inference, weights are not loaded here (because can be merged) +def create_network_from_weights(multiplier, file, vae, text_encoder, unet, weights_sd=None, for_inference=False, **kwargs): + if weights_sd is None: + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file, safe_open + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + # get dim/alpha mapping + modules_dim = {} + modules_alpha = {} + for key, value in weights_sd.items(): + if "." not in key: + continue + + lora_name = key.split(".")[0] + if "alpha" in key: + modules_alpha[lora_name] = value + elif "lora_down" in key: + dim = value.size()[0] + modules_dim[lora_name] = dim + # logger.info(f"{lora_name} {value.size()} {dim}") + + # support old LoRA without alpha + for key in modules_dim.keys(): + if key not in modules_alpha: + modules_alpha = modules_dim[key] + + module_class = DyLoRAModule + + network = DyLoRANetwork( + text_encoder, unet, multiplier=multiplier, modules_dim=modules_dim, modules_alpha=modules_alpha, module_class=module_class + ) + return network, weights_sd + + +class DyLoRANetwork(torch.nn.Module): + UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"] + UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPSdpaAttention", "CLIPMLP"] + LORA_PREFIX_UNET = "lora_unet" + LORA_PREFIX_TEXT_ENCODER = "lora_te" + + def __init__( + self, + text_encoder, + unet, + multiplier=1.0, + lora_dim=4, + alpha=1, + apply_to_conv=False, + modules_dim=None, + modules_alpha=None, + unit=1, + module_class=DyLoRAModule, + varbose=False, + ) -> None: + super().__init__() + self.multiplier = multiplier + + self.lora_dim = lora_dim + self.alpha = alpha + self.apply_to_conv = apply_to_conv + + self.loraplus_lr_ratio = None + self.loraplus_unet_lr_ratio = None + self.loraplus_text_encoder_lr_ratio = None + + if modules_dim is not None: + logger.info("create LoRA network from weights") + else: + logger.info(f"create LoRA network. base dim (rank): {lora_dim}, alpha: {alpha}, unit: {unit}") + if self.apply_to_conv: + logger.info("apply LoRA to Conv2d with kernel size (3,3).") + + # create module instances + def create_modules(is_unet, root_module: torch.nn.Module, target_replace_modules) -> List[DyLoRAModule]: + prefix = DyLoRANetwork.LORA_PREFIX_UNET if is_unet else DyLoRANetwork.LORA_PREFIX_TEXT_ENCODER + loras = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + is_linear = child_module.__class__.__name__ == "Linear" + is_conv2d = child_module.__class__.__name__ == "Conv2d" + is_conv2d_1x1 = is_conv2d and child_module.kernel_size == (1, 1) + + if is_linear or is_conv2d: + lora_name = prefix + "." + name + "." + child_name + lora_name = lora_name.replace(".", "_") + + dim = None + alpha = None + if modules_dim is not None: + if lora_name in modules_dim: + dim = modules_dim[lora_name] + alpha = modules_alpha[lora_name] + else: + if is_linear or is_conv2d_1x1 or apply_to_conv: + dim = self.lora_dim + alpha = self.alpha + + if dim is None or dim == 0: + continue + + # dropout and fan_in_fan_out is default + lora = module_class(lora_name, child_module, self.multiplier, dim, alpha, unit) + loras.append(lora) + return loras + + text_encoders = text_encoder if type(text_encoder) == list else [text_encoder] + + self.text_encoder_loras = [] + for i, text_encoder in enumerate(text_encoders): + if len(text_encoders) > 1: + index = i + 1 + logger.info(f"create LoRA for Text Encoder {index}") + else: + index = None + logger.info("create LoRA for Text Encoder") + + text_encoder_loras = create_modules(False, text_encoder, DyLoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE) + self.text_encoder_loras.extend(text_encoder_loras) + + # self.text_encoder_loras = create_modules(False, text_encoder, DyLoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE) + logger.info(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.") + + # extend U-Net target modules if conv2d 3x3 is enabled, or load from weights + target_modules = DyLoRANetwork.UNET_TARGET_REPLACE_MODULE + if modules_dim is not None or self.apply_to_conv: + target_modules += DyLoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + + self.unet_loras = create_modules(True, unet, target_modules) + logger.info(f"create LoRA for U-Net: {len(self.unet_loras)} modules.") + + def set_loraplus_lr_ratio(self, loraplus_lr_ratio, loraplus_unet_lr_ratio, loraplus_text_encoder_lr_ratio): + self.loraplus_lr_ratio = loraplus_lr_ratio + self.loraplus_unet_lr_ratio = loraplus_unet_lr_ratio + self.loraplus_text_encoder_lr_ratio = loraplus_text_encoder_lr_ratio + + logger.info(f"LoRA+ UNet LR Ratio: {self.loraplus_unet_lr_ratio or self.loraplus_lr_ratio}") + logger.info(f"LoRA+ Text Encoder LR Ratio: {self.loraplus_text_encoder_lr_ratio or self.loraplus_lr_ratio}") + + def set_multiplier(self, multiplier): + self.multiplier = multiplier + for lora in self.text_encoder_loras + self.unet_loras: + lora.multiplier = self.multiplier + + def load_weights(self, file): + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + info = self.load_state_dict(weights_sd, False) + return info + + def apply_to(self, text_encoder, unet, apply_text_encoder=True, apply_unet=True): + if apply_text_encoder: + logger.info("enable LoRA for text encoder") + else: + self.text_encoder_loras = [] + + if apply_unet: + logger.info("enable LoRA for U-Net") + else: + self.unet_loras = [] + + for lora in self.text_encoder_loras + self.unet_loras: + lora.apply_to() + self.add_module(lora.lora_name, lora) + + """ + def merge_to(self, text_encoder, unet, weights_sd, dtype, device): + apply_text_encoder = apply_unet = False + for key in weights_sd.keys(): + if key.startswith(DyLoRANetwork.LORA_PREFIX_TEXT_ENCODER): + apply_text_encoder = True + elif key.startswith(DyLoRANetwork.LORA_PREFIX_UNET): + apply_unet = True + + if apply_text_encoder: + logger.info("enable LoRA for text encoder") + else: + self.text_encoder_loras = [] + + if apply_unet: + logger.info("enable LoRA for U-Net") + else: + self.unet_loras = [] + + for lora in self.text_encoder_loras + self.unet_loras: + sd_for_lora = {} + for key in weights_sd.keys(): + if key.startswith(lora.lora_name): + sd_for_lora[key[len(lora.lora_name) + 1 :]] = weights_sd[key] + lora.merge_to(sd_for_lora, dtype, device) + + logger.info(f"weights are merged") + """ + + # 二つのText Encoderに別々の学習率を設定できるようにするといいかも + def prepare_optimizer_params(self, text_encoder_lr, unet_lr, default_lr): + self.requires_grad_(True) + all_params = [] + + def assemble_params(loras, lr, ratio): + param_groups = {"lora": {}, "plus": {}} + for lora in loras: + for name, param in lora.named_parameters(): + if ratio is not None and "lora_B" in name: + param_groups["plus"][f"{lora.lora_name}.{name}"] = param + else: + param_groups["lora"][f"{lora.lora_name}.{name}"] = param + + params = [] + for key in param_groups.keys(): + param_data = {"params": param_groups[key].values()} + + if len(param_data["params"]) == 0: + continue + + if lr is not None: + if key == "plus": + param_data["lr"] = lr * ratio + else: + param_data["lr"] = lr + + if param_data.get("lr", None) == 0 or param_data.get("lr", None) is None: + continue + + params.append(param_data) + + return params + + if self.text_encoder_loras: + params = assemble_params( + self.text_encoder_loras, + text_encoder_lr if text_encoder_lr is not None else default_lr, + self.loraplus_text_encoder_lr_ratio or self.loraplus_lr_ratio, + ) + all_params.extend(params) + + if self.unet_loras: + params = assemble_params( + self.unet_loras, default_lr if unet_lr is None else unet_lr, self.loraplus_unet_lr_ratio or self.loraplus_lr_ratio + ) + all_params.extend(params) + + return all_params + + def enable_gradient_checkpointing(self): + # not supported + pass + + def prepare_grad_etc(self, text_encoder, unet): + self.requires_grad_(True) + + def on_epoch_start(self, text_encoder, unet): + self.train() + + def get_trainable_params(self): + return self.parameters() + + def save_weights(self, file, dtype, metadata): + if metadata is not None and len(metadata) == 0: + metadata = None + + state_dict = self.state_dict() + + if dtype is not None: + for key in list(state_dict.keys()): + v = state_dict[key] + v = v.detach().clone().to("cpu").to(dtype) + state_dict[key] = v + + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import save_file + from library import train_util + + # Precalculate model hashes to save time on indexing + if metadata is None: + metadata = {} + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + save_file(state_dict, file, metadata) + else: + torch.save(state_dict, file) + + # mask is a tensor with values from 0 to 1 + def set_region(self, sub_prompt_index, is_last_network, mask): + pass + + def set_current_generation(self, batch_size, num_sub_prompts, width, height, shared): + pass diff --git a/extract_lora_from_dylora.py b/extract_lora_from_dylora.py new file mode 100644 index 0000000000000000000000000000000000000000..1184cd8a558bdaa49bb9e98785d753cdd15dd284 --- /dev/null +++ b/extract_lora_from_dylora.py @@ -0,0 +1,128 @@ +# Convert LoRA to different rank approximation (should only be used to go to lower rank) +# This code is based off the extract_lora_from_models.py file which is based on https://github.com/cloneofsimo/lora/blob/develop/lora_diffusion/cli_svd.py +# Thanks to cloneofsimo + +import argparse +import math +import os +import torch +from safetensors.torch import load_file, save_file, safe_open +from tqdm import tqdm +from library import train_util, model_util +import numpy as np +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def load_state_dict(file_name): + if model_util.is_safetensors(file_name): + sd = load_file(file_name) + with safe_open(file_name, framework="pt") as f: + metadata = f.metadata() + else: + sd = torch.load(file_name, map_location="cpu") + metadata = None + + return sd, metadata + + +def save_to_file(file_name, model, metadata): + if model_util.is_safetensors(file_name): + save_file(model, file_name, metadata) + else: + torch.save(model, file_name) + + +def split_lora_model(lora_sd, unit): + max_rank = 0 + + # Extract loaded lora dim and alpha + for key, value in lora_sd.items(): + if "lora_down" in key: + rank = value.size()[0] + if rank > max_rank: + max_rank = rank + logger.info(f"Max rank: {max_rank}") + + rank = unit + split_models = [] + new_alpha = None + while rank < max_rank: + logger.info(f"Splitting rank {rank}") + new_sd = {} + for key, value in lora_sd.items(): + if "lora_down" in key: + new_sd[key] = value[:rank].contiguous() + elif "lora_up" in key: + new_sd[key] = value[:, :rank].contiguous() + else: + # なぜかscaleするとおかしくなる…… + # this_rank = lora_sd[key.replace("alpha", "lora_down.weight")].size()[0] + # scale = math.sqrt(this_rank / rank) # rank is > unit + # logger.info(key, value.size(), this_rank, rank, value, scale) + # new_alpha = value * scale # always same + # new_sd[key] = new_alpha + new_sd[key] = value + + split_models.append((new_sd, rank, new_alpha)) + rank += unit + + return max_rank, split_models + + +def split(args): + logger.info("loading Model...") + lora_sd, metadata = load_state_dict(args.model) + + logger.info("Splitting Model...") + original_rank, split_models = split_lora_model(lora_sd, args.unit) + + comment = metadata.get("ss_training_comment", "") + for state_dict, new_rank, new_alpha in split_models: + # update metadata + if metadata is None: + new_metadata = {} + else: + new_metadata = metadata.copy() + + new_metadata["ss_training_comment"] = f"split from DyLoRA, rank {original_rank} to {new_rank}; {comment}" + new_metadata["ss_network_dim"] = str(new_rank) + # new_metadata["ss_network_alpha"] = str(new_alpha.float().numpy()) + + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + filename, ext = os.path.splitext(args.save_to) + model_file_name = filename + f"-{new_rank:04d}{ext}" + + logger.info(f"saving model to: {model_file_name}") + save_to_file(model_file_name, state_dict, new_metadata) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + + parser.add_argument("--unit", type=int, default=None, help="size of rank to split into / rankを分割するサイズ") + parser.add_argument( + "--save_to", + type=str, + default=None, + help="destination base file name: ckpt or safetensors file / 保存先のファイル名のbase、ckptまたはsafetensors", + ) + parser.add_argument( + "--model", + type=str, + default=None, + help="DyLoRA model to resize at to new rank: ckpt or safetensors file / 読み込むDyLoRAモデル、ckptまたはsafetensors", + ) + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + split(args) diff --git a/extract_lora_from_models.py b/extract_lora_from_models.py new file mode 100644 index 0000000000000000000000000000000000000000..43c1d0058d9390c2a8f0c113464afe42eaaed536 --- /dev/null +++ b/extract_lora_from_models.py @@ -0,0 +1,360 @@ +# extract approximating LoRA by svd from two SD models +# The code is based on https://github.com/cloneofsimo/lora/blob/develop/lora_diffusion/cli_svd.py +# Thanks to cloneofsimo! + +import argparse +import json +import os +import time +import torch +from safetensors.torch import load_file, save_file +from tqdm import tqdm +from library import sai_model_spec, model_util, sdxl_model_util +import lora +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +# CLAMP_QUANTILE = 0.99 +# MIN_DIFF = 1e-1 + + +def save_to_file(file_name, model, state_dict, dtype): + if dtype is not None: + for key in list(state_dict.keys()): + if type(state_dict[key]) == torch.Tensor: + state_dict[key] = state_dict[key].to(dtype) + + if os.path.splitext(file_name)[1] == ".safetensors": + save_file(model, file_name) + else: + torch.save(model, file_name) + + +def svd( + model_org=None, + model_tuned=None, + save_to=None, + dim=4, + v2=None, + sdxl=None, + conv_dim=None, + v_parameterization=None, + device=None, + save_precision=None, + clamp_quantile=0.99, + min_diff=0.01, + no_metadata=False, + load_precision=None, + load_original_model_to=None, + load_tuned_model_to=None, +): + def str_to_dtype(p): + if p == "float": + return torch.float + if p == "fp16": + return torch.float16 + if p == "bf16": + return torch.bfloat16 + return None + + assert v2 != sdxl or (not v2 and not sdxl), "v2 and sdxl cannot be specified at the same time / v2とsdxlは同時に指定できません" + if v_parameterization is None: + v_parameterization = v2 + + load_dtype = str_to_dtype(load_precision) if load_precision else None + save_dtype = str_to_dtype(save_precision) + work_device = "cpu" + + # load models + if not sdxl: + logger.info(f"loading original SD model : {model_org}") + text_encoder_o, _, unet_o = model_util.load_models_from_stable_diffusion_checkpoint(v2, model_org) + text_encoders_o = [text_encoder_o] + if load_dtype is not None: + text_encoder_o = text_encoder_o.to(load_dtype) + unet_o = unet_o.to(load_dtype) + + logger.info(f"loading tuned SD model : {model_tuned}") + text_encoder_t, _, unet_t = model_util.load_models_from_stable_diffusion_checkpoint(v2, model_tuned) + text_encoders_t = [text_encoder_t] + if load_dtype is not None: + text_encoder_t = text_encoder_t.to(load_dtype) + unet_t = unet_t.to(load_dtype) + + model_version = model_util.get_model_version_str_for_sd1_sd2(v2, v_parameterization) + else: + device_org = load_original_model_to if load_original_model_to else "cpu" + device_tuned = load_tuned_model_to if load_tuned_model_to else "cpu" + + logger.info(f"loading original SDXL model : {model_org}") + text_encoder_o1, text_encoder_o2, _, unet_o, _, _ = sdxl_model_util.load_models_from_sdxl_checkpoint( + sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, model_org, device_org + ) + text_encoders_o = [text_encoder_o1, text_encoder_o2] + if load_dtype is not None: + text_encoder_o1 = text_encoder_o1.to(load_dtype) + text_encoder_o2 = text_encoder_o2.to(load_dtype) + unet_o = unet_o.to(load_dtype) + + logger.info(f"loading original SDXL model : {model_tuned}") + text_encoder_t1, text_encoder_t2, _, unet_t, _, _ = sdxl_model_util.load_models_from_sdxl_checkpoint( + sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, model_tuned, device_tuned + ) + text_encoders_t = [text_encoder_t1, text_encoder_t2] + if load_dtype is not None: + text_encoder_t1 = text_encoder_t1.to(load_dtype) + text_encoder_t2 = text_encoder_t2.to(load_dtype) + unet_t = unet_t.to(load_dtype) + + model_version = sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0 + + # create LoRA network to extract weights: Use dim (rank) as alpha + if conv_dim is None: + kwargs = {} + else: + kwargs = {"conv_dim": conv_dim, "conv_alpha": conv_dim} + + lora_network_o = lora.create_network(1.0, dim, dim, None, text_encoders_o, unet_o, **kwargs) + lora_network_t = lora.create_network(1.0, dim, dim, None, text_encoders_t, unet_t, **kwargs) + assert len(lora_network_o.text_encoder_loras) == len( + lora_network_t.text_encoder_loras + ), f"model version is different (SD1.x vs SD2.x) / それぞれのモデルのバージョンが違います（SD1.xベースとSD2.xベース） " + + # get diffs + diffs = {} + text_encoder_different = False + for i, (lora_o, lora_t) in enumerate(zip(lora_network_o.text_encoder_loras, lora_network_t.text_encoder_loras)): + lora_name = lora_o.lora_name + module_o = lora_o.org_module + module_t = lora_t.org_module + diff = module_t.weight.to(work_device) - module_o.weight.to(work_device) + + # clear weight to save memory + module_o.weight = None + module_t.weight = None + + # Text Encoder might be same + if not text_encoder_different and torch.max(torch.abs(diff)) > min_diff: + text_encoder_different = True + logger.info(f"Text encoder is different. {torch.max(torch.abs(diff))} > {min_diff}") + + diffs[lora_name] = diff + + # clear target Text Encoder to save memory + for text_encoder in text_encoders_t: + del text_encoder + + if not text_encoder_different: + logger.warning("Text encoder is same. Extract U-Net only.") + lora_network_o.text_encoder_loras = [] + diffs = {} # clear diffs + + for i, (lora_o, lora_t) in enumerate(zip(lora_network_o.unet_loras, lora_network_t.unet_loras)): + lora_name = lora_o.lora_name + module_o = lora_o.org_module + module_t = lora_t.org_module + diff = module_t.weight.to(work_device) - module_o.weight.to(work_device) + + # clear weight to save memory + module_o.weight = None + module_t.weight = None + + diffs[lora_name] = diff + + # clear LoRA network, target U-Net to save memory + del lora_network_o + del lora_network_t + del unet_t + + # make LoRA with svd + logger.info("calculating by svd") + lora_weights = {} + with torch.no_grad(): + for lora_name, mat in tqdm(list(diffs.items())): + if args.device: + mat = mat.to(args.device) + mat = mat.to(torch.float) # calc by float + + # if conv_dim is None, diffs do not include LoRAs for conv2d-3x3 + conv2d = len(mat.size()) == 4 + kernel_size = None if not conv2d else mat.size()[2:4] + conv2d_3x3 = conv2d and kernel_size != (1, 1) + + rank = dim if not conv2d_3x3 or conv_dim is None else conv_dim + out_dim, in_dim = mat.size()[0:2] + + if device: + mat = mat.to(device) + + # logger.info(lora_name, mat.size(), mat.device, rank, in_dim, out_dim) + rank = min(rank, in_dim, out_dim) # LoRA rank cannot exceed the original dim + + if conv2d: + if conv2d_3x3: + mat = mat.flatten(start_dim=1) + else: + mat = mat.squeeze() + + U, S, Vh = torch.linalg.svd(mat) + + U = U[:, :rank] + S = S[:rank] + U = U @ torch.diag(S) + + Vh = Vh[:rank, :] + + dist = torch.cat([U.flatten(), Vh.flatten()]) + hi_val = torch.quantile(dist, clamp_quantile) + low_val = -hi_val + + U = U.clamp(low_val, hi_val) + Vh = Vh.clamp(low_val, hi_val) + + if conv2d: + U = U.reshape(out_dim, rank, 1, 1) + Vh = Vh.reshape(rank, in_dim, kernel_size[0], kernel_size[1]) + + U = U.to(work_device, dtype=save_dtype).contiguous() + Vh = Vh.to(work_device, dtype=save_dtype).contiguous() + + lora_weights[lora_name] = (U, Vh) + + # make state dict for LoRA + lora_sd = {} + for lora_name, (up_weight, down_weight) in lora_weights.items(): + lora_sd[lora_name + ".lora_up.weight"] = up_weight + lora_sd[lora_name + ".lora_down.weight"] = down_weight + lora_sd[lora_name + ".alpha"] = torch.tensor(down_weight.size()[0]) + + # load state dict to LoRA and save it + lora_network_save, lora_sd = lora.create_network_from_weights(1.0, None, None, text_encoders_o, unet_o, weights_sd=lora_sd) + lora_network_save.apply_to(text_encoders_o, unet_o) # create internal module references for state_dict + + info = lora_network_save.load_state_dict(lora_sd) + logger.info(f"Loading extracted LoRA weights: {info}") + + dir_name = os.path.dirname(save_to) + if dir_name and not os.path.exists(dir_name): + os.makedirs(dir_name, exist_ok=True) + + # minimum metadata + net_kwargs = {} + if conv_dim is not None: + net_kwargs["conv_dim"] = str(conv_dim) + net_kwargs["conv_alpha"] = str(float(conv_dim)) + + metadata = { + "ss_v2": str(v2), + "ss_base_model_version": model_version, + "ss_network_module": "networks.lora", + "ss_network_dim": str(dim), + "ss_network_alpha": str(float(dim)), + "ss_network_args": json.dumps(net_kwargs), + } + + if not no_metadata: + title = os.path.splitext(os.path.basename(save_to))[0] + sai_metadata = sai_model_spec.build_metadata(None, v2, v_parameterization, sdxl, True, False, time.time(), title=title) + metadata.update(sai_metadata) + + lora_network_save.save_weights(save_to, save_dtype, metadata) + logger.info(f"LoRA weights are saved to: {save_to}") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("--v2", action="store_true", help="load Stable Diffusion v2.x model / Stable Diffusion 2.xのモデルを読み込む") + parser.add_argument( + "--v_parameterization", + action="store_true", + default=None, + help="make LoRA metadata for v-parameterization (default is same to v2) / 作成するLoRAのメタデータにv-parameterization用と設定する（省略時はv2と同じ）", + ) + parser.add_argument( + "--sdxl", action="store_true", help="load Stable Diffusion SDXL base model / Stable Diffusion SDXL baseのモデルを読み込む" + ) + parser.add_argument( + "--load_precision", + type=str, + default=None, + choices=[None, "float", "fp16", "bf16"], + help="precision in loading, model default if omitted / 読み込み時に精度を変更して読み込む、省略時はモデルファイルによる" + ) + parser.add_argument( + "--save_precision", + type=str, + default=None, + choices=[None, "float", "fp16", "bf16"], + help="precision in saving, same to merging if omitted / 保存時に精度を変更して保存する、省略時はfloat", + ) + parser.add_argument( + "--model_org", + type=str, + default=None, + required=True, + help="Stable Diffusion original model: ckpt or safetensors file / 元モデル、ckptまたはsafetensors", + ) + parser.add_argument( + "--model_tuned", + type=str, + default=None, + required=True, + help="Stable Diffusion tuned model, LoRA is difference of `original to tuned`: ckpt or safetensors file / 派生モデル（生成されるLoRAは元→派生の差分になります）、ckptまたはsafetensors", + ) + parser.add_argument( + "--save_to", + type=str, + default=None, + required=True, + help="destination file name: ckpt or safetensors file / 保存先のファイル名、ckptまたはsafetensors", + ) + parser.add_argument("--dim", type=int, default=4, help="dimension (rank) of LoRA (default 4) / LoRAの次元数（rank）（デフォルト4）") + parser.add_argument( + "--conv_dim", + type=int, + default=None, + help="dimension (rank) of LoRA for Conv2d-3x3 (default None, disabled) / LoRAのConv2d-3x3の次元数（rank）（デフォルトNone、適用なし）", + ) + parser.add_argument("--device", type=str, default=None, help="device to use, cuda for GPU / 計算を行うデバイス、cuda でGPUを使う") + parser.add_argument( + "--clamp_quantile", + type=float, + default=0.99, + help="Quantile clamping value, float, (0-1). Default = 0.99 / 値をクランプするための分位点、float、(0-1)。デフォルトは0.99", + ) + parser.add_argument( + "--min_diff", + type=float, + default=0.01, + help="Minimum difference between finetuned model and base to consider them different enough to extract, float, (0-1). Default = 0.01 /" + + "LoRAを抽出するために元モデルと派生モデルの差分の最小値、float、(0-1)。デフォルトは0.01", + ) + parser.add_argument( + "--no_metadata", + action="store_true", + help="do not save sai modelspec metadata (minimum ss_metadata for LoRA is saved) / " + + "sai modelspecのメタデータを保存しない（LoRAの最低限のss_metadataは保存される）", + ) + parser.add_argument( + "--load_original_model_to", + type=str, + default=None, + help="location to load original model, cpu or cuda, cuda:0, etc, default is cpu, only for SDXL / 元モデル読み込み先、cpuまたはcuda、cuda:0など、省略時はcpu、SDXLのみ有効", + ) + parser.add_argument( + "--load_tuned_model_to", + type=str, + default=None, + help="location to load tuned model, cpu or cuda, cuda:0, etc, default is cpu, only for SDXL / 派生モデル読み込み先、cpuまたはcuda、cuda:0など、省略時はcpu、SDXLのみ有効", + ) + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + svd(**vars(args)) diff --git a/fine_tune_README_ja.md b/fine_tune_README_ja.md new file mode 100644 index 0000000000000000000000000000000000000000..686947c952b19c016974792cdc5f4f903701cfc9 --- /dev/null +++ b/fine_tune_README_ja.md @@ -0,0 +1,140 @@ +NovelAIの提案した学習手法、自動キャプションニング、タグ付け、Windows＋VRAM 12GB（SD v1.xの場合）環境等に対応したfine tuningです。ここでfine tuningとは、モデルを画像とキャプションで学習することを指します（LoRAやTextual Inversion、Hypernetworksは含みません） + +[学習についての共通ドキュメント](./train_README-ja.md) もあわせてご覧ください。 + +# 概要 + +Diffusersを用いてStable DiffusionのU-Netのfine tuningを行います。NovelAIの記事にある以下の改善に対応しています（Aspect Ratio BucketingについてはNovelAIのコードを参考にしましたが、最終的なコードはすべてオリジナルです）。 + +* CLIP（Text Encoder）の最後の層ではなく最後から二番目の層の出力を用いる。 +* 正方形以外の解像度での学習（Aspect Ratio Bucketing） 。 +* トークン長を75から225に拡張する。 +* BLIPによるキャプショニング（キャプションの自動作成）、DeepDanbooruまたはWD14Taggerによる自動タグ付けを行う。 +* Hypernetworkの学習にも対応する。 +* Stable Diffusion v2.0（baseおよび768/v）に対応。 +* VAEの出力をあらかじめ取得しディスクに保存しておくことで、学習の省メモリ化、高速化を図る。 + +デフォルトではText Encoderの学習は行いません。モデル全体のfine tuningではU-Netだけを学習するのが一般的なようです（NovelAIもそのようです）。オプション指定でText Encoderも学習対象とできます。 + +# 追加機能について + +## CLIPの出力の変更 + +プロンプトを画像に反映するため、テキストの特徴量への変換を行うのがCLIP（Text Encoder）です。Stable DiffusionではCLIPの最後の層の出力を用いていますが、それを最後から二番目の層の出力を用いるよう変更できます。NovelAIによると、これによりより正確にプロンプトが反映されるようになるとのことです。 +元のまま、最後の層の出力を用いることも可能です。 + +※Stable Diffusion 2.0では最後から二番目の層をデフォルトで使います。clip_skipオプションを指定しないでください。 + +## 正方形以外の解像度での学習 + +Stable Diffusionは512\*512で学習されていますが、それに加えて256\*1024や384\*640といった解像度でも学習します。これによりトリミングされる部分が減り、より正しくプロンプトと画像の関係が学習されることが期待されます。 +学習解像度はパラメータとして与えられた解像度の面積（＝メモリ使用量）を超えない範囲で、64ピクセル単位で縦横に調整、作成されます。 + +機械学習では入力サイズをすべて統一するのが一般的ですが、特に制約があるわけではなく、実際は同一のバッチ内で統一されていれば大丈夫です。NovelAIの言うbucketingは、あらかじめ教師データを、アスペクト比に応じた学習解像度ごとに分類しておくことを指しているようです。そしてバッチを各bucket内の画像で作成することで、バッチの画像サイズを統一します。 + +## トークン長の75から225への拡張 + +Stable Diffusionでは最大75トークン（開始・終了を含むと77トークン）ですが、それを225トークンまで拡張します。 +ただしCLIPが受け付ける最大長は75トークンですので、225トークンの場合、単純に三分割してCLIPを呼び出してから結果を連結しています。 + +※これが望ましい実装なのかどうかはいまひとつわかりません。とりあえず動いてはいるようです。特に2.0では何も参考になる実装がないので独自に実装してあります。 + +※Automatic1111氏のWeb UIではカンマを意識して分割、といったこともしているようですが、私の場合はそこまでしておらず単純な分割です。 + +# 学習の手順 + +あらかじめこのリポジトリのREADMEを参照し、環境整備を行ってください。 + +## データの準備 + +[学習データの準備について](./train_README-ja.md) を参照してください。fine tuningではメタデータを用いるfine tuning方式のみ対応しています。 + +## 学習の実行 +たとえば以下のように実行します。以下は省メモリ化のための設定です。それぞれの行を必要に応じて書き換えてください。 + +``` +accelerate launch --num_cpu_threads_per_process 1 fine_tune.py + --pretrained_model_name_or_path=<.ckptまたは.safetensordまたはDiffusers版モデルのディレクトリ> + --output_dir=<学習したモデルの出力先フォルダ> + --output_name=<学習したモデル出力時のファイル名> + --dataset_config=<データ準備で作成した.tomlファイル> + --save_model_as=safetensors + --learning_rate=5e-6 --max_train_steps=10000 + --use_8bit_adam --xformers --gradient_checkpointing + --mixed_precision=fp16 +``` + +`num_cpu_threads_per_process` には通常は1を指定するとよいようです。 + +`pretrained_model_name_or_path` に追加学習を行う元となるモデルを指定します。Stable Diffusionのcheckpointファイル（.ckptまたは.safetensors）、Diffusersのローカルディスクにあるモデルディレクトリ、DiffusersのモデルID（"stabilityai/stable-diffusion-2"など）が指定できます。 + +`output_dir` に学習後のモデルを保存するフォルダを指定します。`output_name` にモデルのファイル名を拡張子を除いて指定します。`save_model_as` でsafetensors形式での保存を指定しています。 + +`dataset_config` に `.toml` ファイルを指定します。ファイル内でのバッチサイズ指定は、当初はメモリ消費を抑えるために `1` としてください。 + +学習させるステップ数 `max_train_steps` を10000とします。学習率 `learning_rate` はここでは5e-6を指定しています。 + +省メモリ化のため `mixed_precision="fp16"` を指定します（RTX30 シリーズ以降では `bf16` も指定できます。環境整備時にaccelerateに行った設定と合わせてください）。また `gradient_checkpointing` を指定します。 + +オプティマイザ（モデルを学習データにあうように最適化＝学習させるクラス）にメモリ消費の少ない 8bit AdamW を使うため、 `optimizer_type="AdamW8bit"` を指定します。 + +`xformers` オプションを指定し、xformersのCrossAttentionを用います。xformersをインストールしていない場合やエラーとなる場合（環境にもよりますが `mixed_precision="no"` の場合など）、代わりに `mem_eff_attn` オプションを指定すると省メモリ版CrossAttentionを使用します（速度は遅くなります）。 + +ある程度メモリがある場合は、`.toml` ファイルを編集してバッチサイズをたとえば `4` くらいに増やしてください（高速化と精度向上の可能性があります）。 + +### よく使われるオプションについて + +以下の場合にはオプションに関するドキュメントを参照してください。 + +- Stable Diffusion 2.xまたはそこからの派生モデルを学習する +- clip skipを2以上を前提としたモデルを学習する +- 75トークンを超えたキャプションで学習する + +### バッチサイズについて + +モデル全体を学習するためLoRA等の学習に比べるとメモリ消費量は多くなります（DreamBoothと同じ）。 + +### 学習率について + +1e-6から5e-6程度が一般的なようです。他のfine tuningの例なども参照してみてください。 + +### 以前の形式のデータセット指定をした場合のコマンドライン + +解像度やバッチサイズをオプションで指定します。コマンドラインの例は以下の通りです。 + +``` +accelerate launch --num_cpu_threads_per_process 1 fine_tune.py + --pretrained_model_name_or_path=model.ckpt + --in_json meta_lat.json + --train_data_dir=train_data + --output_dir=fine_tuned + --shuffle_caption + --train_batch_size=1 --learning_rate=5e-6 --max_train_steps=10000 + --use_8bit_adam --xformers --gradient_checkpointing + --mixed_precision=bf16 + --save_every_n_epochs=4 +``` + + + +# fine tuning特有のその他の主なオプション + +すべてのオプションについては別文書を参照してください。 + +## `train_text_encoder` +Text Encoderも学習対象とします。メモリ使用量が若干増加します。 + +通常のfine tuningではText Encoderは学習対象としませんが（恐らくText Encoderの出力に従うようにU-Netを学習するため）、学習データ数が少ない場合には、DreamBoothのようにText Encoder側に学習させるのも有効的なようです。 + +## `diffusers_xformers` +スクリプト独自のxformers置換機能ではなくDiffusersのxformers機能を利用します。Hypernetworkの学習はできなくなります。 diff --git a/gen_img_README-ja.md b/gen_img_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..8f4442d0074cd69f3218b6a5a0c568b1f0842907 --- /dev/null +++ b/gen_img_README-ja.md @@ -0,0 +1,487 @@ +SD 1.xおよび2.xのモデル、当リポジトリで学習したLoRA、ControlNet（v1.0のみ動作確認）などに対応した、Diffusersベースの推論（画像生成）スクリプトです。コマンドラインから用います。 + +# 概要 + +* Diffusers (v0.10.2) ベースの推論（画像生成）スクリプト。 +* SD 1.xおよび2.x (base/v-parameterization)モデルに対応。 +* txt2img、img2img、inpaintingに対応。 +* 対話モード、およびファイルからのプロンプト読み込み、連続生成に対応。 +* プロンプト1行あたりの生成枚数を指定可能。 +* 全体の繰り返し回数を指定可能。 +* `fp16`だけでなく`bf16`にも対応。 +* xformersに対応し高速生成が可能。 + * xformersにより省メモリ生成を行いますが、Automatic 1111氏のWeb UIほど最適化していないため、512*512の画像生成でおおむね6GB程度のVRAMを使用します。 +* プロンプトの225トークンへの拡張。ネガティブプロンプト、重みづけに対応。 +* Diffusersの各種samplerに対応（Web UIよりもsampler数は少ないです）。 +* Text Encoderのclip skip（最後からn番目の層の出力を用いる）に対応。 +* VAEの別途読み込み。 +* CLIP Guided Stable Diffusion、VGG16 Guided Stable Diffusion、Highres. fix、upscale対応。 + * Highres. fixはWeb UIの実装を全く確認していない独自実装のため、出力結果は異なるかもしれません。 +* LoRA対応。適用率指定、複数LoRA同時利用、重みのマージに対応。 + * Text EncoderとU-Netで別の適用率を指定することはできません。 +* Attention Coupleに対応。 +* ControlNet v1.0に対応。 +* 途中でモデルを切り替えることはできませんが、バッチファイルを組むことで対応できます。 +* 個人的に欲しくなった機能をいろいろ追加。 + +機能追加時にすべてのテストを行っているわけではないため、以前の機能に影響が出て一部機能が動かない可能性があります。何か問題があればお知らせください。 + +# 基本的な使い方 + +## 対話モードでの画像生成 + +以下のように入力してください。 + +```batchfile +python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先> --xformers --fp16 --interactive +``` + +`--ckpt`オプションにモデル（Stable Diffusionのcheckpointファイル、またはDiffusersのモデルフォルダ）、`--outdir`オプションに画像の出力先フォルダを指定します。 + +`--xformers`オプションでxformersの使用を指定します（xformersを使わない場合は外してください）。`--fp16`オプションでfp16（単精度）での推論を行います。RTX 30系のGPUでは `--bf16`オプションでbf16（bfloat16）での推論を行うこともできます。 + +`--interactive`オプションで対話モードを指定しています。 + +Stable Diffusion 2.0（またはそこからの追加学習モデル）を使う場合は`--v2`オプションを追加してください。v-parameterizationを使うモデル（`768-v-ema.ckpt`およびそこからの追加学習モデル）を使う場合はさらに`--v_parameterization`を追加してください。 + +`--v2`の指定有無が間違っているとモデル読み込み時にエラーになります。`--v_parameterization`の指定有無が間違っていると茶色い画像が表示されます。 + +`Type prompt:`と表示されたらプロンプトを入力してください。 + +![image](https://user-images.githubusercontent.com/52813779/235343115-f3b8ac82-456d-4aab-9724-0cc73c4534aa.png) + +※画像が表示されずエラーになる場合、headless（画面表示機能なし）のOpenCVがインストールされているかもしれません。`pip install opencv-python`として通常のOpenCVを入れてください。または`--no_preview`オプションで画像表示を止めてください。 + +画像ウィンドウを選択してから何らかのキーを押すとウィンドウが閉じ、次のプロンプトが入力できます。プロンプトでCtrl+Z、エンターの順に打鍵するとスクリプトを閉じます。 + +## 単一のプロンプトで画像を一括生成 + +以下のように入力します（実際には1行で入力します）。 + +```batchfile +python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先> + --xformers --fp16 --images_per_prompt <生成枚数> --prompt "<プロンプト>" +``` + +`--images_per_prompt`オプションで、プロンプト1件当たりの生成枚数を指定します。`--prompt`オプションでプロンプトを指定します。スペースを含む場合はダブルクォーテーションで囲んでください。 + +`--batch_size`オプションでバッチサイズを指定できます（後述）。 + +## ファイルからプロンプトを読み込み一括生成 + +以下のように入力します。 + +```batchfile +python gen_img_diffusers.py --ckpt <モデル名> --outdir <画像出力先> + --xformers --fp16 --from_file <プロンプトファイル名> +``` + +`--from_file`オプションで、プロンプトが記述されたファイルを指定します。1行1プロンプトで記述してください。`--images_per_prompt`オプションを指定して1行あたり生成枚数を指定できます。 + +## ネガティブプロンプト、重みづけの使用 + +プロンプトオプション（プロンプト内で`--x`のように指定、後述）で`--n`を書くと、以降がネガティブプロンプトとなります。 + +またAUTOMATIC1111氏のWeb UIと同様の `()` や` []` 、`(xxx:1.3)` などによる重みづけが可能です（実装はDiffusersの[Long Prompt Weighting Stable Diffusion](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#long-prompt-weighting-stable-diffusion)からコピーしたものです）。 + +コマンドラインからのプロンプト指定、ファイルからのプロンプト読み込みでも同様に指定できます。 + +![image](https://user-images.githubusercontent.com/52813779/235343128-e79cd768-ec59-46f5-8395-fce9bdc46208.png) + +# 主なオプション + +コマンドラインから指定してください。 + +## モデルの指定 + +- `--ckpt <モデル名>`：モデル名を指定します。`--ckpt`オプションは必須です。Stable Diffusionのcheckpointファイル、またはDiffusersのモデルフォルダ、Hugging FaceのモデルIDを指定できます。 + +- `--v2`：Stable Diffusion 2.x系のモデルを使う場合に指定します。1.x系の場合には指定不要です。 + +- `--v_parameterization`：v-parameterizationを使うモデルを使う場合に指定します（`768-v-ema.ckpt`およびそこからの追加学習モデル、Waifu Diffusion v1.5など）。 + + `--v2`の指定有無が間違っているとモデル読み込み時にエラーになります。`--v_parameterization`の指定有無が間違っていると茶色い画像が表示されます。 + +- `--vae`：使用するVAEを指定します。未指定時はモデル内のVAEを使用します。 + +## 画像生成と出力 + +- `--interactive`：インタラクティブモードで動作します。プロンプトを入力すると画像が生成されます。 + +- `--prompt <プロンプト>`：プロンプトを指定します。スペースを含む場合はダブルクォーテーションで囲んでください。 + +- `--from_file <プロンプトファイル名>`：プロンプトが記述されたファイルを指定します。1行1プロンプトで記述してください。なお画像サイズやguidance scaleはプロンプトオプション（後述）で指定できます。 + +- `--W <画像幅>`：画像の幅を指定します。デフォルトは`512`です。 + +- `--H <画像高さ>`：画像の高さを指定します。デフォルトは`512`です。 + +- `--steps <ステップ数>`：サンプリングステップ数を指定します。デフォルトは`50`です。 + +- `--scale <ガイダンススケール>`：unconditionalガイダンススケールを指定します。デフォルトは`7.5`です。 + +- `--sampler <サンプラー名>`：サンプラーを指定します。デフォルトは`ddim`です。Diffusersで提供されているddim、pndm、dpmsolver、dpmsolver+++、lms、euler、euler_a、が指定可能です（後ろの三つはk_lms、k_euler、k_euler_aでも指定できます）。 + +- `--outdir <画像出力先フォルダ>`：画像の出力先を指定します。 + +- `--images_per_prompt <生成枚数>`：プロンプト1件当たりの生成枚数を指定します。デフォルトは`1`です。 + +- `--clip_skip <スキップ数>`：CLIPの後ろから何番目の層を使うかを指定します。省略時は最後の層を使います。 + +- `--max_embeddings_multiples <倍数>`：CLIPの入出力長をデフォルト（75）の何倍にするかを指定します。未指定時は75のままです。たとえば3を指定すると入出力長が225になります。 + +- `--negative_scale` : uncoditioningのguidance scaleを個別に指定します。[gcem156氏のこちらの記事](https://note.com/gcem156/n/ne9a53e4a6f43)を参考に実装したものです。 + +## メモリ使用量や生成速度の調整 + +- `--batch_size <バッチサイズ>`：バッチサイズを指定します。デフォルトは`1`です。バッチサイズが大きいとメモリを多く消費しますが、生成速度が速くなります。 + +- `--vae_batch_size `：VAEのバッチサイズを指定します。デフォルトはバッチサイズと同じです。 + VAEのほうがメモリを多く消費するため、デノイジング後（stepが100%になった後）でメモリ不足になる場合があります。このような場合にはVAEのバッチサイズを小さくしてください。 + +- `--xformers`：xformersを使う場合に指定します。 + +- `--fp16`：fp16（単精度）での推論を行います。`fp16`と`bf16`をどちらも指定しない場合はfp32（単精度）での推論を行います。 + +- `--bf16`：bf16（bfloat16）での推論を行います。RTX 30系のGPUでのみ指定可能です。`--bf16`オプションはRTX 30系以外のGPUではエラーになります。`fp16`よりも`bf16`のほうが推論結果がNaNになる（真っ黒の画像になる）可能性が低いようです。 + +## 追加ネットワーク（LoRA等）の使用 + +- `--network_module`：使用する追加ネットワークを指定します。LoRAの場合は`--network_module networks.lora`と指定します。複数のLoRAを使用する場合は`--network_module networks.lora networks.lora networks.lora`のように指定します。 + +- `--network_weights`：使用する追加ネットワークの重みファイルを指定します。`--network_weights model.safetensors`のように指定します。複数のLoRAを使用する場合は`--network_weights model1.safetensors model2.safetensors model3.safetensors`のように指定します。引数の数は`--network_module`で指定した数と同じにしてください。 + +- `--network_mul`：使用する追加ネットワークの重みを何倍にするかを指定します。デフォルトは`1`です。`--network_mul 0.8`のように指定します。複数のLoRAを使用する場合は`--network_mul 0.4 0.5 0.7`のように指定します。引数の数は`--network_module`で指定した数と同じにしてください。 + +- `--network_merge`：使用する追加ネットワークの重みを`--network_mul`に指定した重みであらかじめマージします。`--network_pre_calc` と同時に使用できません。プロンプトオプションの`--am`、およびRegional LoRAは使用できなくなりますが、LoRA未使用時と同じ程度まで生成が高速化されます。 + +- `--network_pre_calc`：使用する追加ネットワークの重みを生成ごとにあらかじめ計算します。プロンプトオプションの`--am`が使用できます。LoRA未使用時と同じ程度まで生成は高速化されますが、生成前に重みを計算する時間が必要で、またメモリ使用量も若干増加します。Regional LoRA使用時は無効になります 。 + +# 主なオプションの指定例 + +次は同一プロンプトで64枚をバッチサイズ4で一括生成する例です。 + +```batchfile +python gen_img_diffusers.py --ckpt model.ckpt --outdir outputs + --xformers --fp16 --W 512 --H 704 --scale 12.5 --sampler k_euler_a + --steps 32 --batch_size 4 --images_per_prompt 64 + --prompt "beautiful flowers --n monochrome" +``` + +次はファイルに書かれたプロンプトを、それぞれ10枚ずつ、バッチサイズ4で一括生成する例です。 + +```batchfile +python gen_img_diffusers.py --ckpt model.ckpt --outdir outputs + --xformers --fp16 --W 512 --H 704 --scale 12.5 --sampler k_euler_a + --steps 32 --batch_size 4 --images_per_prompt 10 + --from_file prompts.txt +``` + +Textual Inversion（後述）およびLoRAの使用例です。 + +```batchfile +python gen_img_diffusers.py --ckpt model.safetensors + --scale 8 --steps 48 --outdir txt2img --xformers + --W 512 --H 768 --fp16 --sampler k_euler_a + --textual_inversion_embeddings goodembed.safetensors negprompt.pt + --network_module networks.lora networks.lora + --network_weights model1.safetensors model2.safetensors + --network_mul 0.4 0.8 + --clip_skip 2 --max_embeddings_multiples 1 + --batch_size 8 --images_per_prompt 1 --interactive +``` + +# プロンプトオプション + +プロンプト内で、`--n`のように「ハイフンふたつ+アルファベットn文字」でプロンプトから各種オプションの指定が可能です。対話モード、コマンドライン、ファイル、いずれからプロンプトを指定する場合でも有効です。 + +プロンプトのオプション指定`--n`の前後にはスペースを入れてください。 + +- `--n`：ネガティブプロンプトを指定します。 + +- `--w`：画像幅を指定します。コマンドラインからの指定を上書きします。 + +- `--h`：画像高さを指定します。コマンドラインからの指定を上書きします。 + +- `--s`：ステップ数を指定します。コマンドラインからの指定を上書きします。 + +- `--d`：この画像の乱数seedを指定します。`--images_per_prompt`を指定している場合は「--d 1,2,3,4」のようにカンマ区切りで複数指定してください。 + ※様々な理由により、Web UIとは同じ乱数seedでも生成される画像が異なる場合があります。 + +- `--l`：guidance scaleを指定します。コマンドラインからの指定を上書きします。 + +- `--t`：img2img（後述）のstrengthを指定します。コマンドラインからの指定を上書きします。 + +- `--nl`：ネガティブプロンプトのguidance scaleを指定します（後述）。コマンドラインからの指定を上書きします。 + +- `--am`：追加ネットワークの重みを指定します。コマンドラインからの指定を上書きします。複数の追加ネットワークを使用する場合は`--am 0.8,0.5,0.3`のように __カンマ区切りで__ 指定します。 + +※これらのオプションを指定すると、バッチサイズよりも小さいサイズでバッチが実行される場合があります（これらの値が異なると一括生成できないため）。（あまり気にしなくて大丈夫ですが、ファイルからプロンプトを読み込み生成する場合は、これらの値が同一のプロンプトを並べておくと効率が良くなります。） + +例： +``` +(masterpiece, best quality), 1girl, in shirt and plated skirt, standing at street under cherry blossoms, upper body, [from below], kind smile, looking at another, [goodembed] --n realistic, real life, (negprompt), (lowres:1.1), (worst quality:1.2), (low quality:1.1), bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, normal quality, jpeg artifacts, signature, watermark, username, blurry --w 960 --h 640 --s 28 --d 1 +``` + +![image](https://user-images.githubusercontent.com/52813779/235343446-25654172-fff4-4aaf-977a-20d262b51676.png) + +# img2img + +## オプション + +- `--image_path`：img2imgに利用する画像を指定します。`--image_path template.png`のように指定します。フォルダを指定すると、そのフォルダの画像を順次利用します。 + +- `--strength`：img2imgのstrengthを指定します。`--strength 0.8`のように指定します。デフォルトは`0.8`です。 + +- `--sequential_file_name`：ファイル名を連番にするかどうかを指定します。指定すると生成されるファイル名が`im_000001.png`からの連番になります。 + +- `--use_original_file_name`：指定すると生成ファイル名がオリジナルのファイル名と同じになります。 + +## コマンドラインからの実行例 + +```batchfile +python gen_img_diffusers.py --ckpt trinart_characters_it4_v1_vae_merged.ckpt + --outdir outputs --xformers --fp16 --scale 12.5 --sampler k_euler --steps 32 + --image_path template.png --strength 0.8 + --prompt "1girl, cowboy shot, brown hair, pony tail, brown eyes, + sailor school uniform, outdoors + --n lowres, bad anatomy, bad hands, error, missing fingers, cropped, + worst quality, low quality, normal quality, jpeg artifacts, (blurry), + hair ornament, glasses" + --batch_size 8 --images_per_prompt 32 +``` + +`--image_path`オプションにフォルダを指定すると、そのフォルダの画像を順次読み込みます。生成される枚数は画像枚数ではなく、プロンプト数になりますので、`--images_per_promptPPオプションを指定してimg2imgする画像の枚数とプロンプト数を合わせてください。 + +ファイルはファイル名でソートして読み込みます。なおソート順は文字列順となりますので（`1.jpg→2.jpg→10.jpg`ではなく`1.jpg→10.jpg→2.jpg`の順）、頭を0埋めするなどしてご対応ください（`01.jpg→02.jpg→10.jpg`）。 + +## img2imgを利用したupscale + +img2img時にコマンドラインオプションの`--W`と`--H`で生成画像サイズを指定すると、元画像をそのサイズにリサイズしてからimg2imgを行います。 + +またimg2imgの元画像がこのスクリプトで生成した画像の場合、プロンプトを省略すると、元画像のメタデータからプロンプトを取得しそのまま用います。これによりHighres. fixの2nd stageの動作だけを行うことができます。 + +## img2img時のinpainting + +画像およびマスク画像を指定してinpaintingできます（inpaintingモデルには対応しておらず、単にマスク領域を対象にimg2imgするだけです）。 + +オプションは以下の通りです。 + +- `--mask_image`：マスク画像を指定します。`--img_path`と同様にフォルダを指定すると、そのフォルダの画像を順次利用します。 + +マスク画像はグレースケール画像で、白の部分がinpaintingされます。境界をグラデーションしておくとなんとなく滑らかになりますのでお勧めです。 + +![image](https://user-images.githubusercontent.com/52813779/235343795-9eaa6d98-02ff-4f32-b089-80d1fc482453.png) + +# その他の機能 + +## Textual Inversion + +`--textual_inversion_embeddings`オプションで使用するembeddingsを指定します（複数指定可）。拡張子を除いたファイル名をプロンプト内で使用することで、そのembeddingsを利用します（Web UIと同様の使用法です）。ネガティブプロンプト内でも使用できます。 + +モデルとして、当リポジトリで学習したTextual Inversionモデル、およびWeb UIで学習したTextual Inversionモデル（画像埋め込みは非対応）を利用できます + +## Extended Textual Inversion + +`--textual_inversion_embeddings`の代わりに`--XTI_embeddings`オプションを指定してください。使用法は`--textual_inversion_embeddings`と同じです。 + +## Highres. fix + +AUTOMATIC1111氏のWeb UIにある機能の類似機能です（独自実装のためもしかしたらいろいろ異なるかもしれません）。最初に小さめの画像を生成し、その画像を元にimg2imgすることで、画像全体の破綻を防ぎつつ大きな解像度の画像を生成します。 + +2nd stageのstep数は`--steps` と`--strength`オプションの値から計算されます（`steps*strength`）。 + +img2imgと併用できません。 + +以下のオプションがあります。 + +- `--highres_fix_scale`：Highres. fixを有効にして、1st stageで生成する画像のサイズを、倍率で指定します。最終出力が1024x1024で、最初に512x512の画像を生成する場合は`--highres_fix_scale 0.5`のように指定します。Web UI出の指定の逆数になっていますのでご注意ください。 + +- `--highres_fix_steps`：1st stageの画像のステップ数を指定します。デフォルトは`28`です。 + +- `--highres_fix_save_1st`：1st stageの画像を保存するかどうかを指定します。 + +- `--highres_fix_latents_upscaling`：指定すると2nd stageの画像生成時に1st stageの画像をlatentベースでupscalingします（bilinearのみ対応）。未指定時は画像をLANCZOS4でupscalingします。 + +- `--highres_fix_upscaler`：2nd stageに任意のupscalerを利用します。現在は`--highres_fix_upscaler tools.latent_upscaler` のみ対応しています。 + +- `--highres_fix_upscaler_args`：`--highres_fix_upscaler`で指定したupscalerに渡す引数を指定します。 + `tools.latent_upscaler`の場合は、`--highres_fix_upscaler_args "weights=D:\Work\SD\Models\others\etc\upscaler-v1-e100-220.safetensors"`のように重みファイルを指定します。 + +コマンドラインの例です。 + +```batchfile +python gen_img_diffusers.py --ckpt trinart_characters_it4_v1_vae_merged.ckpt + --n_iter 1 --scale 7.5 --W 1024 --H 1024 --batch_size 1 --outdir ../txt2img + --steps 48 --sampler ddim --fp16 + --xformers + --images_per_prompt 1 --interactive + --highres_fix_scale 0.5 --highres_fix_steps 28 --strength 0.5 +``` + +## ControlNet + +現在はControlNet 1.0のみ動作確認しています。プリプロセスはCannyのみサポートしています。 + +以下のオプションがあります。 + +- `--control_net_models`：ControlNetのモデルファイルを指定します。 + 複数指定すると、それらをstepごとに切り替えて利用します（Web UIのControlNet拡張の実装と異なります）。diffと通常の両方をサポートします。 + +- `--guide_image_path`：ControlNetに使うヒント画像を指定します。`--img_path`と同様にフォルダを指定すると、そのフォルダの画像を順次利用します。Canny以外のモデルの場合には、あらかじめプリプロセスを行っておいてください。 + +- `--control_net_preps`：ControlNetのプリプロセスを指定します。`--control_net_models`と同様に複数指定可能です。現在はcannyのみ対応しています。対象モデルでプリプロセスを使用しない場合は `none` を指定します。 + cannyの場合 `--control_net_preps canny_63_191`のように、閾値1と2を'_'で区切って指定できます。 + +- `--control_net_weights`：ControlNetの適用時の重みを指定します（`1.0`で通常、`0.5`なら半分の影響力で適用）。`--control_net_models`と同様に複数指定可能です。 + +- `--control_net_ratios`：ControlNetを適用するstepの範囲を指定します。`0.5`の場合は、step数の半分までControlNetを適用します。`--control_net_models`と同様に複数指定可能です。 + +コマンドラインの例です。 + +```batchfile +python gen_img_diffusers.py --ckpt model_ckpt --scale 8 --steps 48 --outdir txt2img --xformers + --W 512 --H 768 --bf16 --sampler k_euler_a + --control_net_models diff_control_sd15_canny.safetensors --control_net_weights 1.0 + --guide_image_path guide.png --control_net_ratios 1.0 --interactive +``` + +## Attention Couple + Reginal LoRA + +プロンプトをいくつかの部分に分割し、それぞれのプロンプトを画像内のどの領域に適用するかを指定できる機能です。個別のオプションはありませんが、`mask_path`とプロンプトで指定します。 + +まず、プロンプトで` AND `を利用して、複数部分を定義します。最初の3つに対して領域指定ができ、以降の部分は画像全体へ適用されます。ネガティブプロンプトは画像全体に適用されます。 + +以下ではANDで3つの部分を定義しています。 + +``` +shs 2girls, looking at viewer, smile AND bsb 2girls, looking back AND 2girls --n bad quality, worst quality +``` + +次にマスク画像を用意します。マスク画像はカラーの画像で、RGBの各チャネルがプロンプトのANDで区切られた部分に対応します。またあるチャネルの値がすべて0の場合、画像全体に適用されます。 + +上記の例では、Rチャネルが`shs 2girls, looking at viewer, smile`、Gチャネルが`bsb 2girls, looking back`に、Bチャネルが`2girls`に対応します。次のようなマスク画像を使用すると、Bチャネルに指定がありませんので、`2girls`は画像全体に適用されます。 + +![image](https://user-images.githubusercontent.com/52813779/235343061-b4dc9392-3dae-4831-8347-1e9ae5054251.png) + +マスク画像は`--mask_path`で指定します。現在は1枚のみ対応しています。指定した画像サイズに自動的にリサイズされ適用されます。 + +ControlNetと組み合わせることも可能です（細かい位置指定にはControlNetとの組み合わせを推奨します）。 + +LoRAを指定すると、`--network_weights`で指定した複数のLoRAがそれぞれANDの各部分に対応します。現在の制約として、LoRAの数はANDの部分の数と同じである必要があります。 + +## CLIP Guided Stable Diffusion + +DiffusersのCommunity Examplesの[こちらのcustom pipeline](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#clip-guided-stable-diffusion)からソースをコピー、変更したものです。 + +通常のプロンプトによる生成指定に加えて、追加でより大規模のCLIPでプロンプトのテキストの特徴量を取得し、生成中の画像の特徴量がそのテキストの特徴量に近づくよう、生成される画像をコントロールします（私のざっくりとした理解です）。大きめのCLIPを使いますのでVRAM使用量はかなり増加し（VRAM 8GBでは512*512でも厳しいかもしれません）、生成時間も掛かります。 + +なお選択できるサンプラーはDDIM、PNDM、LMSのみとなります。 + +`--clip_guidance_scale`オプションにどの程度、CLIPの特徴量を反映するかを数値で指定します。先のサンプルでは100になっていますので、そのあたりから始めて増減すると良いようです。 + +デフォルトではプロンプトの先頭75トークン（重みづけの特殊文字を除く）がCLIPに渡されます。プロンプトの`--c`オプションで、通常のプロンプトではなく、CLIPに渡すテキストを別に指定できます（たとえばCLIPはDreamBoothのidentifier（識別子）や「1girl」などのモデル特有の単語は認識できないと思われますので、それらを省いたテキストが良いと思われます）。 + +コマンドラインの例です。 + +```batchfile +python gen_img_diffusers.py --ckpt v1-5-pruned-emaonly.ckpt --n_iter 1 + --scale 2.5 --W 512 --H 512 --batch_size 1 --outdir ../txt2img --steps 36 + --sampler ddim --fp16 --opt_channels_last --xformers --images_per_prompt 1 + --interactive --clip_guidance_scale 100 +``` + +## CLIP Image Guided Stable Diffusion + +テキストではなくCLIPに別の画像を渡し、その特徴量に近づくよう生成をコントロールする機能です。`--clip_image_guidance_scale`オプションで適用量の数値を、`--guide_image_path`オプションでguideに使用する画像（ファイルまたはフォルダ）を指定してください。 + +コマンドラインの例です。 + +```batchfile +python gen_img_diffusers.py --ckpt trinart_characters_it4_v1_vae_merged.ckpt + --n_iter 1 --scale 7.5 --W 512 --H 512 --batch_size 1 --outdir ../txt2img + --steps 80 --sampler ddim --fp16 --opt_channels_last --xformers + --images_per_prompt 1 --interactive --clip_image_guidance_scale 100 + --guide_image_path YUKA160113420I9A4104_TP_V.jpg +``` + +### VGG16 Guided Stable Diffusion + +指定した画像に近づくように画像生成する機能です。通常のプロンプトによる生成指定に加えて、追加でVGG16の特徴量を取得し、生成中の画像が指定したガイド画像に近づくよう、生成される画像をコントロールします。img2imgでの使用をお勧めします（通常の生成では画像がぼやけた感じになります）。CLIP Guided Stable Diffusionの仕組みを流用した独自の機能です。またアイデアはVGGを利用したスタイル変換から拝借しています。 + +なお選択できるサンプラーはDDIM、PNDM、LMSのみとなります。 + +`--vgg16_guidance_scale`オプションにどの程度、VGG16特徴量を反映するかを数値で指定します。試した感じでは100くらいから始めて増減すると良いようです。`--guide_image_path`オプションでguideに使用する画像（ファイルまたはフォルダ）を指定してください。 + +複数枚の画像を一括でimg2img変換し、元画像をガイド画像とする場合、`--guide_image_path`と`--image_path`に同じ値を指定すればOKです。 + +コマンドラインの例です。 + +```batchfile +python gen_img_diffusers.py --ckpt wd-v1-3-full-pruned-half.ckpt + --n_iter 1 --scale 5.5 --steps 60 --outdir ../txt2img + --xformers --sampler ddim --fp16 --W 512 --H 704 + --batch_size 1 --images_per_prompt 1 + --prompt "picturesque, 1girl, solo, anime face, skirt, beautiful face + --n lowres, bad anatomy, bad hands, error, missing fingers, + cropped, worst quality, low quality, normal quality, + jpeg artifacts, blurry, 3d, bad face, monochrome --d 1" + --strength 0.8 --image_path ..\src_image + --vgg16_guidance_scale 100 --guide_image_path ..\src_image +``` + +`--vgg16_guidance_layerPで特徴量取得に使用するVGG16のレイヤー番号を指定できます（デフォルトは20でconv4-2のReLUです）。上の層ほど画風を表現し、下の層ほどコンテンツを表現するといわれています。 + +![image](https://user-images.githubusercontent.com/52813779/235343813-3c1f0d7a-4fb3-4274-98e4-b92d76b551df.png) + +# その他のオプション + +- `--no_preview` : 対話モードでプレビュー画像を表示しません。OpenCVがインストールされていない場合や、出力されたファイルを直接確認する場合に指定してください。 + +- `--n_iter` : 生成を繰り返す回数を指定します。デフォルトは1です。プロンプトをファイルから読み込むとき、複数回の生成を行いたい場合に指定します。 + +- `--tokenizer_cache_dir` : トークナイザーのキャッシュディレクトリを指定します。（作業中） + +- `--seed` : 乱数seedを指定します。1枚生成時はその画像のseed、複数枚生成時は各画像のseedを生成するための乱数のseedになります（`--from_file`で複数画像生成するとき、`--seed`オプションを指定すると複数回実行したときに各画像が同じseedになります）。 + +- `--iter_same_seed` : プロンプトに乱数seedの指定がないとき、`--n_iter`の繰り返し内ではすべて同じseedを使います。`--from_file`で指定した複数のプロンプト間でseedを統一して比較するときに使います。 + +- `--diffusers_xformers` : Diffuserのxformersを使用します。 + +- `--opt_channels_last` : 推論時にテンソルのチャンネルを最後に配置します。場合によっては高速化されることがあります。 + +- `--network_show_meta` : 追加ネットワークのメタデータを表示します。 + + +--- + +# About Gradual Latent + +Gradual Latent is a Hires fix that gradually increases the size of the latent. `gen_img.py`, `sdxl_gen_img.py`, and `gen_img_diffusers.py` have the following options. + +- `--gradual_latent_timesteps`: Specifies the timestep to start increasing the size of the latent. The default is None, which means Gradual Latent is not used. Please try around 750 at first. +- `--gradual_latent_ratio`: Specifies the initial size of the latent. The default is 0.5, which means it starts with half the default latent size. +- `--gradual_latent_ratio_step`: Specifies the ratio to increase the size of the latent. The default is 0.125, which means the latent size is gradually increased to 0.625, 0.75, 0.875, 1.0. +- `--gradual_latent_ratio_every_n_steps`: Specifies the interval to increase the size of the latent. The default is 3, which means the latent size is increased every 3 steps. + +Each option can also be specified with prompt options, `--glt`, `--glr`, `--gls`, `--gle`. + +__Please specify `euler_a` for the sampler.__ Because the source code of the sampler is modified. It will not work with other samplers. + +It is more effective with SD 1.5. It is quite subtle with SDXL. + +# Gradual Latent について + +latentのサイズを徐々に大きくしていくHires fixです。`gen_img.py` 、``sdxl_gen_img.py` 、`gen_img_diffusers.py` に以下のオプションが追加されています。 + +- `--gradual_latent_timesteps` : latentのサイズを大きくし始めるタイムステップを指定します。デフォルトは None で、Gradual Latentを使用しません。750 くらいから始めてみてください。 +- `--gradual_latent_ratio` : latentの初期サイズを指定します。デフォルトは 0.5 で、デフォルトの latent サイズの半分のサイズから始めます。 +- `--gradual_latent_ratio_step`: latentのサイズを大きくする割合を指定します。デフォルトは 0.125 で、latentのサイズを 0.625, 0.75, 0.875, 1.0 と徐々に大きくします。 +- `--gradual_latent_ratio_every_n_steps`: latentのサイズを大きくする間隔を指定します。デフォルトは 3 で、3ステップごとに latent のサイズを大きくします。 + +それぞれのオプションは、プロンプトオプション、`--glt`、`--glr`、`--gls`、`--gle` でも指定できます。 + +サンプラーに手を加えているため、__サンプラーに `euler_a` を指定してください。__ 他のサンプラーでは動作しません。 + +SD 1.5 のほうが効果があります。SDXL ではかなり微妙です。 + diff --git a/gradscaler.py b/gradscaler.py new file mode 100644 index 0000000000000000000000000000000000000000..0a86100958c9254efdc3b6103db7c7cf3693ff59 --- /dev/null +++ b/gradscaler.py @@ -0,0 +1,183 @@ +from collections import defaultdict +import torch +import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import +import intel_extension_for_pytorch._C as core # pylint: disable=import-error, unused-import + +# pylint: disable=protected-access, missing-function-docstring, line-too-long + +device_supports_fp64 = torch.xpu.has_fp64_dtype() if hasattr(torch.xpu, "has_fp64_dtype") else torch.xpu.get_device_properties("xpu").has_fp64 +OptState = ipex.cpu.autocast._grad_scaler.OptState +_MultiDeviceReplicator = ipex.cpu.autocast._grad_scaler._MultiDeviceReplicator +_refresh_per_optimizer_state = ipex.cpu.autocast._grad_scaler._refresh_per_optimizer_state + +def _unscale_grads_(self, optimizer, inv_scale, found_inf, allow_fp16): # pylint: disable=unused-argument + per_device_inv_scale = _MultiDeviceReplicator(inv_scale) + per_device_found_inf = _MultiDeviceReplicator(found_inf) + + # To set up _amp_foreach_non_finite_check_and_unscale_, split grads by device and dtype. + # There could be hundreds of grads, so we'd like to iterate through them just once. + # However, we don't know their devices or dtypes in advance. + + # https://stackoverflow.com/questions/5029934/defaultdict-of-defaultdict + # Google says mypy struggles with defaultdicts type annotations. + per_device_and_dtype_grads = defaultdict(lambda: defaultdict(list)) # type: ignore[var-annotated] + # sync grad to master weight + if hasattr(optimizer, "sync_grad"): + optimizer.sync_grad() + with torch.no_grad(): + for group in optimizer.param_groups: + for param in group["params"]: + if param.grad is None: + continue + if (not allow_fp16) and param.grad.dtype == torch.float16: + raise ValueError("Attempting to unscale FP16 gradients.") + if param.grad.is_sparse: + # is_coalesced() == False means the sparse grad has values with duplicate indices. + # coalesce() deduplicates indices and adds all values that have the same index. + # For scaled fp16 values, there's a good chance coalescing will cause overflow, + # so we should check the coalesced _values(). + if param.grad.dtype is torch.float16: + param.grad = param.grad.coalesce() + to_unscale = param.grad._values() + else: + to_unscale = param.grad + + # -: is there a way to split by device and dtype without appending in the inner loop? + to_unscale = to_unscale.to("cpu") + per_device_and_dtype_grads[to_unscale.device][ + to_unscale.dtype + ].append(to_unscale) + + for _, per_dtype_grads in per_device_and_dtype_grads.items(): + for grads in per_dtype_grads.values(): + core._amp_foreach_non_finite_check_and_unscale_( + grads, + per_device_found_inf.get("cpu"), + per_device_inv_scale.get("cpu"), + ) + + return per_device_found_inf._per_device_tensors + +def unscale_(self, optimizer): + """ + Divides ("unscales") the optimizer's gradient tensors by the scale factor. + :meth:`unscale_` is optional, serving cases where you need to + :ref:`modify or inspect gradients` + between the backward pass(es) and :meth:`step`. + If :meth:`unscale_` is not called explicitly, gradients will be unscaled automatically during :meth:`step`. + Simple example, using :meth:`unscale_` to enable clipping of unscaled gradients:: + ... + scaler.scale(loss).backward() + scaler.unscale_(optimizer) + torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm) + scaler.step(optimizer) + scaler.update() + Args: + optimizer (torch.optim.Optimizer): Optimizer that owns the gradients to be unscaled. + .. warning:: + :meth:`unscale_` should only be called once per optimizer per :meth:`step` call, + and only after all gradients for that optimizer's assigned parameters have been accumulated. + Calling :meth:`unscale_` twice for a given optimizer between each :meth:`step` triggers a RuntimeError. + .. warning:: + :meth:`unscale_` may unscale sparse gradients out of place, replacing the ``.grad`` attribute. + """ + if not self._enabled: + return + + self._check_scale_growth_tracker("unscale_") + + optimizer_state = self._per_optimizer_states[id(optimizer)] + + if optimizer_state["stage"] is OptState.UNSCALED: # pylint: disable=no-else-raise + raise RuntimeError( + "unscale_() has already been called on this optimizer since the last update()." + ) + elif optimizer_state["stage"] is OptState.STEPPED: + raise RuntimeError("unscale_() is being called after step().") + + # FP32 division can be imprecise for certain compile options, so we carry out the reciprocal in FP64. + assert self._scale is not None + if device_supports_fp64: + inv_scale = self._scale.double().reciprocal().float() + else: + inv_scale = self._scale.to("cpu").double().reciprocal().float().to(self._scale.device) + found_inf = torch.full( + (1,), 0.0, dtype=torch.float32, device=self._scale.device + ) + + optimizer_state["found_inf_per_device"] = self._unscale_grads_( + optimizer, inv_scale, found_inf, False + ) + optimizer_state["stage"] = OptState.UNSCALED + +def update(self, new_scale=None): + """ + Updates the scale factor. + If any optimizer steps were skipped the scale is multiplied by ``backoff_factor`` + to reduce it. If ``growth_interval`` unskipped iterations occurred consecutively, + the scale is multiplied by ``growth_factor`` to increase it. + Passing ``new_scale`` sets the new scale value manually. (``new_scale`` is not + used directly, it's used to fill GradScaler's internal scale tensor. So if + ``new_scale`` was a tensor, later in-place changes to that tensor will not further + affect the scale GradScaler uses internally.) + Args: + new_scale (float or :class:`torch.FloatTensor`, optional, default=None): New scale factor. + .. warning:: + :meth:`update` should only be called at the end of the iteration, after ``scaler.step(optimizer)`` has + been invoked for all optimizers used this iteration. + """ + if not self._enabled: + return + + _scale, _growth_tracker = self._check_scale_growth_tracker("update") + + if new_scale is not None: + # Accept a new user-defined scale. + if isinstance(new_scale, float): + self._scale.fill_(new_scale) # type: ignore[union-attr] + else: + reason = "new_scale should be a float or a 1-element torch.FloatTensor with requires_grad=False." + assert isinstance(new_scale, torch.FloatTensor), reason # type: ignore[attr-defined] + assert new_scale.numel() == 1, reason + assert new_scale.requires_grad is False, reason + self._scale.copy_(new_scale) # type: ignore[union-attr] + else: + # Consume shared inf/nan data collected from optimizers to update the scale. + # If all found_inf tensors are on the same device as self._scale, this operation is asynchronous. + found_infs = [ + found_inf.to(device="cpu", non_blocking=True) + for state in self._per_optimizer_states.values() + for found_inf in state["found_inf_per_device"].values() + ] + + assert len(found_infs) > 0, "No inf checks were recorded prior to update." + + found_inf_combined = found_infs[0] + if len(found_infs) > 1: + for i in range(1, len(found_infs)): + found_inf_combined += found_infs[i] + + to_device = _scale.device + _scale = _scale.to("cpu") + _growth_tracker = _growth_tracker.to("cpu") + + core._amp_update_scale_( + _scale, + _growth_tracker, + found_inf_combined, + self._growth_factor, + self._backoff_factor, + self._growth_interval, + ) + + _scale = _scale.to(to_device) + _growth_tracker = _growth_tracker.to(to_device) + # To prepare for next iteration, clear the data collected from optimizers this iteration. + self._per_optimizer_states = defaultdict(_refresh_per_optimizer_state) + +def gradscaler_init(): + torch.xpu.amp.GradScaler = ipex.cpu.autocast._grad_scaler.GradScaler + torch.xpu.amp.GradScaler._unscale_grads_ = _unscale_grads_ + torch.xpu.amp.GradScaler.unscale_ = unscale_ + torch.xpu.amp.GradScaler.update = update + return torch.xpu.amp.GradScaler diff --git a/hijacks.py b/hijacks.py new file mode 100644 index 0000000000000000000000000000000000000000..91569746ad5b193ca46adf2282a2126130675ab2 --- /dev/null +++ b/hijacks.py @@ -0,0 +1,367 @@ +import os +from functools import wraps +from contextlib import nullcontext +import torch +import numpy as np + +device_supports_fp64 = torch.xpu.has_fp64_dtype() if hasattr(torch.xpu, "has_fp64_dtype") else torch.xpu.get_device_properties("xpu").has_fp64 +if os.environ.get('IPEX_FORCE_ATTENTION_SLICE', '0') == '0' and (torch.xpu.get_device_properties("xpu").total_memory / 1024 / 1024 / 1024) > 4.1: + try: + x = torch.ones((33000,33000), dtype=torch.float32, device="xpu") + del x + torch.xpu.empty_cache() + can_allocate_plus_4gb = True + except Exception: + can_allocate_plus_4gb = False +else: + can_allocate_plus_4gb = bool(os.environ.get('IPEX_FORCE_ATTENTION_SLICE', '0') == '-1') + +# pylint: disable=protected-access, missing-function-docstring, line-too-long, unnecessary-lambda, no-else-return + +class DummyDataParallel(torch.nn.Module): # pylint: disable=missing-class-docstring, unused-argument, too-few-public-methods + def __new__(cls, module, device_ids=None, output_device=None, dim=0): # pylint: disable=unused-argument + if isinstance(device_ids, list) and len(device_ids) > 1: + print("IPEX backend doesn't support DataParallel on multiple XPU devices") + return module.to("xpu") + +def return_null_context(*args, **kwargs): # pylint: disable=unused-argument + return nullcontext() + +@property +def is_cuda(self): + return self.device.type == 'xpu' or self.device.type == 'cuda' + +def check_device(device): + return bool((isinstance(device, torch.device) and device.type == "cuda") or (isinstance(device, str) and "cuda" in device) or isinstance(device, int)) + +def return_xpu(device): + return f"xpu:{device.split(':')[-1]}" if isinstance(device, str) and ":" in device else f"xpu:{device}" if isinstance(device, int) else torch.device(f"xpu:{device.index}" if device.index is not None else "xpu") if isinstance(device, torch.device) else "xpu" + + +# Autocast +original_autocast_init = torch.amp.autocast_mode.autocast.__init__ +@wraps(torch.amp.autocast_mode.autocast.__init__) +def autocast_init(self, device_type, dtype=None, enabled=True, cache_enabled=None): + if device_type == "cuda": + return original_autocast_init(self, device_type="xpu", dtype=dtype, enabled=enabled, cache_enabled=cache_enabled) + else: + return original_autocast_init(self, device_type=device_type, dtype=dtype, enabled=enabled, cache_enabled=cache_enabled) + +# Latent Antialias CPU Offload: +original_interpolate = torch.nn.functional.interpolate +@wraps(torch.nn.functional.interpolate) +def interpolate(tensor, size=None, scale_factor=None, mode='nearest', align_corners=None, recompute_scale_factor=None, antialias=False): # pylint: disable=too-many-arguments + if mode in {'bicubic', 'bilinear'}: + return_device = tensor.device + return_dtype = tensor.dtype + return original_interpolate(tensor.to("cpu", dtype=torch.float32), size=size, scale_factor=scale_factor, mode=mode, + align_corners=align_corners, recompute_scale_factor=recompute_scale_factor, antialias=antialias).to(return_device, dtype=return_dtype) + else: + return original_interpolate(tensor, size=size, scale_factor=scale_factor, mode=mode, + align_corners=align_corners, recompute_scale_factor=recompute_scale_factor, antialias=antialias) + + +# Diffusers Float64 (Alchemist GPUs doesn't support 64 bit): +original_from_numpy = torch.from_numpy +@wraps(torch.from_numpy) +def from_numpy(ndarray): + if ndarray.dtype == float: + return original_from_numpy(ndarray.astype('float32')) + else: + return original_from_numpy(ndarray) + +original_as_tensor = torch.as_tensor +@wraps(torch.as_tensor) +def as_tensor(data, dtype=None, device=None): + if check_device(device): + device = return_xpu(device) + if isinstance(data, np.ndarray) and data.dtype == float and not ( + (isinstance(device, torch.device) and device.type == "cpu") or (isinstance(device, str) and "cpu" in device)): + return original_as_tensor(data, dtype=torch.float32, device=device) + else: + return original_as_tensor(data, dtype=dtype, device=device) + + +if can_allocate_plus_4gb: + original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention +else: + # 32 bit attention workarounds for Alchemist: + try: + from .attention import dynamic_scaled_dot_product_attention as original_scaled_dot_product_attention + except Exception: # pylint: disable=broad-exception-caught + original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention + +@wraps(torch.nn.functional.scaled_dot_product_attention) +def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, **kwargs): + if query.dtype != key.dtype: + key = key.to(dtype=query.dtype) + if query.dtype != value.dtype: + value = value.to(dtype=query.dtype) + if attn_mask is not None and query.dtype != attn_mask.dtype: + attn_mask = attn_mask.to(dtype=query.dtype) + return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs) + +# Data Type Errors: +original_torch_bmm = torch.bmm +@wraps(torch.bmm) +def torch_bmm(input, mat2, *, out=None): + if input.dtype != mat2.dtype: + mat2 = mat2.to(input.dtype) + return original_torch_bmm(input, mat2, out=out) + +# Diffusers FreeU +original_fft_fftn = torch.fft.fftn +@wraps(torch.fft.fftn) +def fft_fftn(input, s=None, dim=None, norm=None, *, out=None): + return_dtype = input.dtype + return original_fft_fftn(input.to(dtype=torch.float32), s=s, dim=dim, norm=norm, out=out).to(dtype=return_dtype) + +# Diffusers FreeU +original_fft_ifftn = torch.fft.ifftn +@wraps(torch.fft.ifftn) +def fft_ifftn(input, s=None, dim=None, norm=None, *, out=None): + return_dtype = input.dtype + return original_fft_ifftn(input.to(dtype=torch.float32), s=s, dim=dim, norm=norm, out=out).to(dtype=return_dtype) + +# A1111 FP16 +original_functional_group_norm = torch.nn.functional.group_norm +@wraps(torch.nn.functional.group_norm) +def functional_group_norm(input, num_groups, weight=None, bias=None, eps=1e-05): + if weight is not None and input.dtype != weight.data.dtype: + input = input.to(dtype=weight.data.dtype) + if bias is not None and weight is not None and bias.data.dtype != weight.data.dtype: + bias.data = bias.data.to(dtype=weight.data.dtype) + return original_functional_group_norm(input, num_groups, weight=weight, bias=bias, eps=eps) + +# A1111 BF16 +original_functional_layer_norm = torch.nn.functional.layer_norm +@wraps(torch.nn.functional.layer_norm) +def functional_layer_norm(input, normalized_shape, weight=None, bias=None, eps=1e-05): + if weight is not None and input.dtype != weight.data.dtype: + input = input.to(dtype=weight.data.dtype) + if bias is not None and weight is not None and bias.data.dtype != weight.data.dtype: + bias.data = bias.data.to(dtype=weight.data.dtype) + return original_functional_layer_norm(input, normalized_shape, weight=weight, bias=bias, eps=eps) + +# Training +original_functional_linear = torch.nn.functional.linear +@wraps(torch.nn.functional.linear) +def functional_linear(input, weight, bias=None): + if input.dtype != weight.data.dtype: + input = input.to(dtype=weight.data.dtype) + if bias is not None and bias.data.dtype != weight.data.dtype: + bias.data = bias.data.to(dtype=weight.data.dtype) + return original_functional_linear(input, weight, bias=bias) + +original_functional_conv1d = torch.nn.functional.conv1d +@wraps(torch.nn.functional.conv1d) +def functional_conv1d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1): + if input.dtype != weight.data.dtype: + input = input.to(dtype=weight.data.dtype) + if bias is not None and bias.data.dtype != weight.data.dtype: + bias.data = bias.data.to(dtype=weight.data.dtype) + return original_functional_conv1d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups) + +original_functional_conv2d = torch.nn.functional.conv2d +@wraps(torch.nn.functional.conv2d) +def functional_conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1): + if input.dtype != weight.data.dtype: + input = input.to(dtype=weight.data.dtype) + if bias is not None and bias.data.dtype != weight.data.dtype: + bias.data = bias.data.to(dtype=weight.data.dtype) + return original_functional_conv2d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups) + +# LTX Video +original_functional_conv3d = torch.nn.functional.conv3d +@wraps(torch.nn.functional.conv3d) +def functional_conv3d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1): + if input.dtype != weight.data.dtype: + input = input.to(dtype=weight.data.dtype) + if bias is not None and bias.data.dtype != weight.data.dtype: + bias.data = bias.data.to(dtype=weight.data.dtype) + return original_functional_conv3d(input, weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups) + +# SwinIR BF16: +original_functional_pad = torch.nn.functional.pad +@wraps(torch.nn.functional.pad) +def functional_pad(input, pad, mode='constant', value=None): + if mode == 'reflect' and input.dtype == torch.bfloat16: + return original_functional_pad(input.to(torch.float32), pad, mode=mode, value=value).to(dtype=torch.bfloat16) + else: + return original_functional_pad(input, pad, mode=mode, value=value) + + +original_torch_tensor = torch.tensor +@wraps(torch.tensor) +def torch_tensor(data, *args, dtype=None, device=None, **kwargs): + global device_supports_fp64 + if check_device(device): + device = return_xpu(device) + if not device_supports_fp64: + if (isinstance(device, torch.device) and device.type == "xpu") or (isinstance(device, str) and "xpu" in device): + if dtype == torch.float64: + dtype = torch.float32 + elif dtype is None and (hasattr(data, "dtype") and (data.dtype == torch.float64 or data.dtype == float)): + dtype = torch.float32 + return original_torch_tensor(data, *args, dtype=dtype, device=device, **kwargs) + +original_Tensor_to = torch.Tensor.to +@wraps(torch.Tensor.to) +def Tensor_to(self, device=None, *args, **kwargs): + if check_device(device): + return original_Tensor_to(self, return_xpu(device), *args, **kwargs) + else: + return original_Tensor_to(self, device, *args, **kwargs) + +original_Tensor_cuda = torch.Tensor.cuda +@wraps(torch.Tensor.cuda) +def Tensor_cuda(self, device=None, *args, **kwargs): + if check_device(device): + return original_Tensor_cuda(self, return_xpu(device), *args, **kwargs) + else: + return original_Tensor_cuda(self, device, *args, **kwargs) + +original_Tensor_pin_memory = torch.Tensor.pin_memory +@wraps(torch.Tensor.pin_memory) +def Tensor_pin_memory(self, device=None, *args, **kwargs): + if device is None: + device = "xpu" + if check_device(device): + return original_Tensor_pin_memory(self, return_xpu(device), *args, **kwargs) + else: + return original_Tensor_pin_memory(self, device, *args, **kwargs) + +original_UntypedStorage_init = torch.UntypedStorage.__init__ +@wraps(torch.UntypedStorage.__init__) +def UntypedStorage_init(*args, device=None, **kwargs): + if check_device(device): + return original_UntypedStorage_init(*args, device=return_xpu(device), **kwargs) + else: + return original_UntypedStorage_init(*args, device=device, **kwargs) + +original_UntypedStorage_cuda = torch.UntypedStorage.cuda +@wraps(torch.UntypedStorage.cuda) +def UntypedStorage_cuda(self, device=None, *args, **kwargs): + if check_device(device): + return original_UntypedStorage_cuda(self, return_xpu(device), *args, **kwargs) + else: + return original_UntypedStorage_cuda(self, device, *args, **kwargs) + +original_torch_empty = torch.empty +@wraps(torch.empty) +def torch_empty(*args, device=None, **kwargs): + if check_device(device): + return original_torch_empty(*args, device=return_xpu(device), **kwargs) + else: + return original_torch_empty(*args, device=device, **kwargs) + +original_torch_randn = torch.randn +@wraps(torch.randn) +def torch_randn(*args, device=None, dtype=None, **kwargs): + if dtype is bytes: + dtype = None + if check_device(device): + return original_torch_randn(*args, device=return_xpu(device), **kwargs) + else: + return original_torch_randn(*args, device=device, **kwargs) + +original_torch_ones = torch.ones +@wraps(torch.ones) +def torch_ones(*args, device=None, **kwargs): + if check_device(device): + return original_torch_ones(*args, device=return_xpu(device), **kwargs) + else: + return original_torch_ones(*args, device=device, **kwargs) + +original_torch_zeros = torch.zeros +@wraps(torch.zeros) +def torch_zeros(*args, device=None, **kwargs): + if check_device(device): + return original_torch_zeros(*args, device=return_xpu(device), **kwargs) + else: + return original_torch_zeros(*args, device=device, **kwargs) + +original_torch_full = torch.full +@wraps(torch.full) +def torch_full(*args, device=None, **kwargs): + if check_device(device): + return original_torch_full(*args, device=return_xpu(device), **kwargs) + else: + return original_torch_full(*args, device=device, **kwargs) + +original_torch_linspace = torch.linspace +@wraps(torch.linspace) +def torch_linspace(*args, device=None, **kwargs): + if check_device(device): + return original_torch_linspace(*args, device=return_xpu(device), **kwargs) + else: + return original_torch_linspace(*args, device=device, **kwargs) + +original_torch_load = torch.load +@wraps(torch.load) +def torch_load(f, map_location=None, *args, **kwargs): + if map_location is None: + map_location = "xpu" + if check_device(map_location): + return original_torch_load(f, *args, map_location=return_xpu(map_location), **kwargs) + else: + return original_torch_load(f, *args, map_location=map_location, **kwargs) + +original_torch_Generator = torch.Generator +@wraps(torch.Generator) +def torch_Generator(device=None): + if check_device(device): + return original_torch_Generator(return_xpu(device)) + else: + return original_torch_Generator(device) + +@wraps(torch.cuda.synchronize) +def torch_cuda_synchronize(device=None): + if check_device(device): + return torch.xpu.synchronize(return_xpu(device)) + else: + return torch.xpu.synchronize(device) + + +# Hijack Functions: +def ipex_hijacks(legacy=True): + global device_supports_fp64, can_allocate_plus_4gb + if legacy and float(torch.__version__[:3]) < 2.5: + torch.nn.functional.interpolate = interpolate + torch.tensor = torch_tensor + torch.Tensor.to = Tensor_to + torch.Tensor.cuda = Tensor_cuda + torch.Tensor.pin_memory = Tensor_pin_memory + torch.UntypedStorage.__init__ = UntypedStorage_init + torch.UntypedStorage.cuda = UntypedStorage_cuda + torch.empty = torch_empty + torch.randn = torch_randn + torch.ones = torch_ones + torch.zeros = torch_zeros + torch.full = torch_full + torch.linspace = torch_linspace + torch.load = torch_load + torch.Generator = torch_Generator + torch.cuda.synchronize = torch_cuda_synchronize + + torch.backends.cuda.sdp_kernel = return_null_context + torch.nn.DataParallel = DummyDataParallel + torch.UntypedStorage.is_cuda = is_cuda + torch.amp.autocast_mode.autocast.__init__ = autocast_init + + torch.nn.functional.scaled_dot_product_attention = scaled_dot_product_attention + torch.nn.functional.group_norm = functional_group_norm + torch.nn.functional.layer_norm = functional_layer_norm + torch.nn.functional.linear = functional_linear + torch.nn.functional.conv1d = functional_conv1d + torch.nn.functional.conv2d = functional_conv2d + torch.nn.functional.conv3d = functional_conv3d + torch.nn.functional.pad = functional_pad + + torch.bmm = torch_bmm + torch.fft.fftn = fft_fftn + torch.fft.ifftn = fft_ifftn + if not device_supports_fp64: + torch.from_numpy = from_numpy + torch.as_tensor = as_tensor + return device_supports_fp64, can_allocate_plus_4gb diff --git a/huggingface_util.py b/huggingface_util.py new file mode 100644 index 0000000000000000000000000000000000000000..57b19d982c40414f6626aa8ed85280b716044716 --- /dev/null +++ b/huggingface_util.py @@ -0,0 +1,84 @@ +from typing import Union, BinaryIO +from huggingface_hub import HfApi +from pathlib import Path +import argparse +import os +from library.utils import fire_in_thread +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def exists_repo(repo_id: str, repo_type: str, revision: str = "main", token: str = None): + api = HfApi( + token=token, + ) + try: + api.repo_info(repo_id=repo_id, revision=revision, repo_type=repo_type) + return True + except: + return False + + +def upload( + args: argparse.Namespace, + src: Union[str, Path, bytes, BinaryIO], + dest_suffix: str = "", + force_sync_upload: bool = False, +): + repo_id = args.huggingface_repo_id + repo_type = args.huggingface_repo_type + token = args.huggingface_token + path_in_repo = args.huggingface_path_in_repo + dest_suffix if args.huggingface_path_in_repo is not None else None + private = args.huggingface_repo_visibility is None or args.huggingface_repo_visibility != "public" + api = HfApi(token=token) + if not exists_repo(repo_id=repo_id, repo_type=repo_type, token=token): + try: + api.create_repo(repo_id=repo_id, repo_type=repo_type, private=private) + except Exception as e: # とりあえずRepositoryNotFoundErrorは確認したが他にあると困るので + logger.error("===========================================") + logger.error(f"failed to create HuggingFace repo / HuggingFaceのリポジトリの作成に失敗しました : {e}") + logger.error("===========================================") + + is_folder = (type(src) == str and os.path.isdir(src)) or (isinstance(src, Path) and src.is_dir()) + + def uploader(): + try: + if is_folder: + api.upload_folder( + repo_id=repo_id, + repo_type=repo_type, + folder_path=src, + path_in_repo=path_in_repo, + ) + else: + api.upload_file( + repo_id=repo_id, + repo_type=repo_type, + path_or_fileobj=src, + path_in_repo=path_in_repo, + ) + except Exception as e: # RuntimeErrorを確認済みだが他にあると困るので + logger.error("===========================================") + logger.error(f"failed to upload to HuggingFace / HuggingFaceへのアップロードに失敗しました : {e}") + logger.error("===========================================") + + if args.async_upload and not force_sync_upload: + fire_in_thread(uploader) + else: + uploader() + + +def list_dir( + repo_id: str, + subfolder: str, + repo_type: str, + revision: str = "main", + token: str = None, +): + api = HfApi( + token=token, + ) + repo_info = api.repo_info(repo_id=repo_id, revision=revision, repo_type=repo_type) + file_list = [file for file in repo_info.siblings if file.rfilename.startswith(subfolder)] + return file_list diff --git a/hypernetwork.py b/hypernetwork.py new file mode 100644 index 0000000000000000000000000000000000000000..fbd3fb24e1a5bc314b282407d1c6282a197d96a3 --- /dev/null +++ b/hypernetwork.py @@ -0,0 +1,223 @@ +import torch +import torch.nn.functional as F +from diffusers.models.attention_processor import ( + Attention, + AttnProcessor2_0, + SlicedAttnProcessor, + XFormersAttnProcessor +) + +try: + import xformers.ops +except: + xformers = None + + +loaded_networks = [] + + +def apply_single_hypernetwork( + hypernetwork, hidden_states, encoder_hidden_states +): + context_k, context_v = hypernetwork.forward(hidden_states, encoder_hidden_states) + return context_k, context_v + + +def apply_hypernetworks(context_k, context_v, layer=None): + if len(loaded_networks) == 0: + return context_v, context_v + for hypernetwork in loaded_networks: + context_k, context_v = hypernetwork.forward(context_k, context_v) + + context_k = context_k.to(dtype=context_k.dtype) + context_v = context_v.to(dtype=context_k.dtype) + + return context_k, context_v + + + +def xformers_forward( + self: XFormersAttnProcessor, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor = None, + attention_mask: torch.Tensor = None, +): + batch_size, sequence_length, _ = ( + hidden_states.shape + if encoder_hidden_states is None + else encoder_hidden_states.shape + ) + + attention_mask = attn.prepare_attention_mask( + attention_mask, sequence_length, batch_size + ) + + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + context_k, context_v = apply_hypernetworks(hidden_states, encoder_hidden_states) + + key = attn.to_k(context_k) + value = attn.to_v(context_v) + + query = attn.head_to_batch_dim(query).contiguous() + key = attn.head_to_batch_dim(key).contiguous() + value = attn.head_to_batch_dim(value).contiguous() + + hidden_states = xformers.ops.memory_efficient_attention( + query, + key, + value, + attn_bias=attention_mask, + op=self.attention_op, + scale=attn.scale, + ) + hidden_states = hidden_states.to(query.dtype) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + return hidden_states + + +def sliced_attn_forward( + self: SlicedAttnProcessor, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor = None, + attention_mask: torch.Tensor = None, +): + batch_size, sequence_length, _ = ( + hidden_states.shape + if encoder_hidden_states is None + else encoder_hidden_states.shape + ) + attention_mask = attn.prepare_attention_mask( + attention_mask, sequence_length, batch_size + ) + + query = attn.to_q(hidden_states) + dim = query.shape[-1] + query = attn.head_to_batch_dim(query) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + context_k, context_v = apply_hypernetworks(hidden_states, encoder_hidden_states) + + key = attn.to_k(context_k) + value = attn.to_v(context_v) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + batch_size_attention, query_tokens, _ = query.shape + hidden_states = torch.zeros( + (batch_size_attention, query_tokens, dim // attn.heads), + device=query.device, + dtype=query.dtype, + ) + + for i in range(batch_size_attention // self.slice_size): + start_idx = i * self.slice_size + end_idx = (i + 1) * self.slice_size + + query_slice = query[start_idx:end_idx] + key_slice = key[start_idx:end_idx] + attn_mask_slice = ( + attention_mask[start_idx:end_idx] if attention_mask is not None else None + ) + + attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice) + + attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx]) + + hidden_states[start_idx:end_idx] = attn_slice + + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + return hidden_states + + +def v2_0_forward( + self: AttnProcessor2_0, + attn: Attention, + hidden_states, + encoder_hidden_states=None, + attention_mask=None, +): + batch_size, sequence_length, _ = ( + hidden_states.shape + if encoder_hidden_states is None + else encoder_hidden_states.shape + ) + inner_dim = hidden_states.shape[-1] + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask( + attention_mask, sequence_length, batch_size + ) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view( + batch_size, attn.heads, -1, attention_mask.shape[-1] + ) + + query = attn.to_q(hidden_states) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + context_k, context_v = apply_hypernetworks(hidden_states, encoder_hidden_states) + + key = attn.to_k(context_k) + value = attn.to_v(context_v) + + head_dim = inner_dim // attn.heads + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states = hidden_states.transpose(1, 2).reshape( + batch_size, -1, attn.heads * head_dim + ) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states) + # dropout + hidden_states = attn.to_out[1](hidden_states) + return hidden_states + + +def replace_attentions_for_hypernetwork(): + import diffusers.models.attention_processor + + diffusers.models.attention_processor.XFormersAttnProcessor.__call__ = ( + xformers_forward + ) + diffusers.models.attention_processor.SlicedAttnProcessor.__call__ = ( + sliced_attn_forward + ) + diffusers.models.attention_processor.AttnProcessor2_0.__call__ = v2_0_forward diff --git a/hypernetwork_nai.py b/hypernetwork_nai.py new file mode 100644 index 0000000000000000000000000000000000000000..dcaaa714a08bb2cfc417d827e8bdd01c8c1ad367 --- /dev/null +++ b/hypernetwork_nai.py @@ -0,0 +1,96 @@ +# NAI compatible + +import torch + + +class HypernetworkModule(torch.nn.Module): + def __init__(self, dim, multiplier=1.0): + super().__init__() + + linear1 = torch.nn.Linear(dim, dim * 2) + linear2 = torch.nn.Linear(dim * 2, dim) + linear1.weight.data.normal_(mean=0.0, std=0.01) + linear1.bias.data.zero_() + linear2.weight.data.normal_(mean=0.0, std=0.01) + linear2.bias.data.zero_() + linears = [linear1, linear2] + + self.linear = torch.nn.Sequential(*linears) + self.multiplier = multiplier + + def forward(self, x): + return x + self.linear(x) * self.multiplier + + +class Hypernetwork(torch.nn.Module): + enable_sizes = [320, 640, 768, 1280] + # return self.modules[Hypernetwork.enable_sizes.index(size)] + + def __init__(self, multiplier=1.0) -> None: + super().__init__() + self.modules = [] + for size in Hypernetwork.enable_sizes: + self.modules.append((HypernetworkModule(size, multiplier), HypernetworkModule(size, multiplier))) + self.register_module(f"{size}_0", self.modules[-1][0]) + self.register_module(f"{size}_1", self.modules[-1][1]) + + def apply_to_stable_diffusion(self, text_encoder, vae, unet): + blocks = unet.input_blocks + [unet.middle_block] + unet.output_blocks + for block in blocks: + for subblk in block: + if 'SpatialTransformer' in str(type(subblk)): + for tf_block in subblk.transformer_blocks: + for attn in [tf_block.attn1, tf_block.attn2]: + size = attn.context_dim + if size in Hypernetwork.enable_sizes: + attn.hypernetwork = self + else: + attn.hypernetwork = None + + def apply_to_diffusers(self, text_encoder, vae, unet): + blocks = unet.down_blocks + [unet.mid_block] + unet.up_blocks + for block in blocks: + if hasattr(block, 'attentions'): + for subblk in block.attentions: + if 'SpatialTransformer' in str(type(subblk)) or 'Transformer2DModel' in str(type(subblk)): # 0.6.0 and 0.7~ + for tf_block in subblk.transformer_blocks: + for attn in [tf_block.attn1, tf_block.attn2]: + size = attn.to_k.in_features + if size in Hypernetwork.enable_sizes: + attn.hypernetwork = self + else: + attn.hypernetwork = None + return True # TODO error checking + + def forward(self, x, context): + size = context.shape[-1] + assert size in Hypernetwork.enable_sizes + module = self.modules[Hypernetwork.enable_sizes.index(size)] + return module[0].forward(context), module[1].forward(context) + + def load_from_state_dict(self, state_dict): + # old ver to new ver + changes = { + 'linear1.bias': 'linear.0.bias', + 'linear1.weight': 'linear.0.weight', + 'linear2.bias': 'linear.1.bias', + 'linear2.weight': 'linear.1.weight', + } + for key_from, key_to in changes.items(): + if key_from in state_dict: + state_dict[key_to] = state_dict[key_from] + del state_dict[key_from] + + for size, sd in state_dict.items(): + if type(size) == int: + self.modules[Hypernetwork.enable_sizes.index(size)][0].load_state_dict(sd[0], strict=True) + self.modules[Hypernetwork.enable_sizes.index(size)][1].load_state_dict(sd[1], strict=True) + return True + + def get_state_dict(self): + state_dict = {} + for i, size in enumerate(Hypernetwork.enable_sizes): + sd0 = self.modules[i][0].state_dict() + sd1 = self.modules[i][1].state_dict() + state_dict[size] = [sd0, sd1] + return state_dict diff --git a/latent_upscaler.py b/latent_upscaler.py new file mode 100644 index 0000000000000000000000000000000000000000..f05cf71942613b58f6d657d63a12fc14fc081ba0 --- /dev/null +++ b/latent_upscaler.py @@ -0,0 +1,354 @@ +# 外部から簡単にupscalerを呼ぶためのスクリプト +# 単体で動くようにモデル定義も含めている + +import argparse +import glob +import os +import cv2 +from diffusers import AutoencoderKL + +from typing import Dict, List +import numpy as np + +import torch +from library.device_utils import init_ipex, get_preferred_device +init_ipex() + +from torch import nn +from tqdm import tqdm +from PIL import Image +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +class ResidualBlock(nn.Module): + def __init__(self, in_channels, out_channels=None, kernel_size=3, stride=1, padding=1): + super(ResidualBlock, self).__init__() + + if out_channels is None: + out_channels = in_channels + + self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, bias=False) + self.bn1 = nn.BatchNorm2d(out_channels) + self.relu1 = nn.ReLU(inplace=True) + + self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size, stride, padding, bias=False) + self.bn2 = nn.BatchNorm2d(out_channels) + + self.relu2 = nn.ReLU(inplace=True) # このReLUはresidualに足す前にかけるほうがいいかも + + # initialize weights + self._initialize_weights() + + def _initialize_weights(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu") + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + nn.init.constant_(m.bias, 0) + + def forward(self, x): + residual = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu1(out) + + out = self.conv2(out) + out = self.bn2(out) + + out += residual + + out = self.relu2(out) + + return out + + +class Upscaler(nn.Module): + def __init__(self): + super(Upscaler, self).__init__() + + # define layers + # latent has 4 channels + + self.conv1 = nn.Conv2d(4, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) + self.bn1 = nn.BatchNorm2d(128) + self.relu1 = nn.ReLU(inplace=True) + + # resblocks + # 数の暴力で20個：次元数を増やすよりもブロックを増やしたほうがreceptive fieldが広がるはずだぞ + self.resblock1 = ResidualBlock(128) + self.resblock2 = ResidualBlock(128) + self.resblock3 = ResidualBlock(128) + self.resblock4 = ResidualBlock(128) + self.resblock5 = ResidualBlock(128) + self.resblock6 = ResidualBlock(128) + self.resblock7 = ResidualBlock(128) + self.resblock8 = ResidualBlock(128) + self.resblock9 = ResidualBlock(128) + self.resblock10 = ResidualBlock(128) + self.resblock11 = ResidualBlock(128) + self.resblock12 = ResidualBlock(128) + self.resblock13 = ResidualBlock(128) + self.resblock14 = ResidualBlock(128) + self.resblock15 = ResidualBlock(128) + self.resblock16 = ResidualBlock(128) + self.resblock17 = ResidualBlock(128) + self.resblock18 = ResidualBlock(128) + self.resblock19 = ResidualBlock(128) + self.resblock20 = ResidualBlock(128) + + # last convs + self.conv2 = nn.Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) + self.bn2 = nn.BatchNorm2d(64) + self.relu2 = nn.ReLU(inplace=True) + + self.conv3 = nn.Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) + self.bn3 = nn.BatchNorm2d(64) + self.relu3 = nn.ReLU(inplace=True) + + # final conv: output 4 channels + self.conv_final = nn.Conv2d(64, 4, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0)) + + # initialize weights + self._initialize_weights() + + def _initialize_weights(self): + for m in self.modules(): + if isinstance(m, nn.Conv2d): + nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu") + if m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.BatchNorm2d): + nn.init.constant_(m.weight, 1) + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.Linear): + nn.init.normal_(m.weight, 0, 0.01) + nn.init.constant_(m.bias, 0) + + # initialize final conv weights to 0: 流行りのzero conv + nn.init.constant_(self.conv_final.weight, 0) + + def forward(self, x): + inp = x + + x = self.conv1(x) + x = self.bn1(x) + x = self.relu1(x) + + # いくつかのresblockを通した後に、residualを足すことで精度向上と学習速度向上が見込めるはず + residual = x + x = self.resblock1(x) + x = self.resblock2(x) + x = self.resblock3(x) + x = self.resblock4(x) + x = x + residual + residual = x + x = self.resblock5(x) + x = self.resblock6(x) + x = self.resblock7(x) + x = self.resblock8(x) + x = x + residual + residual = x + x = self.resblock9(x) + x = self.resblock10(x) + x = self.resblock11(x) + x = self.resblock12(x) + x = x + residual + residual = x + x = self.resblock13(x) + x = self.resblock14(x) + x = self.resblock15(x) + x = self.resblock16(x) + x = x + residual + residual = x + x = self.resblock17(x) + x = self.resblock18(x) + x = self.resblock19(x) + x = self.resblock20(x) + x = x + residual + + x = self.conv2(x) + x = self.bn2(x) + x = self.relu2(x) + x = self.conv3(x) + x = self.bn3(x) + + # ここにreluを入れないほうがいい気がする + + x = self.conv_final(x) + + # network estimates the difference between the input and the output + x = x + inp + + return x + + def support_latents(self) -> bool: + return False + + def upscale( + self, + vae: AutoencoderKL, + lowreso_images: List[Image.Image], + lowreso_latents: torch.Tensor, + dtype: torch.dtype, + width: int, + height: int, + batch_size: int = 1, + vae_batch_size: int = 1, + ): + # assertion + assert lowreso_images is not None, "Upscaler requires lowreso image" + + # make upsampled image with lanczos4 + upsampled_images = [] + for lowreso_image in lowreso_images: + upsampled_image = np.array(lowreso_image.resize((width, height), Image.LANCZOS)) + upsampled_images.append(upsampled_image) + + # convert to tensor: this tensor is too large to be converted to cuda + upsampled_images = [torch.from_numpy(upsampled_image).permute(2, 0, 1).float() for upsampled_image in upsampled_images] + upsampled_images = torch.stack(upsampled_images, dim=0) + upsampled_images = upsampled_images.to(dtype) + + # normalize to [-1, 1] + upsampled_images = upsampled_images / 127.5 - 1.0 + + # convert upsample images to latents with batch size + # logger.info("Encoding upsampled (LANCZOS4) images...") + upsampled_latents = [] + for i in tqdm(range(0, upsampled_images.shape[0], vae_batch_size)): + batch = upsampled_images[i : i + vae_batch_size].to(vae.device) + with torch.no_grad(): + batch = vae.encode(batch).latent_dist.sample() + upsampled_latents.append(batch) + + upsampled_latents = torch.cat(upsampled_latents, dim=0) + + # upscale (refine) latents with this model with batch size + logger.info("Upscaling latents...") + upscaled_latents = [] + for i in range(0, upsampled_latents.shape[0], batch_size): + with torch.no_grad(): + upscaled_latents.append(self.forward(upsampled_latents[i : i + batch_size])) + upscaled_latents = torch.cat(upscaled_latents, dim=0) + + return upscaled_latents * 0.18215 + + +# external interface: returns a model +def create_upscaler(**kwargs): + weights = kwargs["weights"] + model = Upscaler() + + logger.info(f"Loading weights from {weights}...") + if os.path.splitext(weights)[1] == ".safetensors": + from safetensors.torch import load_file + + sd = load_file(weights) + else: + sd = torch.load(weights, map_location=torch.device("cpu")) + model.load_state_dict(sd) + return model + + +# another interface: upscale images with a model for given images from command line +def upscale_images(args: argparse.Namespace): + DEVICE = get_preferred_device() + us_dtype = torch.float16 # TODO: support fp32/bf16 + os.makedirs(args.output_dir, exist_ok=True) + + # load VAE with Diffusers + assert args.vae_path is not None, "VAE path is required" + logger.info(f"Loading VAE from {args.vae_path}...") + vae = AutoencoderKL.from_pretrained(args.vae_path, subfolder="vae") + vae.to(DEVICE, dtype=us_dtype) + + # prepare model + logger.info("Preparing model...") + upscaler: Upscaler = create_upscaler(weights=args.weights) + # logger.info("Loading weights from", args.weights) + # upscaler.load_state_dict(torch.load(args.weights)) + upscaler.eval() + upscaler.to(DEVICE, dtype=us_dtype) + + # load images + image_paths = glob.glob(args.image_pattern) + images = [] + for image_path in image_paths: + image = Image.open(image_path) + image = image.convert("RGB") + + # make divisible by 8 + width = image.width + height = image.height + if width % 8 != 0: + width = width - (width % 8) + if height % 8 != 0: + height = height - (height % 8) + if width != image.width or height != image.height: + image = image.crop((0, 0, width, height)) + + images.append(image) + + # debug output + if args.debug: + for image, image_path in zip(images, image_paths): + image_debug = image.resize((image.width * 2, image.height * 2), Image.LANCZOS) + + basename = os.path.basename(image_path) + basename_wo_ext, ext = os.path.splitext(basename) + dest_file_name = os.path.join(args.output_dir, f"{basename_wo_ext}_lanczos4{ext}") + image_debug.save(dest_file_name) + + # upscale + logger.info("Upscaling...") + upscaled_latents = upscaler.upscale( + vae, images, None, us_dtype, width * 2, height * 2, batch_size=args.batch_size, vae_batch_size=args.vae_batch_size + ) + upscaled_latents /= 0.18215 + + # decode with batch + logger.info("Decoding...") + upscaled_images = [] + for i in tqdm(range(0, upscaled_latents.shape[0], args.vae_batch_size)): + with torch.no_grad(): + batch = vae.decode(upscaled_latents[i : i + args.vae_batch_size]).sample + batch = batch.to("cpu") + upscaled_images.append(batch) + upscaled_images = torch.cat(upscaled_images, dim=0) + + # tensor to numpy + upscaled_images = upscaled_images.permute(0, 2, 3, 1).numpy() + upscaled_images = (upscaled_images + 1.0) * 127.5 + upscaled_images = upscaled_images.clip(0, 255).astype(np.uint8) + + upscaled_images = upscaled_images[..., ::-1] + + # save images + for i, image in enumerate(upscaled_images): + basename = os.path.basename(image_paths[i]) + basename_wo_ext, ext = os.path.splitext(basename) + dest_file_name = os.path.join(args.output_dir, f"{basename_wo_ext}_upscaled{ext}") + cv2.imwrite(dest_file_name, image) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--vae_path", type=str, default=None, help="VAE path") + parser.add_argument("--weights", type=str, default=None, help="Weights path") + parser.add_argument("--image_pattern", type=str, default=None, help="Image pattern") + parser.add_argument("--output_dir", type=str, default=".", help="Output directory") + parser.add_argument("--batch_size", type=int, default=4, help="Batch size") + parser.add_argument("--vae_batch_size", type=int, default=1, help="VAE batch size") + parser.add_argument("--debug", action="store_true", help="Debug mode") + + args = parser.parse_args() + upscale_images(args) diff --git a/libbitsandbytes_cpu.dll b/libbitsandbytes_cpu.dll new file mode 100644 index 0000000000000000000000000000000000000000..b733af475eb02eb04f5ad8cbf0530a78a58bc758 Binary files /dev/null and b/libbitsandbytes_cpu.dll differ diff --git a/libbitsandbytes_cuda116.dll b/libbitsandbytes_cuda116.dll new file mode 100644 index 0000000000000000000000000000000000000000..84ef81251d1c54b7c62a6856856332811f6c80ba --- /dev/null +++ b/libbitsandbytes_cuda116.dll @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88f7bd2916ca3effc43f88492f1e1b9088d13cb5be3b4a3a4aede6aa3bf8d412 +size 4724224 diff --git a/libbitsandbytes_cuda118.dll b/libbitsandbytes_cuda118.dll new file mode 100644 index 0000000000000000000000000000000000000000..13b1423b30bb5ee3f434a99cee0d54a6ba3452ff --- /dev/null +++ b/libbitsandbytes_cuda118.dll @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dc34709b8dcb078cbcdd65e5684f116cb395644d12b9c9fb144af5455bb1c18 +size 14026752 diff --git a/logo_aihub.png b/logo_aihub.png new file mode 100644 index 0000000000000000000000000000000000000000..662b331a9674bd9a34d3bfa782dba427871e7de0 Binary files /dev/null and b/logo_aihub.png differ diff --git a/lora.py b/lora.py new file mode 100644 index 0000000000000000000000000000000000000000..1699a60ff0832e0718dfee7df975150107de63cc --- /dev/null +++ b/lora.py @@ -0,0 +1,1410 @@ +# LoRA network module +# reference: +# https://github.com/microsoft/LoRA/blob/main/loralib/layers.py +# https://github.com/cloneofsimo/lora/blob/master/lora_diffusion/lora.py + +import math +import os +from typing import Dict, List, Optional, Tuple, Type, Union +from diffusers import AutoencoderKL +from transformers import CLIPTextModel +import numpy as np +import torch +import re +from library.utils import setup_logging +from library.sdxl_original_unet import SdxlUNet2DConditionModel + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +RE_UPDOWN = re.compile(r"(up|down)_blocks_(\d+)_(resnets|upsamplers|downsamplers|attentions)_(\d+)_") + + +class LoRAModule(torch.nn.Module): + """ + replaces forward method of the original Linear, instead of replacing the original Linear module. + """ + + def __init__( + self, + lora_name, + org_module: torch.nn.Module, + multiplier=1.0, + lora_dim=4, + alpha=1, + dropout=None, + rank_dropout=None, + module_dropout=None, + ): + """if alpha == 0 or None, alpha is rank (no scaling).""" + super().__init__() + self.lora_name = lora_name + + if org_module.__class__.__name__ == "Conv2d": + in_dim = org_module.in_channels + out_dim = org_module.out_channels + else: + in_dim = org_module.in_features + out_dim = org_module.out_features + + # if limit_rank: + # self.lora_dim = min(lora_dim, in_dim, out_dim) + # if self.lora_dim != lora_dim: + # logger.info(f"{lora_name} dim (rank) is changed to: {self.lora_dim}") + # else: + self.lora_dim = lora_dim + + if org_module.__class__.__name__ == "Conv2d": + kernel_size = org_module.kernel_size + stride = org_module.stride + padding = org_module.padding + self.lora_down = torch.nn.Conv2d(in_dim, self.lora_dim, kernel_size, stride, padding, bias=False) + self.lora_up = torch.nn.Conv2d(self.lora_dim, out_dim, (1, 1), (1, 1), bias=False) + else: + self.lora_down = torch.nn.Linear(in_dim, self.lora_dim, bias=False) + self.lora_up = torch.nn.Linear(self.lora_dim, out_dim, bias=False) + + if type(alpha) == torch.Tensor: + alpha = alpha.detach().float().numpy() # without casting, bf16 causes error + alpha = self.lora_dim if alpha is None or alpha == 0 else alpha + self.scale = alpha / self.lora_dim + self.register_buffer("alpha", torch.tensor(alpha)) # 定数として扱える + + # same as microsoft's + torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5)) + torch.nn.init.zeros_(self.lora_up.weight) + + self.multiplier = multiplier + self.org_module = org_module # remove in applying + self.dropout = dropout + self.rank_dropout = rank_dropout + self.module_dropout = module_dropout + + def apply_to(self): + self.org_forward = self.org_module.forward + self.org_module.forward = self.forward + del self.org_module + + def forward(self, x): + org_forwarded = self.org_forward(x) + + # module dropout + if self.module_dropout is not None and self.training: + if torch.rand(1) < self.module_dropout: + return org_forwarded + + lx = self.lora_down(x) + + # normal dropout + if self.dropout is not None and self.training: + lx = torch.nn.functional.dropout(lx, p=self.dropout) + + # rank dropout + if self.rank_dropout is not None and self.training: + mask = torch.rand((lx.size(0), self.lora_dim), device=lx.device) > self.rank_dropout + if len(lx.size()) == 3: + mask = mask.unsqueeze(1) # for Text Encoder + elif len(lx.size()) == 4: + mask = mask.unsqueeze(-1).unsqueeze(-1) # for Conv2d + lx = lx * mask + + # scaling for rank dropout: treat as if the rank is changed + # maskから計算することも考えられるが、augmentation的な効果を期待してrank_dropoutを用いる + scale = self.scale * (1.0 / (1.0 - self.rank_dropout)) # redundant for readability + else: + scale = self.scale + + lx = self.lora_up(lx) + + return org_forwarded + lx * self.multiplier * scale + + +class LoRAInfModule(LoRAModule): + def __init__( + self, + lora_name, + org_module: torch.nn.Module, + multiplier=1.0, + lora_dim=4, + alpha=1, + **kwargs, + ): + # no dropout for inference + super().__init__(lora_name, org_module, multiplier, lora_dim, alpha) + + self.org_module_ref = [org_module] # 後から参照できるように + self.enabled = True + + # check regional or not by lora_name + self.text_encoder = False + if lora_name.startswith("lora_te_"): + self.regional = False + self.use_sub_prompt = True + self.text_encoder = True + elif "attn2_to_k" in lora_name or "attn2_to_v" in lora_name: + self.regional = False + self.use_sub_prompt = True + elif "time_emb" in lora_name: + self.regional = False + self.use_sub_prompt = False + else: + self.regional = True + self.use_sub_prompt = False + + self.network: LoRANetwork = None + + def set_network(self, network): + self.network = network + + # freezeしてマージする + def merge_to(self, sd, dtype, device): + # get up/down weight + up_weight = sd["lora_up.weight"].to(torch.float).to(device) + down_weight = sd["lora_down.weight"].to(torch.float).to(device) + + # extract weight from org_module + org_sd = self.org_module.state_dict() + weight = org_sd["weight"].to(torch.float) + + # merge weight + if len(weight.size()) == 2: + # linear + weight = weight + self.multiplier * (up_weight @ down_weight) * self.scale + elif down_weight.size()[2:4] == (1, 1): + # conv2d 1x1 + weight = ( + weight + + self.multiplier + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * self.scale + ) + else: + # conv2d 3x3 + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + # logger.info(conved.size(), weight.size(), module.stride, module.padding) + weight = weight + self.multiplier * conved * self.scale + + # set weight to org_module + org_sd["weight"] = weight.to(dtype) + self.org_module.load_state_dict(org_sd) + + # 復元できるマージのため、このモジュールのweightを返す + def get_weight(self, multiplier=None): + if multiplier is None: + multiplier = self.multiplier + + # get up/down weight from module + up_weight = self.lora_up.weight.to(torch.float) + down_weight = self.lora_down.weight.to(torch.float) + + # pre-calculated weight + if len(down_weight.size()) == 2: + # linear + weight = self.multiplier * (up_weight @ down_weight) * self.scale + elif down_weight.size()[2:4] == (1, 1): + # conv2d 1x1 + weight = ( + self.multiplier + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * self.scale + ) + else: + # conv2d 3x3 + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + weight = self.multiplier * conved * self.scale + + return weight + + def set_region(self, region): + self.region = region + self.region_mask = None + + def default_forward(self, x): + # logger.info(f"default_forward {self.lora_name} {x.size()}") + return self.org_forward(x) + self.lora_up(self.lora_down(x)) * self.multiplier * self.scale + + def forward(self, x): + if not self.enabled: + return self.org_forward(x) + + if self.network is None or self.network.sub_prompt_index is None: + return self.default_forward(x) + if not self.regional and not self.use_sub_prompt: + return self.default_forward(x) + + if self.regional: + return self.regional_forward(x) + else: + return self.sub_prompt_forward(x) + + def get_mask_for_x(self, x): + # calculate size from shape of x + if len(x.size()) == 4: + h, w = x.size()[2:4] + area = h * w + else: + area = x.size()[1] + + mask = self.network.mask_dic.get(area, None) + if mask is None or len(x.size()) == 2: + # emb_layers in SDXL doesn't have mask + # if "emb" not in self.lora_name: + # print(f"mask is None for resolution {self.lora_name}, {area}, {x.size()}") + mask_size = (1, x.size()[1]) if len(x.size()) == 2 else (1, *x.size()[1:-1], 1) + return torch.ones(mask_size, dtype=x.dtype, device=x.device) / self.network.num_sub_prompts + if len(x.size()) == 3: + mask = torch.reshape(mask, (1, -1, 1)) + return mask + + def regional_forward(self, x): + if "attn2_to_out" in self.lora_name: + return self.to_out_forward(x) + + if self.network.mask_dic is None: # sub_prompt_index >= 3 + return self.default_forward(x) + + # apply mask for LoRA result + lx = self.lora_up(self.lora_down(x)) * self.multiplier * self.scale + mask = self.get_mask_for_x(lx) + # print("regional", self.lora_name, self.network.sub_prompt_index, lx.size(), mask.size()) + # if mask.ndim > lx.ndim: # in some resolution, lx is 2d and mask is 3d (the reason is not checked) + # mask = mask.squeeze(-1) + lx = lx * mask + + x = self.org_forward(x) + x = x + lx + + if "attn2_to_q" in self.lora_name and self.network.is_last_network: + x = self.postp_to_q(x) + + return x + + def postp_to_q(self, x): + # repeat x to num_sub_prompts + has_real_uncond = x.size()[0] // self.network.batch_size == 3 + qc = self.network.batch_size # uncond + qc += self.network.batch_size * self.network.num_sub_prompts # cond + if has_real_uncond: + qc += self.network.batch_size # real_uncond + + query = torch.zeros((qc, x.size()[1], x.size()[2]), device=x.device, dtype=x.dtype) + query[: self.network.batch_size] = x[: self.network.batch_size] + + for i in range(self.network.batch_size): + qi = self.network.batch_size + i * self.network.num_sub_prompts + query[qi : qi + self.network.num_sub_prompts] = x[self.network.batch_size + i] + + if has_real_uncond: + query[-self.network.batch_size :] = x[-self.network.batch_size :] + + # logger.info(f"postp_to_q {self.lora_name} {x.size()} {query.size()} {self.network.num_sub_prompts}") + return query + + def sub_prompt_forward(self, x): + if x.size()[0] == self.network.batch_size: # if uncond in text_encoder, do not apply LoRA + return self.org_forward(x) + + emb_idx = self.network.sub_prompt_index + if not self.text_encoder: + emb_idx += self.network.batch_size + + # apply sub prompt of X + lx = x[emb_idx :: self.network.num_sub_prompts] + lx = self.lora_up(self.lora_down(lx)) * self.multiplier * self.scale + + # logger.info(f"sub_prompt_forward {self.lora_name} {x.size()} {lx.size()} {emb_idx}") + + x = self.org_forward(x) + x[emb_idx :: self.network.num_sub_prompts] += lx + + return x + + def to_out_forward(self, x): + # logger.info(f"to_out_forward {self.lora_name} {x.size()} {self.network.is_last_network}") + + if self.network.is_last_network: + masks = [None] * self.network.num_sub_prompts + self.network.shared[self.lora_name] = (None, masks) + else: + lx, masks = self.network.shared[self.lora_name] + + # call own LoRA + x1 = x[self.network.batch_size + self.network.sub_prompt_index :: self.network.num_sub_prompts] + lx1 = self.lora_up(self.lora_down(x1)) * self.multiplier * self.scale + + if self.network.is_last_network: + lx = torch.zeros( + (self.network.num_sub_prompts * self.network.batch_size, *lx1.size()[1:]), device=lx1.device, dtype=lx1.dtype + ) + self.network.shared[self.lora_name] = (lx, masks) + + # logger.info(f"to_out_forward {lx.size()} {lx1.size()} {self.network.sub_prompt_index} {self.network.num_sub_prompts}") + lx[self.network.sub_prompt_index :: self.network.num_sub_prompts] += lx1 + masks[self.network.sub_prompt_index] = self.get_mask_for_x(lx1) + + # if not last network, return x and masks + x = self.org_forward(x) + if not self.network.is_last_network: + return x + + lx, masks = self.network.shared.pop(self.lora_name) + + # if last network, combine separated x with mask weighted sum + has_real_uncond = x.size()[0] // self.network.batch_size == self.network.num_sub_prompts + 2 + + out = torch.zeros((self.network.batch_size * (3 if has_real_uncond else 2), *x.size()[1:]), device=x.device, dtype=x.dtype) + out[: self.network.batch_size] = x[: self.network.batch_size] # uncond + if has_real_uncond: + out[-self.network.batch_size :] = x[-self.network.batch_size :] # real_uncond + + # logger.info(f"to_out_forward {self.lora_name} {self.network.sub_prompt_index} {self.network.num_sub_prompts}") + # if num_sub_prompts > num of LoRAs, fill with zero + for i in range(len(masks)): + if masks[i] is None: + masks[i] = torch.zeros_like(masks[0]) + + mask = torch.cat(masks) + mask_sum = torch.sum(mask, dim=0) + 1e-4 + for i in range(self.network.batch_size): + # 1枚の画像ごとに処理する + lx1 = lx[i * self.network.num_sub_prompts : (i + 1) * self.network.num_sub_prompts] + lx1 = lx1 * mask + lx1 = torch.sum(lx1, dim=0) + + xi = self.network.batch_size + i * self.network.num_sub_prompts + x1 = x[xi : xi + self.network.num_sub_prompts] + x1 = x1 * mask + x1 = torch.sum(x1, dim=0) + x1 = x1 / mask_sum + + x1 = x1 + lx1 + out[self.network.batch_size + i] = x1 + + # logger.info(f"to_out_forward {x.size()} {out.size()} {has_real_uncond}") + return out + + +def parse_block_lr_kwargs(is_sdxl: bool, nw_kwargs: Dict) -> Optional[List[float]]: + down_lr_weight = nw_kwargs.get("down_lr_weight", None) + mid_lr_weight = nw_kwargs.get("mid_lr_weight", None) + up_lr_weight = nw_kwargs.get("up_lr_weight", None) + + # 以上のいずれにも設定がない場合は無効としてNoneを返す + if down_lr_weight is None and mid_lr_weight is None and up_lr_weight is None: + return None + + # extract learning rate weight for each block + if down_lr_weight is not None: + # if some parameters are not set, use zero + if "," in down_lr_weight: + down_lr_weight = [(float(s) if s else 0.0) for s in down_lr_weight.split(",")] + + if mid_lr_weight is not None: + mid_lr_weight = [(float(s) if s else 0.0) for s in mid_lr_weight.split(",")] + + if up_lr_weight is not None: + if "," in up_lr_weight: + up_lr_weight = [(float(s) if s else 0.0) for s in up_lr_weight.split(",")] + + return get_block_lr_weight( + is_sdxl, down_lr_weight, mid_lr_weight, up_lr_weight, float(nw_kwargs.get("block_lr_zero_threshold", 0.0)) + ) + + +def create_network( + multiplier: float, + network_dim: Optional[int], + network_alpha: Optional[float], + vae: AutoencoderKL, + text_encoder: Union[CLIPTextModel, List[CLIPTextModel]], + unet, + neuron_dropout: Optional[float] = None, + **kwargs, +): + # if unet is an instance of SdxlUNet2DConditionModel or subclass, set is_sdxl to True + is_sdxl = unet is not None and issubclass(unet.__class__, SdxlUNet2DConditionModel) + + if network_dim is None: + network_dim = 4 # default + if network_alpha is None: + network_alpha = 1.0 + + # extract dim/alpha for conv2d, and block dim + conv_dim = kwargs.get("conv_dim", None) + conv_alpha = kwargs.get("conv_alpha", None) + if conv_dim is not None: + conv_dim = int(conv_dim) + if conv_alpha is None: + conv_alpha = 1.0 + else: + conv_alpha = float(conv_alpha) + + # block dim/alpha/lr + block_dims = kwargs.get("block_dims", None) + block_lr_weight = parse_block_lr_kwargs(is_sdxl, kwargs) + + # 以上のいずれかに指定があればblockごとのdim(rank)を有効にする + if block_dims is not None or block_lr_weight is not None: + block_alphas = kwargs.get("block_alphas", None) + conv_block_dims = kwargs.get("conv_block_dims", None) + conv_block_alphas = kwargs.get("conv_block_alphas", None) + + block_dims, block_alphas, conv_block_dims, conv_block_alphas = get_block_dims_and_alphas( + is_sdxl, block_dims, block_alphas, network_dim, network_alpha, conv_block_dims, conv_block_alphas, conv_dim, conv_alpha + ) + + # remove block dim/alpha without learning rate + block_dims, block_alphas, conv_block_dims, conv_block_alphas = remove_block_dims_and_alphas( + is_sdxl, block_dims, block_alphas, conv_block_dims, conv_block_alphas, block_lr_weight + ) + + else: + block_alphas = None + conv_block_dims = None + conv_block_alphas = None + + # rank/module dropout + rank_dropout = kwargs.get("rank_dropout", None) + if rank_dropout is not None: + rank_dropout = float(rank_dropout) + module_dropout = kwargs.get("module_dropout", None) + if module_dropout is not None: + module_dropout = float(module_dropout) + + # すごく引数が多いな ( ^ω^)･･･ + network = LoRANetwork( + text_encoder, + unet, + multiplier=multiplier, + lora_dim=network_dim, + alpha=network_alpha, + dropout=neuron_dropout, + rank_dropout=rank_dropout, + module_dropout=module_dropout, + conv_lora_dim=conv_dim, + conv_alpha=conv_alpha, + block_dims=block_dims, + block_alphas=block_alphas, + conv_block_dims=conv_block_dims, + conv_block_alphas=conv_block_alphas, + varbose=True, + is_sdxl=is_sdxl, + ) + + loraplus_lr_ratio = kwargs.get("loraplus_lr_ratio", None) + loraplus_unet_lr_ratio = kwargs.get("loraplus_unet_lr_ratio", None) + loraplus_text_encoder_lr_ratio = kwargs.get("loraplus_text_encoder_lr_ratio", None) + loraplus_lr_ratio = float(loraplus_lr_ratio) if loraplus_lr_ratio is not None else None + loraplus_unet_lr_ratio = float(loraplus_unet_lr_ratio) if loraplus_unet_lr_ratio is not None else None + loraplus_text_encoder_lr_ratio = float(loraplus_text_encoder_lr_ratio) if loraplus_text_encoder_lr_ratio is not None else None + if loraplus_lr_ratio is not None or loraplus_unet_lr_ratio is not None or loraplus_text_encoder_lr_ratio is not None: + network.set_loraplus_lr_ratio(loraplus_lr_ratio, loraplus_unet_lr_ratio, loraplus_text_encoder_lr_ratio) + + if block_lr_weight is not None: + network.set_block_lr_weight(block_lr_weight) + + return network + + +# このメソッドは外部から呼び出される可能性を考慮しておく +# network_dim, network_alpha にはデフォルト値が入っている。 +# block_dims, block_alphas は両方ともNoneまたは両方とも値が入っている +# conv_dim, conv_alpha は両方ともNoneまたは両方とも値が入っている +def get_block_dims_and_alphas( + is_sdxl, block_dims, block_alphas, network_dim, network_alpha, conv_block_dims, conv_block_alphas, conv_dim, conv_alpha +): + if not is_sdxl: + num_total_blocks = LoRANetwork.NUM_OF_BLOCKS * 2 + LoRANetwork.NUM_OF_MID_BLOCKS + else: + # 1+9+3+9+1=23, no LoRA for emb_layers (0) + num_total_blocks = 1 + LoRANetwork.SDXL_NUM_OF_BLOCKS * 2 + LoRANetwork.SDXL_NUM_OF_MID_BLOCKS + 1 + + def parse_ints(s): + return [int(i) for i in s.split(",")] + + def parse_floats(s): + return [float(i) for i in s.split(",")] + + # block_dimsとblock_alphasをパースする。必ず値が入る + if block_dims is not None: + block_dims = parse_ints(block_dims) + assert len(block_dims) == num_total_blocks, ( + f"block_dims must have {num_total_blocks} elements but {len(block_dims)} elements are given" + + f" / block_dimsは{num_total_blocks}個指定してください（指定された個数: {len(block_dims)}）" + ) + else: + logger.warning( + f"block_dims is not specified. all dims are set to {network_dim} / block_dimsが指定されていません。すべてのdimは{network_dim}になります" + ) + block_dims = [network_dim] * num_total_blocks + + if block_alphas is not None: + block_alphas = parse_floats(block_alphas) + assert ( + len(block_alphas) == num_total_blocks + ), f"block_alphas must have {num_total_blocks} elements / block_alphasは{num_total_blocks}個指定してください" + else: + logger.warning( + f"block_alphas is not specified. all alphas are set to {network_alpha} / block_alphasが指定されていません。すべてのalphaは{network_alpha}になります" + ) + block_alphas = [network_alpha] * num_total_blocks + + # conv_block_dimsとconv_block_alphasを、指定がある場合のみパースする。指定がなければconv_dimとconv_alphaを使う + if conv_block_dims is not None: + conv_block_dims = parse_ints(conv_block_dims) + assert ( + len(conv_block_dims) == num_total_blocks + ), f"conv_block_dims must have {num_total_blocks} elements / conv_block_dimsは{num_total_blocks}個指定してください" + + if conv_block_alphas is not None: + conv_block_alphas = parse_floats(conv_block_alphas) + assert ( + len(conv_block_alphas) == num_total_blocks + ), f"conv_block_alphas must have {num_total_blocks} elements / conv_block_alphasは{num_total_blocks}個指定してください" + else: + if conv_alpha is None: + conv_alpha = 1.0 + logger.warning( + f"conv_block_alphas is not specified. all alphas are set to {conv_alpha} / conv_block_alphasが指定されていません。すべてのalphaは{conv_alpha}になります" + ) + conv_block_alphas = [conv_alpha] * num_total_blocks + else: + if conv_dim is not None: + logger.warning( + f"conv_dim/alpha for all blocks are set to {conv_dim} and {conv_alpha} / すべてのブロックのconv_dimとalphaは{conv_dim}および{conv_alpha}になります" + ) + conv_block_dims = [conv_dim] * num_total_blocks + conv_block_alphas = [conv_alpha] * num_total_blocks + else: + conv_block_dims = None + conv_block_alphas = None + + return block_dims, block_alphas, conv_block_dims, conv_block_alphas + + +# 層別学習率用に層ごとの学習率に対する倍率を定義する、外部から呼び出せるようにclass外に出しておく +# 戻り値は block ごとの倍率のリスト +def get_block_lr_weight( + is_sdxl, + down_lr_weight: Union[str, List[float]], + mid_lr_weight: List[float], + up_lr_weight: Union[str, List[float]], + zero_threshold: float, +) -> Optional[List[float]]: + # パラメータ未指定時は何もせず、今までと同じ動作とする + if up_lr_weight is None and mid_lr_weight is None and down_lr_weight is None: + return None + + if not is_sdxl: + max_len_for_down_or_up = LoRANetwork.NUM_OF_BLOCKS + max_len_for_mid = LoRANetwork.NUM_OF_MID_BLOCKS + else: + max_len_for_down_or_up = LoRANetwork.SDXL_NUM_OF_BLOCKS + max_len_for_mid = LoRANetwork.SDXL_NUM_OF_MID_BLOCKS + + def get_list(name_with_suffix) -> List[float]: + import math + + tokens = name_with_suffix.split("+") + name = tokens[0] + base_lr = float(tokens[1]) if len(tokens) > 1 else 0.0 + + if name == "cosine": + return [ + math.sin(math.pi * (i / (max_len_for_down_or_up - 1)) / 2) + base_lr + for i in reversed(range(max_len_for_down_or_up)) + ] + elif name == "sine": + return [math.sin(math.pi * (i / (max_len_for_down_or_up - 1)) / 2) + base_lr for i in range(max_len_for_down_or_up)] + elif name == "linear": + return [i / (max_len_for_down_or_up - 1) + base_lr for i in range(max_len_for_down_or_up)] + elif name == "reverse_linear": + return [i / (max_len_for_down_or_up - 1) + base_lr for i in reversed(range(max_len_for_down_or_up))] + elif name == "zeros": + return [0.0 + base_lr] * max_len_for_down_or_up + else: + logger.error( + "Unknown lr_weight argument %s is used. Valid arguments: / 不明なlr_weightの引数 %s が使われました。有効な引数:\n\tcosine, sine, linear, reverse_linear, zeros" + % (name) + ) + return None + + if type(down_lr_weight) == str: + down_lr_weight = get_list(down_lr_weight) + if type(up_lr_weight) == str: + up_lr_weight = get_list(up_lr_weight) + + if (up_lr_weight != None and len(up_lr_weight) > max_len_for_down_or_up) or ( + down_lr_weight != None and len(down_lr_weight) > max_len_for_down_or_up + ): + logger.warning("down_weight or up_weight is too long. Parameters after %d-th are ignored." % max_len_for_down_or_up) + logger.warning("down_weightもしくはup_weightが長すぎます。%d個目以降のパラメータは無視されます。" % max_len_for_down_or_up) + up_lr_weight = up_lr_weight[:max_len_for_down_or_up] + down_lr_weight = down_lr_weight[:max_len_for_down_or_up] + + if mid_lr_weight != None and len(mid_lr_weight) > max_len_for_mid: + logger.warning("mid_weight is too long. Parameters after %d-th are ignored." % max_len_for_mid) + logger.warning("mid_weightが長すぎます。%d個目以降のパラメータは無視されます。" % max_len_for_mid) + mid_lr_weight = mid_lr_weight[:max_len_for_mid] + + if (up_lr_weight != None and len(up_lr_weight) < max_len_for_down_or_up) or ( + down_lr_weight != None and len(down_lr_weight) < max_len_for_down_or_up + ): + logger.warning("down_weight or up_weight is too short. Parameters after %d-th are filled with 1." % max_len_for_down_or_up) + logger.warning( + "down_weightもしくはup_weightが短すぎます。%d個目までの不足したパラメータは1で補われます。" % max_len_for_down_or_up + ) + + if down_lr_weight != None and len(down_lr_weight) < max_len_for_down_or_up: + down_lr_weight = down_lr_weight + [1.0] * (max_len_for_down_or_up - len(down_lr_weight)) + if up_lr_weight != None and len(up_lr_weight) < max_len_for_down_or_up: + up_lr_weight = up_lr_weight + [1.0] * (max_len_for_down_or_up - len(up_lr_weight)) + + if mid_lr_weight != None and len(mid_lr_weight) < max_len_for_mid: + logger.warning("mid_weight is too short. Parameters after %d-th are filled with 1." % max_len_for_mid) + logger.warning("mid_weightが短すぎます。%d個目までの不足したパラメータは1で補われます。" % max_len_for_mid) + mid_lr_weight = mid_lr_weight + [1.0] * (max_len_for_mid - len(mid_lr_weight)) + + if (up_lr_weight != None) or (mid_lr_weight != None) or (down_lr_weight != None): + logger.info("apply block learning rate / 階層別学習率を適用します。") + if down_lr_weight != None: + down_lr_weight = [w if w > zero_threshold else 0 for w in down_lr_weight] + logger.info(f"down_lr_weight (shallower -> deeper, 浅い層->深い層): {down_lr_weight}") + else: + down_lr_weight = [1.0] * max_len_for_down_or_up + logger.info("down_lr_weight: all 1.0, すべて1.0") + + if mid_lr_weight != None: + mid_lr_weight = [w if w > zero_threshold else 0 for w in mid_lr_weight] + logger.info(f"mid_lr_weight: {mid_lr_weight}") + else: + mid_lr_weight = [1.0] * max_len_for_mid + logger.info("mid_lr_weight: all 1.0, すべて1.0") + + if up_lr_weight != None: + up_lr_weight = [w if w > zero_threshold else 0 for w in up_lr_weight] + logger.info(f"up_lr_weight (deeper -> shallower, 深い層->浅い層): {up_lr_weight}") + else: + up_lr_weight = [1.0] * max_len_for_down_or_up + logger.info("up_lr_weight: all 1.0, すべて1.0") + + lr_weight = down_lr_weight + mid_lr_weight + up_lr_weight + + if is_sdxl: + lr_weight = [1.0] + lr_weight + [1.0] # add 1.0 for emb_layers and out + + assert (not is_sdxl and len(lr_weight) == LoRANetwork.NUM_OF_BLOCKS * 2 + LoRANetwork.NUM_OF_MID_BLOCKS) or ( + is_sdxl and len(lr_weight) == 1 + LoRANetwork.SDXL_NUM_OF_BLOCKS * 2 + LoRANetwork.SDXL_NUM_OF_MID_BLOCKS + 1 + ), f"lr_weight length is invalid: {len(lr_weight)}" + + return lr_weight + + +# lr_weightが0のblockをblock_dimsから除外する、外部から呼び出す可能性を考慮しておく +def remove_block_dims_and_alphas( + is_sdxl, block_dims, block_alphas, conv_block_dims, conv_block_alphas, block_lr_weight: Optional[List[float]] +): + if block_lr_weight is not None: + for i, lr in enumerate(block_lr_weight): + if lr == 0: + block_dims[i] = 0 + if conv_block_dims is not None: + conv_block_dims[i] = 0 + return block_dims, block_alphas, conv_block_dims, conv_block_alphas + + +# 外部から呼び出す可能性を考慮しておく +def get_block_index(lora_name: str, is_sdxl: bool = False) -> int: + block_idx = -1 # invalid lora name + if not is_sdxl: + m = RE_UPDOWN.search(lora_name) + if m: + g = m.groups() + i = int(g[1]) + j = int(g[3]) + if g[2] == "resnets": + idx = 3 * i + j + elif g[2] == "attentions": + idx = 3 * i + j + elif g[2] == "upsamplers" or g[2] == "downsamplers": + idx = 3 * i + 2 + + if g[0] == "down": + block_idx = 1 + idx # 0に該当するLoRAは存在しない + elif g[0] == "up": + block_idx = LoRANetwork.NUM_OF_BLOCKS + 1 + idx + elif "mid_block_" in lora_name: + block_idx = LoRANetwork.NUM_OF_BLOCKS # idx=12 + else: + # copy from sdxl_train + if lora_name.startswith("lora_unet_"): + name = lora_name[len("lora_unet_") :] + if name.startswith("time_embed_") or name.startswith("label_emb_"): # No LoRA + block_idx = 0 # 0 + elif name.startswith("input_blocks_"): # 1-9 + block_idx = 1 + int(name.split("_")[2]) + elif name.startswith("middle_block_"): # 10-12 + block_idx = 10 + int(name.split("_")[2]) + elif name.startswith("output_blocks_"): # 13-21 + block_idx = 13 + int(name.split("_")[2]) + elif name.startswith("out_"): # 22, out, no LoRA + block_idx = 22 + + return block_idx + + +def convert_diffusers_to_sai_if_needed(weights_sd): + # only supports U-Net LoRA modules + + found_up_down_blocks = False + for k in list(weights_sd.keys()): + if "down_blocks" in k: + found_up_down_blocks = True + break + if "up_blocks" in k: + found_up_down_blocks = True + break + if not found_up_down_blocks: + return + + from library.sdxl_model_util import make_unet_conversion_map + + unet_conversion_map = make_unet_conversion_map() + unet_conversion_map = {hf.replace(".", "_")[:-1]: sd.replace(".", "_")[:-1] for sd, hf in unet_conversion_map} + + # # add extra conversion + # unet_conversion_map["up_blocks_1_upsamplers_0"] = "lora_unet_output_blocks_2_2_conv" + + logger.info(f"Converting LoRA keys from Diffusers to SAI") + lora_unet_prefix = "lora_unet_" + for k in list(weights_sd.keys()): + if not k.startswith(lora_unet_prefix): + continue + + unet_module_name = k[len(lora_unet_prefix) :].split(".")[0] + + # search for conversion: this is slow because the algorithm is O(n^2), but the number of keys is small + for hf_module_name, sd_module_name in unet_conversion_map.items(): + if hf_module_name in unet_module_name: + new_key = ( + lora_unet_prefix + + unet_module_name.replace(hf_module_name, sd_module_name) + + k[len(lora_unet_prefix) + len(unet_module_name) :] + ) + weights_sd[new_key] = weights_sd.pop(k) + found = True + break + + if not found: + logger.warning(f"Key {k} is not found in unet_conversion_map") + + +# Create network from weights for inference, weights are not loaded here (because can be merged) +def create_network_from_weights(multiplier, file, vae, text_encoder, unet, weights_sd=None, for_inference=False, **kwargs): + # if unet is an instance of SdxlUNet2DConditionModel or subclass, set is_sdxl to True + is_sdxl = unet is not None and issubclass(unet.__class__, SdxlUNet2DConditionModel) + + if weights_sd is None: + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file, safe_open + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + # if keys are Diffusers based, convert to SAI based + if is_sdxl: + convert_diffusers_to_sai_if_needed(weights_sd) + + # get dim/alpha mapping + modules_dim = {} + modules_alpha = {} + for key, value in weights_sd.items(): + if "." not in key: + continue + + lora_name = key.split(".")[0] + if "alpha" in key: + modules_alpha[lora_name] = value + elif "lora_down" in key: + dim = value.size()[0] + modules_dim[lora_name] = dim + # logger.info(lora_name, value.size(), dim) + + # support old LoRA without alpha + for key in modules_dim.keys(): + if key not in modules_alpha: + modules_alpha[key] = modules_dim[key] + + module_class = LoRAInfModule if for_inference else LoRAModule + + network = LoRANetwork( + text_encoder, + unet, + multiplier=multiplier, + modules_dim=modules_dim, + modules_alpha=modules_alpha, + module_class=module_class, + is_sdxl=is_sdxl, + ) + + # block lr + block_lr_weight = parse_block_lr_kwargs(is_sdxl, kwargs) + if block_lr_weight is not None: + network.set_block_lr_weight(block_lr_weight) + + return network, weights_sd + + +class LoRANetwork(torch.nn.Module): + NUM_OF_BLOCKS = 12 # フルモデル相当でのup,downの層の数 + NUM_OF_MID_BLOCKS = 1 + SDXL_NUM_OF_BLOCKS = 9 # SDXLのモデルでのinput/outputの層の数 total=1(base) 9(input) + 3(mid) + 9(output) + 1(out) = 23 + SDXL_NUM_OF_MID_BLOCKS = 3 + + UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"] + UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPSdpaAttention", "CLIPMLP"] + LORA_PREFIX_UNET = "lora_unet" + LORA_PREFIX_TEXT_ENCODER = "lora_te" + + # SDXL: must starts with LORA_PREFIX_TEXT_ENCODER + LORA_PREFIX_TEXT_ENCODER1 = "lora_te1" + LORA_PREFIX_TEXT_ENCODER2 = "lora_te2" + + def __init__( + self, + text_encoder: Union[List[CLIPTextModel], CLIPTextModel], + unet, + multiplier: float = 1.0, + lora_dim: int = 4, + alpha: float = 1, + dropout: Optional[float] = None, + rank_dropout: Optional[float] = None, + module_dropout: Optional[float] = None, + conv_lora_dim: Optional[int] = None, + conv_alpha: Optional[float] = None, + block_dims: Optional[List[int]] = None, + block_alphas: Optional[List[float]] = None, + conv_block_dims: Optional[List[int]] = None, + conv_block_alphas: Optional[List[float]] = None, + modules_dim: Optional[Dict[str, int]] = None, + modules_alpha: Optional[Dict[str, int]] = None, + module_class: Type[object] = LoRAModule, + varbose: Optional[bool] = False, + is_sdxl: Optional[bool] = False, + ) -> None: + """ + LoRA network: すごく引数が多いが、パターンは以下の通り + 1. lora_dimとalphaを指定 + 2. lora_dim、alpha、conv_lora_dim、conv_alphaを指定 + 3. block_dimsとblock_alphasを指定 : Conv2d3x3には適用しない + 4. block_dims、block_alphas、conv_block_dims、conv_block_alphasを指定 : Conv2d3x3にも適用する + 5. modules_dimとmodules_alphaを指定 (推論用) + """ + super().__init__() + self.multiplier = multiplier + + self.lora_dim = lora_dim + self.alpha = alpha + self.conv_lora_dim = conv_lora_dim + self.conv_alpha = conv_alpha + self.dropout = dropout + self.rank_dropout = rank_dropout + self.module_dropout = module_dropout + + self.loraplus_lr_ratio = None + self.loraplus_unet_lr_ratio = None + self.loraplus_text_encoder_lr_ratio = None + + if modules_dim is not None: + logger.info(f"create LoRA network from weights") + elif block_dims is not None: + logger.info(f"create LoRA network from block_dims") + logger.info( + f"neuron dropout: p={self.dropout}, rank dropout: p={self.rank_dropout}, module dropout: p={self.module_dropout}" + ) + logger.info(f"block_dims: {block_dims}") + logger.info(f"block_alphas: {block_alphas}") + if conv_block_dims is not None: + logger.info(f"conv_block_dims: {conv_block_dims}") + logger.info(f"conv_block_alphas: {conv_block_alphas}") + else: + logger.info(f"create LoRA network. base dim (rank): {lora_dim}, alpha: {alpha}") + logger.info( + f"neuron dropout: p={self.dropout}, rank dropout: p={self.rank_dropout}, module dropout: p={self.module_dropout}" + ) + if self.conv_lora_dim is not None: + logger.info( + f"apply LoRA to Conv2d with kernel size (3,3). dim (rank): {self.conv_lora_dim}, alpha: {self.conv_alpha}" + ) + + # create module instances + def create_modules( + is_unet: bool, + text_encoder_idx: Optional[int], # None, 1, 2 + root_module: torch.nn.Module, + target_replace_modules: List[torch.nn.Module], + ) -> List[LoRAModule]: + prefix = ( + self.LORA_PREFIX_UNET + if is_unet + else ( + self.LORA_PREFIX_TEXT_ENCODER + if text_encoder_idx is None + else (self.LORA_PREFIX_TEXT_ENCODER1 if text_encoder_idx == 1 else self.LORA_PREFIX_TEXT_ENCODER2) + ) + ) + loras = [] + skipped = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + is_linear = child_module.__class__.__name__ == "Linear" + is_conv2d = child_module.__class__.__name__ == "Conv2d" + is_conv2d_1x1 = is_conv2d and child_module.kernel_size == (1, 1) + + if is_linear or is_conv2d: + lora_name = prefix + "." + name + "." + child_name + lora_name = lora_name.replace(".", "_") + + dim = None + alpha = None + + if modules_dim is not None: + # モジュール指定あり + if lora_name in modules_dim: + dim = modules_dim[lora_name] + alpha = modules_alpha[lora_name] + elif is_unet and block_dims is not None: + # U-Netでblock_dims指定あり + block_idx = get_block_index(lora_name, is_sdxl) + if is_linear or is_conv2d_1x1: + dim = block_dims[block_idx] + alpha = block_alphas[block_idx] + elif conv_block_dims is not None: + dim = conv_block_dims[block_idx] + alpha = conv_block_alphas[block_idx] + else: + # 通常、すべて対象とする + if is_linear or is_conv2d_1x1: + dim = self.lora_dim + alpha = self.alpha + elif self.conv_lora_dim is not None: + dim = self.conv_lora_dim + alpha = self.conv_alpha + + if dim is None or dim == 0: + # skipした情報を出力 + if is_linear or is_conv2d_1x1 or (self.conv_lora_dim is not None or conv_block_dims is not None): + skipped.append(lora_name) + continue + + lora = module_class( + lora_name, + child_module, + self.multiplier, + dim, + alpha, + dropout=dropout, + rank_dropout=rank_dropout, + module_dropout=module_dropout, + ) + loras.append(lora) + return loras, skipped + + text_encoders = text_encoder if type(text_encoder) == list else [text_encoder] + + # create LoRA for text encoder + # 毎回すべてのモジュールを作るのは無駄なので要検討 + self.text_encoder_loras = [] + skipped_te = [] + for i, text_encoder in enumerate(text_encoders): + if len(text_encoders) > 1: + index = i + 1 + logger.info(f"create LoRA for Text Encoder {index}:") + else: + index = None + logger.info(f"create LoRA for Text Encoder:") + + text_encoder_loras, skipped = create_modules(False, index, text_encoder, LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE) + self.text_encoder_loras.extend(text_encoder_loras) + skipped_te += skipped + logger.info(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.") + + # extend U-Net target modules if conv2d 3x3 is enabled, or load from weights + target_modules = LoRANetwork.UNET_TARGET_REPLACE_MODULE + if modules_dim is not None or self.conv_lora_dim is not None or conv_block_dims is not None: + target_modules += LoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + + self.unet_loras, skipped_un = create_modules(True, None, unet, target_modules) + logger.info(f"create LoRA for U-Net: {len(self.unet_loras)} modules.") + + skipped = skipped_te + skipped_un + if varbose and len(skipped) > 0: + logger.warning( + f"because block_lr_weight is 0 or dim (rank) is 0, {len(skipped)} LoRA modules are skipped / block_lr_weightまたはdim (rank)が0の為、次の{len(skipped)}個のLoRAモジュールはスキップされます:" + ) + for name in skipped: + logger.info(f"\t{name}") + + self.block_lr_weight = None + self.block_lr = False + + # assertion + names = set() + for lora in self.text_encoder_loras + self.unet_loras: + assert lora.lora_name not in names, f"duplicated lora name: {lora.lora_name}" + names.add(lora.lora_name) + + def set_multiplier(self, multiplier): + self.multiplier = multiplier + for lora in self.text_encoder_loras + self.unet_loras: + lora.multiplier = self.multiplier + + def set_enabled(self, is_enabled): + for lora in self.text_encoder_loras + self.unet_loras: + lora.enabled = is_enabled + + def load_weights(self, file): + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + info = self.load_state_dict(weights_sd, False) + return info + + def apply_to(self, text_encoder, unet, apply_text_encoder=True, apply_unet=True): + if apply_text_encoder: + logger.info(f"enable LoRA for text encoder: {len(self.text_encoder_loras)} modules") + else: + self.text_encoder_loras = [] + + if apply_unet: + logger.info(f"enable LoRA for U-Net: {len(self.unet_loras)} modules") + else: + self.unet_loras = [] + + for lora in self.text_encoder_loras + self.unet_loras: + lora.apply_to() + self.add_module(lora.lora_name, lora) + + # マージできるかどうかを返す + def is_mergeable(self): + return True + + # TODO refactor to common function with apply_to + def merge_to(self, text_encoder, unet, weights_sd, dtype, device): + apply_text_encoder = apply_unet = False + for key in weights_sd.keys(): + if key.startswith(LoRANetwork.LORA_PREFIX_TEXT_ENCODER): + apply_text_encoder = True + elif key.startswith(LoRANetwork.LORA_PREFIX_UNET): + apply_unet = True + + if apply_text_encoder: + logger.info("enable LoRA for text encoder") + else: + self.text_encoder_loras = [] + + if apply_unet: + logger.info("enable LoRA for U-Net") + else: + self.unet_loras = [] + + for lora in self.text_encoder_loras + self.unet_loras: + sd_for_lora = {} + for key in weights_sd.keys(): + if key.startswith(lora.lora_name): + sd_for_lora[key[len(lora.lora_name) + 1 :]] = weights_sd[key] + lora.merge_to(sd_for_lora, dtype, device) + + logger.info(f"weights are merged") + + # 層別学習率用に層ごとの学習率に対する倍率を定義する　引数の順番が逆だがとりあえず気にしない + def set_block_lr_weight(self, block_lr_weight: Optional[List[float]]): + self.block_lr = True + self.block_lr_weight = block_lr_weight + + def get_lr_weight(self, block_idx: int) -> float: + if not self.block_lr or self.block_lr_weight is None: + return 1.0 + return self.block_lr_weight[block_idx] + + def set_loraplus_lr_ratio(self, loraplus_lr_ratio, loraplus_unet_lr_ratio, loraplus_text_encoder_lr_ratio): + self.loraplus_lr_ratio = loraplus_lr_ratio + self.loraplus_unet_lr_ratio = loraplus_unet_lr_ratio + self.loraplus_text_encoder_lr_ratio = loraplus_text_encoder_lr_ratio + + logger.info(f"LoRA+ UNet LR Ratio: {self.loraplus_unet_lr_ratio or self.loraplus_lr_ratio}") + logger.info(f"LoRA+ Text Encoder LR Ratio: {self.loraplus_text_encoder_lr_ratio or self.loraplus_lr_ratio}") + + # 二つのText Encoderに別々の学習率を設定できるようにするといいかも + def prepare_optimizer_params(self, text_encoder_lr, unet_lr, default_lr): + # TODO warn if optimizer is not compatible with LoRA+ (but it will cause error so we don't need to check it here?) + # if ( + # self.loraplus_lr_ratio is not None + # or self.loraplus_text_encoder_lr_ratio is not None + # or self.loraplus_unet_lr_ratio is not None + # ): + # assert ( + # optimizer_type.lower() != "prodigy" and "dadapt" not in optimizer_type.lower() + # ), "LoRA+ and Prodigy/DAdaptation is not supported / LoRA+とProdigy/DAdaptationの組み合わせはサポートされていません" + + self.requires_grad_(True) + + all_params = [] + lr_descriptions = [] + + def assemble_params(loras, lr, ratio): + param_groups = {"lora": {}, "plus": {}} + for lora in loras: + for name, param in lora.named_parameters(): + if ratio is not None and "lora_up" in name: + param_groups["plus"][f"{lora.lora_name}.{name}"] = param + else: + param_groups["lora"][f"{lora.lora_name}.{name}"] = param + + params = [] + descriptions = [] + for key in param_groups.keys(): + param_data = {"params": param_groups[key].values()} + + if len(param_data["params"]) == 0: + continue + + if lr is not None: + if key == "plus": + param_data["lr"] = lr * ratio + else: + param_data["lr"] = lr + + if param_data.get("lr", None) == 0 or param_data.get("lr", None) is None: + logger.info("NO LR skipping!") + continue + + params.append(param_data) + descriptions.append("plus" if key == "plus" else "") + + return params, descriptions + + if self.text_encoder_loras: + params, descriptions = assemble_params( + self.text_encoder_loras, + text_encoder_lr if text_encoder_lr is not None else default_lr, + self.loraplus_text_encoder_lr_ratio or self.loraplus_lr_ratio, + ) + all_params.extend(params) + lr_descriptions.extend(["textencoder" + (" " + d if d else "") for d in descriptions]) + + if self.unet_loras: + if self.block_lr: + is_sdxl = False + for lora in self.unet_loras: + if "input_blocks" in lora.lora_name or "output_blocks" in lora.lora_name: + is_sdxl = True + break + + # 学習率のグラフをblockごとにしたいので、blockごとにloraを分類 + block_idx_to_lora = {} + for lora in self.unet_loras: + idx = get_block_index(lora.lora_name, is_sdxl) + if idx not in block_idx_to_lora: + block_idx_to_lora[idx] = [] + block_idx_to_lora[idx].append(lora) + + # blockごとにパラメータを設定する + for idx, block_loras in block_idx_to_lora.items(): + params, descriptions = assemble_params( + block_loras, + (unet_lr if unet_lr is not None else default_lr) * self.get_lr_weight(idx), + self.loraplus_unet_lr_ratio or self.loraplus_lr_ratio, + ) + all_params.extend(params) + lr_descriptions.extend([f"unet_block{idx}" + (" " + d if d else "") for d in descriptions]) + + else: + params, descriptions = assemble_params( + self.unet_loras, + unet_lr if unet_lr is not None else default_lr, + self.loraplus_unet_lr_ratio or self.loraplus_lr_ratio, + ) + all_params.extend(params) + lr_descriptions.extend(["unet" + (" " + d if d else "") for d in descriptions]) + + return all_params, lr_descriptions + + def enable_gradient_checkpointing(self): + # not supported + pass + + def prepare_grad_etc(self, text_encoder, unet): + self.requires_grad_(True) + + def on_epoch_start(self, text_encoder, unet): + self.train() + + def get_trainable_params(self): + return self.parameters() + + def save_weights(self, file, dtype, metadata): + if metadata is not None and len(metadata) == 0: + metadata = None + + state_dict = self.state_dict() + + if dtype is not None: + for key in list(state_dict.keys()): + v = state_dict[key] + v = v.detach().clone().to("cpu").to(dtype) + state_dict[key] = v + + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import save_file + from library import train_util + + # Precalculate model hashes to save time on indexing + if metadata is None: + metadata = {} + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + save_file(state_dict, file, metadata) + else: + torch.save(state_dict, file) + + # mask is a tensor with values from 0 to 1 + def set_region(self, sub_prompt_index, is_last_network, mask): + if mask.max() == 0: + mask = torch.ones_like(mask) + + self.mask = mask + self.sub_prompt_index = sub_prompt_index + self.is_last_network = is_last_network + + for lora in self.text_encoder_loras + self.unet_loras: + lora.set_network(self) + + def set_current_generation(self, batch_size, num_sub_prompts, width, height, shared, ds_ratio=None): + self.batch_size = batch_size + self.num_sub_prompts = num_sub_prompts + self.current_size = (height, width) + self.shared = shared + + # create masks + mask = self.mask + mask_dic = {} + mask = mask.unsqueeze(0).unsqueeze(1) # b(1),c(1),h,w + ref_weight = self.text_encoder_loras[0].lora_down.weight if self.text_encoder_loras else self.unet_loras[0].lora_down.weight + dtype = ref_weight.dtype + device = ref_weight.device + + def resize_add(mh, mw): + # logger.info(mh, mw, mh * mw) + m = torch.nn.functional.interpolate(mask, (mh, mw), mode="bilinear") # doesn't work in bf16 + m = m.to(device, dtype=dtype) + mask_dic[mh * mw] = m + + h = height // 8 + w = width // 8 + for _ in range(4): + resize_add(h, w) + if h % 2 == 1 or w % 2 == 1: # add extra shape if h/w is not divisible by 2 + resize_add(h + h % 2, w + w % 2) + + # deep shrink + if ds_ratio is not None: + hd = int(h * ds_ratio) + wd = int(w * ds_ratio) + resize_add(hd, wd) + + h = (h + 1) // 2 + w = (w + 1) // 2 + + self.mask_dic = mask_dic + + def backup_weights(self): + # 重みのバックアップを行う + loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras + for lora in loras: + org_module = lora.org_module_ref[0] + if not hasattr(org_module, "_lora_org_weight"): + sd = org_module.state_dict() + org_module._lora_org_weight = sd["weight"].detach().clone() + org_module._lora_restored = True + + def restore_weights(self): + # 重みのリストアを行う + loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras + for lora in loras: + org_module = lora.org_module_ref[0] + if not org_module._lora_restored: + sd = org_module.state_dict() + sd["weight"] = org_module._lora_org_weight + org_module.load_state_dict(sd) + org_module._lora_restored = True + + def pre_calculation(self): + # 事前計算を行う + loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras + for lora in loras: + org_module = lora.org_module_ref[0] + sd = org_module.state_dict() + + org_weight = sd["weight"] + lora_weight = lora.get_weight().to(org_weight.device, dtype=org_weight.dtype) + sd["weight"] = org_weight + lora_weight + assert sd["weight"].shape == org_weight.shape + org_module.load_state_dict(sd) + + org_module._lora_restored = False + lora.enabled = False + + def apply_max_norm_regularization(self, max_norm_value, device): + downkeys = [] + upkeys = [] + alphakeys = [] + norms = [] + keys_scaled = 0 + + state_dict = self.state_dict() + for key in state_dict.keys(): + if "lora_down" in key and "weight" in key: + downkeys.append(key) + upkeys.append(key.replace("lora_down", "lora_up")) + alphakeys.append(key.replace("lora_down.weight", "alpha")) + + for i in range(len(downkeys)): + down = state_dict[downkeys[i]].to(device) + up = state_dict[upkeys[i]].to(device) + alpha = state_dict[alphakeys[i]].to(device) + dim = down.shape[0] + scale = alpha / dim + + if up.shape[2:] == (1, 1) and down.shape[2:] == (1, 1): + updown = (up.squeeze(2).squeeze(2) @ down.squeeze(2).squeeze(2)).unsqueeze(2).unsqueeze(3) + elif up.shape[2:] == (3, 3) or down.shape[2:] == (3, 3): + updown = torch.nn.functional.conv2d(down.permute(1, 0, 2, 3), up).permute(1, 0, 2, 3) + else: + updown = up @ down + + updown *= scale + + norm = updown.norm().clamp(min=max_norm_value / 2) + desired = torch.clamp(norm, max=max_norm_value) + ratio = desired.cpu() / norm.cpu() + sqrt_ratio = ratio**0.5 + if ratio != 1: + keys_scaled += 1 + state_dict[upkeys[i]] *= sqrt_ratio + state_dict[downkeys[i]] *= sqrt_ratio + scalednorm = updown.norm() * ratio + norms.append(scalednorm.item()) + + return keys_scaled, sum(norms) / len(norms), max(norms) diff --git a/lora_diffusers.py b/lora_diffusers.py new file mode 100644 index 0000000000000000000000000000000000000000..56b74d10387179cfbef9e8d0bc70520f27e574c3 --- /dev/null +++ b/lora_diffusers.py @@ -0,0 +1,616 @@ +# Diffusersで動くLoRA。このファイル単独で完結する。 +# LoRA module for Diffusers. This file works independently. + +import bisect +import math +import random +from typing import Any, Dict, List, Mapping, Optional, Union +from diffusers import UNet2DConditionModel +import numpy as np +from tqdm import tqdm +from transformers import CLIPTextModel + +import torch +from library.device_utils import init_ipex, get_preferred_device +init_ipex() + +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def make_unet_conversion_map() -> Dict[str, str]: + unet_conversion_map_layer = [] + + for i in range(3): # num_blocks is 3 in sdxl + # loop over downblocks/upblocks + for j in range(2): + # loop over resnets/attentions for downblocks + hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}." + sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0." + unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix)) + + if i < 3: + # no attention layers in down_blocks.3 + hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}." + sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1." + unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix)) + + for j in range(3): + # loop over resnets/attentions for upblocks + hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}." + sd_up_res_prefix = f"output_blocks.{3*i + j}.0." + unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix)) + + # if i > 0: commentout for sdxl + # no attention layers in up_blocks.0 + hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}." + sd_up_atn_prefix = f"output_blocks.{3*i + j}.1." + unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix)) + + if i < 3: + # no downsample in down_blocks.3 + hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv." + sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op." + unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix)) + + # no upsample in up_blocks.3 + hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0." + sd_upsample_prefix = f"output_blocks.{3*i + 2}.{2}." # change for sdxl + unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix)) + + hf_mid_atn_prefix = "mid_block.attentions.0." + sd_mid_atn_prefix = "middle_block.1." + unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix)) + + for j in range(2): + hf_mid_res_prefix = f"mid_block.resnets.{j}." + sd_mid_res_prefix = f"middle_block.{2*j}." + unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix)) + + unet_conversion_map_resnet = [ + # (stable-diffusion, HF Diffusers) + ("in_layers.0.", "norm1."), + ("in_layers.2.", "conv1."), + ("out_layers.0.", "norm2."), + ("out_layers.3.", "conv2."), + ("emb_layers.1.", "time_emb_proj."), + ("skip_connection.", "conv_shortcut."), + ] + + unet_conversion_map = [] + for sd, hf in unet_conversion_map_layer: + if "resnets" in hf: + for sd_res, hf_res in unet_conversion_map_resnet: + unet_conversion_map.append((sd + sd_res, hf + hf_res)) + else: + unet_conversion_map.append((sd, hf)) + + for j in range(2): + hf_time_embed_prefix = f"time_embedding.linear_{j+1}." + sd_time_embed_prefix = f"time_embed.{j*2}." + unet_conversion_map.append((sd_time_embed_prefix, hf_time_embed_prefix)) + + for j in range(2): + hf_label_embed_prefix = f"add_embedding.linear_{j+1}." + sd_label_embed_prefix = f"label_emb.0.{j*2}." + unet_conversion_map.append((sd_label_embed_prefix, hf_label_embed_prefix)) + + unet_conversion_map.append(("input_blocks.0.0.", "conv_in.")) + unet_conversion_map.append(("out.0.", "conv_norm_out.")) + unet_conversion_map.append(("out.2.", "conv_out.")) + + sd_hf_conversion_map = {sd.replace(".", "_")[:-1]: hf.replace(".", "_")[:-1] for sd, hf in unet_conversion_map} + return sd_hf_conversion_map + + +UNET_CONVERSION_MAP = make_unet_conversion_map() + + +class LoRAModule(torch.nn.Module): + """ + replaces forward method of the original Linear, instead of replacing the original Linear module. + """ + + def __init__( + self, + lora_name, + org_module: torch.nn.Module, + multiplier=1.0, + lora_dim=4, + alpha=1, + ): + """if alpha == 0 or None, alpha is rank (no scaling).""" + super().__init__() + self.lora_name = lora_name + + if org_module.__class__.__name__ == "Conv2d" or org_module.__class__.__name__ == "LoRACompatibleConv": + in_dim = org_module.in_channels + out_dim = org_module.out_channels + else: + in_dim = org_module.in_features + out_dim = org_module.out_features + + self.lora_dim = lora_dim + + if org_module.__class__.__name__ == "Conv2d" or org_module.__class__.__name__ == "LoRACompatibleConv": + kernel_size = org_module.kernel_size + stride = org_module.stride + padding = org_module.padding + self.lora_down = torch.nn.Conv2d(in_dim, self.lora_dim, kernel_size, stride, padding, bias=False) + self.lora_up = torch.nn.Conv2d(self.lora_dim, out_dim, (1, 1), (1, 1), bias=False) + else: + self.lora_down = torch.nn.Linear(in_dim, self.lora_dim, bias=False) + self.lora_up = torch.nn.Linear(self.lora_dim, out_dim, bias=False) + + if type(alpha) == torch.Tensor: + alpha = alpha.detach().float().numpy() # without casting, bf16 causes error + alpha = self.lora_dim if alpha is None or alpha == 0 else alpha + self.scale = alpha / self.lora_dim + self.register_buffer("alpha", torch.tensor(alpha)) # 勾配計算に含めない / not included in gradient calculation + + # same as microsoft's + torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5)) + torch.nn.init.zeros_(self.lora_up.weight) + + self.multiplier = multiplier + self.org_module = [org_module] + self.enabled = True + self.network: LoRANetwork = None + self.org_forward = None + + # override org_module's forward method + def apply_to(self, multiplier=None): + if multiplier is not None: + self.multiplier = multiplier + if self.org_forward is None: + self.org_forward = self.org_module[0].forward + self.org_module[0].forward = self.forward + + # restore org_module's forward method + def unapply_to(self): + if self.org_forward is not None: + self.org_module[0].forward = self.org_forward + + # forward with lora + # scale is used LoRACompatibleConv, but we ignore it because we have multiplier + def forward(self, x, scale=1.0): + if not self.enabled: + return self.org_forward(x) + return self.org_forward(x) + self.lora_up(self.lora_down(x)) * self.multiplier * self.scale + + def set_network(self, network): + self.network = network + + # merge lora weight to org weight + def merge_to(self, multiplier=1.0): + # get lora weight + lora_weight = self.get_weight(multiplier) + + # get org weight + org_sd = self.org_module[0].state_dict() + org_weight = org_sd["weight"] + weight = org_weight + lora_weight.to(org_weight.device, dtype=org_weight.dtype) + + # set weight to org_module + org_sd["weight"] = weight + self.org_module[0].load_state_dict(org_sd) + + # restore org weight from lora weight + def restore_from(self, multiplier=1.0): + # get lora weight + lora_weight = self.get_weight(multiplier) + + # get org weight + org_sd = self.org_module[0].state_dict() + org_weight = org_sd["weight"] + weight = org_weight - lora_weight.to(org_weight.device, dtype=org_weight.dtype) + + # set weight to org_module + org_sd["weight"] = weight + self.org_module[0].load_state_dict(org_sd) + + # return lora weight + def get_weight(self, multiplier=None): + if multiplier is None: + multiplier = self.multiplier + + # get up/down weight from module + up_weight = self.lora_up.weight.to(torch.float) + down_weight = self.lora_down.weight.to(torch.float) + + # pre-calculated weight + if len(down_weight.size()) == 2: + # linear + weight = self.multiplier * (up_weight @ down_weight) * self.scale + elif down_weight.size()[2:4] == (1, 1): + # conv2d 1x1 + weight = ( + self.multiplier + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * self.scale + ) + else: + # conv2d 3x3 + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + weight = self.multiplier * conved * self.scale + + return weight + + +# Create network from weights for inference, weights are not loaded here +def create_network_from_weights( + text_encoder: Union[CLIPTextModel, List[CLIPTextModel]], unet: UNet2DConditionModel, weights_sd: Dict, multiplier: float = 1.0 +): + # get dim/alpha mapping + modules_dim = {} + modules_alpha = {} + for key, value in weights_sd.items(): + if "." not in key: + continue + + lora_name = key.split(".")[0] + if "alpha" in key: + modules_alpha[lora_name] = value + elif "lora_down" in key: + dim = value.size()[0] + modules_dim[lora_name] = dim + # logger.info(f"{lora_name} {value.size()} {dim}") + + # support old LoRA without alpha + for key in modules_dim.keys(): + if key not in modules_alpha: + modules_alpha[key] = modules_dim[key] + + return LoRANetwork(text_encoder, unet, multiplier=multiplier, modules_dim=modules_dim, modules_alpha=modules_alpha) + + +def merge_lora_weights(pipe, weights_sd: Dict, multiplier: float = 1.0): + text_encoders = [pipe.text_encoder, pipe.text_encoder_2] if hasattr(pipe, "text_encoder_2") else [pipe.text_encoder] + unet = pipe.unet + + lora_network = create_network_from_weights(text_encoders, unet, weights_sd, multiplier=multiplier) + lora_network.load_state_dict(weights_sd) + lora_network.merge_to(multiplier=multiplier) + + +# block weightや学習に対応しない簡易版 / simple version without block weight and training +class LoRANetwork(torch.nn.Module): + UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"] + UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPSdpaAttention", "CLIPMLP"] + LORA_PREFIX_UNET = "lora_unet" + LORA_PREFIX_TEXT_ENCODER = "lora_te" + + # SDXL: must starts with LORA_PREFIX_TEXT_ENCODER + LORA_PREFIX_TEXT_ENCODER1 = "lora_te1" + LORA_PREFIX_TEXT_ENCODER2 = "lora_te2" + + def __init__( + self, + text_encoder: Union[List[CLIPTextModel], CLIPTextModel], + unet: UNet2DConditionModel, + multiplier: float = 1.0, + modules_dim: Optional[Dict[str, int]] = None, + modules_alpha: Optional[Dict[str, int]] = None, + varbose: Optional[bool] = False, + ) -> None: + super().__init__() + self.multiplier = multiplier + + logger.info("create LoRA network from weights") + + # convert SDXL Stability AI's U-Net modules to Diffusers + converted = self.convert_unet_modules(modules_dim, modules_alpha) + if converted: + logger.info(f"converted {converted} Stability AI's U-Net LoRA modules to Diffusers (SDXL)") + + # create module instances + def create_modules( + is_unet: bool, + text_encoder_idx: Optional[int], # None, 1, 2 + root_module: torch.nn.Module, + target_replace_modules: List[torch.nn.Module], + ) -> List[LoRAModule]: + prefix = ( + self.LORA_PREFIX_UNET + if is_unet + else ( + self.LORA_PREFIX_TEXT_ENCODER + if text_encoder_idx is None + else (self.LORA_PREFIX_TEXT_ENCODER1 if text_encoder_idx == 1 else self.LORA_PREFIX_TEXT_ENCODER2) + ) + ) + loras = [] + skipped = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + is_linear = ( + child_module.__class__.__name__ == "Linear" or child_module.__class__.__name__ == "LoRACompatibleLinear" + ) + is_conv2d = ( + child_module.__class__.__name__ == "Conv2d" or child_module.__class__.__name__ == "LoRACompatibleConv" + ) + + if is_linear or is_conv2d: + lora_name = prefix + "." + name + "." + child_name + lora_name = lora_name.replace(".", "_") + + if lora_name not in modules_dim: + # logger.info(f"skipped {lora_name} (not found in modules_dim)") + skipped.append(lora_name) + continue + + dim = modules_dim[lora_name] + alpha = modules_alpha[lora_name] + lora = LoRAModule( + lora_name, + child_module, + self.multiplier, + dim, + alpha, + ) + loras.append(lora) + return loras, skipped + + text_encoders = text_encoder if type(text_encoder) == list else [text_encoder] + + # create LoRA for text encoder + # 毎回すべてのモジュールを作るのは無駄なので要検討 / it is wasteful to create all modules every time, need to consider + self.text_encoder_loras: List[LoRAModule] = [] + skipped_te = [] + for i, text_encoder in enumerate(text_encoders): + if len(text_encoders) > 1: + index = i + 1 + else: + index = None + + text_encoder_loras, skipped = create_modules(False, index, text_encoder, LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE) + self.text_encoder_loras.extend(text_encoder_loras) + skipped_te += skipped + logger.info(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.") + if len(skipped_te) > 0: + logger.warning(f"skipped {len(skipped_te)} modules because of missing weight for text encoder.") + + # extend U-Net target modules to include Conv2d 3x3 + target_modules = LoRANetwork.UNET_TARGET_REPLACE_MODULE + LoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + + self.unet_loras: List[LoRAModule] + self.unet_loras, skipped_un = create_modules(True, None, unet, target_modules) + logger.info(f"create LoRA for U-Net: {len(self.unet_loras)} modules.") + if len(skipped_un) > 0: + logger.warning(f"skipped {len(skipped_un)} modules because of missing weight for U-Net.") + + # assertion + names = set() + for lora in self.text_encoder_loras + self.unet_loras: + names.add(lora.lora_name) + for lora_name in modules_dim.keys(): + assert lora_name in names, f"{lora_name} is not found in created LoRA modules." + + # make to work load_state_dict + for lora in self.text_encoder_loras + self.unet_loras: + self.add_module(lora.lora_name, lora) + + # SDXL: convert SDXL Stability AI's U-Net modules to Diffusers + def convert_unet_modules(self, modules_dim, modules_alpha): + converted_count = 0 + not_converted_count = 0 + + map_keys = list(UNET_CONVERSION_MAP.keys()) + map_keys.sort() + + for key in list(modules_dim.keys()): + if key.startswith(LoRANetwork.LORA_PREFIX_UNET + "_"): + search_key = key.replace(LoRANetwork.LORA_PREFIX_UNET + "_", "") + position = bisect.bisect_right(map_keys, search_key) + map_key = map_keys[position - 1] + if search_key.startswith(map_key): + new_key = key.replace(map_key, UNET_CONVERSION_MAP[map_key]) + modules_dim[new_key] = modules_dim[key] + modules_alpha[new_key] = modules_alpha[key] + del modules_dim[key] + del modules_alpha[key] + converted_count += 1 + else: + not_converted_count += 1 + assert ( + converted_count == 0 or not_converted_count == 0 + ), f"some modules are not converted: {converted_count} converted, {not_converted_count} not converted" + return converted_count + + def set_multiplier(self, multiplier): + self.multiplier = multiplier + for lora in self.text_encoder_loras + self.unet_loras: + lora.multiplier = self.multiplier + + def apply_to(self, multiplier=1.0, apply_text_encoder=True, apply_unet=True): + if apply_text_encoder: + logger.info("enable LoRA for text encoder") + for lora in self.text_encoder_loras: + lora.apply_to(multiplier) + if apply_unet: + logger.info("enable LoRA for U-Net") + for lora in self.unet_loras: + lora.apply_to(multiplier) + + def unapply_to(self): + for lora in self.text_encoder_loras + self.unet_loras: + lora.unapply_to() + + def merge_to(self, multiplier=1.0): + logger.info("merge LoRA weights to original weights") + for lora in tqdm(self.text_encoder_loras + self.unet_loras): + lora.merge_to(multiplier) + logger.info(f"weights are merged") + + def restore_from(self, multiplier=1.0): + logger.info("restore LoRA weights from original weights") + for lora in tqdm(self.text_encoder_loras + self.unet_loras): + lora.restore_from(multiplier) + logger.info(f"weights are restored") + + def load_state_dict(self, state_dict: Mapping[str, Any], strict: bool = True): + # convert SDXL Stability AI's state dict to Diffusers' based state dict + map_keys = list(UNET_CONVERSION_MAP.keys()) # prefix of U-Net modules + map_keys.sort() + for key in list(state_dict.keys()): + if key.startswith(LoRANetwork.LORA_PREFIX_UNET + "_"): + search_key = key.replace(LoRANetwork.LORA_PREFIX_UNET + "_", "") + position = bisect.bisect_right(map_keys, search_key) + map_key = map_keys[position - 1] + if search_key.startswith(map_key): + new_key = key.replace(map_key, UNET_CONVERSION_MAP[map_key]) + state_dict[new_key] = state_dict[key] + del state_dict[key] + + # in case of V2, some weights have different shape, so we need to convert them + # because V2 LoRA is based on U-Net created by use_linear_projection=False + my_state_dict = self.state_dict() + for key in state_dict.keys(): + if state_dict[key].size() != my_state_dict[key].size(): + # logger.info(f"convert {key} from {state_dict[key].size()} to {my_state_dict[key].size()}") + state_dict[key] = state_dict[key].view(my_state_dict[key].size()) + + return super().load_state_dict(state_dict, strict) + + +if __name__ == "__main__": + # sample code to use LoRANetwork + import os + import argparse + from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline + import torch + + device = get_preferred_device() + + parser = argparse.ArgumentParser() + parser.add_argument("--model_id", type=str, default=None, help="model id for huggingface") + parser.add_argument("--lora_weights", type=str, default=None, help="path to LoRA weights") + parser.add_argument("--sdxl", action="store_true", help="use SDXL model") + parser.add_argument("--prompt", type=str, default="A photo of cat", help="prompt text") + parser.add_argument("--negative_prompt", type=str, default="", help="negative prompt text") + parser.add_argument("--seed", type=int, default=0, help="random seed") + args = parser.parse_args() + + image_prefix = args.model_id.replace("/", "_") + "_" + + # load Diffusers model + logger.info(f"load model from {args.model_id}") + pipe: Union[StableDiffusionPipeline, StableDiffusionXLPipeline] + if args.sdxl: + # use_safetensors=True does not work with 0.18.2 + pipe = StableDiffusionXLPipeline.from_pretrained(args.model_id, variant="fp16", torch_dtype=torch.float16) + else: + pipe = StableDiffusionPipeline.from_pretrained(args.model_id, variant="fp16", torch_dtype=torch.float16) + pipe.to(device) + pipe.set_use_memory_efficient_attention_xformers(True) + + text_encoders = [pipe.text_encoder, pipe.text_encoder_2] if args.sdxl else [pipe.text_encoder] + + # load LoRA weights + logger.info(f"load LoRA weights from {args.lora_weights}") + if os.path.splitext(args.lora_weights)[1] == ".safetensors": + from safetensors.torch import load_file + + lora_sd = load_file(args.lora_weights) + else: + lora_sd = torch.load(args.lora_weights) + + # create by LoRA weights and load weights + logger.info(f"create LoRA network") + lora_network: LoRANetwork = create_network_from_weights(text_encoders, pipe.unet, lora_sd, multiplier=1.0) + + logger.info(f"load LoRA network weights") + lora_network.load_state_dict(lora_sd) + + lora_network.to(device, dtype=pipe.unet.dtype) # required to apply_to. merge_to works without this + + # 必要があれば、元のモデルの重みをバックアップしておく + # back-up unet/text encoder weights if necessary + def detach_and_move_to_cpu(state_dict): + for k, v in state_dict.items(): + state_dict[k] = v.detach().cpu() + return state_dict + + org_unet_sd = pipe.unet.state_dict() + detach_and_move_to_cpu(org_unet_sd) + + org_text_encoder_sd = pipe.text_encoder.state_dict() + detach_and_move_to_cpu(org_text_encoder_sd) + + if args.sdxl: + org_text_encoder_2_sd = pipe.text_encoder_2.state_dict() + detach_and_move_to_cpu(org_text_encoder_2_sd) + + def seed_everything(seed): + torch.manual_seed(seed) + torch.cuda.manual_seed_all(seed) + np.random.seed(seed) + random.seed(seed) + + # create image with original weights + logger.info(f"create image with original weights") + seed_everything(args.seed) + image = pipe(args.prompt, negative_prompt=args.negative_prompt).images[0] + image.save(image_prefix + "original.png") + + # apply LoRA network to the model: slower than merge_to, but can be reverted easily + logger.info(f"apply LoRA network to the model") + lora_network.apply_to(multiplier=1.0) + + logger.info(f"create image with applied LoRA") + seed_everything(args.seed) + image = pipe(args.prompt, negative_prompt=args.negative_prompt).images[0] + image.save(image_prefix + "applied_lora.png") + + # unapply LoRA network to the model + logger.info(f"unapply LoRA network to the model") + lora_network.unapply_to() + + logger.info(f"create image with unapplied LoRA") + seed_everything(args.seed) + image = pipe(args.prompt, negative_prompt=args.negative_prompt).images[0] + image.save(image_prefix + "unapplied_lora.png") + + # merge LoRA network to the model: faster than apply_to, but requires back-up of original weights (or unmerge_to) + logger.info(f"merge LoRA network to the model") + lora_network.merge_to(multiplier=1.0) + + logger.info(f"create image with LoRA") + seed_everything(args.seed) + image = pipe(args.prompt, negative_prompt=args.negative_prompt).images[0] + image.save(image_prefix + "merged_lora.png") + + # restore (unmerge) LoRA weights: numerically unstable + # マージされた重みを元に戻す。計算誤差のため、元の重みと完全に一致しないことがあるかもしれない + # 保存したstate_dictから元の重みを復元するのが確実 + logger.info(f"restore (unmerge) LoRA weights") + lora_network.restore_from(multiplier=1.0) + + logger.info(f"create image without LoRA") + seed_everything(args.seed) + image = pipe(args.prompt, negative_prompt=args.negative_prompt).images[0] + image.save(image_prefix + "unmerged_lora.png") + + # restore original weights + logger.info(f"restore original weights") + pipe.unet.load_state_dict(org_unet_sd) + pipe.text_encoder.load_state_dict(org_text_encoder_sd) + if args.sdxl: + pipe.text_encoder_2.load_state_dict(org_text_encoder_2_sd) + + logger.info(f"create image with restored original weights") + seed_everything(args.seed) + image = pipe(args.prompt, negative_prompt=args.negative_prompt).images[0] + image.save(image_prefix + "restore_original.png") + + # use convenience function to merge LoRA weights + logger.info(f"merge LoRA weights with convenience function") + merge_lora_weights(pipe, lora_sd, multiplier=1.0) + + logger.info(f"create image with merged LoRA weights") + seed_everything(args.seed) + image = pipe(args.prompt, negative_prompt=args.negative_prompt).images[0] + image.save(image_prefix + "convenience_merged_lora.png") diff --git a/lora_fa.py b/lora_fa.py new file mode 100644 index 0000000000000000000000000000000000000000..5fe778b404788f9d48230f0cb9c8fe96da5d0cc9 --- /dev/null +++ b/lora_fa.py @@ -0,0 +1,1244 @@ +# LoRA network module +# reference: +# https://github.com/microsoft/LoRA/blob/main/loralib/layers.py +# https://github.com/cloneofsimo/lora/blob/master/lora_diffusion/lora.py + +# temporary implementation of LoRA-FA: https://arxiv.org/abs/2308.03303 +# need to be refactored and merged to lora.py + +import math +import os +from typing import Dict, List, Optional, Tuple, Type, Union +from diffusers import AutoencoderKL +from transformers import CLIPTextModel +import numpy as np +import torch +import re +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +RE_UPDOWN = re.compile(r"(up|down)_blocks_(\d+)_(resnets|upsamplers|downsamplers|attentions)_(\d+)_") + + +class LoRAModule(torch.nn.Module): + """ + replaces forward method of the original Linear, instead of replacing the original Linear module. + """ + + def __init__( + self, + lora_name, + org_module: torch.nn.Module, + multiplier=1.0, + lora_dim=4, + alpha=1, + dropout=None, + rank_dropout=None, + module_dropout=None, + ): + """if alpha == 0 or None, alpha is rank (no scaling).""" + super().__init__() + self.lora_name = lora_name + + if org_module.__class__.__name__ == "Conv2d": + in_dim = org_module.in_channels + out_dim = org_module.out_channels + else: + in_dim = org_module.in_features + out_dim = org_module.out_features + + # if limit_rank: + # self.lora_dim = min(lora_dim, in_dim, out_dim) + # if self.lora_dim != lora_dim: + # logger.info(f"{lora_name} dim (rank) is changed to: {self.lora_dim}") + # else: + self.lora_dim = lora_dim + + if org_module.__class__.__name__ == "Conv2d": + kernel_size = org_module.kernel_size + stride = org_module.stride + padding = org_module.padding + self.lora_down = torch.nn.Conv2d(in_dim, self.lora_dim, kernel_size, stride, padding, bias=False) + self.lora_up = torch.nn.Conv2d(self.lora_dim, out_dim, (1, 1), (1, 1), bias=False) + else: + self.lora_down = torch.nn.Linear(in_dim, self.lora_dim, bias=False) + self.lora_up = torch.nn.Linear(self.lora_dim, out_dim, bias=False) + + if type(alpha) == torch.Tensor: + alpha = alpha.detach().float().numpy() # without casting, bf16 causes error + alpha = self.lora_dim if alpha is None or alpha == 0 else alpha + self.scale = alpha / self.lora_dim + self.register_buffer("alpha", torch.tensor(alpha)) # 定数として扱える + + # # same as microsoft's + # torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5)) + + # according to the paper, initialize LoRA-A (down) as normal distribution + torch.nn.init.normal_(self.lora_down.weight, std=math.sqrt(2.0 / (in_dim + self.lora_dim))) + + torch.nn.init.zeros_(self.lora_up.weight) + + self.multiplier = multiplier + self.org_module = org_module # remove in applying + self.dropout = dropout + self.rank_dropout = rank_dropout + self.module_dropout = module_dropout + + def get_trainable_params(self): + params = self.named_parameters() + trainable_params = [] + for param in params: + if param[0] == "lora_up.weight": # up only + trainable_params.append(param[1]) + return trainable_params + + def requires_grad_(self, requires_grad: bool = True): + self.lora_up.requires_grad_(requires_grad) + self.lora_down.requires_grad_(False) + return self + + def apply_to(self): + self.org_forward = self.org_module.forward + self.org_module.forward = self.forward + del self.org_module + + def forward(self, x): + org_forwarded = self.org_forward(x) + + # module dropout + if self.module_dropout is not None and self.training: + if torch.rand(1) < self.module_dropout: + return org_forwarded + + lx = self.lora_down(x) + + # normal dropout + if self.dropout is not None and self.training: + lx = torch.nn.functional.dropout(lx, p=self.dropout) + + # rank dropout + if self.rank_dropout is not None and self.training: + mask = torch.rand((lx.size(0), self.lora_dim), device=lx.device) > self.rank_dropout + if len(lx.size()) == 3: + mask = mask.unsqueeze(1) # for Text Encoder + elif len(lx.size()) == 4: + mask = mask.unsqueeze(-1).unsqueeze(-1) # for Conv2d + lx = lx * mask + + # scaling for rank dropout: treat as if the rank is changed + # maskから計算することも考えられるが、augmentation的な効果を期待してrank_dropoutを用いる + scale = self.scale * (1.0 / (1.0 - self.rank_dropout)) # redundant for readability + else: + scale = self.scale + + lx = self.lora_up(lx) + + return org_forwarded + lx * self.multiplier * scale + + +class LoRAInfModule(LoRAModule): + def __init__( + self, + lora_name, + org_module: torch.nn.Module, + multiplier=1.0, + lora_dim=4, + alpha=1, + **kwargs, + ): + # no dropout for inference + super().__init__(lora_name, org_module, multiplier, lora_dim, alpha) + + self.org_module_ref = [org_module] # 後から参照できるように + self.enabled = True + + # check regional or not by lora_name + self.text_encoder = False + if lora_name.startswith("lora_te_"): + self.regional = False + self.use_sub_prompt = True + self.text_encoder = True + elif "attn2_to_k" in lora_name or "attn2_to_v" in lora_name: + self.regional = False + self.use_sub_prompt = True + elif "time_emb" in lora_name: + self.regional = False + self.use_sub_prompt = False + else: + self.regional = True + self.use_sub_prompt = False + + self.network: LoRANetwork = None + + def set_network(self, network): + self.network = network + + # freezeしてマージする + def merge_to(self, sd, dtype, device): + # get up/down weight + up_weight = sd["lora_up.weight"].to(torch.float).to(device) + down_weight = sd["lora_down.weight"].to(torch.float).to(device) + + # extract weight from org_module + org_sd = self.org_module.state_dict() + weight = org_sd["weight"].to(torch.float) + + # merge weight + if len(weight.size()) == 2: + # linear + weight = weight + self.multiplier * (up_weight @ down_weight) * self.scale + elif down_weight.size()[2:4] == (1, 1): + # conv2d 1x1 + weight = ( + weight + + self.multiplier + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * self.scale + ) + else: + # conv2d 3x3 + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + # logger.info(conved.size(), weight.size(), module.stride, module.padding) + weight = weight + self.multiplier * conved * self.scale + + # set weight to org_module + org_sd["weight"] = weight.to(dtype) + self.org_module.load_state_dict(org_sd) + + # 復元できるマージのため、このモジュールのweightを返す + def get_weight(self, multiplier=None): + if multiplier is None: + multiplier = self.multiplier + + # get up/down weight from module + up_weight = self.lora_up.weight.to(torch.float) + down_weight = self.lora_down.weight.to(torch.float) + + # pre-calculated weight + if len(down_weight.size()) == 2: + # linear + weight = self.multiplier * (up_weight @ down_weight) * self.scale + elif down_weight.size()[2:4] == (1, 1): + # conv2d 1x1 + weight = ( + self.multiplier + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * self.scale + ) + else: + # conv2d 3x3 + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + weight = self.multiplier * conved * self.scale + + return weight + + def set_region(self, region): + self.region = region + self.region_mask = None + + def default_forward(self, x): + # logger.info("default_forward", self.lora_name, x.size()) + return self.org_forward(x) + self.lora_up(self.lora_down(x)) * self.multiplier * self.scale + + def forward(self, x): + if not self.enabled: + return self.org_forward(x) + + if self.network is None or self.network.sub_prompt_index is None: + return self.default_forward(x) + if not self.regional and not self.use_sub_prompt: + return self.default_forward(x) + + if self.regional: + return self.regional_forward(x) + else: + return self.sub_prompt_forward(x) + + def get_mask_for_x(self, x): + # calculate size from shape of x + if len(x.size()) == 4: + h, w = x.size()[2:4] + area = h * w + else: + area = x.size()[1] + + mask = self.network.mask_dic[area] + if mask is None: + raise ValueError(f"mask is None for resolution {area}") + if len(x.size()) != 4: + mask = torch.reshape(mask, (1, -1, 1)) + return mask + + def regional_forward(self, x): + if "attn2_to_out" in self.lora_name: + return self.to_out_forward(x) + + if self.network.mask_dic is None: # sub_prompt_index >= 3 + return self.default_forward(x) + + # apply mask for LoRA result + lx = self.lora_up(self.lora_down(x)) * self.multiplier * self.scale + mask = self.get_mask_for_x(lx) + # logger.info("regional", self.lora_name, self.network.sub_prompt_index, lx.size(), mask.size()) + lx = lx * mask + + x = self.org_forward(x) + x = x + lx + + if "attn2_to_q" in self.lora_name and self.network.is_last_network: + x = self.postp_to_q(x) + + return x + + def postp_to_q(self, x): + # repeat x to num_sub_prompts + has_real_uncond = x.size()[0] // self.network.batch_size == 3 + qc = self.network.batch_size # uncond + qc += self.network.batch_size * self.network.num_sub_prompts # cond + if has_real_uncond: + qc += self.network.batch_size # real_uncond + + query = torch.zeros((qc, x.size()[1], x.size()[2]), device=x.device, dtype=x.dtype) + query[: self.network.batch_size] = x[: self.network.batch_size] + + for i in range(self.network.batch_size): + qi = self.network.batch_size + i * self.network.num_sub_prompts + query[qi : qi + self.network.num_sub_prompts] = x[self.network.batch_size + i] + + if has_real_uncond: + query[-self.network.batch_size :] = x[-self.network.batch_size :] + + # logger.info("postp_to_q", self.lora_name, x.size(), query.size(), self.network.num_sub_prompts) + return query + + def sub_prompt_forward(self, x): + if x.size()[0] == self.network.batch_size: # if uncond in text_encoder, do not apply LoRA + return self.org_forward(x) + + emb_idx = self.network.sub_prompt_index + if not self.text_encoder: + emb_idx += self.network.batch_size + + # apply sub prompt of X + lx = x[emb_idx :: self.network.num_sub_prompts] + lx = self.lora_up(self.lora_down(lx)) * self.multiplier * self.scale + + # logger.info("sub_prompt_forward", self.lora_name, x.size(), lx.size(), emb_idx) + + x = self.org_forward(x) + x[emb_idx :: self.network.num_sub_prompts] += lx + + return x + + def to_out_forward(self, x): + # logger.info("to_out_forward", self.lora_name, x.size(), self.network.is_last_network) + + if self.network.is_last_network: + masks = [None] * self.network.num_sub_prompts + self.network.shared[self.lora_name] = (None, masks) + else: + lx, masks = self.network.shared[self.lora_name] + + # call own LoRA + x1 = x[self.network.batch_size + self.network.sub_prompt_index :: self.network.num_sub_prompts] + lx1 = self.lora_up(self.lora_down(x1)) * self.multiplier * self.scale + + if self.network.is_last_network: + lx = torch.zeros( + (self.network.num_sub_prompts * self.network.batch_size, *lx1.size()[1:]), device=lx1.device, dtype=lx1.dtype + ) + self.network.shared[self.lora_name] = (lx, masks) + + # logger.info("to_out_forward", lx.size(), lx1.size(), self.network.sub_prompt_index, self.network.num_sub_prompts) + lx[self.network.sub_prompt_index :: self.network.num_sub_prompts] += lx1 + masks[self.network.sub_prompt_index] = self.get_mask_for_x(lx1) + + # if not last network, return x and masks + x = self.org_forward(x) + if not self.network.is_last_network: + return x + + lx, masks = self.network.shared.pop(self.lora_name) + + # if last network, combine separated x with mask weighted sum + has_real_uncond = x.size()[0] // self.network.batch_size == self.network.num_sub_prompts + 2 + + out = torch.zeros((self.network.batch_size * (3 if has_real_uncond else 2), *x.size()[1:]), device=x.device, dtype=x.dtype) + out[: self.network.batch_size] = x[: self.network.batch_size] # uncond + if has_real_uncond: + out[-self.network.batch_size :] = x[-self.network.batch_size :] # real_uncond + + # logger.info("to_out_forward", self.lora_name, self.network.sub_prompt_index, self.network.num_sub_prompts) + # for i in range(len(masks)): + # if masks[i] is None: + # masks[i] = torch.zeros_like(masks[-1]) + + mask = torch.cat(masks) + mask_sum = torch.sum(mask, dim=0) + 1e-4 + for i in range(self.network.batch_size): + # 1枚の画像ごとに処理する + lx1 = lx[i * self.network.num_sub_prompts : (i + 1) * self.network.num_sub_prompts] + lx1 = lx1 * mask + lx1 = torch.sum(lx1, dim=0) + + xi = self.network.batch_size + i * self.network.num_sub_prompts + x1 = x[xi : xi + self.network.num_sub_prompts] + x1 = x1 * mask + x1 = torch.sum(x1, dim=0) + x1 = x1 / mask_sum + + x1 = x1 + lx1 + out[self.network.batch_size + i] = x1 + + # logger.info("to_out_forward", x.size(), out.size(), has_real_uncond) + return out + + +def parse_block_lr_kwargs(nw_kwargs): + down_lr_weight = nw_kwargs.get("down_lr_weight", None) + mid_lr_weight = nw_kwargs.get("mid_lr_weight", None) + up_lr_weight = nw_kwargs.get("up_lr_weight", None) + + # 以上のいずれにも設定がない場合は無効としてNoneを返す + if down_lr_weight is None and mid_lr_weight is None and up_lr_weight is None: + return None, None, None + + # extract learning rate weight for each block + if down_lr_weight is not None: + # if some parameters are not set, use zero + if "," in down_lr_weight: + down_lr_weight = [(float(s) if s else 0.0) for s in down_lr_weight.split(",")] + + if mid_lr_weight is not None: + mid_lr_weight = float(mid_lr_weight) + + if up_lr_weight is not None: + if "," in up_lr_weight: + up_lr_weight = [(float(s) if s else 0.0) for s in up_lr_weight.split(",")] + + down_lr_weight, mid_lr_weight, up_lr_weight = get_block_lr_weight( + down_lr_weight, mid_lr_weight, up_lr_weight, float(nw_kwargs.get("block_lr_zero_threshold", 0.0)) + ) + + return down_lr_weight, mid_lr_weight, up_lr_weight + + +def create_network( + multiplier: float, + network_dim: Optional[int], + network_alpha: Optional[float], + vae: AutoencoderKL, + text_encoder: Union[CLIPTextModel, List[CLIPTextModel]], + unet, + neuron_dropout: Optional[float] = None, + **kwargs, +): + if network_dim is None: + network_dim = 4 # default + if network_alpha is None: + network_alpha = 1.0 + + # extract dim/alpha for conv2d, and block dim + conv_dim = kwargs.get("conv_dim", None) + conv_alpha = kwargs.get("conv_alpha", None) + if conv_dim is not None: + conv_dim = int(conv_dim) + if conv_alpha is None: + conv_alpha = 1.0 + else: + conv_alpha = float(conv_alpha) + + # block dim/alpha/lr + block_dims = kwargs.get("block_dims", None) + down_lr_weight, mid_lr_weight, up_lr_weight = parse_block_lr_kwargs(kwargs) + + # 以上のいずれかに指定があればblockごとのdim(rank)を有効にする + if block_dims is not None or down_lr_weight is not None or mid_lr_weight is not None or up_lr_weight is not None: + block_alphas = kwargs.get("block_alphas", None) + conv_block_dims = kwargs.get("conv_block_dims", None) + conv_block_alphas = kwargs.get("conv_block_alphas", None) + + block_dims, block_alphas, conv_block_dims, conv_block_alphas = get_block_dims_and_alphas( + block_dims, block_alphas, network_dim, network_alpha, conv_block_dims, conv_block_alphas, conv_dim, conv_alpha + ) + + # remove block dim/alpha without learning rate + block_dims, block_alphas, conv_block_dims, conv_block_alphas = remove_block_dims_and_alphas( + block_dims, block_alphas, conv_block_dims, conv_block_alphas, down_lr_weight, mid_lr_weight, up_lr_weight + ) + + else: + block_alphas = None + conv_block_dims = None + conv_block_alphas = None + + # rank/module dropout + rank_dropout = kwargs.get("rank_dropout", None) + if rank_dropout is not None: + rank_dropout = float(rank_dropout) + module_dropout = kwargs.get("module_dropout", None) + if module_dropout is not None: + module_dropout = float(module_dropout) + + # すごく引数が多いな ( ^ω^)･･･ + network = LoRANetwork( + text_encoder, + unet, + multiplier=multiplier, + lora_dim=network_dim, + alpha=network_alpha, + dropout=neuron_dropout, + rank_dropout=rank_dropout, + module_dropout=module_dropout, + conv_lora_dim=conv_dim, + conv_alpha=conv_alpha, + block_dims=block_dims, + block_alphas=block_alphas, + conv_block_dims=conv_block_dims, + conv_block_alphas=conv_block_alphas, + varbose=True, + ) + + if up_lr_weight is not None or mid_lr_weight is not None or down_lr_weight is not None: + network.set_block_lr_weight(up_lr_weight, mid_lr_weight, down_lr_weight) + + return network + + +# このメソッドは外部から呼び出される可能性を考慮しておく +# network_dim, network_alpha にはデフォルト値が入っている。 +# block_dims, block_alphas は両方ともNoneまたは両方とも値が入っている +# conv_dim, conv_alpha は両方ともNoneまたは両方とも値が入っている +def get_block_dims_and_alphas( + block_dims, block_alphas, network_dim, network_alpha, conv_block_dims, conv_block_alphas, conv_dim, conv_alpha +): + num_total_blocks = LoRANetwork.NUM_OF_BLOCKS * 2 + 1 + + def parse_ints(s): + return [int(i) for i in s.split(",")] + + def parse_floats(s): + return [float(i) for i in s.split(",")] + + # block_dimsとblock_alphasをパースする。必ず値が入る + if block_dims is not None: + block_dims = parse_ints(block_dims) + assert ( + len(block_dims) == num_total_blocks + ), f"block_dims must have {num_total_blocks} elements / block_dimsは{num_total_blocks}個指定してください" + else: + logger.warning(f"block_dims is not specified. all dims are set to {network_dim} / block_dimsが指定されていません。すべてのdimは{network_dim}になります") + block_dims = [network_dim] * num_total_blocks + + if block_alphas is not None: + block_alphas = parse_floats(block_alphas) + assert ( + len(block_alphas) == num_total_blocks + ), f"block_alphas must have {num_total_blocks} elements / block_alphasは{num_total_blocks}個指定してください" + else: + logger.warning( + f"block_alphas is not specified. all alphas are set to {network_alpha} / block_alphasが指定されていません。すべてのalphaは{network_alpha}になります" + ) + block_alphas = [network_alpha] * num_total_blocks + + # conv_block_dimsとconv_block_alphasを、指定がある場合のみパースする。指定がなければconv_dimとconv_alphaを使う + if conv_block_dims is not None: + conv_block_dims = parse_ints(conv_block_dims) + assert ( + len(conv_block_dims) == num_total_blocks + ), f"conv_block_dims must have {num_total_blocks} elements / conv_block_dimsは{num_total_blocks}個指定してください" + + if conv_block_alphas is not None: + conv_block_alphas = parse_floats(conv_block_alphas) + assert ( + len(conv_block_alphas) == num_total_blocks + ), f"conv_block_alphas must have {num_total_blocks} elements / conv_block_alphasは{num_total_blocks}個指定してください" + else: + if conv_alpha is None: + conv_alpha = 1.0 + logger.warning( + f"conv_block_alphas is not specified. all alphas are set to {conv_alpha} / conv_block_alphasが指定されていません。すべてのalphaは{conv_alpha}になります" + ) + conv_block_alphas = [conv_alpha] * num_total_blocks + else: + if conv_dim is not None: + logger.warning( + f"conv_dim/alpha for all blocks are set to {conv_dim} and {conv_alpha} / すべてのブロックのconv_dimとalphaは{conv_dim}および{conv_alpha}になります" + ) + conv_block_dims = [conv_dim] * num_total_blocks + conv_block_alphas = [conv_alpha] * num_total_blocks + else: + conv_block_dims = None + conv_block_alphas = None + + return block_dims, block_alphas, conv_block_dims, conv_block_alphas + + +# 層別学習率用に層ごとの学習率に対する倍率を定義する、外部から呼び出される可能性を考慮しておく +def get_block_lr_weight( + down_lr_weight, mid_lr_weight, up_lr_weight, zero_threshold +) -> Tuple[List[float], List[float], List[float]]: + # パラメータ未指定時は何もせず、今までと同じ動作とする + if up_lr_weight is None and mid_lr_weight is None and down_lr_weight is None: + return None, None, None + + max_len = LoRANetwork.NUM_OF_BLOCKS # フルモデル相当でのup,downの層の数 + + def get_list(name_with_suffix) -> List[float]: + import math + + tokens = name_with_suffix.split("+") + name = tokens[0] + base_lr = float(tokens[1]) if len(tokens) > 1 else 0.0 + + if name == "cosine": + return [math.sin(math.pi * (i / (max_len - 1)) / 2) + base_lr for i in reversed(range(max_len))] + elif name == "sine": + return [math.sin(math.pi * (i / (max_len - 1)) / 2) + base_lr for i in range(max_len)] + elif name == "linear": + return [i / (max_len - 1) + base_lr for i in range(max_len)] + elif name == "reverse_linear": + return [i / (max_len - 1) + base_lr for i in reversed(range(max_len))] + elif name == "zeros": + return [0.0 + base_lr] * max_len + else: + logger.error( + "Unknown lr_weight argument %s is used. Valid arguments: / 不明なlr_weightの引数 %s が使われました。有効な引数:\n\tcosine, sine, linear, reverse_linear, zeros" + % (name) + ) + return None + + if type(down_lr_weight) == str: + down_lr_weight = get_list(down_lr_weight) + if type(up_lr_weight) == str: + up_lr_weight = get_list(up_lr_weight) + + if (up_lr_weight != None and len(up_lr_weight) > max_len) or (down_lr_weight != None and len(down_lr_weight) > max_len): + logger.warning("down_weight or up_weight is too long. Parameters after %d-th are ignored." % max_len) + logger.warning("down_weightもしくはup_weightが長すぎます。%d個目以降のパラメータは無視されます。" % max_len) + up_lr_weight = up_lr_weight[:max_len] + down_lr_weight = down_lr_weight[:max_len] + + if (up_lr_weight != None and len(up_lr_weight) < max_len) or (down_lr_weight != None and len(down_lr_weight) < max_len): + logger.warning("down_weight or up_weight is too short. Parameters after %d-th are filled with 1." % max_len) + logger.warning("down_weightもしくはup_weightが短すぎます。%d個目までの不足したパラメータは1で補われます。" % max_len) + + if down_lr_weight != None and len(down_lr_weight) < max_len: + down_lr_weight = down_lr_weight + [1.0] * (max_len - len(down_lr_weight)) + if up_lr_weight != None and len(up_lr_weight) < max_len: + up_lr_weight = up_lr_weight + [1.0] * (max_len - len(up_lr_weight)) + + if (up_lr_weight != None) or (mid_lr_weight != None) or (down_lr_weight != None): + logger.info("apply block learning rate / 階層別学習率を適用します。") + if down_lr_weight != None: + down_lr_weight = [w if w > zero_threshold else 0 for w in down_lr_weight] + logger.info(f"down_lr_weight (shallower -> deeper, 浅い層->深い層): {down_lr_weight}") + else: + logger.info("down_lr_weight: all 1.0, すべて1.0") + + if mid_lr_weight != None: + mid_lr_weight = mid_lr_weight if mid_lr_weight > zero_threshold else 0 + logger.info(f"mid_lr_weight: {mid_lr_weight}") + else: + logger.info("mid_lr_weight: 1.0") + + if up_lr_weight != None: + up_lr_weight = [w if w > zero_threshold else 0 for w in up_lr_weight] + logger.info(f"up_lr_weight (deeper -> shallower, 深い層->浅い層): {up_lr_weight}") + else: + logger.info("up_lr_weight: all 1.0, すべて1.0") + + return down_lr_weight, mid_lr_weight, up_lr_weight + + +# lr_weightが0のblockをblock_dimsから除外する、外部から呼び出す可能性を考慮しておく +def remove_block_dims_and_alphas( + block_dims, block_alphas, conv_block_dims, conv_block_alphas, down_lr_weight, mid_lr_weight, up_lr_weight +): + # set 0 to block dim without learning rate to remove the block + if down_lr_weight != None: + for i, lr in enumerate(down_lr_weight): + if lr == 0: + block_dims[i] = 0 + if conv_block_dims is not None: + conv_block_dims[i] = 0 + if mid_lr_weight != None: + if mid_lr_weight == 0: + block_dims[LoRANetwork.NUM_OF_BLOCKS] = 0 + if conv_block_dims is not None: + conv_block_dims[LoRANetwork.NUM_OF_BLOCKS] = 0 + if up_lr_weight != None: + for i, lr in enumerate(up_lr_weight): + if lr == 0: + block_dims[LoRANetwork.NUM_OF_BLOCKS + 1 + i] = 0 + if conv_block_dims is not None: + conv_block_dims[LoRANetwork.NUM_OF_BLOCKS + 1 + i] = 0 + + return block_dims, block_alphas, conv_block_dims, conv_block_alphas + + +# 外部から呼び出す可能性を考慮しておく +def get_block_index(lora_name: str) -> int: + block_idx = -1 # invalid lora name + + m = RE_UPDOWN.search(lora_name) + if m: + g = m.groups() + i = int(g[1]) + j = int(g[3]) + if g[2] == "resnets": + idx = 3 * i + j + elif g[2] == "attentions": + idx = 3 * i + j + elif g[2] == "upsamplers" or g[2] == "downsamplers": + idx = 3 * i + 2 + + if g[0] == "down": + block_idx = 1 + idx # 0に該当するLoRAは存在しない + elif g[0] == "up": + block_idx = LoRANetwork.NUM_OF_BLOCKS + 1 + idx + + elif "mid_block_" in lora_name: + block_idx = LoRANetwork.NUM_OF_BLOCKS # idx=12 + + return block_idx + + +# Create network from weights for inference, weights are not loaded here (because can be merged) +def create_network_from_weights(multiplier, file, vae, text_encoder, unet, weights_sd=None, for_inference=False, **kwargs): + if weights_sd is None: + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file, safe_open + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + # get dim/alpha mapping + modules_dim = {} + modules_alpha = {} + for key, value in weights_sd.items(): + if "." not in key: + continue + + lora_name = key.split(".")[0] + if "alpha" in key: + modules_alpha[lora_name] = value + elif "lora_down" in key: + dim = value.size()[0] + modules_dim[lora_name] = dim + # logger.info(lora_name, value.size(), dim) + + # support old LoRA without alpha + for key in modules_dim.keys(): + if key not in modules_alpha: + modules_alpha[key] = modules_dim[key] + + module_class = LoRAInfModule if for_inference else LoRAModule + + network = LoRANetwork( + text_encoder, unet, multiplier=multiplier, modules_dim=modules_dim, modules_alpha=modules_alpha, module_class=module_class + ) + + # block lr + down_lr_weight, mid_lr_weight, up_lr_weight = parse_block_lr_kwargs(kwargs) + if up_lr_weight is not None or mid_lr_weight is not None or down_lr_weight is not None: + network.set_block_lr_weight(up_lr_weight, mid_lr_weight, down_lr_weight) + + return network, weights_sd + + +class LoRANetwork(torch.nn.Module): + NUM_OF_BLOCKS = 12 # フルモデル相当でのup,downの層の数 + + UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"] + UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPSdpaAttention", "CLIPMLP"] + LORA_PREFIX_UNET = "lora_unet" + LORA_PREFIX_TEXT_ENCODER = "lora_te" + + # SDXL: must starts with LORA_PREFIX_TEXT_ENCODER + LORA_PREFIX_TEXT_ENCODER1 = "lora_te1" + LORA_PREFIX_TEXT_ENCODER2 = "lora_te2" + + def __init__( + self, + text_encoder: Union[List[CLIPTextModel], CLIPTextModel], + unet, + multiplier: float = 1.0, + lora_dim: int = 4, + alpha: float = 1, + dropout: Optional[float] = None, + rank_dropout: Optional[float] = None, + module_dropout: Optional[float] = None, + conv_lora_dim: Optional[int] = None, + conv_alpha: Optional[float] = None, + block_dims: Optional[List[int]] = None, + block_alphas: Optional[List[float]] = None, + conv_block_dims: Optional[List[int]] = None, + conv_block_alphas: Optional[List[float]] = None, + modules_dim: Optional[Dict[str, int]] = None, + modules_alpha: Optional[Dict[str, int]] = None, + module_class: Type[object] = LoRAModule, + varbose: Optional[bool] = False, + ) -> None: + """ + LoRA network: すごく引数が多いが、パターンは以下の通り + 1. lora_dimとalphaを指定 + 2. lora_dim、alpha、conv_lora_dim、conv_alphaを指定 + 3. block_dimsとblock_alphasを指定 : Conv2d3x3には適用しない + 4. block_dims、block_alphas、conv_block_dims、conv_block_alphasを指定 : Conv2d3x3にも適用する + 5. modules_dimとmodules_alphaを指定 (推論用) + """ + super().__init__() + self.multiplier = multiplier + + self.lora_dim = lora_dim + self.alpha = alpha + self.conv_lora_dim = conv_lora_dim + self.conv_alpha = conv_alpha + self.dropout = dropout + self.rank_dropout = rank_dropout + self.module_dropout = module_dropout + + if modules_dim is not None: + logger.info(f"create LoRA network from weights") + elif block_dims is not None: + logger.info(f"create LoRA network from block_dims") + logger.info(f"neuron dropout: p={self.dropout}, rank dropout: p={self.rank_dropout}, module dropout: p={self.module_dropout}") + logger.info(f"block_dims: {block_dims}") + logger.info(f"block_alphas: {block_alphas}") + if conv_block_dims is not None: + logger.info(f"conv_block_dims: {conv_block_dims}") + logger.info(f"conv_block_alphas: {conv_block_alphas}") + else: + logger.info(f"create LoRA network. base dim (rank): {lora_dim}, alpha: {alpha}") + logger.info(f"neuron dropout: p={self.dropout}, rank dropout: p={self.rank_dropout}, module dropout: p={self.module_dropout}") + if self.conv_lora_dim is not None: + logger.info(f"apply LoRA to Conv2d with kernel size (3,3). dim (rank): {self.conv_lora_dim}, alpha: {self.conv_alpha}") + + # create module instances + def create_modules( + is_unet: bool, + text_encoder_idx: Optional[int], # None, 1, 2 + root_module: torch.nn.Module, + target_replace_modules: List[torch.nn.Module], + ) -> List[LoRAModule]: + prefix = ( + self.LORA_PREFIX_UNET + if is_unet + else ( + self.LORA_PREFIX_TEXT_ENCODER + if text_encoder_idx is None + else (self.LORA_PREFIX_TEXT_ENCODER1 if text_encoder_idx == 1 else self.LORA_PREFIX_TEXT_ENCODER2) + ) + ) + loras = [] + skipped = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + is_linear = child_module.__class__.__name__ == "Linear" + is_conv2d = child_module.__class__.__name__ == "Conv2d" + is_conv2d_1x1 = is_conv2d and child_module.kernel_size == (1, 1) + + if is_linear or is_conv2d: + lora_name = prefix + "." + name + "." + child_name + lora_name = lora_name.replace(".", "_") + + dim = None + alpha = None + + if modules_dim is not None: + # モジュール指定あり + if lora_name in modules_dim: + dim = modules_dim[lora_name] + alpha = modules_alpha[lora_name] + elif is_unet and block_dims is not None: + # U-Netでblock_dims指定あり + block_idx = get_block_index(lora_name) + if is_linear or is_conv2d_1x1: + dim = block_dims[block_idx] + alpha = block_alphas[block_idx] + elif conv_block_dims is not None: + dim = conv_block_dims[block_idx] + alpha = conv_block_alphas[block_idx] + else: + # 通常、すべて対象とする + if is_linear or is_conv2d_1x1: + dim = self.lora_dim + alpha = self.alpha + elif self.conv_lora_dim is not None: + dim = self.conv_lora_dim + alpha = self.conv_alpha + + if dim is None or dim == 0: + # skipした情報を出力 + if is_linear or is_conv2d_1x1 or (self.conv_lora_dim is not None or conv_block_dims is not None): + skipped.append(lora_name) + continue + + lora = module_class( + lora_name, + child_module, + self.multiplier, + dim, + alpha, + dropout=dropout, + rank_dropout=rank_dropout, + module_dropout=module_dropout, + ) + loras.append(lora) + return loras, skipped + + text_encoders = text_encoder if type(text_encoder) == list else [text_encoder] + + # create LoRA for text encoder + # 毎回すべてのモジュールを作るのは無駄なので要検討 + self.text_encoder_loras = [] + skipped_te = [] + for i, text_encoder in enumerate(text_encoders): + if len(text_encoders) > 1: + index = i + 1 + logger.info(f"create LoRA for Text Encoder {index}:") + else: + index = None + logger.info(f"create LoRA for Text Encoder:") + + text_encoder_loras, skipped = create_modules(False, index, text_encoder, LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE) + self.text_encoder_loras.extend(text_encoder_loras) + skipped_te += skipped + logger.info(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.") + + # extend U-Net target modules if conv2d 3x3 is enabled, or load from weights + target_modules = LoRANetwork.UNET_TARGET_REPLACE_MODULE + if modules_dim is not None or self.conv_lora_dim is not None or conv_block_dims is not None: + target_modules += LoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + + self.unet_loras, skipped_un = create_modules(True, None, unet, target_modules) + logger.info(f"create LoRA for U-Net: {len(self.unet_loras)} modules.") + + skipped = skipped_te + skipped_un + if varbose and len(skipped) > 0: + logger.warning( + f"because block_lr_weight is 0 or dim (rank) is 0, {len(skipped)} LoRA modules are skipped / block_lr_weightまたはdim (rank)が0の為、次の{len(skipped)}個のLoRAモジュールはスキップされます:" + ) + for name in skipped: + logger.info(f"\t{name}") + + self.up_lr_weight: List[float] = None + self.down_lr_weight: List[float] = None + self.mid_lr_weight: float = None + self.block_lr = False + + # assertion + names = set() + for lora in self.text_encoder_loras + self.unet_loras: + assert lora.lora_name not in names, f"duplicated lora name: {lora.lora_name}" + names.add(lora.lora_name) + + def set_multiplier(self, multiplier): + self.multiplier = multiplier + for lora in self.text_encoder_loras + self.unet_loras: + lora.multiplier = self.multiplier + + def load_weights(self, file): + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + info = self.load_state_dict(weights_sd, False) + return info + + def apply_to(self, text_encoder, unet, apply_text_encoder=True, apply_unet=True): + if apply_text_encoder: + logger.info("enable LoRA for text encoder") + else: + self.text_encoder_loras = [] + + if apply_unet: + logger.info("enable LoRA for U-Net") + else: + self.unet_loras = [] + + for lora in self.text_encoder_loras + self.unet_loras: + lora.apply_to() + self.add_module(lora.lora_name, lora) + + # マージできるかどうかを返す + def is_mergeable(self): + return True + + # TODO refactor to common function with apply_to + def merge_to(self, text_encoder, unet, weights_sd, dtype, device): + apply_text_encoder = apply_unet = False + for key in weights_sd.keys(): + if key.startswith(LoRANetwork.LORA_PREFIX_TEXT_ENCODER): + apply_text_encoder = True + elif key.startswith(LoRANetwork.LORA_PREFIX_UNET): + apply_unet = True + + if apply_text_encoder: + logger.info("enable LoRA for text encoder") + else: + self.text_encoder_loras = [] + + if apply_unet: + logger.info("enable LoRA for U-Net") + else: + self.unet_loras = [] + + for lora in self.text_encoder_loras + self.unet_loras: + sd_for_lora = {} + for key in weights_sd.keys(): + if key.startswith(lora.lora_name): + sd_for_lora[key[len(lora.lora_name) + 1 :]] = weights_sd[key] + lora.merge_to(sd_for_lora, dtype, device) + + logger.info(f"weights are merged") + + # 層別学習率用に層ごとの学習率に対する倍率を定義する　引数の順番が逆だがとりあえず気にしない + def set_block_lr_weight( + self, + up_lr_weight: List[float] = None, + mid_lr_weight: float = None, + down_lr_weight: List[float] = None, + ): + self.block_lr = True + self.down_lr_weight = down_lr_weight + self.mid_lr_weight = mid_lr_weight + self.up_lr_weight = up_lr_weight + + def get_lr_weight(self, lora: LoRAModule) -> float: + lr_weight = 1.0 + block_idx = get_block_index(lora.lora_name) + if block_idx < 0: + return lr_weight + + if block_idx < LoRANetwork.NUM_OF_BLOCKS: + if self.down_lr_weight != None: + lr_weight = self.down_lr_weight[block_idx] + elif block_idx == LoRANetwork.NUM_OF_BLOCKS: + if self.mid_lr_weight != None: + lr_weight = self.mid_lr_weight + elif block_idx > LoRANetwork.NUM_OF_BLOCKS: + if self.up_lr_weight != None: + lr_weight = self.up_lr_weight[block_idx - LoRANetwork.NUM_OF_BLOCKS - 1] + + return lr_weight + + # 二つのText Encoderに別々の学習率を設定できるようにするといいかも + def prepare_optimizer_params(self, text_encoder_lr, unet_lr, default_lr): + self.requires_grad_(True) + all_params = [] + + def enumerate_params(loras: List[LoRAModule]): + params = [] + for lora in loras: + # params.extend(lora.parameters()) + params.extend(lora.get_trainable_params()) + return params + + if self.text_encoder_loras: + param_data = {"params": enumerate_params(self.text_encoder_loras)} + if text_encoder_lr is not None: + param_data["lr"] = text_encoder_lr + all_params.append(param_data) + + if self.unet_loras: + if self.block_lr: + # 学習率のグラフをblockごとにしたいので、blockごとにloraを分類 + block_idx_to_lora = {} + for lora in self.unet_loras: + idx = get_block_index(lora.lora_name) + if idx not in block_idx_to_lora: + block_idx_to_lora[idx] = [] + block_idx_to_lora[idx].append(lora) + + # blockごとにパラメータを設定する + for idx, block_loras in block_idx_to_lora.items(): + param_data = {"params": enumerate_params(block_loras)} + + if unet_lr is not None: + param_data["lr"] = unet_lr * self.get_lr_weight(block_loras[0]) + elif default_lr is not None: + param_data["lr"] = default_lr * self.get_lr_weight(block_loras[0]) + if ("lr" in param_data) and (param_data["lr"] == 0): + continue + all_params.append(param_data) + + else: + param_data = {"params": enumerate_params(self.unet_loras)} + if unet_lr is not None: + param_data["lr"] = unet_lr + all_params.append(param_data) + + return all_params + + def enable_gradient_checkpointing(self): + # not supported + pass + + def prepare_grad_etc(self, text_encoder, unet): + self.requires_grad_(True) + + def on_epoch_start(self, text_encoder, unet): + self.train() + + def get_trainable_params(self): + return self.parameters() + + def save_weights(self, file, dtype, metadata): + if metadata is not None and len(metadata) == 0: + metadata = None + + state_dict = self.state_dict() + + if dtype is not None: + for key in list(state_dict.keys()): + v = state_dict[key] + v = v.detach().clone().to("cpu").to(dtype) + state_dict[key] = v + + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import save_file + from library import train_util + + # Precalculate model hashes to save time on indexing + if metadata is None: + metadata = {} + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + save_file(state_dict, file, metadata) + else: + torch.save(state_dict, file) + + # mask is a tensor with values from 0 to 1 + def set_region(self, sub_prompt_index, is_last_network, mask): + if mask.max() == 0: + mask = torch.ones_like(mask) + + self.mask = mask + self.sub_prompt_index = sub_prompt_index + self.is_last_network = is_last_network + + for lora in self.text_encoder_loras + self.unet_loras: + lora.set_network(self) + + def set_current_generation(self, batch_size, num_sub_prompts, width, height, shared): + self.batch_size = batch_size + self.num_sub_prompts = num_sub_prompts + self.current_size = (height, width) + self.shared = shared + + # create masks + mask = self.mask + mask_dic = {} + mask = mask.unsqueeze(0).unsqueeze(1) # b(1),c(1),h,w + ref_weight = self.text_encoder_loras[0].lora_down.weight if self.text_encoder_loras else self.unet_loras[0].lora_down.weight + dtype = ref_weight.dtype + device = ref_weight.device + + def resize_add(mh, mw): + # logger.info(mh, mw, mh * mw) + m = torch.nn.functional.interpolate(mask, (mh, mw), mode="bilinear") # doesn't work in bf16 + m = m.to(device, dtype=dtype) + mask_dic[mh * mw] = m + + h = height // 8 + w = width // 8 + for _ in range(4): + resize_add(h, w) + if h % 2 == 1 or w % 2 == 1: # add extra shape if h/w is not divisible by 2 + resize_add(h + h % 2, w + w % 2) + h = (h + 1) // 2 + w = (w + 1) // 2 + + self.mask_dic = mask_dic + + def backup_weights(self): + # 重みのバックアップを行う + loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras + for lora in loras: + org_module = lora.org_module_ref[0] + if not hasattr(org_module, "_lora_org_weight"): + sd = org_module.state_dict() + org_module._lora_org_weight = sd["weight"].detach().clone() + org_module._lora_restored = True + + def restore_weights(self): + # 重みのリストアを行う + loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras + for lora in loras: + org_module = lora.org_module_ref[0] + if not org_module._lora_restored: + sd = org_module.state_dict() + sd["weight"] = org_module._lora_org_weight + org_module.load_state_dict(sd) + org_module._lora_restored = True + + def pre_calculation(self): + # 事前計算を行う + loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras + for lora in loras: + org_module = lora.org_module_ref[0] + sd = org_module.state_dict() + + org_weight = sd["weight"] + lora_weight = lora.get_weight().to(org_weight.device, dtype=org_weight.dtype) + sd["weight"] = org_weight + lora_weight + assert sd["weight"].shape == org_weight.shape + org_module.load_state_dict(sd) + + org_module._lora_restored = False + lora.enabled = False + + def apply_max_norm_regularization(self, max_norm_value, device): + downkeys = [] + upkeys = [] + alphakeys = [] + norms = [] + keys_scaled = 0 + + state_dict = self.state_dict() + for key in state_dict.keys(): + if "lora_down" in key and "weight" in key: + downkeys.append(key) + upkeys.append(key.replace("lora_down", "lora_up")) + alphakeys.append(key.replace("lora_down.weight", "alpha")) + + for i in range(len(downkeys)): + down = state_dict[downkeys[i]].to(device) + up = state_dict[upkeys[i]].to(device) + alpha = state_dict[alphakeys[i]].to(device) + dim = down.shape[0] + scale = alpha / dim + + if up.shape[2:] == (1, 1) and down.shape[2:] == (1, 1): + updown = (up.squeeze(2).squeeze(2) @ down.squeeze(2).squeeze(2)).unsqueeze(2).unsqueeze(3) + elif up.shape[2:] == (3, 3) or down.shape[2:] == (3, 3): + updown = torch.nn.functional.conv2d(down.permute(1, 0, 2, 3), up).permute(1, 0, 2, 3) + else: + updown = up @ down + + updown *= scale + + norm = updown.norm().clamp(min=max_norm_value / 2) + desired = torch.clamp(norm, max=max_norm_value) + ratio = desired.cpu() / norm.cpu() + sqrt_ratio = ratio**0.5 + if ratio != 1: + keys_scaled += 1 + state_dict[upkeys[i]] *= sqrt_ratio + state_dict[downkeys[i]] *= sqrt_ratio + scalednorm = updown.norm() * ratio + norms.append(scalednorm.item()) + + return keys_scaled, sum(norms) / len(norms), max(norms) diff --git a/lora_interrogator.py b/lora_interrogator.py new file mode 100644 index 0000000000000000000000000000000000000000..6aaa58107136b19792cf5890d93e0f1ee5dcef88 --- /dev/null +++ b/lora_interrogator.py @@ -0,0 +1,146 @@ + + +from tqdm import tqdm +from library import model_util +import library.train_util as train_util +import argparse +from transformers import CLIPTokenizer + +import torch +from library.device_utils import init_ipex, get_preferred_device +init_ipex() + +import library.model_util as model_util +import lora +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +TOKENIZER_PATH = "openai/clip-vit-large-patch14" +V2_STABLE_DIFFUSION_PATH = "stabilityai/stable-diffusion-2" # ここからtokenizerだけ使う + +DEVICE = get_preferred_device() + + +def interrogate(args): + weights_dtype = torch.float16 + + # いろいろ準備する + logger.info(f"loading SD model: {args.sd_model}") + args.pretrained_model_name_or_path = args.sd_model + args.vae = None + text_encoder, vae, unet, _ = train_util._load_target_model(args,weights_dtype, DEVICE) + + logger.info(f"loading LoRA: {args.model}") + network, weights_sd = lora.create_network_from_weights(1.0, args.model, vae, text_encoder, unet) + + # text encoder向けの重みがあるかチェックする：本当はlora側でやるのがいい + has_te_weight = False + for key in weights_sd.keys(): + if 'lora_te' in key: + has_te_weight = True + break + if not has_te_weight: + logger.error("This LoRA does not have modules for Text Encoder, cannot interrogate / このLoRAはText Encoder向けのモジュールがないため調査できません") + return + del vae + + logger.info("loading tokenizer") + if args.v2: + tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(V2_STABLE_DIFFUSION_PATH, subfolder="tokenizer") + else: + tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(TOKENIZER_PATH) # , model_max_length=max_token_length + 2) + + text_encoder.to(DEVICE, dtype=weights_dtype) + text_encoder.eval() + unet.to(DEVICE, dtype=weights_dtype) + unet.eval() # U-Netは呼び出さないので不要だけど + + # トークンをひとつひとつ当たっていく + token_id_start = 0 + token_id_end = max(tokenizer.all_special_ids) + logger.info(f"interrogate tokens are: {token_id_start} to {token_id_end}") + + def get_all_embeddings(text_encoder): + embs = [] + with torch.no_grad(): + for token_id in tqdm(range(token_id_start, token_id_end + 1, args.batch_size)): + batch = [] + for tid in range(token_id, min(token_id_end + 1, token_id + args.batch_size)): + tokens = [tokenizer.bos_token_id, tid, tokenizer.eos_token_id] + # tokens = [tid] # こちらは結果がいまひとつ + batch.append(tokens) + + # batch_embs = text_encoder(torch.tensor(batch).to(DEVICE))[0].to("cpu") # bos/eosも含めたほうが差が出るようだ [:, 1] + # clip skip対応 + batch = torch.tensor(batch).to(DEVICE) + if args.clip_skip is None: + encoder_hidden_states = text_encoder(batch)[0] + else: + enc_out = text_encoder(batch, output_hidden_states=True, return_dict=True) + encoder_hidden_states = enc_out['hidden_states'][-args.clip_skip] + encoder_hidden_states = text_encoder.text_model.final_layer_norm(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.to("cpu") + + embs.extend(encoder_hidden_states) + return torch.stack(embs) + + logger.info("get original text encoder embeddings.") + orig_embs = get_all_embeddings(text_encoder) + + network.apply_to(text_encoder, unet, True, len(network.unet_loras) > 0) + info = network.load_state_dict(weights_sd, strict=False) + logger.info(f"Loading LoRA weights: {info}") + + network.to(DEVICE, dtype=weights_dtype) + network.eval() + + del unet + + logger.info("You can ignore warning messages start with '_IncompatibleKeys' (LoRA model does not have alpha because trained by older script) / '_IncompatibleKeys'の警告は無視して構いません（以前のスクリプトで学習されたLoRAモデルのためalphaの定義がありません）") + logger.info("get text encoder embeddings with lora.") + lora_embs = get_all_embeddings(text_encoder) + + # 比べる：とりあえず単純に差分の絶対値で + logger.info("comparing...") + diffs = {} + for i, (orig_emb, lora_emb) in enumerate(zip(orig_embs, tqdm(lora_embs))): + diff = torch.mean(torch.abs(orig_emb - lora_emb)) + # diff = torch.mean(torch.cosine_similarity(orig_emb, lora_emb, dim=1)) # うまく検出できない + diff = float(diff.detach().to('cpu').numpy()) + diffs[token_id_start + i] = diff + + diffs_sorted = sorted(diffs.items(), key=lambda x: -x[1]) + + # 結果を表示する + print("top 100:") + for i, (token, diff) in enumerate(diffs_sorted[:100]): + # if diff < 1e-6: + # break + string = tokenizer.convert_tokens_to_string(tokenizer.convert_ids_to_tokens([token])) + print(f"[{i:3d}]: {token:5d} {string:<20s}: {diff:.5f}") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + + parser.add_argument("--v2", action='store_true', + help='load Stable Diffusion v2.x model / Stable Diffusion 2.xのモデルを読み込む') + parser.add_argument("--sd_model", type=str, default=None, + help="Stable Diffusion model to load: ckpt or safetensors file / 読み込むSDのモデル、ckptまたはsafetensors") + parser.add_argument("--model", type=str, default=None, + help="LoRA model to interrogate: ckpt or safetensors file / 調査するLoRAモデル、ckptまたはsafetensors") + parser.add_argument("--batch_size", type=int, default=16, + help="batch size for processing with Text Encoder / Text Encoderで処理するときのバッチサイズ") + parser.add_argument("--clip_skip", type=int, default=None, + help="use output of nth layer from back of text encoder (n>=1) / text encoderの後ろからn番目の層の出力を用いる（nは1以上）") + + return parser + + +if __name__ == '__main__': + parser = setup_parser() + + args = parser.parse_args() + interrogate(args) diff --git a/lpw_stable_diffusion.py b/lpw_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..5717233d47ba82deef7e540b7300e2112ae4e0af --- /dev/null +++ b/lpw_stable_diffusion.py @@ -0,0 +1,1233 @@ +# copy from https://github.com/huggingface/diffusers/blob/main/examples/community/lpw_stable_diffusion.py +# and modify to support SD2.x + +import inspect +import re +from typing import Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +import diffusers +from diffusers import SchedulerMixin, StableDiffusionPipeline +from diffusers.models import AutoencoderKL, UNet2DConditionModel +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker +from diffusers.utils import logging + +try: + from diffusers.utils import PIL_INTERPOLATION +except ImportError: + if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"): + PIL_INTERPOLATION = { + "linear": PIL.Image.Resampling.BILINEAR, + "bilinear": PIL.Image.Resampling.BILINEAR, + "bicubic": PIL.Image.Resampling.BICUBIC, + "lanczos": PIL.Image.Resampling.LANCZOS, + "nearest": PIL.Image.Resampling.NEAREST, + } + else: + PIL_INTERPOLATION = { + "linear": PIL.Image.LINEAR, + "bilinear": PIL.Image.BILINEAR, + "bicubic": PIL.Image.BICUBIC, + "lanczos": PIL.Image.LANCZOS, + "nearest": PIL.Image.NEAREST, + } +# ------------------------------------------------------------------------------ + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +re_attention = re.compile( + r""" +\\\(| +\\\)| +\\\[| +\\]| +\\\\| +\\| +\(| +\[| +:([+-]?[.\d]+)\)| +\)| +]| +[^\\()\[\]:]+| +: +""", + re.X, +) + + +def parse_prompt_attention(text): + """ + Parses a string with attention tokens and returns a list of pairs: text and its associated weight. + Accepted tokens are: + (abc) - increases attention to abc by a multiplier of 1.1 + (abc:3.12) - increases attention to abc by a multiplier of 3.12 + [abc] - decreases attention to abc by a multiplier of 1.1 + \( - literal character '(' + \[ - literal character '[' + \) - literal character ')' + \] - literal character ']' + \\ - literal character '\' + anything else - just text + >>> parse_prompt_attention('normal text') + [['normal text', 1.0]] + >>> parse_prompt_attention('an (important) word') + [['an ', 1.0], ['important', 1.1], [' word', 1.0]] + >>> parse_prompt_attention('(unbalanced') + [['unbalanced', 1.1]] + >>> parse_prompt_attention('\(literal\]') + [['(literal]', 1.0]] + >>> parse_prompt_attention('(unnecessary)(parens)') + [['unnecessaryparens', 1.1]] + >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).') + [['a ', 1.0], + ['house', 1.5730000000000004], + [' ', 1.1], + ['on', 1.0], + [' a ', 1.1], + ['hill', 0.55], + [', sun, ', 1.1], + ['sky', 1.4641000000000006], + ['.', 1.1]] + """ + + res = [] + round_brackets = [] + square_brackets = [] + + round_bracket_multiplier = 1.1 + square_bracket_multiplier = 1 / 1.1 + + def multiply_range(start_position, multiplier): + for p in range(start_position, len(res)): + res[p][1] *= multiplier + + for m in re_attention.finditer(text): + text = m.group(0) + weight = m.group(1) + + if text.startswith("\\"): + res.append([text[1:], 1.0]) + elif text == "(": + round_brackets.append(len(res)) + elif text == "[": + square_brackets.append(len(res)) + elif weight is not None and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), float(weight)) + elif text == ")" and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), round_bracket_multiplier) + elif text == "]" and len(square_brackets) > 0: + multiply_range(square_brackets.pop(), square_bracket_multiplier) + else: + res.append([text, 1.0]) + + for pos in round_brackets: + multiply_range(pos, round_bracket_multiplier) + + for pos in square_brackets: + multiply_range(pos, square_bracket_multiplier) + + if len(res) == 0: + res = [["", 1.0]] + + # merge runs of identical weights + i = 0 + while i + 1 < len(res): + if res[i][1] == res[i + 1][1]: + res[i][0] += res[i + 1][0] + res.pop(i + 1) + else: + i += 1 + + return res + + +def get_prompts_with_weights(pipe: StableDiffusionPipeline, prompt: List[str], max_length: int): + r""" + Tokenize a list of prompts and return its tokens with weights of each token. + + No padding, starting or ending token is included. + """ + tokens = [] + weights = [] + truncated = False + for text in prompt: + texts_and_weights = parse_prompt_attention(text) + text_token = [] + text_weight = [] + for word, weight in texts_and_weights: + # tokenize and discard the starting and the ending token + token = pipe.tokenizer(word).input_ids[1:-1] + text_token += token + # copy the weight by length of token + text_weight += [weight] * len(token) + # stop if the text is too long (longer than truncation limit) + if len(text_token) > max_length: + truncated = True + break + # truncate + if len(text_token) > max_length: + truncated = True + text_token = text_token[:max_length] + text_weight = text_weight[:max_length] + tokens.append(text_token) + weights.append(text_weight) + if truncated: + logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples") + return tokens, weights + + +def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, no_boseos_middle=True, chunk_length=77): + r""" + Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length. + """ + max_embeddings_multiples = (max_length - 2) // (chunk_length - 2) + weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length + for i in range(len(tokens)): + tokens[i] = [bos] + tokens[i] + [eos] * (max_length - 1 - len(tokens[i])) + if no_boseos_middle: + weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i])) + else: + w = [] + if len(weights[i]) == 0: + w = [1.0] * weights_length + else: + for j in range(max_embeddings_multiples): + w.append(1.0) # weight for starting token in this chunk + w += weights[i][j * (chunk_length - 2) : min(len(weights[i]), (j + 1) * (chunk_length - 2))] + w.append(1.0) # weight for ending token in this chunk + w += [1.0] * (weights_length - len(w)) + weights[i] = w[:] + + return tokens, weights + + +def get_unweighted_text_embeddings( + pipe: StableDiffusionPipeline, + text_input: torch.Tensor, + chunk_length: int, + clip_skip: int, + eos: int, + pad: int, + no_boseos_middle: Optional[bool] = True, +): + """ + When the length of tokens is a multiple of the capacity of the text encoder, + it should be split into chunks and sent to the text encoder individually. + """ + max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2) + if max_embeddings_multiples > 1: + text_embeddings = [] + for i in range(max_embeddings_multiples): + # extract the i-th chunk + text_input_chunk = text_input[:, i * (chunk_length - 2) : (i + 1) * (chunk_length - 2) + 2].clone() + + # cover the head and the tail by the starting and the ending tokens + text_input_chunk[:, 0] = text_input[0, 0] + if pad == eos: # v1 + text_input_chunk[:, -1] = text_input[0, -1] + else: # v2 + for j in range(len(text_input_chunk)): + if text_input_chunk[j, -1] != eos and text_input_chunk[j, -1] != pad: # 最後に普通の文字がある + text_input_chunk[j, -1] = eos + if text_input_chunk[j, 1] == pad: # BOSだけであとはPAD + text_input_chunk[j, 1] = eos + + if clip_skip is None or clip_skip == 1: + text_embedding = pipe.text_encoder(text_input_chunk)[0] + else: + enc_out = pipe.text_encoder(text_input_chunk, output_hidden_states=True, return_dict=True) + text_embedding = enc_out["hidden_states"][-clip_skip] + text_embedding = pipe.text_encoder.text_model.final_layer_norm(text_embedding) + + if no_boseos_middle: + if i == 0: + # discard the ending token + text_embedding = text_embedding[:, :-1] + elif i == max_embeddings_multiples - 1: + # discard the starting token + text_embedding = text_embedding[:, 1:] + else: + # discard both starting and ending tokens + text_embedding = text_embedding[:, 1:-1] + + text_embeddings.append(text_embedding) + text_embeddings = torch.concat(text_embeddings, axis=1) + else: + if clip_skip is None or clip_skip == 1: + text_embeddings = pipe.text_encoder(text_input)[0] + else: + enc_out = pipe.text_encoder(text_input, output_hidden_states=True, return_dict=True) + text_embeddings = enc_out["hidden_states"][-clip_skip] + text_embeddings = pipe.text_encoder.text_model.final_layer_norm(text_embeddings) + return text_embeddings + + +def get_weighted_text_embeddings( + pipe: StableDiffusionPipeline, + prompt: Union[str, List[str]], + uncond_prompt: Optional[Union[str, List[str]]] = None, + max_embeddings_multiples: Optional[int] = 3, + no_boseos_middle: Optional[bool] = False, + skip_parsing: Optional[bool] = False, + skip_weighting: Optional[bool] = False, + clip_skip=None, +): + r""" + Prompts can be assigned with local weights using brackets. For example, + prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful', + and the embedding tokens corresponding to the words get multiplied by a constant, 1.1. + + Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean. + + Args: + pipe (`StableDiffusionPipeline`): + Pipe to provide access to the tokenizer and the text encoder. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + uncond_prompt (`str` or `List[str]`): + The unconditional prompt or prompts for guide the image generation. If unconditional prompt + is provided, the embeddings of prompt and uncond_prompt are concatenated. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + no_boseos_middle (`bool`, *optional*, defaults to `False`): + If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and + ending token in each of the chunk in the middle. + skip_parsing (`bool`, *optional*, defaults to `False`): + Skip the parsing of brackets. + skip_weighting (`bool`, *optional*, defaults to `False`): + Skip the weighting. When the parsing is skipped, it is forced True. + """ + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + if isinstance(prompt, str): + prompt = [prompt] + + if not skip_parsing: + prompt_tokens, prompt_weights = get_prompts_with_weights(pipe, prompt, max_length - 2) + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens, uncond_weights = get_prompts_with_weights(pipe, uncond_prompt, max_length - 2) + else: + prompt_tokens = [token[1:-1] for token in pipe.tokenizer(prompt, max_length=max_length, truncation=True).input_ids] + prompt_weights = [[1.0] * len(token) for token in prompt_tokens] + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens = [ + token[1:-1] for token in pipe.tokenizer(uncond_prompt, max_length=max_length, truncation=True).input_ids + ] + uncond_weights = [[1.0] * len(token) for token in uncond_tokens] + + # round up the longest length of tokens to a multiple of (model_max_length - 2) + max_length = max([len(token) for token in prompt_tokens]) + if uncond_prompt is not None: + max_length = max(max_length, max([len(token) for token in uncond_tokens])) + + max_embeddings_multiples = min( + max_embeddings_multiples, + (max_length - 1) // (pipe.tokenizer.model_max_length - 2) + 1, + ) + max_embeddings_multiples = max(1, max_embeddings_multiples) + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + + # pad the length of tokens and weights + bos = pipe.tokenizer.bos_token_id + eos = pipe.tokenizer.eos_token_id + pad = pipe.tokenizer.pad_token_id + prompt_tokens, prompt_weights = pad_tokens_and_weights( + prompt_tokens, + prompt_weights, + max_length, + bos, + eos, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, + ) + prompt_tokens = torch.tensor(prompt_tokens, dtype=torch.long, device=pipe.device) + if uncond_prompt is not None: + uncond_tokens, uncond_weights = pad_tokens_and_weights( + uncond_tokens, + uncond_weights, + max_length, + bos, + eos, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, + ) + uncond_tokens = torch.tensor(uncond_tokens, dtype=torch.long, device=pipe.device) + + # get the embeddings + text_embeddings = get_unweighted_text_embeddings( + pipe, + prompt_tokens, + pipe.tokenizer.model_max_length, + clip_skip, + eos, + pad, + no_boseos_middle=no_boseos_middle, + ) + prompt_weights = torch.tensor(prompt_weights, dtype=text_embeddings.dtype, device=pipe.device) + if uncond_prompt is not None: + uncond_embeddings = get_unweighted_text_embeddings( + pipe, + uncond_tokens, + pipe.tokenizer.model_max_length, + clip_skip, + eos, + pad, + no_boseos_middle=no_boseos_middle, + ) + uncond_weights = torch.tensor(uncond_weights, dtype=uncond_embeddings.dtype, device=pipe.device) + + # assign weights to the prompts and normalize in the sense of mean + # TODO: should we normalize by chunk or in a whole (current implementation)? + if (not skip_parsing) and (not skip_weighting): + previous_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= prompt_weights.unsqueeze(-1) + current_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + if uncond_prompt is not None: + previous_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= uncond_weights.unsqueeze(-1) + current_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + + if uncond_prompt is not None: + return text_embeddings, uncond_embeddings + return text_embeddings, None + + +def preprocess_image(image): + w, h = image.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask, scale_factor=8): + mask = mask.convert("L") + w, h = mask.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"]) + mask = np.array(mask).astype(np.float32) / 255.0 + mask = np.tile(mask, (4, 1, 1)) + mask = mask[None].transpose(0, 1, 2, 3) # what does this step do? + mask = 1 - mask # repaint white, keep black + mask = torch.from_numpy(mask) + return mask + + +def prepare_controlnet_image( + image: PIL.Image.Image, + width: int, + height: int, + batch_size: int, + num_images_per_prompt: int, + device: torch.device, + dtype: torch.dtype, + do_classifier_free_guidance: bool = False, + guess_mode: bool = False, +): + if not isinstance(image, torch.Tensor): + if isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + images = [] + + for image_ in image: + image_ = image_.convert("RGB") + image_ = image_.resize((width, height), resample=PIL_INTERPOLATION["lanczos"]) + image_ = np.array(image_) + image_ = image_[None, :] + images.append(image_) + + image = images + + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + +class StableDiffusionLongPromptWeightingPipeline(StableDiffusionPipeline): + r""" + Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing + weighting in prompt. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + # if version.parse(version.parse(diffusers.__version__).base_version) >= version.parse("0.9.0"): + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: SchedulerMixin, + # clip_skip: int, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPFeatureExtractor, + requires_safety_checker: bool = True, + image_encoder: CLIPVisionModelWithProjection = None, + clip_skip: int = 1, + ): + super().__init__( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + requires_safety_checker=requires_safety_checker, + image_encoder=image_encoder, + ) + self.custom_clip_skip = clip_skip + self.__init__additional__() + + def __init__additional__(self): + if not hasattr(self, "vae_scale_factor"): + setattr(self, "vae_scale_factor", 2 ** (len(self.vae.config.block_out_channels) - 1)) + + @property + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"): + return self.device + for module in self.unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + max_embeddings_multiples, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + """ + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + if negative_prompt is None: + negative_prompt = [""] * batch_size + elif isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + if batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + text_embeddings, uncond_embeddings = get_weighted_text_embeddings( + pipe=self, + prompt=prompt, + uncond_prompt=negative_prompt if do_classifier_free_guidance else None, + max_embeddings_multiples=max_embeddings_multiples, + clip_skip=self.custom_clip_skip, + ) + bs_embed, seq_len, _ = text_embeddings.shape + text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) + text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + bs_embed, seq_len, _ = uncond_embeddings.shape + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + + return text_embeddings + + def check_inputs(self, prompt, height, width, strength, callback_steps): + if not isinstance(prompt, str) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 8 != 0 or width % 8 != 0: + logger.info(f'{height} {width}') + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" f" {type(callback_steps)}." + ) + + def get_timesteps(self, num_inference_steps, strength, device, is_text2img): + if is_text2img: + return self.scheduler.timesteps.to(device), num_inference_steps + else: + # get the original timestep using init_timestep + offset = self.scheduler.config.get("steps_offset", 0) + init_timestep = int(num_inference_steps * strength) + offset + init_timestep = min(init_timestep, num_inference_steps) + + t_start = max(num_inference_steps - init_timestep + offset, 0) + timesteps = self.scheduler.timesteps[t_start:].to(device) + return timesteps, num_inference_steps - t_start + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker(images=image, clip_input=safety_checker_input.pixel_values.to(dtype)) + else: + has_nsfw_concept = None + return image, has_nsfw_concept + + def decode_latents(self, latents): + latents = 1 / 0.18215 * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def prepare_latents(self, image, timestep, batch_size, height, width, dtype, device, generator, latents=None): + if image is None: + shape = ( + batch_size, + self.unet.in_channels, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents, None, None + else: + init_latent_dist = self.vae.encode(image).latent_dist + init_latents = init_latent_dist.sample(generator=generator) + init_latents = 0.18215 * init_latents + init_latents = torch.cat([init_latents] * batch_size, dim=0) + init_latents_orig = init_latents + shape = init_latents.shape + + # add noise to latents using the timesteps + if device.type == "mps": + noise = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + noise = torch.randn(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.add_noise(init_latents, noise, timestep) + return latents, init_latents_orig, noise + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + image: Union[torch.FloatTensor, PIL.Image.Image] = None, + mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + strength: float = 0.8, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + controlnet=None, + controlnet_image=None, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. + `image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + controlnet (`diffusers.ControlNetModel`, *optional*): + A controlnet model to be used for the inference. If not provided, controlnet will be disabled. + controlnet_image (`torch.FloatTensor` or `PIL.Image.Image`, *optional*): + `Image`, or tensor representing an image batch, to be used as the starting point for the controlnet + inference. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + `None` if cancelled by `is_cancelled_callback`, + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + if controlnet is not None and controlnet_image is None: + raise ValueError("controlnet_image must be provided if controlnet is not None.") + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, strength, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_embeddings = self._encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + max_embeddings_multiples, + ) + dtype = text_embeddings.dtype + + # 4. Preprocess image and mask + if isinstance(image, PIL.Image.Image): + image = preprocess_image(image) + if image is not None: + image = image.to(device=self.device, dtype=dtype) + if isinstance(mask_image, PIL.Image.Image): + mask_image = preprocess_mask(mask_image, self.vae_scale_factor) + if mask_image is not None: + mask = mask_image.to(device=self.device, dtype=dtype) + mask = torch.cat([mask] * batch_size * num_images_per_prompt) + else: + mask = None + + if controlnet_image is not None: + controlnet_image = prepare_controlnet_image( + controlnet_image, width, height, batch_size, 1, self.device, controlnet.dtype, do_classifier_free_guidance, False + ) + + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device, image is None) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents, init_latents_orig, noise = self.prepare_latents( + image, + latent_timestep, + batch_size * num_images_per_prompt, + height, + width, + dtype, + device, + generator, + latents, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + unet_additional_args = {} + if controlnet is not None: + down_block_res_samples, mid_block_res_sample = controlnet( + latent_model_input, + t, + encoder_hidden_states=text_embeddings, + controlnet_cond=controlnet_image, + conditioning_scale=1.0, + guess_mode=False, + return_dict=False, + ) + unet_additional_args["down_block_additional_residuals"] = down_block_res_samples + unet_additional_args["mid_block_additional_residual"] = mid_block_res_sample + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings, **unet_additional_args).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + if mask is not None: + # masking + init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t])) + latents = (init_latents_proper * mask) + (latents * (1 - mask)) + + # call the callback, if provided + if i % callback_steps == 0: + if callback is not None: + callback(i, t, latents) + if is_cancelled_callback is not None and is_cancelled_callback(): + return None + + return latents + + def latents_to_image(self, latents): + # 9. Post-processing + image = self.decode_latents(latents.to(self.vae.dtype)) + image = self.numpy_to_pil(image) + return image + + def text2img( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function for text-to-image generation. + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, + ) + + def img2img( + self, + image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[torch.Generator] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function for image-to-image generation. + Args: + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. + `image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + image=image, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + strength=strength, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, + ) + + def inpaint( + self, + image: Union[torch.FloatTensor, PIL.Image.Image], + mask_image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[torch.Generator] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function for inpaint. + Args: + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. This is the image whose masked region will be inpainted. + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength` + is 1, the denoising process will be run on the masked area for the full number of iterations specified + in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more + noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur. + num_inference_steps (`int`, *optional*, defaults to 50): + The reference number of denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. This parameter will be modulated by `strength`, as explained above. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + image=image, + mask_image=mask_image, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + strength=strength, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, + ) diff --git a/main.py b/main.py new file mode 100644 index 0000000000000000000000000000000000000000..380f85aecef2485bc566284da8f41370e25fc686 --- /dev/null +++ b/main.py @@ -0,0 +1,166 @@ +""" +extract factors the build is dependent on: +[X] compute capability + [ ] TODO: Q - What if we have multiple GPUs of different makes? +- CUDA version +- Software: + - CPU-only: only CPU quantization functions (no optimizer, no matrix multiple) + - CuBLAS-LT: full-build 8-bit optimizer + - no CuBLAS-LT: no 8-bit matrix multiplication (`nomatmul`) + +evaluation: + - if paths faulty, return meaningful error + - else: + - determine CUDA version + - determine capabilities + - based on that set the default path +""" + +import ctypes + +from .paths import determine_cuda_runtime_lib_path + + +def check_cuda_result(cuda, result_val): + # 3. Check for CUDA errors + if result_val != 0: + error_str = ctypes.c_char_p() + cuda.cuGetErrorString(result_val, ctypes.byref(error_str)) + print(f"CUDA exception! Error code: {error_str.value.decode()}") + +def get_cuda_version(cuda, cudart_path): + # https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART____VERSION.html#group__CUDART____VERSION + try: + cudart = ctypes.CDLL(cudart_path) + except OSError: + # TODO: shouldn't we error or at least warn here? + print(f'ERROR: libcudart.so could not be read from path: {cudart_path}!') + return None + + version = ctypes.c_int() + check_cuda_result(cuda, cudart.cudaRuntimeGetVersion(ctypes.byref(version))) + version = int(version.value) + major = version//1000 + minor = (version-(major*1000))//10 + + if major < 11: + print('CUDA SETUP: CUDA version lower than 11 are currently not supported for LLM.int8(). You will be only to use 8-bit optimizers and quantization routines!!') + + return f'{major}{minor}' + + +def get_cuda_lib_handle(): + # 1. find libcuda.so library (GPU driver) (/usr/lib) + try: + cuda = ctypes.CDLL("libcuda.so") + except OSError: + # TODO: shouldn't we error or at least warn here? + print('CUDA SETUP: WARNING! libcuda.so not found! Do you have a CUDA driver installed? If you are on a cluster, make sure you are on a CUDA machine!') + return None + check_cuda_result(cuda, cuda.cuInit(0)) + + return cuda + + +def get_compute_capabilities(cuda): + """ + 1. find libcuda.so library (GPU driver) (/usr/lib) + init_device -> init variables -> call function by reference + 2. call extern C function to determine CC + (https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__DEVICE__DEPRECATED.html) + 3. Check for CUDA errors + https://stackoverflow.com/questions/14038589/what-is-the-canonical-way-to-check-for-errors-using-the-cuda-runtime-api + # bits taken from https://gist.github.com/f0k/63a664160d016a491b2cbea15913d549 + """ + + + nGpus = ctypes.c_int() + cc_major = ctypes.c_int() + cc_minor = ctypes.c_int() + + device = ctypes.c_int() + + check_cuda_result(cuda, cuda.cuDeviceGetCount(ctypes.byref(nGpus))) + ccs = [] + for i in range(nGpus.value): + check_cuda_result(cuda, cuda.cuDeviceGet(ctypes.byref(device), i)) + ref_major = ctypes.byref(cc_major) + ref_minor = ctypes.byref(cc_minor) + # 2. call extern C function to determine CC + check_cuda_result( + cuda, cuda.cuDeviceComputeCapability(ref_major, ref_minor, device) + ) + ccs.append(f"{cc_major.value}.{cc_minor.value}") + + return ccs + + +# def get_compute_capability()-> Union[List[str, ...], None]: # FIXME: error +def get_compute_capability(cuda): + """ + Extracts the highest compute capbility from all available GPUs, as compute + capabilities are downwards compatible. If no GPUs are detected, it returns + None. + """ + ccs = get_compute_capabilities(cuda) + if ccs is not None: + # TODO: handle different compute capabilities; for now, take the max + return ccs[-1] + return None + + +def evaluate_cuda_setup(): + print('') + print('='*35 + 'BUG REPORT' + '='*35) + print('Welcome to bitsandbytes. For bug reports, please submit your error trace to: https://github.com/TimDettmers/bitsandbytes/issues') + print('For effortless bug reporting copy-paste your error into this form: https://docs.google.com/forms/d/e/1FAIpQLScPB8emS3Thkp66nvqwmjTEgxp8Y9ufuWTzFyr9kJ5AoI47dQ/viewform?usp=sf_link') + print('='*80) + return "libbitsandbytes_cuda116.dll" # $$$ + + binary_name = "libbitsandbytes_cpu.so" + #if not torch.cuda.is_available(): + #print('No GPU detected. Loading CPU library...') + #return binary_name + + cudart_path = determine_cuda_runtime_lib_path() + if cudart_path is None: + print( + "WARNING: No libcudart.so found! Install CUDA or the cudatoolkit package (anaconda)!" + ) + return binary_name + + print(f"CUDA SETUP: CUDA runtime path found: {cudart_path}") + cuda = get_cuda_lib_handle() + cc = get_compute_capability(cuda) + print(f"CUDA SETUP: Highest compute capability among GPUs detected: {cc}") + cuda_version_string = get_cuda_version(cuda, cudart_path) + + + if cc == '': + print( + "WARNING: No GPU detected! Check your CUDA paths. Processing to load CPU-only library..." + ) + return binary_name + + # 7.5 is the minimum CC vor cublaslt + has_cublaslt = cc in ["7.5", "8.0", "8.6"] + + # TODO: + # (1) CUDA missing cases (no CUDA installed by CUDA driver (nvidia-smi accessible) + # (2) Multiple CUDA versions installed + + # we use ls -l instead of nvcc to determine the cuda version + # since most installations will have the libcudart.so installed, but not the compiler + print(f'CUDA SETUP: Detected CUDA version {cuda_version_string}') + + def get_binary_name(): + "if not has_cublaslt (CC < 7.5), then we have to choose _nocublaslt.so" + bin_base_name = "libbitsandbytes_cuda" + if has_cublaslt: + return f"{bin_base_name}{cuda_version_string}.so" + else: + return f"{bin_base_name}{cuda_version_string}_nocublaslt.so" + + binary_name = get_binary_name() + + return binary_name diff --git a/make_captions.py b/make_captions.py new file mode 100644 index 0000000000000000000000000000000000000000..489bdbcce101ee079a399c886ba2f20ab224b504 --- /dev/null +++ b/make_captions.py @@ -0,0 +1,210 @@ +import argparse +import glob +import os +import json +import random +import sys + +from pathlib import Path +from PIL import Image +from tqdm import tqdm +import numpy as np + +import torch +from library.device_utils import init_ipex, get_preferred_device +init_ipex() + +from torchvision import transforms +from torchvision.transforms.functional import InterpolationMode +sys.path.append(os.path.dirname(__file__)) +from blip.blip import blip_decoder, is_url +import library.train_util as train_util +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +DEVICE = get_preferred_device() + + +IMAGE_SIZE = 384 + +# 正方形でいいのか？　という気がするがソースがそうなので +IMAGE_TRANSFORM = transforms.Compose( + [ + transforms.Resize((IMAGE_SIZE, IMAGE_SIZE), interpolation=InterpolationMode.BICUBIC), + transforms.ToTensor(), + transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)), + ] +) + + +# 共通化したいが微妙に処理が異なる…… +class ImageLoadingTransformDataset(torch.utils.data.Dataset): + def __init__(self, image_paths): + self.images = image_paths + + def __len__(self): + return len(self.images) + + def __getitem__(self, idx): + img_path = self.images[idx] + + try: + image = Image.open(img_path).convert("RGB") + # convert to tensor temporarily so dataloader will accept it + tensor = IMAGE_TRANSFORM(image) + except Exception as e: + logger.error(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}") + return None + + return (tensor, img_path) + + +def collate_fn_remove_corrupted(batch): + """Collate function that allows to remove corrupted examples in the + dataloader. It expects that the dataloader returns 'None' when that occurs. + The 'None's in the batch are removed. + """ + # Filter out all the Nones (corrupted examples) + batch = list(filter(lambda x: x is not None, batch)) + return batch + + +def main(args): + # fix the seed for reproducibility + seed = args.seed # + utils.get_rank() + torch.manual_seed(seed) + np.random.seed(seed) + random.seed(seed) + + if not os.path.exists("blip"): + args.train_data_dir = os.path.abspath(args.train_data_dir) # convert to absolute path + + cwd = os.getcwd() + logger.info(f"Current Working Directory is: {cwd}") + os.chdir("finetune") + if not is_url(args.caption_weights) and not os.path.isfile(args.caption_weights): + args.caption_weights = os.path.join("..", args.caption_weights) + + logger.info(f"load images from {args.train_data_dir}") + train_data_dir_path = Path(args.train_data_dir) + image_paths = train_util.glob_images_pathlib(train_data_dir_path, args.recursive) + logger.info(f"found {len(image_paths)} images.") + + logger.info(f"loading BLIP caption: {args.caption_weights}") + model = blip_decoder(pretrained=args.caption_weights, image_size=IMAGE_SIZE, vit="large", med_config="./blip/med_config.json") + model.eval() + model = model.to(DEVICE) + logger.info("BLIP loaded") + + # captioningする + def run_batch(path_imgs): + imgs = torch.stack([im for _, im in path_imgs]).to(DEVICE) + + with torch.no_grad(): + if args.beam_search: + captions = model.generate( + imgs, sample=False, num_beams=args.num_beams, max_length=args.max_length, min_length=args.min_length + ) + else: + captions = model.generate( + imgs, sample=True, top_p=args.top_p, max_length=args.max_length, min_length=args.min_length + ) + + for (image_path, _), caption in zip(path_imgs, captions): + with open(os.path.splitext(image_path)[0] + args.caption_extension, "wt", encoding="utf-8") as f: + f.write(caption + "\n") + if args.debug: + logger.info(f'{image_path} {caption}') + + # 読み込みの高速化のためにDataLoaderを使うオプション + if args.max_data_loader_n_workers is not None: + dataset = ImageLoadingTransformDataset(image_paths) + data = torch.utils.data.DataLoader( + dataset, + batch_size=args.batch_size, + shuffle=False, + num_workers=args.max_data_loader_n_workers, + collate_fn=collate_fn_remove_corrupted, + drop_last=False, + ) + else: + data = [[(None, ip)] for ip in image_paths] + + b_imgs = [] + for data_entry in tqdm(data, smoothing=0.0): + for data in data_entry: + if data is None: + continue + + img_tensor, image_path = data + if img_tensor is None: + try: + raw_image = Image.open(image_path) + if raw_image.mode != "RGB": + raw_image = raw_image.convert("RGB") + img_tensor = IMAGE_TRANSFORM(raw_image) + except Exception as e: + logger.error(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}") + continue + + b_imgs.append((image_path, img_tensor)) + if len(b_imgs) >= args.batch_size: + run_batch(b_imgs) + b_imgs.clear() + if len(b_imgs) > 0: + run_batch(b_imgs) + + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ") + parser.add_argument( + "--caption_weights", + type=str, + default="https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_large_caption.pth", + help="BLIP caption weights (model_large_caption.pth) / BLIP captionの重みファイル(model_large_caption.pth)", + ) + parser.add_argument( + "--caption_extention", + type=str, + default=None, + help="extension of caption file (for backward compatibility) / 出力されるキャプションファイルの拡張子（スペルミスしていたのを残してあります）", + ) + parser.add_argument("--caption_extension", type=str, default=".caption", help="extension of caption file / 出力されるキャプションファイルの拡張子") + parser.add_argument( + "--beam_search", + action="store_true", + help="use beam search (default Nucleus sampling) / beam searchを使う（このオプション未指定時はNucleus sampling）", + ) + parser.add_argument("--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ") + parser.add_argument( + "--max_data_loader_n_workers", + type=int, + default=None, + help="enable image reading by DataLoader with this number of workers (faster) / DataLoaderによる画像読み込みを有効にしてこのワーカー数を適用する（読み込みを高速化）", + ) + parser.add_argument("--num_beams", type=int, default=1, help="num of beams in beam search /beam search時のビーム数（多いと精度が上がるが時間がかかる）") + parser.add_argument("--top_p", type=float, default=0.9, help="top_p in Nucleus sampling / Nucleus sampling時のtop_p") + parser.add_argument("--max_length", type=int, default=75, help="max length of caption / captionの最大長") + parser.add_argument("--min_length", type=int, default=5, help="min length of caption / captionの最小長") + parser.add_argument("--seed", default=42, type=int, help="seed for reproducibility / 再現性を確保するための乱数seed") + parser.add_argument("--debug", action="store_true", help="debug mode") + parser.add_argument("--recursive", action="store_true", help="search for images in subfolders recursively / サブフォルダを再帰的に検索する") + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + + # スペルミスしていたオプションを復元する + if args.caption_extention is not None: + args.caption_extension = args.caption_extention + + main(args) diff --git a/make_captions_by_git.py b/make_captions_by_git.py new file mode 100644 index 0000000000000000000000000000000000000000..edeebadf3bbd86e8bc40d55bdf9f327e00bb7025 --- /dev/null +++ b/make_captions_by_git.py @@ -0,0 +1,183 @@ +import argparse +import os +import re + +from pathlib import Path +from PIL import Image +from tqdm import tqdm + +import torch +from library.device_utils import init_ipex, get_preferred_device +init_ipex() + +from transformers import AutoProcessor, AutoModelForCausalLM +from transformers.generation.utils import GenerationMixin + +import library.train_util as train_util +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") + +PATTERN_REPLACE = [ + re.compile(r'(has|with|and) the (words?|letters?|name) (" ?[^"]*"|\w+)( ?(is )?(on|in) (the |her |their |him )?\w+)?'), + re.compile(r'(with a sign )?that says ?(" ?[^"]*"|\w+)( ?on it)?'), + re.compile(r"(with a sign )?that says ?(' ?(i'm)?[^']*'|\w+)( ?on it)?"), + re.compile(r"with the number \d+ on (it|\w+ \w+)"), + re.compile(r'with the words "'), + re.compile(r"word \w+ on it"), + re.compile(r"that says the word \w+ on it"), + re.compile("that says'the word \"( on it)?"), +] + +# 誤検知しまくりの with the word xxxx を消す + + +def remove_words(captions, debug): + removed_caps = [] + for caption in captions: + cap = caption + for pat in PATTERN_REPLACE: + cap = pat.sub("", cap) + if debug and cap != caption: + logger.info(caption) + logger.info(cap) + removed_caps.append(cap) + return removed_caps + + +def collate_fn_remove_corrupted(batch): + """Collate function that allows to remove corrupted examples in the + dataloader. It expects that the dataloader returns 'None' when that occurs. + The 'None's in the batch are removed. + """ + # Filter out all the Nones (corrupted examples) + batch = list(filter(lambda x: x is not None, batch)) + return batch + + +def main(args): + r""" + transformers 4.30.2で、バッチサイズ>1でも動くようになったので、以下コメントアウト + + # GITにバッチサイズが1より大きくても動くようにパッチを当てる: transformers 4.26.0用 + org_prepare_input_ids_for_generation = GenerationMixin._prepare_input_ids_for_generation + curr_batch_size = [args.batch_size] # ループの最後で件数がbatch_size未満になるので入れ替えられるように + + # input_idsがバッチサイズと同じ件数である必要がある：バッチサイズはこの関数から参照できないので外から渡す + # ここより上で置き換えようとするとすごく大変 + def _prepare_input_ids_for_generation_patch(self, bos_token_id, encoder_outputs): + input_ids = org_prepare_input_ids_for_generation(self, bos_token_id, encoder_outputs) + if input_ids.size()[0] != curr_batch_size[0]: + input_ids = input_ids.repeat(curr_batch_size[0], 1) + return input_ids + + GenerationMixin._prepare_input_ids_for_generation = _prepare_input_ids_for_generation_patch + """ + + logger.info(f"load images from {args.train_data_dir}") + train_data_dir_path = Path(args.train_data_dir) + image_paths = train_util.glob_images_pathlib(train_data_dir_path, args.recursive) + logger.info(f"found {len(image_paths)} images.") + + # できればcacheに依存せず明示的にダウンロードしたい + logger.info(f"loading GIT: {args.model_id}") + git_processor = AutoProcessor.from_pretrained(args.model_id) + git_model = AutoModelForCausalLM.from_pretrained(args.model_id).to(DEVICE) + logger.info("GIT loaded") + + # captioningする + def run_batch(path_imgs): + imgs = [im for _, im in path_imgs] + + # curr_batch_size[0] = len(path_imgs) + inputs = git_processor(images=imgs, return_tensors="pt").to(DEVICE) # 画像はpil形式 + generated_ids = git_model.generate(pixel_values=inputs.pixel_values, max_length=args.max_length) + captions = git_processor.batch_decode(generated_ids, skip_special_tokens=True) + + if args.remove_words: + captions = remove_words(captions, args.debug) + + for (image_path, _), caption in zip(path_imgs, captions): + with open(os.path.splitext(image_path)[0] + args.caption_extension, "wt", encoding="utf-8") as f: + f.write(caption + "\n") + if args.debug: + logger.info(f"{image_path} {caption}") + + # 読み込みの高速化のためにDataLoaderを使うオプション + if args.max_data_loader_n_workers is not None: + dataset = train_util.ImageLoadingDataset(image_paths) + data = torch.utils.data.DataLoader( + dataset, + batch_size=args.batch_size, + shuffle=False, + num_workers=args.max_data_loader_n_workers, + collate_fn=collate_fn_remove_corrupted, + drop_last=False, + ) + else: + data = [[(None, ip)] for ip in image_paths] + + b_imgs = [] + for data_entry in tqdm(data, smoothing=0.0): + for data in data_entry: + if data is None: + continue + + image, image_path = data + if image is None: + try: + image = Image.open(image_path) + if image.mode != "RGB": + image = image.convert("RGB") + except Exception as e: + logger.error(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}") + continue + + b_imgs.append((image_path, image)) + if len(b_imgs) >= args.batch_size: + run_batch(b_imgs) + b_imgs.clear() + + if len(b_imgs) > 0: + run_batch(b_imgs) + + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ") + parser.add_argument("--caption_extension", type=str, default=".caption", help="extension of caption file / 出力されるキャプションファイルの拡張子") + parser.add_argument( + "--model_id", + type=str, + default="microsoft/git-large-textcaps", + help="model id for GIT in Hugging Face / 使用するGITのHugging FaceのモデルID", + ) + parser.add_argument("--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ") + parser.add_argument( + "--max_data_loader_n_workers", + type=int, + default=None, + help="enable image reading by DataLoader with this number of workers (faster) / DataLoaderによる画像読み込みを有効にしてこのワーカー数を適用する（読み込みを高速化）", + ) + parser.add_argument("--max_length", type=int, default=50, help="max length of caption / captionの最大長") + parser.add_argument( + "--remove_words", + action="store_true", + help="remove like `with the words xxx` from caption / `with the words xxx`のような部分をキャプションから削除する", + ) + parser.add_argument("--debug", action="store_true", help="debug mode") + parser.add_argument("--recursive", action="store_true", help="search for images in subfolders recursively / サブフォルダを再帰的に検索する") + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + main(args) diff --git a/masked_loss_README-ja.md b/masked_loss_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..58f042c3be4414c9c36fddce5fc1da0973d4fb11 --- /dev/null +++ b/masked_loss_README-ja.md @@ -0,0 +1,57 @@ +## マスクロスについて + +マスクロスは、入力画像のマスクで指定された部分だけ損失計算することで、画像の一部分だけを学習することができる機能です。 +たとえばキャラクタを学習したい場合、キャラクタ部分だけをマスクして学習することで、背景を無視して学習することができます。 + +マスクロスのマスクには、二種類の指定方法があります。 + +- マスク画像を用いる方法 +- 透明度（アルファチャネル）を使用する方法 + +なお、サンプルは [ずんずんPJイラスト/3Dデータ](https://zunko.jp/con_illust.html) の「AI画像モデル用学習データ」を使用しています。 + +### マスク画像を用いる方法 + +学習画像それぞれに対応するマスク画像を用意する方法です。学習画像と同じファイル名のマスク画像を用意し、それを学習画像と別のディレクトリに保存します。 + +- 学習画像 + ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/607c5116-5f62-47de-8b66-9c4a597f0441) +- マスク画像 + ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/53e9b0f8-a4bf-49ed-882d-4026f84e8450) + +```.toml +[[datasets.subsets]] +image_dir = "/path/to/a_zundamon" +caption_extension = ".txt" +conditioning_data_dir = "/path/to/a_zundamon_mask" +num_repeats = 8 +``` + +マスク画像は、学習画像と同じサイズで、学習する部分を白、無視する部分を黒で描画します。グレースケールにも対応しています（127 ならロス重みが 0.5 になります）。なお、正確にはマスク画像の R チャネルが用いられます。 + +DreamBooth 方式の dataset で、`conditioning_data_dir` で指定したディレクトリにマスク画像を保存してください。ControlNet のデータセットと同じですので、詳細は [ControlNet-LLLite](train_lllite_README-ja.md#データセットの準備) を参照してください。 + +### 透明度（アルファチャネル）を使用する方法 + +学習画像の透明度（アルファチャネル）がマスクとして使用されます。透明度が 0 の部分は無視され、255 の部分は学習されます。半透明の場合は、その透明度に応じてロス重みが変化します（127 ならおおむね 0.5）。 + +![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/0baa129b-446a-4aac-b98c-7208efb0e75e) + +※それぞれの画像は透過PNG + +学習時のスクリプトのオプションに `--alpha_mask` を指定するか、dataset の設定ファイルの subset で、`alpha_mask` を指定してください。たとえば、以下のようになります。 + +```toml +[[datasets.subsets]] +image_dir = "/path/to/image/dir" +caption_extension = ".txt" +num_repeats = 8 +alpha_mask = true +``` + +## 学習時の注意事項 + +- 現時点では DreamBooth 方式の dataset のみ対応しています。 +- マスクは latents のサイズ、つまり 1/8 に縮小されてから適用されます。そのため、細かい部分（たとえばアホ毛やイヤリングなど）はうまく学習できない可能性があります。マスクをわずかに拡張するなどの工夫が必要かもしれません。 +- マスクロスを用いる場合、学習対象外の部分をキャプションに含める必要はないかもしれません。（要検証） +- `alpha_mask` の場合、マスクの有無を切り替えると latents キャッシュが自動的に再生成されます。 diff --git a/masked_loss_README.md b/masked_loss_README.md new file mode 100644 index 0000000000000000000000000000000000000000..3ac5ad211d5496e1e1ba1b17e18d5b0c36251edd --- /dev/null +++ b/masked_loss_README.md @@ -0,0 +1,56 @@ +## Masked Loss + +Masked loss is a feature that allows you to train only part of an image by calculating the loss only for the part specified by the mask of the input image. For example, if you want to train a character, you can train only the character part by masking it, ignoring the background. + +There are two ways to specify the mask for masked loss. + +- Using a mask image +- Using transparency (alpha channel) of the image + +The sample uses the "AI image model training data" from [ZunZunPJ Illustration/3D Data](https://zunko.jp/con_illust.html). + +### Using a mask image + +This is a method of preparing a mask image corresponding to each training image. Prepare a mask image with the same file name as the training image and save it in a different directory from the training image. + +- Training image + ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/607c5116-5f62-47de-8b66-9c4a597f0441) +- Mask image + ![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/53e9b0f8-a4bf-49ed-882d-4026f84e8450) + +```.toml +[[datasets.subsets]] +image_dir = "/path/to/a_zundamon" +caption_extension = ".txt" +conditioning_data_dir = "/path/to/a_zundamon_mask" +num_repeats = 8 +``` + +The mask image is the same size as the training image, with the part to be trained drawn in white and the part to be ignored in black. It also supports grayscale (127 gives a loss weight of 0.5). The R channel of the mask image is used currently. + +Use the dataset in the DreamBooth method, and save the mask image in the directory specified by `conditioning_data_dir`. It is the same as the ControlNet dataset, so please refer to [ControlNet-LLLite](train_lllite_README.md#Preparing-the-dataset) for details. + +### Using transparency (alpha channel) of the image + +The transparency (alpha channel) of the training image is used as a mask. The part with transparency 0 is ignored, the part with transparency 255 is trained. For semi-transparent parts, the loss weight changes according to the transparency (127 gives a weight of about 0.5). + +![image](https://github.com/kohya-ss/sd-scripts/assets/52813779/0baa129b-446a-4aac-b98c-7208efb0e75e) + +※Each image is a transparent PNG + +Specify `--alpha_mask` in the training script options or specify `alpha_mask` in the subset of the dataset configuration file. For example, it will look like this. + +```toml +[[datasets.subsets]] +image_dir = "/path/to/image/dir" +caption_extension = ".txt" +num_repeats = 8 +alpha_mask = true +``` + +## Notes on training + +- At the moment, only the dataset in the DreamBooth method is supported. +- The mask is applied after the size is reduced to 1/8, which is the size of the latents. Therefore, fine details (such as ahoge or earrings) may not be learned well. Some dilations of the mask may be necessary. +- If using masked loss, it may not be necessary to include parts that are not to be trained in the caption. (To be verified) +- In the case of `alpha_mask`, the latents cache is automatically regenerated when the enable/disable state of the mask is switched. diff --git a/med.py b/med.py new file mode 100644 index 0000000000000000000000000000000000000000..7b00a35450b736180a805d4f4664b4fb95aeba01 --- /dev/null +++ b/med.py @@ -0,0 +1,955 @@ +''' + * Copyright (c) 2022, salesforce.com, inc. + * All rights reserved. + * SPDX-License-Identifier: BSD-3-Clause + * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause + * By Junnan Li + * Based on huggingface code base + * https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert +''' + +import math +import os +import warnings +from dataclasses import dataclass +from typing import Optional, Tuple + +import torch +from torch import Tensor, device, dtype, nn +import torch.utils.checkpoint +from torch import nn +from torch.nn import CrossEntropyLoss +import torch.nn.functional as F + +from transformers.activations import ACT2FN +from transformers.file_utils import ( + ModelOutput, +) +from transformers.modeling_outputs import ( + BaseModelOutputWithPastAndCrossAttentions, + BaseModelOutputWithPoolingAndCrossAttentions, + CausalLMOutputWithCrossAttentions, + MaskedLMOutput, + MultipleChoiceModelOutput, + NextSentencePredictorOutput, + QuestionAnsweringModelOutput, + SequenceClassifierOutput, + TokenClassifierOutput, +) +from transformers.modeling_utils import ( + PreTrainedModel, + apply_chunking_to_forward, + find_pruneable_heads_and_indices, + prune_linear_layer, +) +from transformers.utils import logging +from transformers.models.bert.configuration_bert import BertConfig + + +logger = logging.get_logger(__name__) + + +class BertEmbeddings(nn.Module): + """Construct the embeddings from word and position embeddings.""" + + def __init__(self, config): + super().__init__() + self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id) + self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size) + + # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load + # any TensorFlow checkpoint file + self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.dropout = nn.Dropout(config.hidden_dropout_prob) + + # position_ids (1, len position emb) is contiguous in memory and exported when serialized + self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))) + self.position_embedding_type = getattr(config, "position_embedding_type", "absolute") + + self.config = config + + def forward( + self, input_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0 + ): + if input_ids is not None: + input_shape = input_ids.size() + else: + input_shape = inputs_embeds.size()[:-1] + + seq_length = input_shape[1] + + if position_ids is None: + position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length] + + if inputs_embeds is None: + inputs_embeds = self.word_embeddings(input_ids) + + embeddings = inputs_embeds + + if self.position_embedding_type == "absolute": + position_embeddings = self.position_embeddings(position_ids) + embeddings += position_embeddings + embeddings = self.LayerNorm(embeddings) + embeddings = self.dropout(embeddings) + return embeddings + + +class BertSelfAttention(nn.Module): + def __init__(self, config, is_cross_attention): + super().__init__() + self.config = config + if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"): + raise ValueError( + "The hidden size (%d) is not a multiple of the number of attention " + "heads (%d)" % (config.hidden_size, config.num_attention_heads) + ) + + self.num_attention_heads = config.num_attention_heads + self.attention_head_size = int(config.hidden_size / config.num_attention_heads) + self.all_head_size = self.num_attention_heads * self.attention_head_size + + self.query = nn.Linear(config.hidden_size, self.all_head_size) + if is_cross_attention: + self.key = nn.Linear(config.encoder_width, self.all_head_size) + self.value = nn.Linear(config.encoder_width, self.all_head_size) + else: + self.key = nn.Linear(config.hidden_size, self.all_head_size) + self.value = nn.Linear(config.hidden_size, self.all_head_size) + + self.dropout = nn.Dropout(config.attention_probs_dropout_prob) + self.position_embedding_type = getattr(config, "position_embedding_type", "absolute") + if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query": + self.max_position_embeddings = config.max_position_embeddings + self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size) + self.save_attention = False + + def save_attn_gradients(self, attn_gradients): + self.attn_gradients = attn_gradients + + def get_attn_gradients(self): + return self.attn_gradients + + def save_attention_map(self, attention_map): + self.attention_map = attention_map + + def get_attention_map(self): + return self.attention_map + + def transpose_for_scores(self, x): + new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size) + x = x.view(*new_x_shape) + return x.permute(0, 2, 1, 3) + + def forward( + self, + hidden_states, + attention_mask=None, + head_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + past_key_value=None, + output_attentions=False, + ): + mixed_query_layer = self.query(hidden_states) + + # If this is instantiated as a cross-attention module, the keys + # and values come from an encoder; the attention mask needs to be + # such that the encoder's padding tokens are not attended to. + is_cross_attention = encoder_hidden_states is not None + + if is_cross_attention: + key_layer = self.transpose_for_scores(self.key(encoder_hidden_states)) + value_layer = self.transpose_for_scores(self.value(encoder_hidden_states)) + attention_mask = encoder_attention_mask + elif past_key_value is not None: + key_layer = self.transpose_for_scores(self.key(hidden_states)) + value_layer = self.transpose_for_scores(self.value(hidden_states)) + key_layer = torch.cat([past_key_value[0], key_layer], dim=2) + value_layer = torch.cat([past_key_value[1], value_layer], dim=2) + else: + key_layer = self.transpose_for_scores(self.key(hidden_states)) + value_layer = self.transpose_for_scores(self.value(hidden_states)) + + query_layer = self.transpose_for_scores(mixed_query_layer) + + past_key_value = (key_layer, value_layer) + + # Take the dot product between "query" and "key" to get the raw attention scores. + attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2)) + + if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query": + seq_length = hidden_states.size()[1] + position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1) + position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1) + distance = position_ids_l - position_ids_r + positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1) + positional_embedding = positional_embedding.to(dtype=query_layer.dtype) # fp16 compatibility + + if self.position_embedding_type == "relative_key": + relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding) + attention_scores = attention_scores + relative_position_scores + elif self.position_embedding_type == "relative_key_query": + relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding) + relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding) + attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key + + attention_scores = attention_scores / math.sqrt(self.attention_head_size) + if attention_mask is not None: + # Apply the attention mask is (precomputed for all layers in BertModel forward() function) + attention_scores = attention_scores + attention_mask + + # Normalize the attention scores to probabilities. + attention_probs = nn.Softmax(dim=-1)(attention_scores) + + if is_cross_attention and self.save_attention: + self.save_attention_map(attention_probs) + attention_probs.register_hook(self.save_attn_gradients) + + # This is actually dropping out entire tokens to attend to, which might + # seem a bit unusual, but is taken from the original Transformer paper. + attention_probs_dropped = self.dropout(attention_probs) + + # Mask heads if we want to + if head_mask is not None: + attention_probs_dropped = attention_probs_dropped * head_mask + + context_layer = torch.matmul(attention_probs_dropped, value_layer) + + context_layer = context_layer.permute(0, 2, 1, 3).contiguous() + new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,) + context_layer = context_layer.view(*new_context_layer_shape) + + outputs = (context_layer, attention_probs) if output_attentions else (context_layer,) + + outputs = outputs + (past_key_value,) + return outputs + + +class BertSelfOutput(nn.Module): + def __init__(self, config): + super().__init__() + self.dense = nn.Linear(config.hidden_size, config.hidden_size) + self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.dropout = nn.Dropout(config.hidden_dropout_prob) + + def forward(self, hidden_states, input_tensor): + hidden_states = self.dense(hidden_states) + hidden_states = self.dropout(hidden_states) + hidden_states = self.LayerNorm(hidden_states + input_tensor) + return hidden_states + + +class BertAttention(nn.Module): + def __init__(self, config, is_cross_attention=False): + super().__init__() + self.self = BertSelfAttention(config, is_cross_attention) + self.output = BertSelfOutput(config) + self.pruned_heads = set() + + def prune_heads(self, heads): + if len(heads) == 0: + return + heads, index = find_pruneable_heads_and_indices( + heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads + ) + + # Prune linear layers + self.self.query = prune_linear_layer(self.self.query, index) + self.self.key = prune_linear_layer(self.self.key, index) + self.self.value = prune_linear_layer(self.self.value, index) + self.output.dense = prune_linear_layer(self.output.dense, index, dim=1) + + # Update hyper params and store pruned heads + self.self.num_attention_heads = self.self.num_attention_heads - len(heads) + self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads + self.pruned_heads = self.pruned_heads.union(heads) + + def forward( + self, + hidden_states, + attention_mask=None, + head_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + past_key_value=None, + output_attentions=False, + ): + self_outputs = self.self( + hidden_states, + attention_mask, + head_mask, + encoder_hidden_states, + encoder_attention_mask, + past_key_value, + output_attentions, + ) + attention_output = self.output(self_outputs[0], hidden_states) + outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them + return outputs + + +class BertIntermediate(nn.Module): + def __init__(self, config): + super().__init__() + self.dense = nn.Linear(config.hidden_size, config.intermediate_size) + if isinstance(config.hidden_act, str): + self.intermediate_act_fn = ACT2FN[config.hidden_act] + else: + self.intermediate_act_fn = config.hidden_act + + def forward(self, hidden_states): + hidden_states = self.dense(hidden_states) + hidden_states = self.intermediate_act_fn(hidden_states) + return hidden_states + + +class BertOutput(nn.Module): + def __init__(self, config): + super().__init__() + self.dense = nn.Linear(config.intermediate_size, config.hidden_size) + self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.dropout = nn.Dropout(config.hidden_dropout_prob) + + def forward(self, hidden_states, input_tensor): + hidden_states = self.dense(hidden_states) + hidden_states = self.dropout(hidden_states) + hidden_states = self.LayerNorm(hidden_states + input_tensor) + return hidden_states + + +class BertLayer(nn.Module): + def __init__(self, config, layer_num): + super().__init__() + self.config = config + self.chunk_size_feed_forward = config.chunk_size_feed_forward + self.seq_len_dim = 1 + self.attention = BertAttention(config) + self.layer_num = layer_num + if self.config.add_cross_attention: + self.crossattention = BertAttention(config, is_cross_attention=self.config.add_cross_attention) + self.intermediate = BertIntermediate(config) + self.output = BertOutput(config) + + def forward( + self, + hidden_states, + attention_mask=None, + head_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + past_key_value=None, + output_attentions=False, + mode=None, + ): + # decoder uni-directional self-attention cached key/values tuple is at positions 1,2 + self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None + self_attention_outputs = self.attention( + hidden_states, + attention_mask, + head_mask, + output_attentions=output_attentions, + past_key_value=self_attn_past_key_value, + ) + attention_output = self_attention_outputs[0] + + outputs = self_attention_outputs[1:-1] + present_key_value = self_attention_outputs[-1] + + if mode=='multimodal': + assert encoder_hidden_states is not None, "encoder_hidden_states must be given for cross-attention layers" + + cross_attention_outputs = self.crossattention( + attention_output, + attention_mask, + head_mask, + encoder_hidden_states, + encoder_attention_mask, + output_attentions=output_attentions, + ) + attention_output = cross_attention_outputs[0] + outputs = outputs + cross_attention_outputs[1:-1] # add cross attentions if we output attention weights + layer_output = apply_chunking_to_forward( + self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output + ) + outputs = (layer_output,) + outputs + + outputs = outputs + (present_key_value,) + + return outputs + + def feed_forward_chunk(self, attention_output): + intermediate_output = self.intermediate(attention_output) + layer_output = self.output(intermediate_output, attention_output) + return layer_output + + +class BertEncoder(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.layer = nn.ModuleList([BertLayer(config,i) for i in range(config.num_hidden_layers)]) + self.gradient_checkpointing = False + + def forward( + self, + hidden_states, + attention_mask=None, + head_mask=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + past_key_values=None, + use_cache=None, + output_attentions=False, + output_hidden_states=False, + return_dict=True, + mode='multimodal', + ): + all_hidden_states = () if output_hidden_states else None + all_self_attentions = () if output_attentions else None + all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None + + next_decoder_cache = () if use_cache else None + + for i in range(self.config.num_hidden_layers): + layer_module = self.layer[i] + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + layer_head_mask = head_mask[i] if head_mask is not None else None + past_key_value = past_key_values[i] if past_key_values is not None else None + + if self.gradient_checkpointing and self.training: + + if use_cache: + logger.warn( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs, past_key_value, output_attentions) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(layer_module), + hidden_states, + attention_mask, + layer_head_mask, + encoder_hidden_states, + encoder_attention_mask, + mode=mode, + ) + else: + layer_outputs = layer_module( + hidden_states, + attention_mask, + layer_head_mask, + encoder_hidden_states, + encoder_attention_mask, + past_key_value, + output_attentions, + mode=mode, + ) + + hidden_states = layer_outputs[0] + if use_cache: + next_decoder_cache += (layer_outputs[-1],) + if output_attentions: + all_self_attentions = all_self_attentions + (layer_outputs[1],) + + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if not return_dict: + return tuple( + v + for v in [ + hidden_states, + next_decoder_cache, + all_hidden_states, + all_self_attentions, + all_cross_attentions, + ] + if v is not None + ) + return BaseModelOutputWithPastAndCrossAttentions( + last_hidden_state=hidden_states, + past_key_values=next_decoder_cache, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + cross_attentions=all_cross_attentions, + ) + + +class BertPooler(nn.Module): + def __init__(self, config): + super().__init__() + self.dense = nn.Linear(config.hidden_size, config.hidden_size) + self.activation = nn.Tanh() + + def forward(self, hidden_states): + # We "pool" the model by simply taking the hidden state corresponding + # to the first token. + first_token_tensor = hidden_states[:, 0] + pooled_output = self.dense(first_token_tensor) + pooled_output = self.activation(pooled_output) + return pooled_output + + +class BertPredictionHeadTransform(nn.Module): + def __init__(self, config): + super().__init__() + self.dense = nn.Linear(config.hidden_size, config.hidden_size) + if isinstance(config.hidden_act, str): + self.transform_act_fn = ACT2FN[config.hidden_act] + else: + self.transform_act_fn = config.hidden_act + self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + def forward(self, hidden_states): + hidden_states = self.dense(hidden_states) + hidden_states = self.transform_act_fn(hidden_states) + hidden_states = self.LayerNorm(hidden_states) + return hidden_states + + +class BertLMPredictionHead(nn.Module): + def __init__(self, config): + super().__init__() + self.transform = BertPredictionHeadTransform(config) + + # The output weights are the same as the input embeddings, but there is + # an output-only bias for each token. + self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + self.bias = nn.Parameter(torch.zeros(config.vocab_size)) + + # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings` + self.decoder.bias = self.bias + + def forward(self, hidden_states): + hidden_states = self.transform(hidden_states) + hidden_states = self.decoder(hidden_states) + return hidden_states + + +class BertOnlyMLMHead(nn.Module): + def __init__(self, config): + super().__init__() + self.predictions = BertLMPredictionHead(config) + + def forward(self, sequence_output): + prediction_scores = self.predictions(sequence_output) + return prediction_scores + + +class BertPreTrainedModel(PreTrainedModel): + """ + An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained + models. + """ + + config_class = BertConfig + base_model_prefix = "bert" + _keys_to_ignore_on_load_missing = [r"position_ids"] + + def _init_weights(self, module): + """ Initialize the weights """ + if isinstance(module, (nn.Linear, nn.Embedding)): + # Slightly different from the TF version which uses truncated_normal for initialization + # cf https://github.com/pytorch/pytorch/pull/5617 + module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) + elif isinstance(module, nn.LayerNorm): + module.bias.data.zero_() + module.weight.data.fill_(1.0) + if isinstance(module, nn.Linear) and module.bias is not None: + module.bias.data.zero_() + + +class BertModel(BertPreTrainedModel): + """ + The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of + cross-attention is added between the self-attention layers, following the architecture described in `Attention is + all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, + Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. + argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an + input to the forward pass. + """ + + def __init__(self, config, add_pooling_layer=True): + super().__init__(config) + self.config = config + + self.embeddings = BertEmbeddings(config) + + self.encoder = BertEncoder(config) + + self.pooler = BertPooler(config) if add_pooling_layer else None + + self.init_weights() + + + def get_input_embeddings(self): + return self.embeddings.word_embeddings + + def set_input_embeddings(self, value): + self.embeddings.word_embeddings = value + + def _prune_heads(self, heads_to_prune): + """ + Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base + class PreTrainedModel + """ + for layer, heads in heads_to_prune.items(): + self.encoder.layer[layer].attention.prune_heads(heads) + + + def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple[int], device: device, is_decoder: bool) -> Tensor: + """ + Makes broadcastable attention and causal masks so that future and masked tokens are ignored. + + Arguments: + attention_mask (:obj:`torch.Tensor`): + Mask with ones indicating tokens to attend to, zeros for tokens to ignore. + input_shape (:obj:`Tuple[int]`): + The shape of the input to the model. + device: (:obj:`torch.device`): + The device of the input to the model. + + Returns: + :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`. + """ + # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length] + # ourselves in which case we just need to make it broadcastable to all heads. + if attention_mask.dim() == 3: + extended_attention_mask = attention_mask[:, None, :, :] + elif attention_mask.dim() == 2: + # Provided a padding mask of dimensions [batch_size, seq_length] + # - if the model is a decoder, apply a causal mask in addition to the padding mask + # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length] + if is_decoder: + batch_size, seq_length = input_shape + + seq_ids = torch.arange(seq_length, device=device) + causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None] + # in case past_key_values are used we need to add a prefix ones mask to the causal mask + # causal and attention masks must have same type with pytorch version < 1.3 + causal_mask = causal_mask.to(attention_mask.dtype) + + if causal_mask.shape[1] < attention_mask.shape[1]: + prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1] + causal_mask = torch.cat( + [ + torch.ones((batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype), + causal_mask, + ], + axis=-1, + ) + + extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :] + else: + extended_attention_mask = attention_mask[:, None, None, :] + else: + raise ValueError( + "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format( + input_shape, attention_mask.shape + ) + ) + + # Since attention_mask is 1.0 for positions we want to attend and 0.0 for + # masked positions, this operation will create a tensor which is 0.0 for + # positions we want to attend and -10000.0 for masked positions. + # Since we are adding it to the raw scores before the softmax, this is + # effectively the same as removing these entirely. + extended_attention_mask = extended_attention_mask.to(dtype=self.dtype) # fp16 compatibility + extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0 + return extended_attention_mask + + def forward( + self, + input_ids=None, + attention_mask=None, + position_ids=None, + head_mask=None, + inputs_embeds=None, + encoder_embeds=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + past_key_values=None, + use_cache=None, + output_attentions=None, + output_hidden_states=None, + return_dict=None, + is_decoder=False, + mode='multimodal', + ): + r""" + encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`): + Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if + the model is configured as a decoder. + encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`): + Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in + the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``: + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): + Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. + If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids` + (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)` + instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`. + use_cache (:obj:`bool`, `optional`): + If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up + decoding (see :obj:`past_key_values`). + """ + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if is_decoder: + use_cache = use_cache if use_cache is not None else self.config.use_cache + else: + use_cache = False + + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + elif input_ids is not None: + input_shape = input_ids.size() + batch_size, seq_length = input_shape + device = input_ids.device + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + batch_size, seq_length = input_shape + device = inputs_embeds.device + elif encoder_embeds is not None: + input_shape = encoder_embeds.size()[:-1] + batch_size, seq_length = input_shape + device = encoder_embeds.device + else: + raise ValueError("You have to specify either input_ids or inputs_embeds or encoder_embeds") + + # past_key_values_length + past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 + + if attention_mask is None: + attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device) + + # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length] + # ourselves in which case we just need to make it broadcastable to all heads. + extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, + device, is_decoder) + + # If a 2D or 3D attention mask is provided for the cross-attention + # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length] + if encoder_hidden_states is not None: + if type(encoder_hidden_states) == list: + encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size() + else: + encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size() + encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length) + + if type(encoder_attention_mask) == list: + encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask] + elif encoder_attention_mask is None: + encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device) + encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask) + else: + encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask) + else: + encoder_extended_attention_mask = None + + # Prepare head mask if needed + # 1.0 in head_mask indicate we keep the head + # attention_probs has shape bsz x n_heads x N x N + # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] + # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length] + head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers) + + if encoder_embeds is None: + embedding_output = self.embeddings( + input_ids=input_ids, + position_ids=position_ids, + inputs_embeds=inputs_embeds, + past_key_values_length=past_key_values_length, + ) + else: + embedding_output = encoder_embeds + + encoder_outputs = self.encoder( + embedding_output, + attention_mask=extended_attention_mask, + head_mask=head_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_extended_attention_mask, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + mode=mode, + ) + sequence_output = encoder_outputs[0] + pooled_output = self.pooler(sequence_output) if self.pooler is not None else None + + if not return_dict: + return (sequence_output, pooled_output) + encoder_outputs[1:] + + return BaseModelOutputWithPoolingAndCrossAttentions( + last_hidden_state=sequence_output, + pooler_output=pooled_output, + past_key_values=encoder_outputs.past_key_values, + hidden_states=encoder_outputs.hidden_states, + attentions=encoder_outputs.attentions, + cross_attentions=encoder_outputs.cross_attentions, + ) + + + +class BertLMHeadModel(BertPreTrainedModel): + + _keys_to_ignore_on_load_unexpected = [r"pooler"] + _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"] + + def __init__(self, config): + super().__init__(config) + + self.bert = BertModel(config, add_pooling_layer=False) + self.cls = BertOnlyMLMHead(config) + + self.init_weights() + + def get_output_embeddings(self): + return self.cls.predictions.decoder + + def set_output_embeddings(self, new_embeddings): + self.cls.predictions.decoder = new_embeddings + + def forward( + self, + input_ids=None, + attention_mask=None, + position_ids=None, + head_mask=None, + inputs_embeds=None, + encoder_hidden_states=None, + encoder_attention_mask=None, + labels=None, + past_key_values=None, + use_cache=None, + output_attentions=None, + output_hidden_states=None, + return_dict=None, + return_logits=False, + is_decoder=True, + reduction='mean', + mode='multimodal', + ): + r""" + encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`): + Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if + the model is configured as a decoder. + encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`): + Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in + the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``: + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`): + Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in + ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are + ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]`` + past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): + Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. + If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids` + (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)` + instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`. + use_cache (:obj:`bool`, `optional`): + If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up + decoding (see :obj:`past_key_values`). + Returns: + Example:: + >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig + >>> import torch + >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased') + >>> config = BertConfig.from_pretrained("bert-base-cased") + >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config) + >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt") + >>> outputs = model(**inputs) + >>> prediction_logits = outputs.logits + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + if labels is not None: + use_cache = False + + outputs = self.bert( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + is_decoder=is_decoder, + mode=mode, + ) + + sequence_output = outputs[0] + prediction_scores = self.cls(sequence_output) + + if return_logits: + return prediction_scores[:, :-1, :].contiguous() + + lm_loss = None + if labels is not None: + # we are doing next-token prediction; shift prediction scores and input ids by one + shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous() + labels = labels[:, 1:].contiguous() + loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1) + lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)) + if reduction=='none': + lm_loss = lm_loss.view(prediction_scores.size(0),-1).sum(1) + + if not return_dict: + output = (prediction_scores,) + outputs[2:] + return ((lm_loss,) + output) if lm_loss is not None else output + + return CausalLMOutputWithCrossAttentions( + loss=lm_loss, + logits=prediction_scores, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + cross_attentions=outputs.cross_attentions, + ) + + def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs): + input_shape = input_ids.shape + # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly + if attention_mask is None: + attention_mask = input_ids.new_ones(input_shape) + + # cut decoder_input_ids if past is used + if past is not None: + input_ids = input_ids[:, -1:] + + return { + "input_ids": input_ids, + "attention_mask": attention_mask, + "past_key_values": past, + "encoder_hidden_states": model_kwargs.get("encoder_hidden_states", None), + "encoder_attention_mask": model_kwargs.get("encoder_attention_mask", None), + "is_decoder": True, + } + + def _reorder_cache(self, past, beam_idx): + reordered_past = () + for layer_past in past: + reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),) + return reordered_past diff --git a/med_config.json b/med_config.json new file mode 100644 index 0000000000000000000000000000000000000000..dc12b99cf539b751d442b4ca7785c9f6a4f8306e --- /dev/null +++ b/med_config.json @@ -0,0 +1,22 @@ +{ + "architectures": [ + "BertModel" + ], + "attention_probs_dropout_prob": 0.1, + "hidden_act": "gelu", + "hidden_dropout_prob": 0.1, + "hidden_size": 768, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_norm_eps": 1e-12, + "max_position_embeddings": 512, + "model_type": "bert", + "num_attention_heads": 12, + "num_hidden_layers": 12, + "pad_token_id": 0, + "type_vocab_size": 2, + "vocab_size": 30524, + "encoder_width": 768, + "add_cross_attention": true + } + \ No newline at end of file diff --git a/merge_captions_to_metadata.py b/merge_captions_to_metadata.py new file mode 100644 index 0000000000000000000000000000000000000000..89f717473af765c0f2995d0b736050995006a9fe --- /dev/null +++ b/merge_captions_to_metadata.py @@ -0,0 +1,100 @@ +import argparse +import json +from pathlib import Path +from typing import List +from tqdm import tqdm +import library.train_util as train_util +import os +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +def main(args): + assert not args.recursive or ( + args.recursive and args.full_path + ), "recursive requires full_path / recursiveはfull_pathと同時に指定してください" + + train_data_dir_path = Path(args.train_data_dir) + image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive) + logger.info(f"found {len(image_paths)} images.") + + if args.in_json is None and Path(args.out_json).is_file(): + args.in_json = args.out_json + + if args.in_json is not None: + logger.info(f"loading existing metadata: {args.in_json}") + metadata = json.loads(Path(args.in_json).read_text(encoding="utf-8")) + logger.warning("captions for existing images will be overwritten / 既存の画像のキャプションは上書きされます") + else: + logger.info("new metadata will be created / 新しいメタデータファイルが作成されます") + metadata = {} + + logger.info("merge caption texts to metadata json.") + for image_path in tqdm(image_paths): + caption_path = image_path.with_suffix(args.caption_extension) + caption = caption_path.read_text(encoding="utf-8").strip() + + if not os.path.exists(caption_path): + caption_path = os.path.join(image_path, args.caption_extension) + + image_key = str(image_path) if args.full_path else image_path.stem + if image_key not in metadata: + metadata[image_key] = {} + + metadata[image_key]["caption"] = caption + if args.debug: + logger.info(f"{image_key} {caption}") + + # metadataを書き出して終わり + logger.info(f"writing metadata: {args.out_json}") + Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding="utf-8") + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ") + parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先") + parser.add_argument( + "--in_json", + type=str, + help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）", + ) + parser.add_argument( + "--caption_extention", + type=str, + default=None, + help="extension of caption file (for backward compatibility) / 読み込むキャプションファイルの拡張子（スペルミスしていたのを残してあります）", + ) + parser.add_argument( + "--caption_extension", type=str, default=".caption", help="extension of caption file / 読み込むキャプションファイルの拡張子" + ) + parser.add_argument( + "--full_path", + action="store_true", + help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）", + ) + parser.add_argument( + "--recursive", + action="store_true", + help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す", + ) + parser.add_argument("--debug", action="store_true", help="debug mode") + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + + # スペルミスしていたオプションを復元する + if args.caption_extention is not None: + args.caption_extension = args.caption_extention + + main(args) diff --git a/merge_dd_tags_to_metadata.py b/merge_dd_tags_to_metadata.py new file mode 100644 index 0000000000000000000000000000000000000000..ce22d990e48b3d34905be7b3e168672f395dc3f8 --- /dev/null +++ b/merge_dd_tags_to_metadata.py @@ -0,0 +1,93 @@ +import argparse +import json +from pathlib import Path +from typing import List +from tqdm import tqdm +import library.train_util as train_util +import os +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +def main(args): + assert not args.recursive or ( + args.recursive and args.full_path + ), "recursive requires full_path / recursiveはfull_pathと同時に指定してください" + + train_data_dir_path = Path(args.train_data_dir) + image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive) + logger.info(f"found {len(image_paths)} images.") + + if args.in_json is None and Path(args.out_json).is_file(): + args.in_json = args.out_json + + if args.in_json is not None: + logger.info(f"loading existing metadata: {args.in_json}") + metadata = json.loads(Path(args.in_json).read_text(encoding="utf-8")) + logger.warning("tags data for existing images will be overwritten / 既存の画像のタグは上書きされます") + else: + logger.info("new metadata will be created / 新しいメタデータファイルが作成されます") + metadata = {} + + logger.info("merge tags to metadata json.") + for image_path in tqdm(image_paths): + tags_path = image_path.with_suffix(args.caption_extension) + tags = tags_path.read_text(encoding="utf-8").strip() + + if not os.path.exists(tags_path): + tags_path = os.path.join(image_path, args.caption_extension) + + image_key = str(image_path) if args.full_path else image_path.stem + if image_key not in metadata: + metadata[image_key] = {} + + metadata[image_key]["tags"] = tags + if args.debug: + logger.info(f"{image_key} {tags}") + + # metadataを書き出して終わり + logger.info(f"writing metadata: {args.out_json}") + Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding="utf-8") + + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ") + parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先") + parser.add_argument( + "--in_json", + type=str, + help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）", + ) + parser.add_argument( + "--full_path", + action="store_true", + help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）", + ) + parser.add_argument( + "--recursive", + action="store_true", + help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す", + ) + parser.add_argument( + "--caption_extension", + type=str, + default=".txt", + help="extension of caption (tag) file / 読み込むキャプション（タグ）ファイルの拡張子", + ) + parser.add_argument("--debug", action="store_true", help="debug mode, print tags") + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + main(args) diff --git a/merge_lora.py b/merge_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..fea8a3f3238e42b8f74ae2bfbbc8643f6981810c --- /dev/null +++ b/merge_lora.py @@ -0,0 +1,360 @@ +import math +import argparse +import os +import time +import torch +from safetensors.torch import load_file, save_file +from library import sai_model_spec, train_util +import library.model_util as model_util +import lora +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def load_state_dict(file_name, dtype): + if os.path.splitext(file_name)[1] == ".safetensors": + sd = load_file(file_name) + metadata = train_util.load_metadata_from_safetensors(file_name) + else: + sd = torch.load(file_name, map_location="cpu") + metadata = {} + + for key in list(sd.keys()): + if type(sd[key]) == torch.Tensor: + sd[key] = sd[key].to(dtype) + + return sd, metadata + + +def save_to_file(file_name, model, state_dict, dtype, metadata): + if dtype is not None: + for key in list(state_dict.keys()): + if type(state_dict[key]) == torch.Tensor: + state_dict[key] = state_dict[key].to(dtype) + + if os.path.splitext(file_name)[1] == ".safetensors": + save_file(model, file_name, metadata=metadata) + else: + torch.save(model, file_name) + + +def merge_to_sd_model(text_encoder, unet, models, ratios, merge_dtype): + text_encoder.to(merge_dtype) + unet.to(merge_dtype) + + # create module map + name_to_module = {} + for i, root_module in enumerate([text_encoder, unet]): + if i == 0: + prefix = lora.LoRANetwork.LORA_PREFIX_TEXT_ENCODER + target_replace_modules = lora.LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE + else: + prefix = lora.LoRANetwork.LORA_PREFIX_UNET + target_replace_modules = ( + lora.LoRANetwork.UNET_TARGET_REPLACE_MODULE + lora.LoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + ) + + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + if child_module.__class__.__name__ == "Linear" or child_module.__class__.__name__ == "Conv2d": + lora_name = prefix + "." + name + "." + child_name + lora_name = lora_name.replace(".", "_") + name_to_module[lora_name] = child_module + + for model, ratio in zip(models, ratios): + logger.info(f"loading: {model}") + lora_sd, _ = load_state_dict(model, merge_dtype) + + logger.info(f"merging...") + for key in lora_sd.keys(): + if "lora_down" in key: + up_key = key.replace("lora_down", "lora_up") + alpha_key = key[: key.index("lora_down")] + "alpha" + + # find original module for this lora + module_name = ".".join(key.split(".")[:-2]) # remove trailing ".lora_down.weight" + if module_name not in name_to_module: + logger.info(f"no module found for LoRA weight: {key}") + continue + module = name_to_module[module_name] + # logger.info(f"apply {key} to {module}") + + down_weight = lora_sd[key] + up_weight = lora_sd[up_key] + + dim = down_weight.size()[0] + alpha = lora_sd.get(alpha_key, dim) + scale = alpha / dim + + # W <- W + U * D + weight = module.weight + if len(weight.size()) == 2: + # linear + if len(up_weight.size()) == 4: # use linear projection mismatch + up_weight = up_weight.squeeze(3).squeeze(2) + down_weight = down_weight.squeeze(3).squeeze(2) + weight = weight + ratio * (up_weight @ down_weight) * scale + elif down_weight.size()[2:4] == (1, 1): + # conv2d 1x1 + weight = ( + weight + + ratio + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * scale + ) + else: + # conv2d 3x3 + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + # logger.info(conved.size(), weight.size(), module.stride, module.padding) + weight = weight + ratio * conved * scale + + module.weight = torch.nn.Parameter(weight) + + +def merge_lora_models(models, ratios, merge_dtype, concat=False, shuffle=False): + base_alphas = {} # alpha for merged model + base_dims = {} + + merged_sd = {} + v2 = None + base_model = None + for model, ratio in zip(models, ratios): + logger.info(f"loading: {model}") + lora_sd, lora_metadata = load_state_dict(model, merge_dtype) + + if lora_metadata is not None: + if v2 is None: + v2 = lora_metadata.get(train_util.SS_METADATA_KEY_V2, None) # return string + if base_model is None: + base_model = lora_metadata.get(train_util.SS_METADATA_KEY_BASE_MODEL_VERSION, None) + + # get alpha and dim + alphas = {} # alpha for current model + dims = {} # dims for current model + for key in lora_sd.keys(): + if "alpha" in key: + lora_module_name = key[: key.rfind(".alpha")] + alpha = float(lora_sd[key].detach().numpy()) + alphas[lora_module_name] = alpha + if lora_module_name not in base_alphas: + base_alphas[lora_module_name] = alpha + elif "lora_down" in key: + lora_module_name = key[: key.rfind(".lora_down")] + dim = lora_sd[key].size()[0] + dims[lora_module_name] = dim + if lora_module_name not in base_dims: + base_dims[lora_module_name] = dim + + for lora_module_name in dims.keys(): + if lora_module_name not in alphas: + alpha = dims[lora_module_name] + alphas[lora_module_name] = alpha + if lora_module_name not in base_alphas: + base_alphas[lora_module_name] = alpha + + logger.info(f"dim: {list(set(dims.values()))}, alpha: {list(set(alphas.values()))}") + + # merge + logger.info(f"merging...") + for key in lora_sd.keys(): + if "alpha" in key: + continue + if "lora_up" in key and concat: + concat_dim = 1 + elif "lora_down" in key and concat: + concat_dim = 0 + else: + concat_dim = None + + lora_module_name = key[: key.rfind(".lora_")] + + base_alpha = base_alphas[lora_module_name] + alpha = alphas[lora_module_name] + + scale = math.sqrt(alpha / base_alpha) * ratio + scale = abs(scale) if "lora_up" in key else scale # マイナスの重みに対応する。 + + if key in merged_sd: + assert ( + merged_sd[key].size() == lora_sd[key].size() or concat_dim is not None + ), f"weights shape mismatch merging v1 and v2, different dims? / 重みのサイズが合いません。v1とv2、または次元数の異なるモデルはマージできません" + if concat_dim is not None: + merged_sd[key] = torch.cat([merged_sd[key], lora_sd[key] * scale], dim=concat_dim) + else: + merged_sd[key] = merged_sd[key] + lora_sd[key] * scale + else: + merged_sd[key] = lora_sd[key] * scale + + # set alpha to sd + for lora_module_name, alpha in base_alphas.items(): + key = lora_module_name + ".alpha" + merged_sd[key] = torch.tensor(alpha) + if shuffle: + key_down = lora_module_name + ".lora_down.weight" + key_up = lora_module_name + ".lora_up.weight" + dim = merged_sd[key_down].shape[0] + perm = torch.randperm(dim) + merged_sd[key_down] = merged_sd[key_down][perm] + merged_sd[key_up] = merged_sd[key_up][:,perm] + + logger.info("merged model") + logger.info(f"dim: {list(set(base_dims.values()))}, alpha: {list(set(base_alphas.values()))}") + + # check all dims are same + dims_list = list(set(base_dims.values())) + alphas_list = list(set(base_alphas.values())) + all_same_dims = True + all_same_alphas = True + for dims in dims_list: + if dims != dims_list[0]: + all_same_dims = False + break + for alphas in alphas_list: + if alphas != alphas_list[0]: + all_same_alphas = False + break + + # build minimum metadata + dims = f"{dims_list[0]}" if all_same_dims else "Dynamic" + alphas = f"{alphas_list[0]}" if all_same_alphas else "Dynamic" + metadata = train_util.build_minimum_network_metadata(v2, base_model, "networks.lora", dims, alphas, None) + + return merged_sd, metadata, v2 == "True" + + +def merge(args): + assert len(args.models) == len(args.ratios), f"number of models must be equal to number of ratios / モデルの数と重みの数は合わせてください" + + def str_to_dtype(p): + if p == "float": + return torch.float + if p == "fp16": + return torch.float16 + if p == "bf16": + return torch.bfloat16 + return None + + merge_dtype = str_to_dtype(args.precision) + save_dtype = str_to_dtype(args.save_precision) + if save_dtype is None: + save_dtype = merge_dtype + + if args.sd_model is not None: + logger.info(f"loading SD model: {args.sd_model}") + + text_encoder, vae, unet = model_util.load_models_from_stable_diffusion_checkpoint(args.v2, args.sd_model) + + merge_to_sd_model(text_encoder, unet, args.models, args.ratios, merge_dtype) + + if args.no_metadata: + sai_metadata = None + else: + merged_from = sai_model_spec.build_merged_from([args.sd_model] + args.models) + title = os.path.splitext(os.path.basename(args.save_to))[0] + sai_metadata = sai_model_spec.build_metadata( + None, + args.v2, + args.v2, + False, + False, + False, + time.time(), + title=title, + merged_from=merged_from, + is_stable_diffusion_ckpt=True, + ) + if args.v2: + # TODO read sai modelspec + logger.warning( + "Cannot determine if model is for v-prediction, so save metadata as v-prediction / modelがv-prediction用か否か不明なため、仮にv-prediction用としてmetadataを保存します" + ) + + logger.info(f"saving SD model to: {args.save_to}") + model_util.save_stable_diffusion_checkpoint( + args.v2, args.save_to, text_encoder, unet, args.sd_model, 0, 0, sai_metadata, save_dtype, vae + ) + else: + state_dict, metadata, v2 = merge_lora_models(args.models, args.ratios, merge_dtype, args.concat, args.shuffle) + + logger.info(f"calculating hashes and creating metadata...") + + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + if not args.no_metadata: + merged_from = sai_model_spec.build_merged_from(args.models) + title = os.path.splitext(os.path.basename(args.save_to))[0] + sai_metadata = sai_model_spec.build_metadata( + state_dict, v2, v2, False, True, False, time.time(), title=title, merged_from=merged_from + ) + if v2: + # TODO read sai modelspec + logger.warning( + "Cannot determine if LoRA is for v-prediction, so save metadata as v-prediction / LoRAがv-prediction用か否か不明なため、仮にv-prediction用としてmetadataを保存します" + ) + metadata.update(sai_metadata) + + logger.info(f"saving model to: {args.save_to}") + save_to_file(args.save_to, state_dict, state_dict, save_dtype, metadata) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("--v2", action="store_true", help="load Stable Diffusion v2.x model / Stable Diffusion 2.xのモデルを読み込む") + parser.add_argument( + "--save_precision", + type=str, + default=None, + choices=[None, "float", "fp16", "bf16"], + help="precision in saving, same to merging if omitted / 保存時に精度を変更して保存する、省略時はマージ時の精度と同じ", + ) + parser.add_argument( + "--precision", + type=str, + default="float", + choices=["float", "fp16", "bf16"], + help="precision in merging (float is recommended) / マージの計算時の精度（floatを推奨）", + ) + parser.add_argument( + "--sd_model", + type=str, + default=None, + help="Stable Diffusion model to load: ckpt or safetensors file, merge LoRA models if omitted / 読み込むモデル、ckptまたはsafetensors。省略時はLoRAモデル同士をマージする", + ) + parser.add_argument( + "--save_to", type=str, default=None, help="destination file name: ckpt or safetensors file / 保存先のファイル名、ckptまたはsafetensors" + ) + parser.add_argument( + "--models", type=str, nargs="*", help="LoRA models to merge: ckpt or safetensors file / マージするLoRAモデル、ckptまたはsafetensors" + ) + parser.add_argument("--ratios", type=float, nargs="*", help="ratios for each model / それぞれのLoRAモデルの比率") + parser.add_argument( + "--no_metadata", + action="store_true", + help="do not save sai modelspec metadata (minimum ss_metadata for LoRA is saved) / " + + "sai modelspecのメタデータを保存しない（LoRAの最低限のss_metadataは保存される）", + ) + parser.add_argument( + "--concat", + action="store_true", + help="concat lora instead of merge (The dim(rank) of the output LoRA is the sum of the input dims) / " + + "マージの代わりに結合する（LoRAのdim(rank)は入力dimの合計になる）", + ) + parser.add_argument( + "--shuffle", + action="store_true", + help="shuffle lora weight./ " + + "LoRAの重みをシャッフルする", + ) + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + merge(args) diff --git a/merge_lora_old.py b/merge_lora_old.py new file mode 100644 index 0000000000000000000000000000000000000000..334d127b75598c3ada0d45c94a8c6a1e01bd4711 --- /dev/null +++ b/merge_lora_old.py @@ -0,0 +1,190 @@ + + +import argparse +import os +import torch +from safetensors.torch import load_file, save_file +import library.model_util as model_util +import lora +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def load_state_dict(file_name, dtype): + if os.path.splitext(file_name)[1] == '.safetensors': + sd = load_file(file_name) + else: + sd = torch.load(file_name, map_location='cpu') + for key in list(sd.keys()): + if type(sd[key]) == torch.Tensor: + sd[key] = sd[key].to(dtype) + return sd + + +def save_to_file(file_name, model, state_dict, dtype): + if dtype is not None: + for key in list(state_dict.keys()): + if type(state_dict[key]) == torch.Tensor: + state_dict[key] = state_dict[key].to(dtype) + + if os.path.splitext(file_name)[1] == '.safetensors': + save_file(model, file_name) + else: + torch.save(model, file_name) + + +def merge_to_sd_model(text_encoder, unet, models, ratios, merge_dtype): + text_encoder.to(merge_dtype) + unet.to(merge_dtype) + + # create module map + name_to_module = {} + for i, root_module in enumerate([text_encoder, unet]): + if i == 0: + prefix = lora.LoRANetwork.LORA_PREFIX_TEXT_ENCODER + target_replace_modules = lora.LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE + else: + prefix = lora.LoRANetwork.LORA_PREFIX_UNET + target_replace_modules = lora.LoRANetwork.UNET_TARGET_REPLACE_MODULE + + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + if child_module.__class__.__name__ == "Linear" or (child_module.__class__.__name__ == "Conv2d" and child_module.kernel_size == (1, 1)): + lora_name = prefix + '.' + name + '.' + child_name + lora_name = lora_name.replace('.', '_') + name_to_module[lora_name] = child_module + + for model, ratio in zip(models, ratios): + logger.info(f"loading: {model}") + lora_sd = load_state_dict(model, merge_dtype) + + logger.info(f"merging...") + for key in lora_sd.keys(): + if "lora_down" in key: + up_key = key.replace("lora_down", "lora_up") + alpha_key = key[:key.index("lora_down")] + 'alpha' + + # find original module for this lora + module_name = '.'.join(key.split('.')[:-2]) # remove trailing ".lora_down.weight" + if module_name not in name_to_module: + logger.info(f"no module found for LoRA weight: {key}") + continue + module = name_to_module[module_name] + # logger.info(f"apply {key} to {module}") + + down_weight = lora_sd[key] + up_weight = lora_sd[up_key] + + dim = down_weight.size()[0] + alpha = lora_sd.get(alpha_key, dim) + scale = alpha / dim + + # W <- W + U * D + weight = module.weight + if len(weight.size()) == 2: + # linear + weight = weight + ratio * (up_weight @ down_weight) * scale + else: + # conv2d + weight = weight + ratio * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) * scale + + module.weight = torch.nn.Parameter(weight) + + +def merge_lora_models(models, ratios, merge_dtype): + merged_sd = {} + + alpha = None + dim = None + for model, ratio in zip(models, ratios): + logger.info(f"loading: {model}") + lora_sd = load_state_dict(model, merge_dtype) + + logger.info(f"merging...") + for key in lora_sd.keys(): + if 'alpha' in key: + if key in merged_sd: + assert merged_sd[key] == lora_sd[key], f"alpha mismatch / alphaが異なる場合、現時点ではマージできません" + else: + alpha = lora_sd[key].detach().numpy() + merged_sd[key] = lora_sd[key] + else: + if key in merged_sd: + assert merged_sd[key].size() == lora_sd[key].size( + ), f"weights shape mismatch merging v1 and v2, different dims? / 重みのサイズが合いません。v1とv2、または次元数の異なるモデルはマージできません" + merged_sd[key] = merged_sd[key] + lora_sd[key] * ratio + else: + if "lora_down" in key: + dim = lora_sd[key].size()[0] + merged_sd[key] = lora_sd[key] * ratio + + logger.info(f"dim (rank): {dim}, alpha: {alpha}") + if alpha is None: + alpha = dim + + return merged_sd, dim, alpha + + +def merge(args): + assert len(args.models) == len(args.ratios), f"number of models must be equal to number of ratios / モデルの数と重みの数は合わせてください" + + def str_to_dtype(p): + if p == 'float': + return torch.float + if p == 'fp16': + return torch.float16 + if p == 'bf16': + return torch.bfloat16 + return None + + merge_dtype = str_to_dtype(args.precision) + save_dtype = str_to_dtype(args.save_precision) + if save_dtype is None: + save_dtype = merge_dtype + + if args.sd_model is not None: + logger.info(f"loading SD model: {args.sd_model}") + + text_encoder, vae, unet = model_util.load_models_from_stable_diffusion_checkpoint(args.v2, args.sd_model) + + merge_to_sd_model(text_encoder, unet, args.models, args.ratios, merge_dtype) + + logger.info("") + logger.info(f"saving SD model to: {args.save_to}") + model_util.save_stable_diffusion_checkpoint(args.v2, args.save_to, text_encoder, unet, + args.sd_model, 0, 0, save_dtype, vae) + else: + state_dict, _, _ = merge_lora_models(args.models, args.ratios, merge_dtype) + + logger.info(f"") + logger.info(f"saving model to: {args.save_to}") + save_to_file(args.save_to, state_dict, state_dict, save_dtype) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("--v2", action='store_true', + help='load Stable Diffusion v2.x model / Stable Diffusion 2.xのモデルを読み込む') + parser.add_argument("--save_precision", type=str, default=None, + choices=[None, "float", "fp16", "bf16"], help="precision in saving, same to merging if omitted / 保存時に精度を変更して保存する、省略時はマージ時の精度と同じ") + parser.add_argument("--precision", type=str, default="float", + choices=["float", "fp16", "bf16"], help="precision in merging (float is recommended) / マージの計算時の精度（floatを推奨）") + parser.add_argument("--sd_model", type=str, default=None, + help="Stable Diffusion model to load: ckpt or safetensors file, merge LoRA models if omitted / 読み込むモデル、ckptまたはsafetensors。省略時はLoRAモデル同士をマージする") + parser.add_argument("--save_to", type=str, default=None, + help="destination file name: ckpt or safetensors file / 保存先のファイル名、ckptまたはsafetensors") + parser.add_argument("--models", type=str, nargs='*', + help="LoRA models to merge: ckpt or safetensors file / マージするLoRAモデル、ckptまたはsafetensors") + parser.add_argument("--ratios", type=float, nargs='*', + help="ratios for each model / それぞれのLoRAモデルの比率") + + return parser + + +if __name__ == '__main__': + parser = setup_parser() + + args = parser.parse_args() + merge(args) diff --git a/merge_models.py b/merge_models.py new file mode 100644 index 0000000000000000000000000000000000000000..8f1fbf2f80f8859a13e30e4a35140788969604b6 --- /dev/null +++ b/merge_models.py @@ -0,0 +1,171 @@ +import argparse +import os + +import torch +from safetensors import safe_open +from safetensors.torch import load_file, save_file +from tqdm import tqdm +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def is_unet_key(key): + # VAE or TextEncoder, the last one is for SDXL + return not ("first_stage_model" in key or "cond_stage_model" in key or "conditioner." in key) + + +TEXT_ENCODER_KEY_REPLACEMENTS = [ + ("cond_stage_model.transformer.embeddings.", "cond_stage_model.transformer.text_model.embeddings."), + ("cond_stage_model.transformer.encoder.", "cond_stage_model.transformer.text_model.encoder."), + ("cond_stage_model.transformer.final_layer_norm.", "cond_stage_model.transformer.text_model.final_layer_norm."), +] + + +# support for models with different text encoder keys +def replace_text_encoder_key(key): + for rep_from, rep_to in TEXT_ENCODER_KEY_REPLACEMENTS: + if key.startswith(rep_from): + return True, rep_to + key[len(rep_from) :] + return False, key + + +def merge(args): + if args.precision == "fp16": + dtype = torch.float16 + elif args.precision == "bf16": + dtype = torch.bfloat16 + else: + dtype = torch.float + + if args.saving_precision == "fp16": + save_dtype = torch.float16 + elif args.saving_precision == "bf16": + save_dtype = torch.bfloat16 + else: + save_dtype = torch.float + + # check if all models are safetensors + for model in args.models: + if not model.endswith("safetensors"): + logger.info(f"Model {model} is not a safetensors model") + exit() + if not os.path.isfile(model): + logger.info(f"Model {model} does not exist") + exit() + + assert args.ratios is None or len(args.models) == len(args.ratios), "ratios must be the same length as models" + + # load and merge + ratio = 1.0 / len(args.models) # default + supplementary_key_ratios = {} # [key] = ratio, for keys not in all models, add later + + merged_sd = None + first_model_keys = set() # check missing keys in other models + for i, model in enumerate(args.models): + if args.ratios is not None: + ratio = args.ratios[i] + + if merged_sd is None: + # load first model + logger.info(f"Loading model {model}, ratio = {ratio}...") + merged_sd = {} + with safe_open(model, framework="pt", device=args.device) as f: + for key in tqdm(f.keys()): + value = f.get_tensor(key) + _, key = replace_text_encoder_key(key) + + first_model_keys.add(key) + + if not is_unet_key(key) and args.unet_only: + supplementary_key_ratios[key] = 1.0 # use first model's value for VAE or TextEncoder + continue + + value = ratio * value.to(dtype) # first model's value * ratio + merged_sd[key] = value + + logger.info(f"Model has {len(merged_sd)} keys " + ("(UNet only)" if args.unet_only else "")) + continue + + # load other models + logger.info(f"Loading model {model}, ratio = {ratio}...") + + with safe_open(model, framework="pt", device=args.device) as f: + model_keys = f.keys() + for key in tqdm(model_keys): + _, new_key = replace_text_encoder_key(key) + if new_key not in merged_sd: + if args.show_skipped and new_key not in first_model_keys: + logger.info(f"Skip: {new_key}") + continue + + value = f.get_tensor(key) + merged_sd[new_key] = merged_sd[new_key] + ratio * value.to(dtype) + + # enumerate keys not in this model + model_keys = set(model_keys) + for key in merged_sd.keys(): + if key in model_keys: + continue + logger.warning(f"Key {key} not in model {model}, use first model's value") + if key in supplementary_key_ratios: + supplementary_key_ratios[key] += ratio + else: + supplementary_key_ratios[key] = ratio + + # add supplementary keys' value (including VAE and TextEncoder) + if len(supplementary_key_ratios) > 0: + logger.info("add first model's value") + with safe_open(args.models[0], framework="pt", device=args.device) as f: + for key in tqdm(f.keys()): + _, new_key = replace_text_encoder_key(key) + if new_key not in supplementary_key_ratios: + continue + + if is_unet_key(new_key): # not VAE or TextEncoder + logger.warning(f"Key {new_key} not in all models, ratio = {supplementary_key_ratios[new_key]}") + + value = f.get_tensor(key) # original key + + if new_key not in merged_sd: + merged_sd[new_key] = supplementary_key_ratios[new_key] * value.to(dtype) + else: + merged_sd[new_key] = merged_sd[new_key] + supplementary_key_ratios[new_key] * value.to(dtype) + + # save + output_file = args.output + if not output_file.endswith(".safetensors"): + output_file = output_file + ".safetensors" + + logger.info(f"Saving to {output_file}...") + + # convert to save_dtype + for k in merged_sd.keys(): + merged_sd[k] = merged_sd[k].to(save_dtype) + + save_file(merged_sd, output_file) + + logger.info("Done!") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser(description="Merge models") + parser.add_argument("--models", nargs="+", type=str, help="Models to merge") + parser.add_argument("--output", type=str, help="Output model") + parser.add_argument("--ratios", nargs="+", type=float, help="Ratios of models, default is equal, total = 1.0") + parser.add_argument("--unet_only", action="store_true", help="Only merge unet") + parser.add_argument("--device", type=str, default="cpu", help="Device to use, default is cpu") + parser.add_argument( + "--precision", type=str, default="float", choices=["float", "fp16", "bf16"], help="Calculation precision, default is float" + ) + parser.add_argument( + "--saving_precision", + type=str, + default="float", + choices=["float", "fp16", "bf16"], + help="Saving precision, default is float", + ) + parser.add_argument("--show_skipped", action="store_true", help="Show skipped keys (keys not in first model)") + + args = parser.parse_args() + merge(args) diff --git a/model_util.py b/model_util.py new file mode 100644 index 0000000000000000000000000000000000000000..9918c7b2af38062c239bb4904538098bf711ead2 --- /dev/null +++ b/model_util.py @@ -0,0 +1,1355 @@ +# v1: split from train_db_fixed.py. +# v2: support safetensors + +import math +import os + +import torch +from library.device_utils import init_ipex +init_ipex() + +import diffusers +from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextConfig, logging +from diffusers import AutoencoderKL, DDIMScheduler, StableDiffusionPipeline # , UNet2DConditionModel +from safetensors.torch import load_file, save_file +from library.original_unet import UNet2DConditionModel +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +# DiffUsers版StableDiffusionのモデルパラメータ +NUM_TRAIN_TIMESTEPS = 1000 +BETA_START = 0.00085 +BETA_END = 0.0120 + +UNET_PARAMS_MODEL_CHANNELS = 320 +UNET_PARAMS_CHANNEL_MULT = [1, 2, 4, 4] +UNET_PARAMS_ATTENTION_RESOLUTIONS = [4, 2, 1] +UNET_PARAMS_IMAGE_SIZE = 64 # fixed from old invalid value `32` +UNET_PARAMS_IN_CHANNELS = 4 +UNET_PARAMS_OUT_CHANNELS = 4 +UNET_PARAMS_NUM_RES_BLOCKS = 2 +UNET_PARAMS_CONTEXT_DIM = 768 +UNET_PARAMS_NUM_HEADS = 8 +# UNET_PARAMS_USE_LINEAR_PROJECTION = False + +VAE_PARAMS_Z_CHANNELS = 4 +VAE_PARAMS_RESOLUTION = 256 +VAE_PARAMS_IN_CHANNELS = 3 +VAE_PARAMS_OUT_CH = 3 +VAE_PARAMS_CH = 128 +VAE_PARAMS_CH_MULT = [1, 2, 4, 4] +VAE_PARAMS_NUM_RES_BLOCKS = 2 + +# V2 +V2_UNET_PARAMS_ATTENTION_HEAD_DIM = [5, 10, 20, 20] +V2_UNET_PARAMS_CONTEXT_DIM = 1024 +# V2_UNET_PARAMS_USE_LINEAR_PROJECTION = True + +# Diffusersの設定を読み込むための参照モデル +DIFFUSERS_REF_MODEL_ID_V1 = "runwayml/stable-diffusion-v1-5" +DIFFUSERS_REF_MODEL_ID_V2 = "stabilityai/stable-diffusion-2-1" + + +# region StableDiffusion->Diffusersの変換コード +# convert_original_stable_diffusion_to_diffusers をコピーして修正している（ASL 2.0） + + +def shave_segments(path, n_shave_prefix_segments=1): + """ + Removes segments. Positive values shave the first segments, negative shave the last segments. + """ + if n_shave_prefix_segments >= 0: + return ".".join(path.split(".")[n_shave_prefix_segments:]) + else: + return ".".join(path.split(".")[:n_shave_prefix_segments]) + + +def renew_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item.replace("in_layers.0", "norm1") + new_item = new_item.replace("in_layers.2", "conv1") + + new_item = new_item.replace("out_layers.0", "norm2") + new_item = new_item.replace("out_layers.3", "conv2") + + new_item = new_item.replace("emb_layers.1", "time_emb_proj") + new_item = new_item.replace("skip_connection", "conv_shortcut") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("nin_shortcut", "conv_shortcut") + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + # new_item = new_item.replace('norm.weight', 'group_norm.weight') + # new_item = new_item.replace('norm.bias', 'group_norm.bias') + + # new_item = new_item.replace('proj_out.weight', 'proj_attn.weight') + # new_item = new_item.replace('proj_out.bias', 'proj_attn.bias') + + # new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("norm.weight", "group_norm.weight") + new_item = new_item.replace("norm.bias", "group_norm.bias") + + if diffusers.__version__ < "0.17.0": + new_item = new_item.replace("q.weight", "query.weight") + new_item = new_item.replace("q.bias", "query.bias") + + new_item = new_item.replace("k.weight", "key.weight") + new_item = new_item.replace("k.bias", "key.bias") + + new_item = new_item.replace("v.weight", "value.weight") + new_item = new_item.replace("v.bias", "value.bias") + + new_item = new_item.replace("proj_out.weight", "proj_attn.weight") + new_item = new_item.replace("proj_out.bias", "proj_attn.bias") + else: + new_item = new_item.replace("q.weight", "to_q.weight") + new_item = new_item.replace("q.bias", "to_q.bias") + + new_item = new_item.replace("k.weight", "to_k.weight") + new_item = new_item.replace("k.bias", "to_k.bias") + + new_item = new_item.replace("v.weight", "to_v.weight") + new_item = new_item.replace("v.bias", "to_v.bias") + + new_item = new_item.replace("proj_out.weight", "to_out.0.weight") + new_item = new_item.replace("proj_out.bias", "to_out.0.bias") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def assign_to_checkpoint( + paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None +): + """ + This does the final conversion step: take locally converted weights and apply a global renaming + to them. It splits attention layers, and takes into account additional replacements + that may arise. + + Assigns the weights to the new checkpoint. + """ + assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys." + + # Splits the attention layers into three variables. + if attention_paths_to_split is not None: + for path, path_map in attention_paths_to_split.items(): + old_tensor = old_checkpoint[path] + channels = old_tensor.shape[0] // 3 + + target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1) + + num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3 + + old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:]) + query, key, value = old_tensor.split(channels // num_heads, dim=1) + + checkpoint[path_map["query"]] = query.reshape(target_shape) + checkpoint[path_map["key"]] = key.reshape(target_shape) + checkpoint[path_map["value"]] = value.reshape(target_shape) + + for path in paths: + new_path = path["new"] + + # These have already been assigned + if attention_paths_to_split is not None and new_path in attention_paths_to_split: + continue + + # Global renaming happens here + new_path = new_path.replace("middle_block.0", "mid_block.resnets.0") + new_path = new_path.replace("middle_block.1", "mid_block.attentions.0") + new_path = new_path.replace("middle_block.2", "mid_block.resnets.1") + + if additional_replacements is not None: + for replacement in additional_replacements: + new_path = new_path.replace(replacement["old"], replacement["new"]) + + # proj_attn.weight has to be converted from conv 1D to linear + reshaping = False + if diffusers.__version__ < "0.17.0": + if "proj_attn.weight" in new_path: + reshaping = True + else: + if ".attentions." in new_path and ".0.to_" in new_path and old_checkpoint[path["old"]].ndim > 2: + reshaping = True + + if reshaping: + checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0, 0] + else: + checkpoint[new_path] = old_checkpoint[path["old"]] + + +def conv_attn_to_linear(checkpoint): + keys = list(checkpoint.keys()) + attn_keys = ["query.weight", "key.weight", "value.weight"] + for key in keys: + if ".".join(key.split(".")[-2:]) in attn_keys: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0, 0] + elif "proj_attn.weight" in key: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0] + + +def linear_transformer_to_conv(checkpoint): + keys = list(checkpoint.keys()) + tf_keys = ["proj_in.weight", "proj_out.weight"] + for key in keys: + if ".".join(key.split(".")[-2:]) in tf_keys: + if checkpoint[key].ndim == 2: + checkpoint[key] = checkpoint[key].unsqueeze(2).unsqueeze(2) + + +def convert_ldm_unet_checkpoint(v2, checkpoint, config): + """ + Takes a state dict and a config, and returns a converted checkpoint. + """ + + # extract state_dict for UNet + unet_state_dict = {} + unet_key = "model.diffusion_model." + keys = list(checkpoint.keys()) + for key in keys: + if key.startswith(unet_key): + unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key) + + new_checkpoint = {} + + new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"] + new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"] + new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"] + new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"] + + new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"] + new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"] + + new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"] + new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"] + new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"] + new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"] + + # Retrieves the keys for the input blocks only + num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer}) + input_blocks = { + layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}." in key] for layer_id in range(num_input_blocks) + } + + # Retrieves the keys for the middle blocks only + num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer}) + middle_blocks = { + layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}." in key] for layer_id in range(num_middle_blocks) + } + + # Retrieves the keys for the output blocks only + num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer}) + output_blocks = { + layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}." in key] for layer_id in range(num_output_blocks) + } + + for i in range(1, num_input_blocks): + block_id = (i - 1) // (config["layers_per_block"] + 1) + layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1) + + resnets = [key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key] + attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key] + + if f"input_blocks.{i}.0.op.weight" in unet_state_dict: + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop( + f"input_blocks.{i}.0.op.weight" + ) + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(f"input_blocks.{i}.0.op.bias") + + paths = renew_resnet_paths(resnets) + meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint(paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config) + + if len(attentions): + paths = renew_attention_paths(attentions) + meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"} + assign_to_checkpoint(paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config) + + resnet_0 = middle_blocks[0] + attentions = middle_blocks[1] + resnet_1 = middle_blocks[2] + + resnet_0_paths = renew_resnet_paths(resnet_0) + assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config) + + resnet_1_paths = renew_resnet_paths(resnet_1) + assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config) + + attentions_paths = renew_attention_paths(attentions) + meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config) + + for i in range(num_output_blocks): + block_id = i // (config["layers_per_block"] + 1) + layer_in_block_id = i % (config["layers_per_block"] + 1) + output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]] + output_block_list = {} + + for layer in output_block_layers: + layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1) + if layer_id in output_block_list: + output_block_list[layer_id].append(layer_name) + else: + output_block_list[layer_id] = [layer_name] + + if len(output_block_list) > 1: + resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key] + attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key] + + resnet_0_paths = renew_resnet_paths(resnets) + paths = renew_resnet_paths(resnets) + + meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint(paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config) + + # オリジナル： + # if ["conv.weight", "conv.bias"] in output_block_list.values(): + # index = list(output_block_list.values()).index(["conv.weight", "conv.bias"]) + + # biasとweightの順番に依存しないようにする：もっといいやり方がありそうだが + for l in output_block_list.values(): + l.sort() + + if ["conv.bias", "conv.weight"] in output_block_list.values(): + index = list(output_block_list.values()).index(["conv.bias", "conv.weight"]) + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.bias" + ] + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.weight" + ] + + # Clear attentions as they have been attributed above. + if len(attentions) == 2: + attentions = [] + + if len(attentions): + paths = renew_attention_paths(attentions) + meta_path = { + "old": f"output_blocks.{i}.1", + "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}", + } + assign_to_checkpoint(paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config) + else: + resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1) + for path in resnet_0_paths: + old_path = ".".join(["output_blocks", str(i), path["old"]]) + new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]]) + + new_checkpoint[new_path] = unet_state_dict[old_path] + + # SDのv2では1*1のconv2dがlinearに変わっている + # 誤って Diffusers 側を conv2d のままにしてしまったので、変換必要 + if v2 and not config.get("use_linear_projection", False): + linear_transformer_to_conv(new_checkpoint) + + return new_checkpoint + + +def convert_ldm_vae_checkpoint(checkpoint, config): + # extract state dict for VAE + vae_state_dict = {} + vae_key = "first_stage_model." + keys = list(checkpoint.keys()) + for key in keys: + if key.startswith(vae_key): + vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key) + # if len(vae_state_dict) == 0: + # # 渡されたcheckpointは.ckptから読み込んだcheckpointではなくvaeのstate_dict + # vae_state_dict = checkpoint + + new_checkpoint = {} + + new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"] + new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"] + new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"] + new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"] + new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"] + new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"] + + new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"] + new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"] + new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"] + new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"] + new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"] + new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"] + + new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"] + new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"] + new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"] + new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"] + + # Retrieves the keys for the encoder down blocks only + num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer}) + down_blocks = {layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)} + + # Retrieves the keys for the decoder up blocks only + num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer}) + up_blocks = {layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)} + + for i in range(num_down_blocks): + resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key] + + if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict: + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.weight" + ) + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.bias" + ) + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + + for i in range(num_up_blocks): + block_id = num_up_blocks - 1 - i + resnets = [key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key] + + if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict: + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.weight" + ] + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.bias" + ] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + return new_checkpoint + + +def create_unet_diffusers_config(v2, use_linear_projection_in_v2=False): + """ + Creates a config for the diffusers based on the config of the LDM model. + """ + # unet_params = original_config.model.params.unet_config.params + + block_out_channels = [UNET_PARAMS_MODEL_CHANNELS * mult for mult in UNET_PARAMS_CHANNEL_MULT] + + down_block_types = [] + resolution = 1 + for i in range(len(block_out_channels)): + block_type = "CrossAttnDownBlock2D" if resolution in UNET_PARAMS_ATTENTION_RESOLUTIONS else "DownBlock2D" + down_block_types.append(block_type) + if i != len(block_out_channels) - 1: + resolution *= 2 + + up_block_types = [] + for i in range(len(block_out_channels)): + block_type = "CrossAttnUpBlock2D" if resolution in UNET_PARAMS_ATTENTION_RESOLUTIONS else "UpBlock2D" + up_block_types.append(block_type) + resolution //= 2 + + config = dict( + sample_size=UNET_PARAMS_IMAGE_SIZE, + in_channels=UNET_PARAMS_IN_CHANNELS, + out_channels=UNET_PARAMS_OUT_CHANNELS, + down_block_types=tuple(down_block_types), + up_block_types=tuple(up_block_types), + block_out_channels=tuple(block_out_channels), + layers_per_block=UNET_PARAMS_NUM_RES_BLOCKS, + cross_attention_dim=UNET_PARAMS_CONTEXT_DIM if not v2 else V2_UNET_PARAMS_CONTEXT_DIM, + attention_head_dim=UNET_PARAMS_NUM_HEADS if not v2 else V2_UNET_PARAMS_ATTENTION_HEAD_DIM, + # use_linear_projection=UNET_PARAMS_USE_LINEAR_PROJECTION if not v2 else V2_UNET_PARAMS_USE_LINEAR_PROJECTION, + ) + if v2 and use_linear_projection_in_v2: + config["use_linear_projection"] = True + + return config + + +def create_vae_diffusers_config(): + """ + Creates a config for the diffusers based on the config of the LDM model. + """ + # vae_params = original_config.model.params.first_stage_config.params.ddconfig + # _ = original_config.model.params.first_stage_config.params.embed_dim + block_out_channels = [VAE_PARAMS_CH * mult for mult in VAE_PARAMS_CH_MULT] + down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels) + up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels) + + config = dict( + sample_size=VAE_PARAMS_RESOLUTION, + in_channels=VAE_PARAMS_IN_CHANNELS, + out_channels=VAE_PARAMS_OUT_CH, + down_block_types=tuple(down_block_types), + up_block_types=tuple(up_block_types), + block_out_channels=tuple(block_out_channels), + latent_channels=VAE_PARAMS_Z_CHANNELS, + layers_per_block=VAE_PARAMS_NUM_RES_BLOCKS, + ) + return config + + +def convert_ldm_clip_checkpoint_v1(checkpoint): + keys = list(checkpoint.keys()) + text_model_dict = {} + for key in keys: + if key.startswith("cond_stage_model.transformer"): + text_model_dict[key[len("cond_stage_model.transformer.") :]] = checkpoint[key] + + # remove position_ids for newer transformer, which causes error :( + if "text_model.embeddings.position_ids" in text_model_dict: + text_model_dict.pop("text_model.embeddings.position_ids") + + return text_model_dict + + +def convert_ldm_clip_checkpoint_v2(checkpoint, max_length): + # 嫌になるくらい違うぞ！ + def convert_key(key): + if not key.startswith("cond_stage_model"): + return None + + # common conversion + key = key.replace("cond_stage_model.model.transformer.", "text_model.encoder.") + key = key.replace("cond_stage_model.model.", "text_model.") + + if "resblocks" in key: + # resblocks conversion + key = key.replace(".resblocks.", ".layers.") + if ".ln_" in key: + key = key.replace(".ln_", ".layer_norm") + elif ".mlp." in key: + key = key.replace(".c_fc.", ".fc1.") + key = key.replace(".c_proj.", ".fc2.") + elif ".attn.out_proj" in key: + key = key.replace(".attn.out_proj.", ".self_attn.out_proj.") + elif ".attn.in_proj" in key: + key = None # 特殊なので後で処理する + else: + raise ValueError(f"unexpected key in SD: {key}") + elif ".positional_embedding" in key: + key = key.replace(".positional_embedding", ".embeddings.position_embedding.weight") + elif ".text_projection" in key: + key = None # 使われない??? + elif ".logit_scale" in key: + key = None # 使われない??? + elif ".token_embedding" in key: + key = key.replace(".token_embedding.weight", ".embeddings.token_embedding.weight") + elif ".ln_final" in key: + key = key.replace(".ln_final", ".final_layer_norm") + return key + + keys = list(checkpoint.keys()) + new_sd = {} + for key in keys: + # remove resblocks 23 + if ".resblocks.23." in key: + continue + new_key = convert_key(key) + if new_key is None: + continue + new_sd[new_key] = checkpoint[key] + + # attnの変換 + for key in keys: + if ".resblocks.23." in key: + continue + if ".resblocks" in key and ".attn.in_proj_" in key: + # 三つに分割 + values = torch.chunk(checkpoint[key], 3) + + key_suffix = ".weight" if "weight" in key else ".bias" + key_pfx = key.replace("cond_stage_model.model.transformer.resblocks.", "text_model.encoder.layers.") + key_pfx = key_pfx.replace("_weight", "") + key_pfx = key_pfx.replace("_bias", "") + key_pfx = key_pfx.replace(".attn.in_proj", ".self_attn.") + new_sd[key_pfx + "q_proj" + key_suffix] = values[0] + new_sd[key_pfx + "k_proj" + key_suffix] = values[1] + new_sd[key_pfx + "v_proj" + key_suffix] = values[2] + + # remove position_ids for newer transformer, which causes error :( + ANOTHER_POSITION_IDS_KEY = "text_model.encoder.text_model.embeddings.position_ids" + if ANOTHER_POSITION_IDS_KEY in new_sd: + # waifu diffusion v1.4 + del new_sd[ANOTHER_POSITION_IDS_KEY] + + if "text_model.embeddings.position_ids" in new_sd: + del new_sd["text_model.embeddings.position_ids"] + + return new_sd + + +# endregion + + +# region Diffusers->StableDiffusion の変換コード +# convert_diffusers_to_original_stable_diffusion をコピーして修正している（ASL 2.0） + + +def conv_transformer_to_linear(checkpoint): + keys = list(checkpoint.keys()) + tf_keys = ["proj_in.weight", "proj_out.weight"] + for key in keys: + if ".".join(key.split(".")[-2:]) in tf_keys: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0, 0] + + +def convert_unet_state_dict_to_sd(v2, unet_state_dict): + unet_conversion_map = [ + # (stable-diffusion, HF Diffusers) + ("time_embed.0.weight", "time_embedding.linear_1.weight"), + ("time_embed.0.bias", "time_embedding.linear_1.bias"), + ("time_embed.2.weight", "time_embedding.linear_2.weight"), + ("time_embed.2.bias", "time_embedding.linear_2.bias"), + ("input_blocks.0.0.weight", "conv_in.weight"), + ("input_blocks.0.0.bias", "conv_in.bias"), + ("out.0.weight", "conv_norm_out.weight"), + ("out.0.bias", "conv_norm_out.bias"), + ("out.2.weight", "conv_out.weight"), + ("out.2.bias", "conv_out.bias"), + ] + + unet_conversion_map_resnet = [ + # (stable-diffusion, HF Diffusers) + ("in_layers.0", "norm1"), + ("in_layers.2", "conv1"), + ("out_layers.0", "norm2"), + ("out_layers.3", "conv2"), + ("emb_layers.1", "time_emb_proj"), + ("skip_connection", "conv_shortcut"), + ] + + unet_conversion_map_layer = [] + for i in range(4): + # loop over downblocks/upblocks + + for j in range(2): + # loop over resnets/attentions for downblocks + hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}." + sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0." + unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix)) + + if i < 3: + # no attention layers in down_blocks.3 + hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}." + sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1." + unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix)) + + for j in range(3): + # loop over resnets/attentions for upblocks + hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}." + sd_up_res_prefix = f"output_blocks.{3*i + j}.0." + unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix)) + + if i > 0: + # no attention layers in up_blocks.0 + hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}." + sd_up_atn_prefix = f"output_blocks.{3*i + j}.1." + unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix)) + + if i < 3: + # no downsample in down_blocks.3 + hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv." + sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op." + unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix)) + + # no upsample in up_blocks.3 + hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0." + sd_upsample_prefix = f"output_blocks.{3*i + 2}.{1 if i == 0 else 2}." + unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix)) + + hf_mid_atn_prefix = "mid_block.attentions.0." + sd_mid_atn_prefix = "middle_block.1." + unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix)) + + for j in range(2): + hf_mid_res_prefix = f"mid_block.resnets.{j}." + sd_mid_res_prefix = f"middle_block.{2*j}." + unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix)) + + # buyer beware: this is a *brittle* function, + # and correct output requires that all of these pieces interact in + # the exact order in which I have arranged them. + mapping = {k: k for k in unet_state_dict.keys()} + for sd_name, hf_name in unet_conversion_map: + mapping[hf_name] = sd_name + for k, v in mapping.items(): + if "resnets" in k: + for sd_part, hf_part in unet_conversion_map_resnet: + v = v.replace(hf_part, sd_part) + mapping[k] = v + for k, v in mapping.items(): + for sd_part, hf_part in unet_conversion_map_layer: + v = v.replace(hf_part, sd_part) + mapping[k] = v + new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()} + + if v2: + conv_transformer_to_linear(new_state_dict) + + return new_state_dict + + +def controlnet_conversion_map(): + unet_conversion_map = [ + ("time_embed.0.weight", "time_embedding.linear_1.weight"), + ("time_embed.0.bias", "time_embedding.linear_1.bias"), + ("time_embed.2.weight", "time_embedding.linear_2.weight"), + ("time_embed.2.bias", "time_embedding.linear_2.bias"), + ("input_blocks.0.0.weight", "conv_in.weight"), + ("input_blocks.0.0.bias", "conv_in.bias"), + ("middle_block_out.0.weight", "controlnet_mid_block.weight"), + ("middle_block_out.0.bias", "controlnet_mid_block.bias"), + ] + + unet_conversion_map_resnet = [ + ("in_layers.0", "norm1"), + ("in_layers.2", "conv1"), + ("out_layers.0", "norm2"), + ("out_layers.3", "conv2"), + ("emb_layers.1", "time_emb_proj"), + ("skip_connection", "conv_shortcut"), + ] + + unet_conversion_map_layer = [] + for i in range(4): + for j in range(2): + hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}." + sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0." + unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix)) + + if i < 3: + hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}." + sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1." + unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix)) + + if i < 3: + hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv." + sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op." + unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix)) + + hf_mid_atn_prefix = "mid_block.attentions.0." + sd_mid_atn_prefix = "middle_block.1." + unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix)) + + for j in range(2): + hf_mid_res_prefix = f"mid_block.resnets.{j}." + sd_mid_res_prefix = f"middle_block.{2*j}." + unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix)) + + controlnet_cond_embedding_names = ["conv_in"] + [f"blocks.{i}" for i in range(6)] + ["conv_out"] + for i, hf_prefix in enumerate(controlnet_cond_embedding_names): + hf_prefix = f"controlnet_cond_embedding.{hf_prefix}." + sd_prefix = f"input_hint_block.{i*2}." + unet_conversion_map_layer.append((sd_prefix, hf_prefix)) + + for i in range(12): + hf_prefix = f"controlnet_down_blocks.{i}." + sd_prefix = f"zero_convs.{i}.0." + unet_conversion_map_layer.append((sd_prefix, hf_prefix)) + + return unet_conversion_map, unet_conversion_map_resnet, unet_conversion_map_layer + + +def convert_controlnet_state_dict_to_sd(controlnet_state_dict): + unet_conversion_map, unet_conversion_map_resnet, unet_conversion_map_layer = controlnet_conversion_map() + + mapping = {k: k for k in controlnet_state_dict.keys()} + for sd_name, diffusers_name in unet_conversion_map: + mapping[diffusers_name] = sd_name + for k, v in mapping.items(): + if "resnets" in k: + for sd_part, diffusers_part in unet_conversion_map_resnet: + v = v.replace(diffusers_part, sd_part) + mapping[k] = v + for k, v in mapping.items(): + for sd_part, diffusers_part in unet_conversion_map_layer: + v = v.replace(diffusers_part, sd_part) + mapping[k] = v + new_state_dict = {v: controlnet_state_dict[k] for k, v in mapping.items()} + return new_state_dict + + +def convert_controlnet_state_dict_to_diffusers(controlnet_state_dict): + unet_conversion_map, unet_conversion_map_resnet, unet_conversion_map_layer = controlnet_conversion_map() + + mapping = {k: k for k in controlnet_state_dict.keys()} + for sd_name, diffusers_name in unet_conversion_map: + mapping[sd_name] = diffusers_name + for k, v in mapping.items(): + for sd_part, diffusers_part in unet_conversion_map_layer: + v = v.replace(sd_part, diffusers_part) + mapping[k] = v + for k, v in mapping.items(): + if "resnets" in v: + for sd_part, diffusers_part in unet_conversion_map_resnet: + v = v.replace(sd_part, diffusers_part) + mapping[k] = v + new_state_dict = {v: controlnet_state_dict[k] for k, v in mapping.items()} + return new_state_dict + + +# ================# +# VAE Conversion # +# ================# + + +def reshape_weight_for_sd(w): + # convert HF linear weights to SD conv2d weights + return w.reshape(*w.shape, 1, 1) + + +def convert_vae_state_dict(vae_state_dict): + vae_conversion_map = [ + # (stable-diffusion, HF Diffusers) + ("nin_shortcut", "conv_shortcut"), + ("norm_out", "conv_norm_out"), + ("mid.attn_1.", "mid_block.attentions.0."), + ] + + for i in range(4): + # down_blocks have two resnets + for j in range(2): + hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}." + sd_down_prefix = f"encoder.down.{i}.block.{j}." + vae_conversion_map.append((sd_down_prefix, hf_down_prefix)) + + if i < 3: + hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0." + sd_downsample_prefix = f"down.{i}.downsample." + vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix)) + + hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0." + sd_upsample_prefix = f"up.{3-i}.upsample." + vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix)) + + # up_blocks have three resnets + # also, up blocks in hf are numbered in reverse from sd + for j in range(3): + hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}." + sd_up_prefix = f"decoder.up.{3-i}.block.{j}." + vae_conversion_map.append((sd_up_prefix, hf_up_prefix)) + + # this part accounts for mid blocks in both the encoder and the decoder + for i in range(2): + hf_mid_res_prefix = f"mid_block.resnets.{i}." + sd_mid_res_prefix = f"mid.block_{i+1}." + vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix)) + + if diffusers.__version__ < "0.17.0": + vae_conversion_map_attn = [ + # (stable-diffusion, HF Diffusers) + ("norm.", "group_norm."), + ("q.", "query."), + ("k.", "key."), + ("v.", "value."), + ("proj_out.", "proj_attn."), + ] + else: + vae_conversion_map_attn = [ + # (stable-diffusion, HF Diffusers) + ("norm.", "group_norm."), + ("q.", "to_q."), + ("k.", "to_k."), + ("v.", "to_v."), + ("proj_out.", "to_out.0."), + ] + + mapping = {k: k for k in vae_state_dict.keys()} + for k, v in mapping.items(): + for sd_part, hf_part in vae_conversion_map: + v = v.replace(hf_part, sd_part) + mapping[k] = v + for k, v in mapping.items(): + if "attentions" in k: + for sd_part, hf_part in vae_conversion_map_attn: + v = v.replace(hf_part, sd_part) + mapping[k] = v + new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()} + weights_to_convert = ["q", "k", "v", "proj_out"] + for k, v in new_state_dict.items(): + for weight_name in weights_to_convert: + if f"mid.attn_1.{weight_name}.weight" in k: + # logger.info(f"Reshaping {k} for SD format: shape {v.shape} -> {v.shape} x 1 x 1") + new_state_dict[k] = reshape_weight_for_sd(v) + + return new_state_dict + + +# endregion + +# region 自作のモデル読み書きなど + + +def is_safetensors(path): + return os.path.splitext(path)[1].lower() == ".safetensors" + + +def load_checkpoint_with_text_encoder_conversion(ckpt_path, device="cpu"): + # text encoderの格納形式が違うモデルに対応する ('text_model'がない) + TEXT_ENCODER_KEY_REPLACEMENTS = [ + ("cond_stage_model.transformer.embeddings.", "cond_stage_model.transformer.text_model.embeddings."), + ("cond_stage_model.transformer.encoder.", "cond_stage_model.transformer.text_model.encoder."), + ("cond_stage_model.transformer.final_layer_norm.", "cond_stage_model.transformer.text_model.final_layer_norm."), + ] + + if is_safetensors(ckpt_path): + checkpoint = None + state_dict = load_file(ckpt_path) # , device) # may causes error + else: + checkpoint = torch.load(ckpt_path, map_location=device) + if "state_dict" in checkpoint: + state_dict = checkpoint["state_dict"] + else: + state_dict = checkpoint + checkpoint = None + + key_reps = [] + for rep_from, rep_to in TEXT_ENCODER_KEY_REPLACEMENTS: + for key in state_dict.keys(): + if key.startswith(rep_from): + new_key = rep_to + key[len(rep_from) :] + key_reps.append((key, new_key)) + + for key, new_key in key_reps: + state_dict[new_key] = state_dict[key] + del state_dict[key] + + return checkpoint, state_dict + + +# TODO dtype指定の動作が怪しいので確認する text_encoderを指定形式で作れるか未確認 +def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dtype=None, unet_use_linear_projection_in_v2=True): + _, state_dict = load_checkpoint_with_text_encoder_conversion(ckpt_path, device) + + # Convert the UNet2DConditionModel model. + unet_config = create_unet_diffusers_config(v2, unet_use_linear_projection_in_v2) + converted_unet_checkpoint = convert_ldm_unet_checkpoint(v2, state_dict, unet_config) + + unet = UNet2DConditionModel(**unet_config).to(device) + info = unet.load_state_dict(converted_unet_checkpoint) + logger.info(f"loading u-net: {info}") + + # Convert the VAE model. + vae_config = create_vae_diffusers_config() + converted_vae_checkpoint = convert_ldm_vae_checkpoint(state_dict, vae_config) + + vae = AutoencoderKL(**vae_config).to(device) + info = vae.load_state_dict(converted_vae_checkpoint) + logger.info(f"loading vae: {info}") + + # convert text_model + if v2: + converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v2(state_dict, 77) + cfg = CLIPTextConfig( + vocab_size=49408, + hidden_size=1024, + intermediate_size=4096, + num_hidden_layers=23, + num_attention_heads=16, + max_position_embeddings=77, + hidden_act="gelu", + layer_norm_eps=1e-05, + dropout=0.0, + attention_dropout=0.0, + initializer_range=0.02, + initializer_factor=1.0, + pad_token_id=1, + bos_token_id=0, + eos_token_id=2, + model_type="clip_text_model", + projection_dim=512, + torch_dtype="float32", + transformers_version="4.25.0.dev0", + ) + text_model = CLIPTextModel._from_config(cfg) + info = text_model.load_state_dict(converted_text_encoder_checkpoint) + else: + converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v1(state_dict) + + # logging.set_verbosity_error() # don't show annoying warning + # text_model = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14").to(device) + # logging.set_verbosity_warning() + # logger.info(f"config: {text_model.config}") + cfg = CLIPTextConfig( + vocab_size=49408, + hidden_size=768, + intermediate_size=3072, + num_hidden_layers=12, + num_attention_heads=12, + max_position_embeddings=77, + hidden_act="quick_gelu", + layer_norm_eps=1e-05, + dropout=0.0, + attention_dropout=0.0, + initializer_range=0.02, + initializer_factor=1.0, + pad_token_id=1, + bos_token_id=0, + eos_token_id=2, + model_type="clip_text_model", + projection_dim=768, + torch_dtype="float32", + ) + text_model = CLIPTextModel._from_config(cfg) + info = text_model.load_state_dict(converted_text_encoder_checkpoint) + logger.info(f"loading text encoder: {info}") + + return text_model, vae, unet + + +def get_model_version_str_for_sd1_sd2(v2, v_parameterization): + # only for reference + version_str = "sd" + if v2: + version_str += "_v2" + else: + version_str += "_v1" + if v_parameterization: + version_str += "_v" + return version_str + + +def convert_text_encoder_state_dict_to_sd_v2(checkpoint, make_dummy_weights=False): + def convert_key(key): + # position_idsの除去 + if ".position_ids" in key: + return None + + # common + key = key.replace("text_model.encoder.", "transformer.") + key = key.replace("text_model.", "") + if "layers" in key: + # resblocks conversion + key = key.replace(".layers.", ".resblocks.") + if ".layer_norm" in key: + key = key.replace(".layer_norm", ".ln_") + elif ".mlp." in key: + key = key.replace(".fc1.", ".c_fc.") + key = key.replace(".fc2.", ".c_proj.") + elif ".self_attn.out_proj" in key: + key = key.replace(".self_attn.out_proj.", ".attn.out_proj.") + elif ".self_attn." in key: + key = None # 特殊なので後で処理する + else: + raise ValueError(f"unexpected key in DiffUsers model: {key}") + elif ".position_embedding" in key: + key = key.replace("embeddings.position_embedding.weight", "positional_embedding") + elif ".token_embedding" in key: + key = key.replace("embeddings.token_embedding.weight", "token_embedding.weight") + elif "final_layer_norm" in key: + key = key.replace("final_layer_norm", "ln_final") + return key + + keys = list(checkpoint.keys()) + new_sd = {} + for key in keys: + new_key = convert_key(key) + if new_key is None: + continue + new_sd[new_key] = checkpoint[key] + + # attnの変換 + for key in keys: + if "layers" in key and "q_proj" in key: + # 三つを結合 + key_q = key + key_k = key.replace("q_proj", "k_proj") + key_v = key.replace("q_proj", "v_proj") + + value_q = checkpoint[key_q] + value_k = checkpoint[key_k] + value_v = checkpoint[key_v] + value = torch.cat([value_q, value_k, value_v]) + + new_key = key.replace("text_model.encoder.layers.", "transformer.resblocks.") + new_key = new_key.replace(".self_attn.q_proj.", ".attn.in_proj_") + new_sd[new_key] = value + + # 最後の層などを捏造するか + if make_dummy_weights: + logger.info("make dummy weights for resblock.23, text_projection and logit scale.") + keys = list(new_sd.keys()) + for key in keys: + if key.startswith("transformer.resblocks.22."): + new_sd[key.replace(".22.", ".23.")] = new_sd[key].clone() # copyしないとsafetensorsの保存で落ちる + + # Diffusersに含まれない重みを作っておく + new_sd["text_projection"] = torch.ones((1024, 1024), dtype=new_sd[keys[0]].dtype, device=new_sd[keys[0]].device) + new_sd["logit_scale"] = torch.tensor(1) + + return new_sd + + +def save_stable_diffusion_checkpoint( + v2, output_file, text_encoder, unet, ckpt_path, epochs, steps, metadata, save_dtype=None, vae=None +): + if ckpt_path is not None: + # epoch/stepを参照する。またVAEがメモリ上にないときなど、もう一度VAEを含めて読み込む + checkpoint, state_dict = load_checkpoint_with_text_encoder_conversion(ckpt_path) + if checkpoint is None: # safetensors または state_dictのckpt + checkpoint = {} + strict = False + else: + strict = True + if "state_dict" in state_dict: + del state_dict["state_dict"] + else: + # 新しく作る + assert vae is not None, "VAE is required to save a checkpoint without a given checkpoint" + checkpoint = {} + state_dict = {} + strict = False + + def update_sd(prefix, sd): + for k, v in sd.items(): + key = prefix + k + assert not strict or key in state_dict, f"Illegal key in save SD: {key}" + if save_dtype is not None: + v = v.detach().clone().to("cpu").to(save_dtype) + state_dict[key] = v + + # Convert the UNet model + unet_state_dict = convert_unet_state_dict_to_sd(v2, unet.state_dict()) + update_sd("model.diffusion_model.", unet_state_dict) + + # Convert the text encoder model + if v2: + make_dummy = ckpt_path is None # 参照元のcheckpointがない場合は最後の層を前の層から複製して作るなどダミーの重みを入れる + text_enc_dict = convert_text_encoder_state_dict_to_sd_v2(text_encoder.state_dict(), make_dummy) + update_sd("cond_stage_model.model.", text_enc_dict) + else: + text_enc_dict = text_encoder.state_dict() + update_sd("cond_stage_model.transformer.", text_enc_dict) + + # Convert the VAE + if vae is not None: + vae_dict = convert_vae_state_dict(vae.state_dict()) + update_sd("first_stage_model.", vae_dict) + + # Put together new checkpoint + key_count = len(state_dict.keys()) + new_ckpt = {"state_dict": state_dict} + + # epoch and global_step are sometimes not int + try: + if "epoch" in checkpoint: + epochs += checkpoint["epoch"] + if "global_step" in checkpoint: + steps += checkpoint["global_step"] + except: + pass + + new_ckpt["epoch"] = epochs + new_ckpt["global_step"] = steps + + if is_safetensors(output_file): + # TODO Tensor以外のdictの値を削除したほうがいいか + save_file(state_dict, output_file, metadata) + else: + torch.save(new_ckpt, output_file) + + return key_count + + +def save_diffusers_checkpoint(v2, output_dir, text_encoder, unet, pretrained_model_name_or_path, vae=None, use_safetensors=False): + if pretrained_model_name_or_path is None: + # load default settings for v1/v2 + if v2: + pretrained_model_name_or_path = DIFFUSERS_REF_MODEL_ID_V2 + else: + pretrained_model_name_or_path = DIFFUSERS_REF_MODEL_ID_V1 + + scheduler = DDIMScheduler.from_pretrained(pretrained_model_name_or_path, subfolder="scheduler") + tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer") + if vae is None: + vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae") + + # original U-Net cannot be saved, so we need to convert it to the Diffusers version + # TODO this consumes a lot of memory + diffusers_unet = diffusers.UNet2DConditionModel.from_pretrained(pretrained_model_name_or_path, subfolder="unet") + diffusers_unet.load_state_dict(unet.state_dict()) + + pipeline = StableDiffusionPipeline( + unet=diffusers_unet, + text_encoder=text_encoder, + vae=vae, + scheduler=scheduler, + tokenizer=tokenizer, + safety_checker=None, + feature_extractor=None, + requires_safety_checker=None, + ) + pipeline.save_pretrained(output_dir, safe_serialization=use_safetensors) + + +VAE_PREFIX = "first_stage_model." + + +def load_vae(vae_id, dtype): + logger.info(f"load VAE: {vae_id}") + if os.path.isdir(vae_id) or not os.path.isfile(vae_id): + # Diffusers local/remote + try: + vae = AutoencoderKL.from_pretrained(vae_id, subfolder=None, torch_dtype=dtype) + except EnvironmentError as e: + logger.error(f"exception occurs in loading vae: {e}") + logger.error("retry with subfolder='vae'") + vae = AutoencoderKL.from_pretrained(vae_id, subfolder="vae", torch_dtype=dtype) + return vae + + # local + vae_config = create_vae_diffusers_config() + + if vae_id.endswith(".bin"): + # SD 1.5 VAE on Huggingface + converted_vae_checkpoint = torch.load(vae_id, map_location="cpu") + else: + # StableDiffusion + vae_model = load_file(vae_id, "cpu") if is_safetensors(vae_id) else torch.load(vae_id, map_location="cpu") + vae_sd = vae_model["state_dict"] if "state_dict" in vae_model else vae_model + + # vae only or full model + full_model = False + for vae_key in vae_sd: + if vae_key.startswith(VAE_PREFIX): + full_model = True + break + if not full_model: + sd = {} + for key, value in vae_sd.items(): + sd[VAE_PREFIX + key] = value + vae_sd = sd + del sd + + # Convert the VAE model. + converted_vae_checkpoint = convert_ldm_vae_checkpoint(vae_sd, vae_config) + + vae = AutoencoderKL(**vae_config) + vae.load_state_dict(converted_vae_checkpoint) + return vae + + +# endregion + + +def make_bucket_resolutions(max_reso, min_size=256, max_size=1024, divisible=64): + max_width, max_height = max_reso + max_area = max_width * max_height + + resos = set() + + width = int(math.sqrt(max_area) // divisible) * divisible + resos.add((width, width)) + + width = min_size + while width <= max_size: + height = min(max_size, int((max_area // width) // divisible) * divisible) + if height >= min_size: + resos.add((width, height)) + resos.add((height, width)) + + # # make additional resos + # if width >= height and width - divisible >= min_size: + # resos.add((width - divisible, height)) + # resos.add((height, width - divisible)) + # if height >= width and height - divisible >= min_size: + # resos.add((width, height - divisible)) + # resos.add((height - divisible, width)) + + width += divisible + + resos = list(resos) + resos.sort() + return resos + + +if __name__ == "__main__": + resos = make_bucket_resolutions((512, 768)) + logger.info(f"{len(resos)}") + logger.info(f"{resos}") + aspect_ratios = [w / h for w, h in resos] + logger.info(f"{aspect_ratios}") + + ars = set() + for ar in aspect_ratios: + if ar in ars: + logger.error(f"error! duplicate ar: {ar}") + ars.add(ar) diff --git a/oft.py b/oft.py new file mode 100644 index 0000000000000000000000000000000000000000..6321def3b5800e378d969b43065fe608437588bc --- /dev/null +++ b/oft.py @@ -0,0 +1,459 @@ +# OFT network module + +import math +import os +from typing import Dict, List, Optional, Tuple, Type, Union +from diffusers import AutoencoderKL +import einops +from transformers import CLIPTextModel +import numpy as np +import torch +import torch.nn.functional as F +import re +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +RE_UPDOWN = re.compile(r"(up|down)_blocks_(\d+)_(resnets|upsamplers|downsamplers|attentions)_(\d+)_") + + +class OFTModule(torch.nn.Module): + """ + replaces forward method of the original Linear, instead of replacing the original Linear module. + """ + + def __init__( + self, + oft_name, + org_module: torch.nn.Module, + multiplier=1.0, + dim=4, + alpha=1, + ): + """ + dim -> num blocks + alpha -> constraint + """ + super().__init__() + self.oft_name = oft_name + + self.num_blocks = dim + + if "Linear" in org_module.__class__.__name__: + out_dim = org_module.out_features + elif "Conv" in org_module.__class__.__name__: + out_dim = org_module.out_channels + + if type(alpha) == torch.Tensor: + alpha = alpha.detach().numpy() + + # constraint in original paper is alpha * out_dim * out_dim, but we use alpha * out_dim for backward compatibility + # original alpha is 1e-6, so we use 1e-3 or 1e-4 for alpha + self.constraint = alpha * out_dim + + self.register_buffer("alpha", torch.tensor(alpha)) + + self.block_size = out_dim // self.num_blocks + self.oft_blocks = torch.nn.Parameter(torch.zeros(self.num_blocks, self.block_size, self.block_size)) + self.I = torch.eye(self.block_size).unsqueeze(0).repeat(self.num_blocks, 1, 1) # cpu + + self.out_dim = out_dim + self.shape = org_module.weight.shape + + self.multiplier = multiplier + self.org_module = [org_module] # moduleにならないようにlistに入れる + + def apply_to(self): + self.org_forward = self.org_module[0].forward + self.org_module[0].forward = self.forward + + def get_weight(self, multiplier=None): + if multiplier is None: + multiplier = self.multiplier + + block_Q = self.oft_blocks - self.oft_blocks.transpose(1, 2) + norm_Q = torch.norm(block_Q.flatten()) + new_norm_Q = torch.clamp(norm_Q, max=self.constraint) + block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) + + if self.I.device != block_Q.device: + self.I = self.I.to(block_Q.device) + I = self.I + block_R = torch.matmul(I + block_Q, (I - block_Q).float().inverse()) + block_R_weighted = self.multiplier * (block_R - I) + I + return block_R_weighted + + def forward(self, x, scale=None): + if self.multiplier == 0.0: + return self.org_forward(x) + org_module = self.org_module[0] + org_dtype = x.dtype + + R = self.get_weight().to(torch.float32) + W = org_module.weight.to(torch.float32) + + if len(W.shape) == 4: # Conv2d + W_reshaped = einops.rearrange(W, "(k n) ... -> k n ...", k=self.num_blocks, n=self.block_size) + RW = torch.einsum("k n m, k n ... -> k m ...", R, W_reshaped) + RW = einops.rearrange(RW, "k m ... -> (k m) ...") + result = F.conv2d( + x, RW.to(org_dtype), org_module.bias, org_module.stride, org_module.padding, org_module.dilation, org_module.groups + ) + else: # Linear + W_reshaped = einops.rearrange(W, "(k n) m -> k n m", k=self.num_blocks, n=self.block_size) + RW = torch.einsum("k n m, k n p -> k m p", R, W_reshaped) + RW = einops.rearrange(RW, "k m p -> (k m) p") + result = F.linear(x, RW.to(org_dtype), org_module.bias) + return result + + +class OFTInfModule(OFTModule): + def __init__( + self, + oft_name, + org_module: torch.nn.Module, + multiplier=1.0, + dim=4, + alpha=1, + **kwargs, + ): + # no dropout for inference + super().__init__(oft_name, org_module, multiplier, dim, alpha) + self.enabled = True + self.network: OFTNetwork = None + + def set_network(self, network): + self.network = network + + def forward(self, x, scale=None): + if not self.enabled: + return self.org_forward(x) + return super().forward(x, scale) + + def merge_to(self, multiplier=None): + # get org weight + org_sd = self.org_module[0].state_dict() + org_weight = org_sd["weight"].to(torch.float32) + + R = self.get_weight(multiplier).to(torch.float32) + + weight = org_weight.reshape(self.num_blocks, self.block_size, -1) + weight = torch.einsum("k n m, k n ... -> k m ...", R, weight) + weight = weight.reshape(org_weight.shape) + + # convert back to original dtype + weight = weight.to(org_sd["weight"].dtype) + + # set weight to org_module + org_sd["weight"] = weight + self.org_module[0].load_state_dict(org_sd) + + +def create_network( + multiplier: float, + network_dim: Optional[int], + network_alpha: Optional[float], + vae: AutoencoderKL, + text_encoder: Union[CLIPTextModel, List[CLIPTextModel]], + unet, + neuron_dropout: Optional[float] = None, + **kwargs, +): + if network_dim is None: + network_dim = 4 # default + if network_alpha is None: # should be set + logger.info( + "network_alpha is not set, use default value 1e-3 / network_alphaが設定されていないのでデフォルト値 1e-3 を使用します" + ) + network_alpha = 1e-3 + elif network_alpha >= 1: + logger.warning( + "network_alpha is too large (>=1, maybe default value is too large), please consider to set smaller value like 1e-3" + " / network_alphaが大きすぎるようです(>=1, デフォルト値が大きすぎる可能性があります)。1e-3のような小さな値を推奨" + ) + + enable_all_linear = kwargs.get("enable_all_linear", None) + enable_conv = kwargs.get("enable_conv", None) + if enable_all_linear is not None: + enable_all_linear = bool(enable_all_linear) + if enable_conv is not None: + enable_conv = bool(enable_conv) + + network = OFTNetwork( + text_encoder, + unet, + multiplier=multiplier, + dim=network_dim, + alpha=network_alpha, + enable_all_linear=enable_all_linear, + enable_conv=enable_conv, + varbose=True, + ) + return network + + +# Create network from weights for inference, weights are not loaded here (because can be merged) +def create_network_from_weights(multiplier, file, vae, text_encoder, unet, weights_sd=None, for_inference=False, **kwargs): + if weights_sd is None: + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file, safe_open + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + # check dim, alpha and if weights have for conv2d + dim = None + alpha = None + has_conv2d = None + all_linear = None + for name, param in weights_sd.items(): + if name.endswith(".alpha"): + if alpha is None: + alpha = param.item() + else: + if dim is None: + dim = param.size()[0] + if has_conv2d is None and "in_layers_2" in name: + has_conv2d = True + if all_linear is None and "_ff_" in name: + all_linear = True + if dim is not None and alpha is not None and has_conv2d is not None and all_linear is not None: + break + if has_conv2d is None: + has_conv2d = False + if all_linear is None: + all_linear = False + + module_class = OFTInfModule if for_inference else OFTModule + network = OFTNetwork( + text_encoder, + unet, + multiplier=multiplier, + dim=dim, + alpha=alpha, + enable_all_linear=all_linear, + enable_conv=has_conv2d, + module_class=module_class, + ) + return network, weights_sd + + +class OFTNetwork(torch.nn.Module): + UNET_TARGET_REPLACE_MODULE_ATTN_ONLY = ["CrossAttention"] + UNET_TARGET_REPLACE_MODULE_ALL_LINEAR = ["Transformer2DModel"] + UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"] + OFT_PREFIX_UNET = "oft_unet" # これ変えないほうがいいかな + + def __init__( + self, + text_encoder: Union[List[CLIPTextModel], CLIPTextModel], + unet, + multiplier: float = 1.0, + dim: int = 4, + alpha: float = 1, + enable_all_linear: Optional[bool] = False, + enable_conv: Optional[bool] = False, + module_class: Type[object] = OFTModule, + varbose: Optional[bool] = False, + ) -> None: + super().__init__() + self.multiplier = multiplier + + self.dim = dim + self.alpha = alpha + + logger.info( + f"create OFT network. num blocks: {self.dim}, constraint: {self.alpha}, multiplier: {self.multiplier}, enable_conv: {enable_conv}, enable_all_linear: {enable_all_linear}" + ) + + # create module instances + def create_modules( + root_module: torch.nn.Module, + target_replace_modules: List[torch.nn.Module], + ) -> List[OFTModule]: + prefix = self.OFT_PREFIX_UNET + ofts = [] + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + is_linear = "Linear" in child_module.__class__.__name__ + is_conv2d = "Conv2d" in child_module.__class__.__name__ + is_conv2d_1x1 = is_conv2d and child_module.kernel_size == (1, 1) + + if is_linear or is_conv2d_1x1 or (is_conv2d and enable_conv): + oft_name = prefix + "." + name + "." + child_name + oft_name = oft_name.replace(".", "_") + # logger.info(oft_name) + + oft = module_class( + oft_name, + child_module, + self.multiplier, + dim, + alpha, + ) + ofts.append(oft) + return ofts + + # extend U-Net target modules if conv2d 3x3 is enabled, or load from weights + if enable_all_linear: + target_modules = OFTNetwork.UNET_TARGET_REPLACE_MODULE_ALL_LINEAR + else: + target_modules = OFTNetwork.UNET_TARGET_REPLACE_MODULE_ATTN_ONLY + if enable_conv: + target_modules += OFTNetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + + self.unet_ofts: List[OFTModule] = create_modules(unet, target_modules) + logger.info(f"create OFT for U-Net: {len(self.unet_ofts)} modules.") + + # assertion + names = set() + for oft in self.unet_ofts: + assert oft.oft_name not in names, f"duplicated oft name: {oft.oft_name}" + names.add(oft.oft_name) + + def set_multiplier(self, multiplier): + self.multiplier = multiplier + for oft in self.unet_ofts: + oft.multiplier = self.multiplier + + def load_weights(self, file): + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import load_file + + weights_sd = load_file(file) + else: + weights_sd = torch.load(file, map_location="cpu") + + info = self.load_state_dict(weights_sd, False) + return info + + def apply_to(self, text_encoder, unet, apply_text_encoder=True, apply_unet=True): + assert apply_unet, "apply_unet must be True" + + for oft in self.unet_ofts: + oft.apply_to() + self.add_module(oft.oft_name, oft) + + # マージできるかどうかを返す + def is_mergeable(self): + return True + + # TODO refactor to common function with apply_to + def merge_to(self, text_encoder, unet, weights_sd, dtype, device): + logger.info("enable OFT for U-Net") + + for oft in self.unet_ofts: + sd_for_lora = {} + for key in weights_sd.keys(): + if key.startswith(oft.oft_name): + sd_for_lora[key[len(oft.oft_name) + 1 :]] = weights_sd[key] + oft.load_state_dict(sd_for_lora, False) + oft.merge_to() + + logger.info(f"weights are merged") + + # 二つのText Encoderに別々の学習率を設定できるようにするといいかも + def prepare_optimizer_params(self, text_encoder_lr, unet_lr, default_lr): + self.requires_grad_(True) + all_params = [] + + def enumerate_params(ofts): + params = [] + for oft in ofts: + params.extend(oft.parameters()) + + # logger.info num of params + num_params = 0 + for p in params: + num_params += p.numel() + logger.info(f"OFT params: {num_params}") + return params + + param_data = {"params": enumerate_params(self.unet_ofts)} + if unet_lr is not None: + param_data["lr"] = unet_lr + all_params.append(param_data) + + return all_params + + def enable_gradient_checkpointing(self): + # not supported + pass + + def prepare_grad_etc(self, text_encoder, unet): + self.requires_grad_(True) + + def on_epoch_start(self, text_encoder, unet): + self.train() + + def get_trainable_params(self): + return self.parameters() + + def save_weights(self, file, dtype, metadata): + if metadata is not None and len(metadata) == 0: + metadata = None + + state_dict = self.state_dict() + + if dtype is not None: + for key in list(state_dict.keys()): + v = state_dict[key] + v = v.detach().clone().to("cpu").to(dtype) + state_dict[key] = v + + if os.path.splitext(file)[1] == ".safetensors": + from safetensors.torch import save_file + from library import train_util + + # Precalculate model hashes to save time on indexing + if metadata is None: + metadata = {} + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + save_file(state_dict, file, metadata) + else: + torch.save(state_dict, file) + + def backup_weights(self): + # 重みのバックアップを行う + ofts: List[OFTInfModule] = self.unet_ofts + for oft in ofts: + org_module = oft.org_module[0] + if not hasattr(org_module, "_lora_org_weight"): + sd = org_module.state_dict() + org_module._lora_org_weight = sd["weight"].detach().clone() + org_module._lora_restored = True + + def restore_weights(self): + # 重みのリストアを行う + ofts: List[OFTInfModule] = self.unet_ofts + for oft in ofts: + org_module = oft.org_module[0] + if not org_module._lora_restored: + sd = org_module.state_dict() + sd["weight"] = org_module._lora_org_weight + org_module.load_state_dict(sd) + org_module._lora_restored = True + + def pre_calculation(self): + # 事前計算を行う + ofts: List[OFTInfModule] = self.unet_ofts + for oft in ofts: + org_module = oft.org_module[0] + oft.merge_to() + # sd = org_module.state_dict() + # org_weight = sd["weight"] + # lora_weight = oft.get_weight().to(org_weight.device, dtype=org_weight.dtype) + # sd["weight"] = org_weight + lora_weight + # assert sd["weight"].shape == org_weight.shape + # org_module.load_state_dict(sd) + + org_module._lora_restored = False + oft.enabled = False diff --git a/original_control_net.py b/original_control_net.py new file mode 100644 index 0000000000000000000000000000000000000000..5640d542d9e0b5bf6a519ff00003a376c947f221 --- /dev/null +++ b/original_control_net.py @@ -0,0 +1,353 @@ +from typing import List, NamedTuple, Any +import numpy as np +import cv2 +import torch +from safetensors.torch import load_file + +from library.original_unet import UNet2DConditionModel, SampleOutput + +import library.model_util as model_util +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +class ControlNetInfo(NamedTuple): + unet: Any + net: Any + prep: Any + weight: float + ratio: float + + +class ControlNet(torch.nn.Module): + def __init__(self) -> None: + super().__init__() + + # make control model + self.control_model = torch.nn.Module() + + dims = [320, 320, 320, 320, 640, 640, 640, 1280, 1280, 1280, 1280, 1280] + zero_convs = torch.nn.ModuleList() + for i, dim in enumerate(dims): + sub_list = torch.nn.ModuleList([torch.nn.Conv2d(dim, dim, 1)]) + zero_convs.append(sub_list) + self.control_model.add_module("zero_convs", zero_convs) + + middle_block_out = torch.nn.Conv2d(1280, 1280, 1) + self.control_model.add_module("middle_block_out", torch.nn.ModuleList([middle_block_out])) + + dims = [16, 16, 32, 32, 96, 96, 256, 320] + strides = [1, 1, 2, 1, 2, 1, 2, 1] + prev_dim = 3 + input_hint_block = torch.nn.Sequential() + for i, (dim, stride) in enumerate(zip(dims, strides)): + input_hint_block.append(torch.nn.Conv2d(prev_dim, dim, 3, stride, 1)) + if i < len(dims) - 1: + input_hint_block.append(torch.nn.SiLU()) + prev_dim = dim + self.control_model.add_module("input_hint_block", input_hint_block) + + +def load_control_net(v2, unet, model): + device = unet.device + + # control sdからキー変換しつつU-Netに対応する部分のみ取り出し、DiffusersのU-Netに読み込む + # state dictを読み込む + logger.info(f"ControlNet: loading control SD model : {model}") + + if model_util.is_safetensors(model): + ctrl_sd_sd = load_file(model) + else: + ctrl_sd_sd = torch.load(model, map_location="cpu") + ctrl_sd_sd = ctrl_sd_sd.pop("state_dict", ctrl_sd_sd) + + # 重みをU-Netに読み込めるようにする。ControlNetはSD版のstate dictなので、それを読み込む + is_difference = "difference" in ctrl_sd_sd + logger.info(f"ControlNet: loading difference: {is_difference}") + + # ControlNetには存在しないキーがあるので、まず現在のU-NetでSD版の全keyを作っておく + # またTransfer Controlの元weightとなる + ctrl_unet_sd_sd = model_util.convert_unet_state_dict_to_sd(v2, unet.state_dict()) + + # 元のU-Netに影響しないようにコピーする。またprefixが付いていないので付ける + for key in list(ctrl_unet_sd_sd.keys()): + ctrl_unet_sd_sd["model.diffusion_model." + key] = ctrl_unet_sd_sd.pop(key).clone() + + zero_conv_sd = {} + for key in list(ctrl_sd_sd.keys()): + if key.startswith("control_"): + unet_key = "model.diffusion_" + key[len("control_") :] + if unet_key not in ctrl_unet_sd_sd: # zero conv + zero_conv_sd[key] = ctrl_sd_sd[key] + continue + if is_difference: # Transfer Control + ctrl_unet_sd_sd[unet_key] += ctrl_sd_sd[key].to(device, dtype=unet.dtype) + else: + ctrl_unet_sd_sd[unet_key] = ctrl_sd_sd[key].to(device, dtype=unet.dtype) + + unet_config = model_util.create_unet_diffusers_config(v2) + ctrl_unet_du_sd = model_util.convert_ldm_unet_checkpoint(v2, ctrl_unet_sd_sd, unet_config) # DiffUsers版ControlNetのstate dict + + # ControlNetのU-Netを作成する + ctrl_unet = UNet2DConditionModel(**unet_config) + info = ctrl_unet.load_state_dict(ctrl_unet_du_sd) + logger.info(f"ControlNet: loading Control U-Net: {info}") + + # U-Net以外のControlNetを作成する + # TODO support middle only + ctrl_net = ControlNet() + info = ctrl_net.load_state_dict(zero_conv_sd) + logger.info("ControlNet: loading ControlNet: {info}") + + ctrl_unet.to(unet.device, dtype=unet.dtype) + ctrl_net.to(unet.device, dtype=unet.dtype) + return ctrl_unet, ctrl_net + + +def load_preprocess(prep_type: str): + if prep_type is None or prep_type.lower() == "none": + return None + + if prep_type.startswith("canny"): + args = prep_type.split("_") + th1 = int(args[1]) if len(args) >= 2 else 63 + th2 = int(args[2]) if len(args) >= 3 else 191 + + def canny(img): + img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY) + return cv2.Canny(img, th1, th2) + + return canny + + logger.info(f"Unsupported prep type: {prep_type}") + return None + + +def preprocess_ctrl_net_hint_image(image): + image = np.array(image).astype(np.float32) / 255.0 + # ControlNetのサンプルはcv2を使っているが、読み込みはGradioなので実はRGBになっている + # image = image[:, :, ::-1].copy() # rgb to bgr + image = image[None].transpose(0, 3, 1, 2) # nchw + image = torch.from_numpy(image) + return image # 0 to 1 + + +def get_guided_hints(control_nets: List[ControlNetInfo], num_latent_input, b_size, hints): + guided_hints = [] + for i, cnet_info in enumerate(control_nets): + # hintは 1枚目の画像のcnet1, 1枚目の画像のcnet2, 1枚目の画像のcnet3, 2枚目の画像のcnet1, 2枚目の画像のcnet2 ... と並んでいること + b_hints = [] + if len(hints) == 1: # すべて同じ画像をhintとして使う + hint = hints[0] + if cnet_info.prep is not None: + hint = cnet_info.prep(hint) + hint = preprocess_ctrl_net_hint_image(hint) + b_hints = [hint for _ in range(b_size)] + else: + for bi in range(b_size): + hint = hints[(bi * len(control_nets) + i) % len(hints)] + if cnet_info.prep is not None: + hint = cnet_info.prep(hint) + hint = preprocess_ctrl_net_hint_image(hint) + b_hints.append(hint) + b_hints = torch.cat(b_hints, dim=0) + b_hints = b_hints.to(cnet_info.unet.device, dtype=cnet_info.unet.dtype) + + guided_hint = cnet_info.net.control_model.input_hint_block(b_hints) + guided_hints.append(guided_hint) + return guided_hints + + +def call_unet_and_control_net( + step, + num_latent_input, + original_unet, + control_nets: List[ControlNetInfo], + guided_hints, + current_ratio, + sample, + timestep, + encoder_hidden_states, + encoder_hidden_states_for_control_net, +): + # ControlNet + # 複数のControlNetの場合は、出力をマージするのではなく交互に適用する + cnet_cnt = len(control_nets) + cnet_idx = step % cnet_cnt + cnet_info = control_nets[cnet_idx] + + # logger.info(current_ratio, cnet_info.prep, cnet_info.weight, cnet_info.ratio) + if cnet_info.ratio < current_ratio: + return original_unet(sample, timestep, encoder_hidden_states) + + guided_hint = guided_hints[cnet_idx] + + # gradual latent support: match the size of guided_hint to the size of sample + if guided_hint.shape[-2:] != sample.shape[-2:]: + # print(f"guided_hint.shape={guided_hint.shape}, sample.shape={sample.shape}") + org_dtype = guided_hint.dtype + if org_dtype == torch.bfloat16: + guided_hint = guided_hint.to(torch.float32) + guided_hint = torch.nn.functional.interpolate(guided_hint, size=sample.shape[-2:], mode="bicubic") + if org_dtype == torch.bfloat16: + guided_hint = guided_hint.to(org_dtype) + + guided_hint = guided_hint.repeat((num_latent_input, 1, 1, 1)) + outs = unet_forward( + True, cnet_info.net, cnet_info.unet, guided_hint, None, sample, timestep, encoder_hidden_states_for_control_net + ) + outs = [o * cnet_info.weight for o in outs] + + # U-Net + return unet_forward(False, cnet_info.net, original_unet, None, outs, sample, timestep, encoder_hidden_states) + + +""" + # これはmergeのバージョン + # ControlNet + cnet_outs_list = [] + for i, cnet_info in enumerate(control_nets): + # logger.info(current_ratio, cnet_info.prep, cnet_info.weight, cnet_info.ratio) + if cnet_info.ratio < current_ratio: + continue + guided_hint = guided_hints[i] + outs = unet_forward(True, cnet_info.net, cnet_info.unet, guided_hint, None, sample, timestep, encoder_hidden_states) + for i in range(len(outs)): + outs[i] *= cnet_info.weight + + cnet_outs_list.append(outs) + + count = len(cnet_outs_list) + if count == 0: + return original_unet(sample, timestep, encoder_hidden_states) + + # sum of controlnets + for i in range(1, count): + cnet_outs_list[0] += cnet_outs_list[i] + + # U-Net + return unet_forward(False, cnet_info.net, original_unet, None, cnet_outs_list[0], sample, timestep, encoder_hidden_states) +""" + + +def unet_forward( + is_control_net, + control_net: ControlNet, + unet: UNet2DConditionModel, + guided_hint, + ctrl_outs, + sample, + timestep, + encoder_hidden_states, +): + # copy from UNet2DConditionModel + default_overall_up_factor = 2**unet.num_upsamplers + + forward_upsample_size = False + upsample_size = None + + if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]): + logger.info("Forward upsample size to force interpolation output size.") + forward_upsample_size = True + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = unet.time_proj(timesteps) + + # timesteps does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=unet.dtype) + emb = unet.time_embedding(t_emb) + + outs = [] # output of ControlNet + zc_idx = 0 + + # 2. pre-process + sample = unet.conv_in(sample) + if is_control_net: + sample += guided_hint + outs.append(control_net.control_model.zero_convs[zc_idx][0](sample)) # , emb, encoder_hidden_states)) + zc_idx += 1 + + # 3. down + down_block_res_samples = (sample,) + for downsample_block in unet.down_blocks: + if downsample_block.has_cross_attention: + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + if is_control_net: + for rs in res_samples: + outs.append(control_net.control_model.zero_convs[zc_idx][0](rs)) # , emb, encoder_hidden_states)) + zc_idx += 1 + + down_block_res_samples += res_samples + + # 4. mid + sample = unet.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states) + if is_control_net: + outs.append(control_net.control_model.middle_block_out[0](sample)) + return outs + + if not is_control_net: + sample += ctrl_outs.pop() + + # 5. up + for i, upsample_block in enumerate(unet.up_blocks): + is_final_block = i == len(unet.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + if not is_control_net and len(ctrl_outs) > 0: + res_samples = list(res_samples) + apply_ctrl_outs = ctrl_outs[-len(res_samples) :] + ctrl_outs = ctrl_outs[: -len(res_samples)] + for j in range(len(res_samples)): + res_samples[j] = res_samples[j] + apply_ctrl_outs[j] + res_samples = tuple(res_samples) + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + upsample_size=upsample_size, + ) + else: + sample = upsample_block( + hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size + ) + # 6. post-process + sample = unet.conv_norm_out(sample) + sample = unet.conv_act(sample) + sample = unet.conv_out(sample) + + return SampleOutput(sample=sample) diff --git a/original_unet.py b/original_unet.py new file mode 100644 index 0000000000000000000000000000000000000000..e944ff22b3bfdd45d164d677e04ceba0c2440d04 --- /dev/null +++ b/original_unet.py @@ -0,0 +1,1919 @@ +# Diffusers 0.10.2からStable Diffusionに必要な部分だけを持ってくる +# 条件分岐等で不要な部分は削除している +# コードの多くはDiffusersからコピーしている +# 制約として、モデルのstate_dictがDiffusers 0.10.2のものと同じ形式である必要がある + +# Copy from Diffusers 0.10.2 for Stable Diffusion. Most of the code is copied from Diffusers. +# Unnecessary parts are deleted by condition branching. +# As a constraint, the state_dict of the model must be in the same format as that of Diffusers 0.10.2 + +""" +v1.5とv2.1の相違点は +- attention_head_dimがintかlist[int]か +- cross_attention_dimが768か1024か +- use_linear_projection: trueがない（=False, 1.5）かあるか +- upcast_attentionがFalse(1.5)かTrue(2.1)か +- （以下は多分無視していい） +- sample_sizeが64か96か +- dual_cross_attentionがあるかないか +- num_class_embedsがあるかないか +- only_cross_attentionがあるかないか + +v1.5 +{ + "_class_name": "UNet2DConditionModel", + "_diffusers_version": "0.6.0", + "act_fn": "silu", + "attention_head_dim": 8, + "block_out_channels": [ + 320, + 640, + 1280, + 1280 + ], + "center_input_sample": false, + "cross_attention_dim": 768, + "down_block_types": [ + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D" + ], + "downsample_padding": 1, + "flip_sin_to_cos": true, + "freq_shift": 0, + "in_channels": 4, + "layers_per_block": 2, + "mid_block_scale_factor": 1, + "norm_eps": 1e-05, + "norm_num_groups": 32, + "out_channels": 4, + "sample_size": 64, + "up_block_types": [ + "UpBlock2D", + "CrossAttnUpBlock2D", + "CrossAttnUpBlock2D", + "CrossAttnUpBlock2D" + ] +} + +v2.1 +{ + "_class_name": "UNet2DConditionModel", + "_diffusers_version": "0.10.0.dev0", + "act_fn": "silu", + "attention_head_dim": [ + 5, + 10, + 20, + 20 + ], + "block_out_channels": [ + 320, + 640, + 1280, + 1280 + ], + "center_input_sample": false, + "cross_attention_dim": 1024, + "down_block_types": [ + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D" + ], + "downsample_padding": 1, + "dual_cross_attention": false, + "flip_sin_to_cos": true, + "freq_shift": 0, + "in_channels": 4, + "layers_per_block": 2, + "mid_block_scale_factor": 1, + "norm_eps": 1e-05, + "norm_num_groups": 32, + "num_class_embeds": null, + "only_cross_attention": false, + "out_channels": 4, + "sample_size": 96, + "up_block_types": [ + "UpBlock2D", + "CrossAttnUpBlock2D", + "CrossAttnUpBlock2D", + "CrossAttnUpBlock2D" + ], + "use_linear_projection": true, + "upcast_attention": true +} +""" + +import math +from types import SimpleNamespace +from typing import Dict, Optional, Tuple, Union +import torch +from torch import nn +from torch.nn import functional as F +from einops import rearrange +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +BLOCK_OUT_CHANNELS: Tuple[int] = (320, 640, 1280, 1280) +TIMESTEP_INPUT_DIM = BLOCK_OUT_CHANNELS[0] +TIME_EMBED_DIM = BLOCK_OUT_CHANNELS[0] * 4 +IN_CHANNELS: int = 4 +OUT_CHANNELS: int = 4 +LAYERS_PER_BLOCK: int = 2 +LAYERS_PER_BLOCK_UP: int = LAYERS_PER_BLOCK + 1 +TIME_EMBED_FLIP_SIN_TO_COS: bool = True +TIME_EMBED_FREQ_SHIFT: int = 0 +NORM_GROUPS: int = 32 +NORM_EPS: float = 1e-5 +TRANSFORMER_NORM_NUM_GROUPS = 32 + +DOWN_BLOCK_TYPES = ["CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D"] +UP_BLOCK_TYPES = ["UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"] + + +# region memory efficient attention + +# FlashAttentionを使うCrossAttention +# based on https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/memory_efficient_attention_pytorch/flash_attention.py +# LICENSE MIT https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/LICENSE + +# constants + +EPSILON = 1e-6 + +# helper functions + + +def exists(val): + return val is not None + + +def default(val, d): + return val if exists(val) else d + + +# flash attention forwards and backwards + +# https://arxiv.org/abs/2205.14135 + + +class FlashAttentionFunction(torch.autograd.Function): + @staticmethod + @torch.no_grad() + def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size): + """Algorithm 2 in the paper""" + + device = q.device + dtype = q.dtype + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + o = torch.zeros_like(q) + all_row_sums = torch.zeros((*q.shape[:-1], 1), dtype=dtype, device=device) + all_row_maxes = torch.full((*q.shape[:-1], 1), max_neg_value, dtype=dtype, device=device) + + scale = q.shape[-1] ** -0.5 + + if not exists(mask): + mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size) + else: + mask = rearrange(mask, "b n -> b 1 1 n") + mask = mask.split(q_bucket_size, dim=-1) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + mask, + all_row_sums.split(q_bucket_size, dim=-2), + all_row_maxes.split(q_bucket_size, dim=-2), + ) + + for ind, (qc, oc, row_mask, row_sums, row_maxes) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale + + if exists(row_mask): + attn_weights.masked_fill_(~row_mask, max_neg_value) + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device).triu( + q_start_index - k_start_index + 1 + ) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + block_row_maxes = attn_weights.amax(dim=-1, keepdims=True) + attn_weights -= block_row_maxes + exp_weights = torch.exp(attn_weights) + + if exists(row_mask): + exp_weights.masked_fill_(~row_mask, 0.0) + + block_row_sums = exp_weights.sum(dim=-1, keepdims=True).clamp(min=EPSILON) + + new_row_maxes = torch.maximum(block_row_maxes, row_maxes) + + exp_values = torch.einsum("... i j, ... j d -> ... i d", exp_weights, vc) + + exp_row_max_diff = torch.exp(row_maxes - new_row_maxes) + exp_block_row_max_diff = torch.exp(block_row_maxes - new_row_maxes) + + new_row_sums = exp_row_max_diff * row_sums + exp_block_row_max_diff * block_row_sums + + oc.mul_((row_sums / new_row_sums) * exp_row_max_diff).add_((exp_block_row_max_diff / new_row_sums) * exp_values) + + row_maxes.copy_(new_row_maxes) + row_sums.copy_(new_row_sums) + + ctx.args = (causal, scale, mask, q_bucket_size, k_bucket_size) + ctx.save_for_backward(q, k, v, o, all_row_sums, all_row_maxes) + + return o + + @staticmethod + @torch.no_grad() + def backward(ctx, do): + """Algorithm 4 in the paper""" + + causal, scale, mask, q_bucket_size, k_bucket_size = ctx.args + q, k, v, o, l, m = ctx.saved_tensors + + device = q.device + + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + dq = torch.zeros_like(q) + dk = torch.zeros_like(k) + dv = torch.zeros_like(v) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + do.split(q_bucket_size, dim=-2), + mask, + l.split(q_bucket_size, dim=-2), + m.split(q_bucket_size, dim=-2), + dq.split(q_bucket_size, dim=-2), + ) + + for ind, (qc, oc, doc, row_mask, lc, mc, dqc) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + dk.split(k_bucket_size, dim=-2), + dv.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc, dkc, dvc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device).triu( + q_start_index - k_start_index + 1 + ) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + exp_attn_weights = torch.exp(attn_weights - mc) + + if exists(row_mask): + exp_attn_weights.masked_fill_(~row_mask, 0.0) + + p = exp_attn_weights / lc + + dv_chunk = torch.einsum("... i j, ... i d -> ... j d", p, doc) + dp = torch.einsum("... i d, ... j d -> ... i j", doc, vc) + + D = (doc * oc).sum(dim=-1, keepdims=True) + ds = p * scale * (dp - D) + + dq_chunk = torch.einsum("... i j, ... j d -> ... i d", ds, kc) + dk_chunk = torch.einsum("... i j, ... i d -> ... j d", ds, qc) + + dqc.add_(dq_chunk) + dkc.add_(dk_chunk) + dvc.add_(dv_chunk) + + return dq, dk, dv, None, None, None, None + + +# endregion + + +def get_parameter_dtype(parameter: torch.nn.Module): + return next(parameter.parameters()).dtype + + +def get_parameter_device(parameter: torch.nn.Module): + return next(parameter.parameters()).device + + +def get_timestep_embedding( + timesteps: torch.Tensor, + embedding_dim: int, + flip_sin_to_cos: bool = False, + downscale_freq_shift: float = 1, + scale: float = 1, + max_period: int = 10000, +): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings. + + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the + embeddings. :return: an [N x dim] Tensor of positional embeddings. + """ + assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" + + half_dim = embedding_dim // 2 + exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device) + exponent = exponent / (half_dim - downscale_freq_shift) + + emb = torch.exp(exponent) + emb = timesteps[:, None].float() * emb[None, :] + + # scale embeddings + emb = scale * emb + + # concat sine and cosine embeddings + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) + + # flip sine and cosine embeddings + if flip_sin_to_cos: + emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1) + + # zero pad + if embedding_dim % 2 == 1: + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +# Deep Shrink: We do not common this function, because minimize dependencies. +def resize_like(x, target, mode="bicubic", align_corners=False): + org_dtype = x.dtype + if org_dtype == torch.bfloat16: + x = x.to(torch.float32) + + if x.shape[-2:] != target.shape[-2:]: + if mode == "nearest": + x = F.interpolate(x, size=target.shape[-2:], mode=mode) + else: + x = F.interpolate(x, size=target.shape[-2:], mode=mode, align_corners=align_corners) + + if org_dtype == torch.bfloat16: + x = x.to(org_dtype) + return x + + +class SampleOutput: + def __init__(self, sample): + self.sample = sample + + +class TimestepEmbedding(nn.Module): + def __init__(self, in_channels: int, time_embed_dim: int, act_fn: str = "silu", out_dim: int = None): + super().__init__() + + self.linear_1 = nn.Linear(in_channels, time_embed_dim) + self.act = None + if act_fn == "silu": + self.act = nn.SiLU() + elif act_fn == "mish": + self.act = nn.Mish() + + if out_dim is not None: + time_embed_dim_out = out_dim + else: + time_embed_dim_out = time_embed_dim + self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out) + + def forward(self, sample): + sample = self.linear_1(sample) + + if self.act is not None: + sample = self.act(sample) + + sample = self.linear_2(sample) + return sample + + +class Timesteps(nn.Module): + def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float): + super().__init__() + self.num_channels = num_channels + self.flip_sin_to_cos = flip_sin_to_cos + self.downscale_freq_shift = downscale_freq_shift + + def forward(self, timesteps): + t_emb = get_timestep_embedding( + timesteps, + self.num_channels, + flip_sin_to_cos=self.flip_sin_to_cos, + downscale_freq_shift=self.downscale_freq_shift, + ) + return t_emb + + +class ResnetBlock2D(nn.Module): + def __init__( + self, + in_channels, + out_channels, + ): + super().__init__() + self.in_channels = in_channels + self.out_channels = out_channels + + self.norm1 = torch.nn.GroupNorm(num_groups=NORM_GROUPS, num_channels=in_channels, eps=NORM_EPS, affine=True) + + self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) + + self.time_emb_proj = torch.nn.Linear(TIME_EMBED_DIM, out_channels) + + self.norm2 = torch.nn.GroupNorm(num_groups=NORM_GROUPS, num_channels=out_channels, eps=NORM_EPS, affine=True) + self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1) + + # if non_linearity == "swish": + self.nonlinearity = lambda x: F.silu(x) + + self.use_in_shortcut = self.in_channels != self.out_channels + + self.conv_shortcut = None + if self.use_in_shortcut: + self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0) + + def forward(self, input_tensor, temb): + hidden_states = input_tensor + + hidden_states = self.norm1(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + + hidden_states = self.conv1(hidden_states) + + temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None] + hidden_states = hidden_states + temb + + hidden_states = self.norm2(hidden_states) + hidden_states = self.nonlinearity(hidden_states) + + hidden_states = self.conv2(hidden_states) + + if self.conv_shortcut is not None: + input_tensor = self.conv_shortcut(input_tensor) + + output_tensor = input_tensor + hidden_states + + return output_tensor + + +class DownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + add_downsample=True, + ): + super().__init__() + + self.has_cross_attention = False + resnets = [] + + for i in range(LAYERS_PER_BLOCK): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + ) + ) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = [Downsample2D(out_channels, out_channels=out_channels)] + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + pass + + def set_use_sdpa(self, sdpa): + pass + + def forward(self, hidden_states, temb=None): + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + else: + hidden_states = resnet(hidden_states, temb) + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +class Downsample2D(nn.Module): + def __init__(self, channels, out_channels): + super().__init__() + + self.channels = channels + self.out_channels = out_channels + + self.conv = nn.Conv2d(self.channels, self.out_channels, 3, stride=2, padding=1) + + def forward(self, hidden_states): + assert hidden_states.shape[1] == self.channels + hidden_states = self.conv(hidden_states) + + return hidden_states + + +class CrossAttention(nn.Module): + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + dim_head: int = 64, + upcast_attention: bool = False, + ): + super().__init__() + inner_dim = dim_head * heads + cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.upcast_attention = upcast_attention + + self.scale = dim_head**-0.5 + self.heads = heads + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=False) + self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=False) + + self.to_out = nn.ModuleList([]) + self.to_out.append(nn.Linear(inner_dim, query_dim)) + # no dropout here + + self.use_memory_efficient_attention_xformers = False + self.use_memory_efficient_attention_mem_eff = False + self.use_sdpa = False + + # Attention processor + self.processor = None + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + self.use_memory_efficient_attention_xformers = xformers + self.use_memory_efficient_attention_mem_eff = mem_eff + + def set_use_sdpa(self, sdpa): + self.use_sdpa = sdpa + + def reshape_heads_to_batch_dim(self, tensor): + batch_size, seq_len, dim = tensor.shape + head_size = self.heads + tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size) + tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size) + return tensor + + def reshape_batch_dim_to_heads(self, tensor): + batch_size, seq_len, dim = tensor.shape + head_size = self.heads + tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim) + tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size) + return tensor + + def set_processor(self): + return self.processor + + def get_processor(self): + return self.processor + + def forward(self, hidden_states, context=None, mask=None, **kwargs): + if self.processor is not None: + ( + hidden_states, + encoder_hidden_states, + attention_mask, + ) = translate_attention_names_from_diffusers( + hidden_states=hidden_states, context=context, mask=mask, **kwargs + ) + return self.processor( + attn=self, + hidden_states=hidden_states, + encoder_hidden_states=context, + attention_mask=mask, + **kwargs + ) + if self.use_memory_efficient_attention_xformers: + return self.forward_memory_efficient_xformers(hidden_states, context, mask) + if self.use_memory_efficient_attention_mem_eff: + return self.forward_memory_efficient_mem_eff(hidden_states, context, mask) + if self.use_sdpa: + return self.forward_sdpa(hidden_states, context, mask) + + query = self.to_q(hidden_states) + context = context if context is not None else hidden_states + key = self.to_k(context) + value = self.to_v(context) + + query = self.reshape_heads_to_batch_dim(query) + key = self.reshape_heads_to_batch_dim(key) + value = self.reshape_heads_to_batch_dim(value) + + hidden_states = self._attention(query, key, value) + + # linear proj + hidden_states = self.to_out[0](hidden_states) + # hidden_states = self.to_out[1](hidden_states) # no dropout + return hidden_states + + def _attention(self, query, key, value): + if self.upcast_attention: + query = query.float() + key = key.float() + + attention_scores = torch.baddbmm( + torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device), + query, + key.transpose(-1, -2), + beta=0, + alpha=self.scale, + ) + attention_probs = attention_scores.softmax(dim=-1) + + # cast back to the original dtype + attention_probs = attention_probs.to(value.dtype) + + # compute attention output + hidden_states = torch.bmm(attention_probs, value) + + # reshape hidden_states + hidden_states = self.reshape_batch_dim_to_heads(hidden_states) + return hidden_states + + # TODO support Hypernetworks + def forward_memory_efficient_xformers(self, x, context=None, mask=None): + import xformers.ops + + h = self.heads + q_in = self.to_q(x) + context = context if context is not None else x + context = context.to(x.dtype) + k_in = self.to_k(context) + v_in = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b n h d", h=h), (q_in, k_in, v_in)) + del q_in, k_in, v_in + + q = q.contiguous() + k = k.contiguous() + v = v.contiguous() + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None) # 最適なのを選んでくれる + + out = rearrange(out, "b n h d -> b n (h d)", h=h) + + out = self.to_out[0](out) + return out + + def forward_memory_efficient_mem_eff(self, x, context=None, mask=None): + flash_func = FlashAttentionFunction + + q_bucket_size = 512 + k_bucket_size = 1024 + + h = self.heads + q = self.to_q(x) + context = context if context is not None else x + context = context.to(x.dtype) + k = self.to_k(context) + v = self.to_v(context) + del context, x + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v)) + + out = flash_func.apply(q, k, v, mask, False, q_bucket_size, k_bucket_size) + + out = rearrange(out, "b h n d -> b n (h d)") + + out = self.to_out[0](out) + return out + + def forward_sdpa(self, x, context=None, mask=None): + h = self.heads + q_in = self.to_q(x) + context = context if context is not None else x + context = context.to(x.dtype) + k_in = self.to_k(context) + v_in = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q_in, k_in, v_in)) + del q_in, k_in, v_in + + out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False) + + out = rearrange(out, "b h n d -> b n (h d)", h=h) + + out = self.to_out[0](out) + return out + +def translate_attention_names_from_diffusers( + hidden_states: torch.FloatTensor, + context: Optional[torch.FloatTensor] = None, + mask: Optional[torch.FloatTensor] = None, + # HF naming + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None +): + # translate from hugging face diffusers + context = context if context is not None else encoder_hidden_states + + # translate from hugging face diffusers + mask = mask if mask is not None else attention_mask + + return hidden_states, context, mask + +# feedforward +class GEGLU(nn.Module): + r""" + A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + """ + + def __init__(self, dim_in: int, dim_out: int): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def gelu(self, gate): + if gate.device.type != "mps": + return F.gelu(gate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype) + + def forward(self, hidden_states): + hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1) + return hidden_states * self.gelu(gate) + + +class FeedForward(nn.Module): + def __init__( + self, + dim: int, + ): + super().__init__() + inner_dim = int(dim * 4) # mult is always 4 + + self.net = nn.ModuleList([]) + # project in + self.net.append(GEGLU(dim, inner_dim)) + # project dropout + self.net.append(nn.Identity()) # nn.Dropout(0)) # dummy for dropout with 0 + # project out + self.net.append(nn.Linear(inner_dim, dim)) + + def forward(self, hidden_states): + for module in self.net: + hidden_states = module(hidden_states) + return hidden_states + + +class BasicTransformerBlock(nn.Module): + def __init__( + self, dim: int, num_attention_heads: int, attention_head_dim: int, cross_attention_dim: int, upcast_attention: bool = False + ): + super().__init__() + + # 1. Self-Attn + self.attn1 = CrossAttention( + query_dim=dim, + cross_attention_dim=None, + heads=num_attention_heads, + dim_head=attention_head_dim, + upcast_attention=upcast_attention, + ) + self.ff = FeedForward(dim) + + # 2. Cross-Attn + self.attn2 = CrossAttention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + upcast_attention=upcast_attention, + ) + + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(dim) + + def set_use_memory_efficient_attention(self, xformers: bool, mem_eff: bool): + self.attn1.set_use_memory_efficient_attention(xformers, mem_eff) + self.attn2.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa: bool): + self.attn1.set_use_sdpa(sdpa) + self.attn2.set_use_sdpa(sdpa) + + def forward(self, hidden_states, context=None, timestep=None): + # 1. Self-Attention + norm_hidden_states = self.norm1(hidden_states) + + hidden_states = self.attn1(norm_hidden_states) + hidden_states + + # 2. Cross-Attention + norm_hidden_states = self.norm2(hidden_states) + hidden_states = self.attn2(norm_hidden_states, context=context) + hidden_states + + # 3. Feed-forward + hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states + + return hidden_states + + +class Transformer2DModel(nn.Module): + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + use_linear_projection: bool = False, + upcast_attention: bool = False, + ): + super().__init__() + self.in_channels = in_channels + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + self.use_linear_projection = use_linear_projection + + self.norm = torch.nn.GroupNorm(num_groups=TRANSFORMER_NORM_NUM_GROUPS, num_channels=in_channels, eps=1e-6, affine=True) + + if use_linear_projection: + self.proj_in = nn.Linear(in_channels, inner_dim) + else: + self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + upcast_attention=upcast_attention, + ) + ] + ) + + if use_linear_projection: + self.proj_out = nn.Linear(in_channels, inner_dim) + else: + self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + for transformer in self.transformer_blocks: + transformer.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa): + for transformer in self.transformer_blocks: + transformer.set_use_sdpa(sdpa) + + def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True): + # 1. Input + batch, _, height, weight = hidden_states.shape + residual = hidden_states + + hidden_states = self.norm(hidden_states) + if not self.use_linear_projection: + hidden_states = self.proj_in(hidden_states) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) + hidden_states = self.proj_in(hidden_states) + + # 2. Blocks + for block in self.transformer_blocks: + hidden_states = block(hidden_states, context=encoder_hidden_states, timestep=timestep) + + # 3. Output + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous() + hidden_states = self.proj_out(hidden_states) + else: + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous() + + output = hidden_states + residual + + if not return_dict: + return (output,) + + return SampleOutput(sample=output) + + +class CrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + add_downsample=True, + cross_attention_dim=1280, + attn_num_head_channels=1, + use_linear_projection=False, + upcast_attention=False, + ): + super().__init__() + self.has_cross_attention = True + resnets = [] + attentions = [] + + self.attn_num_head_channels = attn_num_head_channels + + for i in range(LAYERS_PER_BLOCK): + in_channels = in_channels if i == 0 else out_channels + + resnets.append(ResnetBlock2D(in_channels=in_channels, out_channels=out_channels)) + attentions.append( + Transformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + cross_attention_dim=cross_attention_dim, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList([Downsample2D(out_channels, out_channels)]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + for attn in self.attentions: + attn.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa): + for attn in self.attentions: + attn.set_use_sdpa(sdpa) + + def forward(self, hidden_states, temb=None, encoder_hidden_states=None): + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states + )[0] + else: + hidden_states = resnet(hidden_states, temb) + hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +class UNetMidBlock2DCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + attn_num_head_channels=1, + cross_attention_dim=1280, + use_linear_projection=False, + ): + super().__init__() + + self.has_cross_attention = True + self.attn_num_head_channels = attn_num_head_channels + + # Middle block has two resnets and one attention + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + ), + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + ), + ] + attentions = [ + Transformer2DModel( + attn_num_head_channels, + in_channels // attn_num_head_channels, + in_channels=in_channels, + cross_attention_dim=cross_attention_dim, + use_linear_projection=use_linear_projection, + ) + ] + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + for attn in self.attentions: + attn.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa): + for attn in self.attentions: + attn.set_use_sdpa(sdpa) + + def forward(self, hidden_states, temb=None, encoder_hidden_states=None): + for i, resnet in enumerate(self.resnets): + attn = None if i == 0 else self.attentions[i - 1] + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + if attn is not None: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states + )[0] + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + else: + if attn is not None: + hidden_states = attn(hidden_states, encoder_hidden_states).sample + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class Upsample2D(nn.Module): + def __init__(self, channels, out_channels): + super().__init__() + self.channels = channels + self.out_channels = out_channels + self.conv = nn.Conv2d(self.channels, self.out_channels, 3, padding=1) + + def forward(self, hidden_states, output_size): + assert hidden_states.shape[1] == self.channels + + # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16 + # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch + # https://github.com/pytorch/pytorch/issues/86679 + dtype = hidden_states.dtype + if dtype == torch.bfloat16: + hidden_states = hidden_states.to(torch.float32) + + # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984 + if hidden_states.shape[0] >= 64: + hidden_states = hidden_states.contiguous() + + # if `output_size` is passed we force the interpolation output size and do not make use of `scale_factor=2` + if output_size is None: + hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest") + else: + hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest") + + # If the input is bfloat16, we cast back to bfloat16 + if dtype == torch.bfloat16: + hidden_states = hidden_states.to(dtype) + + hidden_states = self.conv(hidden_states) + + return hidden_states + + +class UpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + add_upsample=True, + ): + super().__init__() + + self.has_cross_attention = False + resnets = [] + + for i in range(LAYERS_PER_BLOCK_UP): + res_skip_channels = in_channels if (i == LAYERS_PER_BLOCK_UP - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + pass + + def set_use_sdpa(self, sdpa): + pass + + def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None): + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + else: + hidden_states = resnet(hidden_states, temb) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + +class CrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + attn_num_head_channels=1, + cross_attention_dim=1280, + add_upsample=True, + use_linear_projection=False, + upcast_attention=False, + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.attn_num_head_channels = attn_num_head_channels + + for i in range(LAYERS_PER_BLOCK_UP): + res_skip_channels = in_channels if (i == LAYERS_PER_BLOCK_UP - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + ) + ) + attentions.append( + Transformer2DModel( + attn_num_head_channels, + out_channels // attn_num_head_channels, + in_channels=out_channels, + cross_attention_dim=cross_attention_dim, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + for attn in self.attentions: + attn.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa): + for attn in self.attentions: + attn.set_use_sdpa(sdpa) + + def forward( + self, + hidden_states, + res_hidden_states_tuple, + temb=None, + encoder_hidden_states=None, + upsample_size=None, + ): + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(attn, return_dict=False), hidden_states, encoder_hidden_states + )[0] + else: + hidden_states = resnet(hidden_states, temb) + hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + +def get_down_block( + down_block_type, + in_channels, + out_channels, + add_downsample, + attn_num_head_channels, + cross_attention_dim, + use_linear_projection, + upcast_attention, +): + if down_block_type == "DownBlock2D": + return DownBlock2D( + in_channels=in_channels, + out_channels=out_channels, + add_downsample=add_downsample, + ) + elif down_block_type == "CrossAttnDownBlock2D": + return CrossAttnDownBlock2D( + in_channels=in_channels, + out_channels=out_channels, + add_downsample=add_downsample, + cross_attention_dim=cross_attention_dim, + attn_num_head_channels=attn_num_head_channels, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + ) + + +def get_up_block( + up_block_type, + in_channels, + out_channels, + prev_output_channel, + add_upsample, + attn_num_head_channels, + cross_attention_dim=None, + use_linear_projection=False, + upcast_attention=False, +): + if up_block_type == "UpBlock2D": + return UpBlock2D( + in_channels=in_channels, + prev_output_channel=prev_output_channel, + out_channels=out_channels, + add_upsample=add_upsample, + ) + elif up_block_type == "CrossAttnUpBlock2D": + return CrossAttnUpBlock2D( + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + attn_num_head_channels=attn_num_head_channels, + cross_attention_dim=cross_attention_dim, + add_upsample=add_upsample, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + ) + + +class UNet2DConditionModel(nn.Module): + _supports_gradient_checkpointing = True + + def __init__( + self, + sample_size: Optional[int] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + cross_attention_dim: int = 1280, + use_linear_projection: bool = False, + upcast_attention: bool = False, + **kwargs, + ): + super().__init__() + assert sample_size is not None, "sample_size must be specified" + logger.info( + f"UNet2DConditionModel: {sample_size}, {attention_head_dim}, {cross_attention_dim}, {use_linear_projection}, {upcast_attention}" + ) + + # 外部からの参照用に定義しておく + self.in_channels = IN_CHANNELS + self.out_channels = OUT_CHANNELS + + self.sample_size = sample_size + self.prepare_config(sample_size=sample_size) + + # state_dictの書式が変わるのでmoduleの持ち方は変えられない + + # input + self.conv_in = nn.Conv2d(IN_CHANNELS, BLOCK_OUT_CHANNELS[0], kernel_size=3, padding=(1, 1)) + + # time + self.time_proj = Timesteps(BLOCK_OUT_CHANNELS[0], TIME_EMBED_FLIP_SIN_TO_COS, TIME_EMBED_FREQ_SHIFT) + + self.time_embedding = TimestepEmbedding(TIMESTEP_INPUT_DIM, TIME_EMBED_DIM) + + self.down_blocks = nn.ModuleList([]) + self.mid_block = None + self.up_blocks = nn.ModuleList([]) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * 4 + + # down + output_channel = BLOCK_OUT_CHANNELS[0] + for i, down_block_type in enumerate(DOWN_BLOCK_TYPES): + input_channel = output_channel + output_channel = BLOCK_OUT_CHANNELS[i] + is_final_block = i == len(BLOCK_OUT_CHANNELS) - 1 + + down_block = get_down_block( + down_block_type, + in_channels=input_channel, + out_channels=output_channel, + add_downsample=not is_final_block, + attn_num_head_channels=attention_head_dim[i], + cross_attention_dim=cross_attention_dim, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + ) + self.down_blocks.append(down_block) + + # mid + self.mid_block = UNetMidBlock2DCrossAttn( + in_channels=BLOCK_OUT_CHANNELS[-1], + attn_num_head_channels=attention_head_dim[-1], + cross_attention_dim=cross_attention_dim, + use_linear_projection=use_linear_projection, + ) + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(BLOCK_OUT_CHANNELS)) + reversed_attention_head_dim = list(reversed(attention_head_dim)) + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(UP_BLOCK_TYPES): + is_final_block = i == len(BLOCK_OUT_CHANNELS) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(BLOCK_OUT_CHANNELS) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + add_upsample=add_upsample, + attn_num_head_channels=reversed_attention_head_dim[i], + cross_attention_dim=cross_attention_dim, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + self.conv_norm_out = nn.GroupNorm(num_channels=BLOCK_OUT_CHANNELS[0], num_groups=NORM_GROUPS, eps=NORM_EPS) + self.conv_act = nn.SiLU() + self.conv_out = nn.Conv2d(BLOCK_OUT_CHANNELS[0], OUT_CHANNELS, kernel_size=3, padding=1) + + # region diffusers compatibility + def prepare_config(self, *args, **kwargs): + self.config = SimpleNamespace(**kwargs) + + @property + def dtype(self) -> torch.dtype: + # `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype). + return get_parameter_dtype(self) + + @property + def device(self) -> torch.device: + # `torch.device`: The device on which the module is (assuming that all the module parameters are on the same device). + return get_parameter_device(self) + + def set_attention_slice(self, slice_size): + raise NotImplementedError("Attention slicing is not supported for this model.") + + def is_gradient_checkpointing(self) -> bool: + return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for m in self.modules()) + + def enable_gradient_checkpointing(self): + self.set_gradient_checkpointing(value=True) + + def disable_gradient_checkpointing(self): + self.set_gradient_checkpointing(value=False) + + def set_use_memory_efficient_attention(self, xformers: bool, mem_eff: bool) -> None: + modules = self.down_blocks + [self.mid_block] + self.up_blocks + for module in modules: + module.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa: bool) -> None: + modules = self.down_blocks + [self.mid_block] + self.up_blocks + for module in modules: + module.set_use_sdpa(sdpa) + + def set_gradient_checkpointing(self, value=False): + modules = self.down_blocks + [self.mid_block] + self.up_blocks + for module in modules: + logger.info(f"{module.__class__.__name__} {module.gradient_checkpointing} -> {value}") + module.gradient_checkpointing = value + + # endregion + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + return_dict: bool = True, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + ) -> Union[Dict, Tuple]: + r""" + Args: + sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor + timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps + encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a dict instead of a plain tuple. + + Returns: + `SampleOutput` or `tuple`: + `SampleOutput` if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layears). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + # デフォルトではサンプルは「2^アップサンプルの数」、つまり64の倍数である必要がある + # ただそれ以外のサイズにも対応できるように、必要ならアップサンプルのサイズを変更する + # 多分画質が悪くなるので、64で割り切れるようにしておくのが良い + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + # 64で割り切れないときはupsamplerにサイズを伝える + forward_upsample_size = False + upsample_size = None + + if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]): + # logger.info("Forward upsample size to force interpolation output size.") + forward_upsample_size = True + + # 1. time + timesteps = timestep + timesteps = self.handle_unusual_timesteps(sample, timesteps) # 変な時だけ処理 + + t_emb = self.time_proj(timesteps) + + # timesteps does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + # timestepsは重みを含まないので常にfloat32のテンソルを返す + # しかしtime_embeddingはfp16で動いているかもしれないので、ここでキャストする必要がある + # time_projでキャストしておけばいいんじゃね？ + t_emb = t_emb.to(dtype=self.dtype) + emb = self.time_embedding(t_emb) + + # 2. pre-process + sample = self.conv_in(sample) + + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + # downblockはforwardで必ずencoder_hidden_statesを受け取るようにしても良さそうだけど、 + # まあこちらのほうがわかりやすいかもしれない + if downsample_block.has_cross_attention: + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + + down_block_res_samples += res_samples + + # skip connectionにControlNetの出力を追加する + if down_block_additional_residuals is not None: + down_block_res_samples = list(down_block_res_samples) + for i in range(len(down_block_res_samples)): + down_block_res_samples[i] += down_block_additional_residuals[i] + down_block_res_samples = tuple(down_block_res_samples) + + # 4. mid + sample = self.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states) + + # ControlNetの出力を追加する + if mid_block_additional_residual is not None: + sample += mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] # skip connection + + # if we have not reached the final block and need to forward the upsample size, we do it here + # 前述のように最後のブロック以外ではupsample_sizeを伝える + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + upsample_size=upsample_size, + ) + else: + sample = upsample_block( + hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size + ) + + # 6. post-process + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if not return_dict: + return (sample,) + + return SampleOutput(sample=sample) + + def handle_unusual_timesteps(self, sample, timesteps): + r""" + timestampsがTensorでない場合、Tensorに変換する。またOnnx/Core MLと互換性のあるようにbatchサイズまでbroadcastする。 + """ + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timesteps, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + return timesteps + + +class InferUNet2DConditionModel: + def __init__(self, original_unet: UNet2DConditionModel): + self.delegate = original_unet + + # override original model's forward method: because forward is not called by `__call__` + # overriding `__call__` is not enough, because nn.Module.forward has a special handling + self.delegate.forward = self.forward + + # override original model's up blocks' forward method + for up_block in self.delegate.up_blocks: + if up_block.__class__.__name__ == "UpBlock2D": + + def resnet_wrapper(func, block): + def forward(*args, **kwargs): + return func(block, *args, **kwargs) + + return forward + + up_block.forward = resnet_wrapper(self.up_block_forward, up_block) + + elif up_block.__class__.__name__ == "CrossAttnUpBlock2D": + + def cross_attn_up_wrapper(func, block): + def forward(*args, **kwargs): + return func(block, *args, **kwargs) + + return forward + + up_block.forward = cross_attn_up_wrapper(self.cross_attn_up_block_forward, up_block) + + # Deep Shrink + self.ds_depth_1 = None + self.ds_depth_2 = None + self.ds_timesteps_1 = None + self.ds_timesteps_2 = None + self.ds_ratio = None + + # call original model's methods + def __getattr__(self, name): + return getattr(self.delegate, name) + + def __call__(self, *args, **kwargs): + return self.delegate(*args, **kwargs) + + def set_deep_shrink(self, ds_depth_1, ds_timesteps_1=650, ds_depth_2=None, ds_timesteps_2=None, ds_ratio=0.5): + if ds_depth_1 is None: + logger.info("Deep Shrink is disabled.") + self.ds_depth_1 = None + self.ds_timesteps_1 = None + self.ds_depth_2 = None + self.ds_timesteps_2 = None + self.ds_ratio = None + else: + logger.info( + f"Deep Shrink is enabled: [depth={ds_depth_1}/{ds_depth_2}, timesteps={ds_timesteps_1}/{ds_timesteps_2}, ratio={ds_ratio}]" + ) + self.ds_depth_1 = ds_depth_1 + self.ds_timesteps_1 = ds_timesteps_1 + self.ds_depth_2 = ds_depth_2 if ds_depth_2 is not None else -1 + self.ds_timesteps_2 = ds_timesteps_2 if ds_timesteps_2 is not None else 1000 + self.ds_ratio = ds_ratio + + def up_block_forward(self, _self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None): + for resnet in _self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # Deep Shrink + if res_hidden_states.shape[-2:] != hidden_states.shape[-2:]: + hidden_states = resize_like(hidden_states, res_hidden_states) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + hidden_states = resnet(hidden_states, temb) + + if _self.upsamplers is not None: + for upsampler in _self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + def cross_attn_up_block_forward( + self, + _self, + hidden_states, + res_hidden_states_tuple, + temb=None, + encoder_hidden_states=None, + upsample_size=None, + ): + for resnet, attn in zip(_self.resnets, _self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # Deep Shrink + if res_hidden_states.shape[-2:] != hidden_states.shape[-2:]: + hidden_states = resize_like(hidden_states, res_hidden_states) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + hidden_states = resnet(hidden_states, temb) + hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample + + if _self.upsamplers is not None: + for upsampler in _self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + return_dict: bool = True, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + ) -> Union[Dict, Tuple]: + r""" + current implementation is a copy of `UNet2DConditionModel.forward()` with Deep Shrink. + """ + + r""" + Args: + sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor + timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps + encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a dict instead of a plain tuple. + + Returns: + `SampleOutput` or `tuple`: + `SampleOutput` if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. + """ + + _self = self.delegate + + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layears). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + # デフォルトではサンプルは「2^アップサンプルの数」、つまり64の倍数である必要がある + # ただそれ以外のサイズにも対応できるように、必要ならアップサンプルのサイズを変更する + # 多分画質が悪くなるので、64で割り切れるようにしておくのが良い + default_overall_up_factor = 2**_self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + # 64で割り切れないときはupsamplerにサイズを伝える + forward_upsample_size = False + upsample_size = None + + if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]): + # logger.info("Forward upsample size to force interpolation output size.") + forward_upsample_size = True + + # 1. time + timesteps = timestep + timesteps = _self.handle_unusual_timesteps(sample, timesteps) # 変な時だけ処理 + + t_emb = _self.time_proj(timesteps) + + # timesteps does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + # timestepsは重みを含まないので常にfloat32のテンソルを返す + # しかしtime_embeddingはfp16で動いているかもしれないので、ここでキャストする必要がある + # time_projでキャストしておけばいいんじゃね？ + t_emb = t_emb.to(dtype=_self.dtype) + emb = _self.time_embedding(t_emb) + + # 2. pre-process + sample = _self.conv_in(sample) + + down_block_res_samples = (sample,) + for depth, downsample_block in enumerate(_self.down_blocks): + # Deep Shrink + if self.ds_depth_1 is not None: + if (depth == self.ds_depth_1 and timesteps[0] >= self.ds_timesteps_1) or ( + self.ds_depth_2 is not None + and depth == self.ds_depth_2 + and timesteps[0] < self.ds_timesteps_1 + and timesteps[0] >= self.ds_timesteps_2 + ): + org_dtype = sample.dtype + if org_dtype == torch.bfloat16: + sample = sample.to(torch.float32) + sample = F.interpolate(sample, scale_factor=self.ds_ratio, mode="bicubic", align_corners=False).to(org_dtype) + + # downblockはforwardで必ずencoder_hidden_statesを受け取るようにしても良さそうだけど、 + # まあこちらのほうがわかりやすいかもしれない + if downsample_block.has_cross_attention: + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + + down_block_res_samples += res_samples + + # skip connectionにControlNetの出力を追加する + if down_block_additional_residuals is not None: + down_block_res_samples = list(down_block_res_samples) + for i in range(len(down_block_res_samples)): + down_block_res_samples[i] += down_block_additional_residuals[i] + down_block_res_samples = tuple(down_block_res_samples) + + # 4. mid + sample = _self.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states) + + # ControlNetの出力を追加する + if mid_block_additional_residual is not None: + sample += mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(_self.up_blocks): + is_final_block = i == len(_self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] # skip connection + + # if we have not reached the final block and need to forward the upsample size, we do it here + # 前述のように最後のブロック以外ではupsample_sizeを伝える + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + upsample_size=upsample_size, + ) + else: + sample = upsample_block( + hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size + ) + + # 6. post-process + sample = _self.conv_norm_out(sample) + sample = _self.conv_act(sample) + sample = _self.conv_out(sample) + + if not return_dict: + return (sample,) + + return SampleOutput(sample=sample) diff --git a/prepare_buckets_latents.py b/prepare_buckets_latents.py new file mode 100644 index 0000000000000000000000000000000000000000..019c737a629fcdfe6246a548e469ef3a44ac456f --- /dev/null +++ b/prepare_buckets_latents.py @@ -0,0 +1,286 @@ +import argparse +import os +import json + +from pathlib import Path +from typing import List +from tqdm import tqdm +import numpy as np +from PIL import Image +import cv2 + +import torch +from library.device_utils import init_ipex, get_preferred_device + +init_ipex() + +from torchvision import transforms + +import library.model_util as model_util +import library.train_util as train_util +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +DEVICE = get_preferred_device() + +IMAGE_TRANSFORMS = transforms.Compose( + [ + transforms.ToTensor(), + transforms.Normalize([0.5], [0.5]), + ] +) + + +def collate_fn_remove_corrupted(batch): + """Collate function that allows to remove corrupted examples in the + dataloader. It expects that the dataloader returns 'None' when that occurs. + The 'None's in the batch are removed. + """ + # Filter out all the Nones (corrupted examples) + batch = list(filter(lambda x: x is not None, batch)) + return batch + + +def get_npz_filename(data_dir, image_key, is_full_path, recursive): + if is_full_path: + base_name = os.path.splitext(os.path.basename(image_key))[0] + relative_path = os.path.relpath(os.path.dirname(image_key), data_dir) + else: + base_name = image_key + relative_path = "" + + if recursive and relative_path: + return os.path.join(data_dir, relative_path, base_name) + ".npz" + else: + return os.path.join(data_dir, base_name) + ".npz" + + +def main(args): + # assert args.bucket_reso_steps % 8 == 0, f"bucket_reso_steps must be divisible by 8 / bucket_reso_stepは8で割り切れる必要があります" + if args.bucket_reso_steps % 8 > 0: + logger.warning(f"resolution of buckets in training time is a multiple of 8 / 学習時の各bucketの解像度は8単位になります") + if args.bucket_reso_steps % 32 > 0: + logger.warning( + f"WARNING: bucket_reso_steps is not divisible by 32. It is not working with SDXL / bucket_reso_stepsが32で割り切れません。SDXLでは動作しません" + ) + + train_data_dir_path = Path(args.train_data_dir) + image_paths: List[str] = [str(p) for p in train_util.glob_images_pathlib(train_data_dir_path, args.recursive)] + logger.info(f"found {len(image_paths)} images.") + + if os.path.exists(args.in_json): + logger.info(f"loading existing metadata: {args.in_json}") + with open(args.in_json, "rt", encoding="utf-8") as f: + metadata = json.load(f) + else: + logger.error(f"no metadata / メタデータファイルがありません: {args.in_json}") + return + + weight_dtype = torch.float32 + if args.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif args.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + vae = model_util.load_vae(args.model_name_or_path, weight_dtype) + vae.eval() + vae.to(DEVICE, dtype=weight_dtype) + + # bucketのサイズを計算する + max_reso = tuple([int(t) for t in args.max_resolution.split(",")]) + assert ( + len(max_reso) == 2 + ), f"illegal resolution (not 'width,height') / 画像サイズに誤りがあります。'幅,高さ'で指定してください: {args.max_resolution}" + + bucket_manager = train_util.BucketManager( + args.bucket_no_upscale, max_reso, args.min_bucket_reso, args.max_bucket_reso, args.bucket_reso_steps + ) + if not args.bucket_no_upscale: + bucket_manager.make_buckets() + else: + logger.warning( + "min_bucket_reso and max_bucket_reso are ignored if bucket_no_upscale is set, because bucket reso is defined by image size automatically / bucket_no_upscaleが指定された場合は、bucketの解像度は画像サイズから自動計算されるため、min_bucket_resoとmax_bucket_resoは無視されます" + ) + + # 画像をひとつずつ適切なbucketに割り当てながらlatentを計算する + img_ar_errors = [] + + def process_batch(is_last): + for bucket in bucket_manager.buckets: + if (is_last and len(bucket) > 0) or len(bucket) >= args.batch_size: + train_util.cache_batch_latents(vae, True, bucket, args.flip_aug, args.alpha_mask, False) + bucket.clear() + + # 読み込みの高速化のためにDataLoaderを使うオプション + if args.max_data_loader_n_workers is not None: + dataset = train_util.ImageLoadingDataset(image_paths) + data = torch.utils.data.DataLoader( + dataset, + batch_size=1, + shuffle=False, + num_workers=args.max_data_loader_n_workers, + collate_fn=collate_fn_remove_corrupted, + drop_last=False, + ) + else: + data = [[(None, ip)] for ip in image_paths] + + bucket_counts = {} + for data_entry in tqdm(data, smoothing=0.0): + if data_entry[0] is None: + continue + + img_tensor, image_path = data_entry[0] + if img_tensor is not None: + image = transforms.functional.to_pil_image(img_tensor) + else: + try: + image = Image.open(image_path) + if image.mode != "RGB": + image = image.convert("RGB") + except Exception as e: + logger.error(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}") + continue + + image_key = image_path if args.full_path else os.path.splitext(os.path.basename(image_path))[0] + if image_key not in metadata: + metadata[image_key] = {} + + # 本当はこのあとの部分もDataSetに持っていけば高速化できるがいろいろ大変 + + reso, resized_size, ar_error = bucket_manager.select_bucket(image.width, image.height) + img_ar_errors.append(abs(ar_error)) + bucket_counts[reso] = bucket_counts.get(reso, 0) + 1 + + # メタデータに記録する解像度はlatent単位とするので、8単位で切り捨て + metadata[image_key]["train_resolution"] = (reso[0] - reso[0] % 8, reso[1] - reso[1] % 8) + + if not args.bucket_no_upscale: + # upscaleを行わないときには、resize後のサイズは、bucketのサイズと、縦横どちらかが同じであることを確認する + assert ( + resized_size[0] == reso[0] or resized_size[1] == reso[1] + ), f"internal error, resized size not match: {reso}, {resized_size}, {image.width}, {image.height}" + assert ( + resized_size[0] >= reso[0] and resized_size[1] >= reso[1] + ), f"internal error, resized size too small: {reso}, {resized_size}, {image.width}, {image.height}" + + assert ( + resized_size[0] >= reso[0] and resized_size[1] >= reso[1] + ), f"internal error resized size is small: {resized_size}, {reso}" + + # 既に存在するファイルがあればshape等を確認して同じならskipする + npz_file_name = get_npz_filename(args.train_data_dir, image_key, args.full_path, args.recursive) + if args.skip_existing: + if train_util.is_disk_cached_latents_is_expected(reso, npz_file_name, args.flip_aug): + continue + + # バッチへ追加 + image_info = train_util.ImageInfo(image_key, 1, "", False, image_path) + image_info.latents_npz = npz_file_name + image_info.bucket_reso = reso + image_info.resized_size = resized_size + image_info.image = image + bucket_manager.add_image(reso, image_info) + + # バッチを推論するか判定して推論する + process_batch(False) + + # 残りを処理する + process_batch(True) + + bucket_manager.sort() + for i, reso in enumerate(bucket_manager.resos): + count = bucket_counts.get(reso, 0) + if count > 0: + logger.info(f"bucket {i} {reso}: {count}") + img_ar_errors = np.array(img_ar_errors) + logger.info(f"mean ar error: {np.mean(img_ar_errors)}") + + # metadataを書き出して終わり + logger.info(f"writing metadata: {args.out_json}") + with open(args.out_json, "wt", encoding="utf-8") as f: + json.dump(metadata, f, indent=2) + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ") + parser.add_argument("in_json", type=str, help="metadata file to input / 読み込むメタデータファイル") + parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先") + parser.add_argument("model_name_or_path", type=str, help="model name or path to encode latents / latentを取得するためのモデル") + parser.add_argument( + "--v2", action="store_true", help="not used (for backward compatibility) / 使用されません（互換性のため残してあります）" + ) + parser.add_argument("--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ") + parser.add_argument( + "--max_data_loader_n_workers", + type=int, + default=None, + help="enable image reading by DataLoader with this number of workers (faster) / DataLoaderによる画像読み込みを有効にしてこのワーカー数を適用する（読み込みを高速化）", + ) + parser.add_argument( + "--max_resolution", + type=str, + default="512,512", + help="max resolution in fine tuning (width,height) / fine tuning時の最大画像サイズ 「幅,高さ」（使用メモリ量に関係します）", + ) + parser.add_argument("--min_bucket_reso", type=int, default=256, help="minimum resolution for buckets / bucketの最小解像度") + parser.add_argument("--max_bucket_reso", type=int, default=1024, help="maximum resolution for buckets / bucketの最大解像度") + parser.add_argument( + "--bucket_reso_steps", + type=int, + default=64, + help="steps of resolution for buckets, divisible by 8 is recommended / bucketの解像度の単位、8で割り切れる値を推奨します", + ) + parser.add_argument( + "--bucket_no_upscale", + action="store_true", + help="make bucket for each image without upscaling / 画像を拡大せずbucketを作成します", + ) + parser.add_argument( + "--mixed_precision", + type=str, + default="no", + choices=["no", "fp16", "bf16"], + help="use mixed precision / 混合精度を使う場合、その精度", + ) + parser.add_argument( + "--full_path", + action="store_true", + help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）", + ) + parser.add_argument( + "--flip_aug", + action="store_true", + help="flip augmentation, save latents for flipped images / 左右反転した画像もlatentを取得、保存する", + ) + parser.add_argument( + "--alpha_mask", + type=str, + default="", + help="save alpha mask for images for loss calculation / 損失計算用に画像のアルファマスクを保存する", + ) + parser.add_argument( + "--skip_existing", + action="store_true", + help="skip images if npz already exists (both normal and flipped exists if flip_aug is enabled) / npzが既に存在する画像をスキップする（flip_aug有効時は通常、反転の両方が存在する画像をスキップ）", + ) + parser.add_argument( + "--recursive", + action="store_true", + help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す", + ) + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + main(args) diff --git a/resize_images_to_resolution.py b/resize_images_to_resolution.py new file mode 100644 index 0000000000000000000000000000000000000000..0f9e00b1e328fed7166f0c68c4bf3f11dfd5fc02 --- /dev/null +++ b/resize_images_to_resolution.py @@ -0,0 +1,134 @@ +import glob +import os +import cv2 +import argparse +import shutil +import math +from PIL import Image +import numpy as np +from library.utils import setup_logging, pil_resize +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def resize_images(src_img_folder, dst_img_folder, max_resolution="512x512", divisible_by=2, interpolation=None, save_as_png=False, copy_associated_files=False): + # Split the max_resolution string by "," and strip any whitespaces + max_resolutions = [res.strip() for res in max_resolution.split(',')] + + # # Calculate max_pixels from max_resolution string + # max_pixels = int(max_resolution.split("x")[0]) * int(max_resolution.split("x")[1]) + + # Create destination folder if it does not exist + if not os.path.exists(dst_img_folder): + os.makedirs(dst_img_folder) + + # Select interpolation method + if interpolation == 'lanczos4': + pil_interpolation = Image.LANCZOS + elif interpolation == 'cubic': + pil_interpolation = Image.BICUBIC + else: + cv2_interpolation = cv2.INTER_AREA + + # Iterate through all files in src_img_folder + img_exts = (".png", ".jpg", ".jpeg", ".webp", ".bmp") # copy from train_util.py + for filename in os.listdir(src_img_folder): + # Check if the image is png, jpg or webp etc... + if not filename.endswith(img_exts): + # Copy the file to the destination folder if not png, jpg or webp etc (.txt or .caption or etc.) + shutil.copy(os.path.join(src_img_folder, filename), os.path.join(dst_img_folder, filename)) + continue + + # Load image + # img = cv2.imread(os.path.join(src_img_folder, filename)) + image = Image.open(os.path.join(src_img_folder, filename)) + if not image.mode == "RGB": + image = image.convert("RGB") + img = np.array(image, np.uint8) + + base, _ = os.path.splitext(filename) + for max_resolution in max_resolutions: + # Calculate max_pixels from max_resolution string + max_pixels = int(max_resolution.split("x")[0]) * int(max_resolution.split("x")[1]) + + # Calculate current number of pixels + current_pixels = img.shape[0] * img.shape[1] + + # Check if the image needs resizing + if current_pixels > max_pixels: + # Calculate scaling factor + scale_factor = max_pixels / current_pixels + + # Calculate new dimensions + new_height = int(img.shape[0] * math.sqrt(scale_factor)) + new_width = int(img.shape[1] * math.sqrt(scale_factor)) + + # Resize image + if cv2_interpolation: + img = cv2.resize(img, (new_width, new_height), interpolation=cv2_interpolation) + else: + img = pil_resize(img, (new_width, new_height), interpolation=pil_interpolation) + else: + new_height, new_width = img.shape[0:2] + + # Calculate the new height and width that are divisible by divisible_by (with/without resizing) + new_height = new_height if new_height % divisible_by == 0 else new_height - new_height % divisible_by + new_width = new_width if new_width % divisible_by == 0 else new_width - new_width % divisible_by + + # Center crop the image to the calculated dimensions + y = int((img.shape[0] - new_height) / 2) + x = int((img.shape[1] - new_width) / 2) + img = img[y:y + new_height, x:x + new_width] + + # Split filename into base and extension + new_filename = base + '+' + max_resolution + ('.png' if save_as_png else '.jpg') + + # Save resized image in dst_img_folder + # cv2.imwrite(os.path.join(dst_img_folder, new_filename), img, [cv2.IMWRITE_JPEG_QUALITY, 100]) + image = Image.fromarray(img) + image.save(os.path.join(dst_img_folder, new_filename), quality=100) + + proc = "Resized" if current_pixels > max_pixels else "Saved" + logger.info(f"{proc} image: {filename} with size {img.shape[0]}x{img.shape[1]} as {new_filename}") + + # If other files with same basename, copy them with resolution suffix + if copy_associated_files: + asoc_files = glob.glob(os.path.join(src_img_folder, base + ".*")) + for asoc_file in asoc_files: + ext = os.path.splitext(asoc_file)[1] + if ext in img_exts: + continue + for max_resolution in max_resolutions: + new_asoc_file = base + '+' + max_resolution + ext + logger.info(f"Copy {asoc_file} as {new_asoc_file}") + shutil.copy(os.path.join(src_img_folder, asoc_file), os.path.join(dst_img_folder, new_asoc_file)) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser( + description='Resize images in a folder to a specified max resolution(s) / 指定されたフォルダ内の画像を指定した最大画像サイズ（面積）以下にアスペクト比を維持したままリサイズします') + parser.add_argument('src_img_folder', type=str, help='Source folder containing the images / 元画像のフォルダ') + parser.add_argument('dst_img_folder', type=str, help='Destination folder to save the resized images / リサイズ後の画像を保存するフォルダ') + parser.add_argument('--max_resolution', type=str, + help='Maximum resolution(s) in the format "512x512,384x384, etc, etc" / 最大画像サイズをカンマ区切りで指定 ("512x512,384x384, etc, etc" など)', default="512x512,384x384,256x256,128x128") + parser.add_argument('--divisible_by', type=int, + help='Ensure new dimensions are divisible by this value / リサイズ後の画像のサイズをこの値で割り切れるようにします', default=1) + parser.add_argument('--interpolation', type=str, choices=['area', 'cubic', 'lanczos4'], + default='area', help='Interpolation method for resizing / リサイズ時の補完方法') + parser.add_argument('--save_as_png', action='store_true', help='Save as png format / png形式で保存') + parser.add_argument('--copy_associated_files', action='store_true', + help='Copy files with same base name to images (captions etc) / 画像と同じファイル名（拡張子を除く）のファイルもコピーする') + + return parser + + +def main(): + parser = setup_parser() + + args = parser.parse_args() + resize_images(args.src_img_folder, args.dst_img_folder, args.max_resolution, + args.divisible_by, args.interpolation, args.save_as_png, args.copy_associated_files) + + +if __name__ == '__main__': + main() diff --git a/resize_lora.py b/resize_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..18326437303c163dacf44c981d1871b9d8750d11 --- /dev/null +++ b/resize_lora.py @@ -0,0 +1,425 @@ +# Convert LoRA to different rank approximation (should only be used to go to lower rank) +# This code is based off the extract_lora_from_models.py file which is based on https://github.com/cloneofsimo/lora/blob/develop/lora_diffusion/cli_svd.py +# Thanks to cloneofsimo + +import os +import argparse +import torch +from safetensors.torch import load_file, save_file, safe_open +from tqdm import tqdm +import numpy as np + +from library import train_util +from library import model_util +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +MIN_SV = 1e-6 + +LORA_DOWN_UP_FORMATS = [ + ("lora_down", "lora_up"), # sd-scripts LoRA + ("lora_A", "lora_B"), # PEFT LoRA + ("down", "up"), # ControlLoRA +] + + +# Model save and load functions + + +def load_state_dict(file_name, dtype): + if model_util.is_safetensors(file_name): + sd = load_file(file_name) + with safe_open(file_name, framework="pt") as f: + metadata = f.metadata() + else: + sd = torch.load(file_name, map_location="cpu") + metadata = None + + for key in list(sd.keys()): + if type(sd[key]) == torch.Tensor: + sd[key] = sd[key].to(dtype) + + return sd, metadata + + +def save_to_file(file_name, state_dict, metadata): + if model_util.is_safetensors(file_name): + save_file(state_dict, file_name, metadata) + else: + torch.save(state_dict, file_name) + + +# Indexing functions + + +def index_sv_cumulative(S, target): + original_sum = float(torch.sum(S)) + cumulative_sums = torch.cumsum(S, dim=0) / original_sum + index = int(torch.searchsorted(cumulative_sums, target)) + 1 + index = max(1, min(index, len(S) - 1)) + + return index + + +def index_sv_fro(S, target): + S_squared = S.pow(2) + S_fro_sq = float(torch.sum(S_squared)) + sum_S_squared = torch.cumsum(S_squared, dim=0) / S_fro_sq + index = int(torch.searchsorted(sum_S_squared, target**2)) + 1 + index = max(1, min(index, len(S) - 1)) + + return index + + +def index_sv_ratio(S, target): + max_sv = S[0] + min_sv = max_sv / target + index = int(torch.sum(S > min_sv).item()) + index = max(1, min(index, len(S) - 1)) + + return index + + +# Modified from Kohaku-blueleaf's extract/merge functions +def extract_conv(weight, lora_rank, dynamic_method, dynamic_param, device, scale=1): + out_size, in_size, kernel_size, _ = weight.size() + U, S, Vh = torch.linalg.svd(weight.reshape(out_size, -1).to(device)) + + param_dict = rank_resize(S, lora_rank, dynamic_method, dynamic_param, scale) + lora_rank = param_dict["new_rank"] + + U = U[:, :lora_rank] + S = S[:lora_rank] + U = U @ torch.diag(S) + Vh = Vh[:lora_rank, :] + + param_dict["lora_down"] = Vh.reshape(lora_rank, in_size, kernel_size, kernel_size).cpu() + param_dict["lora_up"] = U.reshape(out_size, lora_rank, 1, 1).cpu() + del U, S, Vh, weight + return param_dict + + +def extract_linear(weight, lora_rank, dynamic_method, dynamic_param, device, scale=1): + out_size, in_size = weight.size() + + U, S, Vh = torch.linalg.svd(weight.to(device)) + + param_dict = rank_resize(S, lora_rank, dynamic_method, dynamic_param, scale) + lora_rank = param_dict["new_rank"] + + U = U[:, :lora_rank] + S = S[:lora_rank] + U = U @ torch.diag(S) + Vh = Vh[:lora_rank, :] + + param_dict["lora_down"] = Vh.reshape(lora_rank, in_size).cpu() + param_dict["lora_up"] = U.reshape(out_size, lora_rank).cpu() + del U, S, Vh, weight + return param_dict + + +def merge_conv(lora_down, lora_up, device): + in_rank, in_size, kernel_size, k_ = lora_down.shape + out_size, out_rank, _, _ = lora_up.shape + assert in_rank == out_rank and kernel_size == k_, f"rank {in_rank} {out_rank} or kernel {kernel_size} {k_} mismatch" + + lora_down = lora_down.to(device) + lora_up = lora_up.to(device) + + merged = lora_up.reshape(out_size, -1) @ lora_down.reshape(in_rank, -1) + weight = merged.reshape(out_size, in_size, kernel_size, kernel_size) + del lora_up, lora_down + return weight + + +def merge_linear(lora_down, lora_up, device): + in_rank, in_size = lora_down.shape + out_size, out_rank = lora_up.shape + assert in_rank == out_rank, f"rank {in_rank} {out_rank} mismatch" + + lora_down = lora_down.to(device) + lora_up = lora_up.to(device) + + weight = lora_up @ lora_down + del lora_up, lora_down + return weight + + +# Calculate new rank + + +def rank_resize(S, rank, dynamic_method, dynamic_param, scale=1): + param_dict = {} + + if dynamic_method == "sv_ratio": + # Calculate new dim and alpha based off ratio + new_rank = index_sv_ratio(S, dynamic_param) + 1 + new_alpha = float(scale * new_rank) + + elif dynamic_method == "sv_cumulative": + # Calculate new dim and alpha based off cumulative sum + new_rank = index_sv_cumulative(S, dynamic_param) + 1 + new_alpha = float(scale * new_rank) + + elif dynamic_method == "sv_fro": + # Calculate new dim and alpha based off sqrt sum of squares + new_rank = index_sv_fro(S, dynamic_param) + 1 + new_alpha = float(scale * new_rank) + else: + new_rank = rank + new_alpha = float(scale * new_rank) + + if S[0] <= MIN_SV: # Zero matrix, set dim to 1 + new_rank = 1 + new_alpha = float(scale * new_rank) + elif new_rank > rank: # cap max rank at rank + new_rank = rank + new_alpha = float(scale * new_rank) + + # Calculate resize info + s_sum = torch.sum(torch.abs(S)) + s_rank = torch.sum(torch.abs(S[:new_rank])) + + S_squared = S.pow(2) + s_fro = torch.sqrt(torch.sum(S_squared)) + s_red_fro = torch.sqrt(torch.sum(S_squared[:new_rank])) + fro_percent = float(s_red_fro / s_fro) + + param_dict["new_rank"] = new_rank + param_dict["new_alpha"] = new_alpha + param_dict["sum_retained"] = (s_rank) / s_sum + param_dict["fro_retained"] = fro_percent + param_dict["max_ratio"] = S[0] / S[new_rank - 1] + + return param_dict + + +def resize_lora_model(lora_sd, new_rank, new_conv_rank, save_dtype, device, dynamic_method, dynamic_param, verbose): + max_old_rank = None + new_alpha = None + verbose_str = "\n" + fro_list = [] + + if dynamic_method: + logger.info( + f"Dynamically determining new alphas and dims based off {dynamic_method}: {dynamic_param}, max rank is {new_rank}" + ) + + lora_down_weight = None + lora_up_weight = None + + o_lora_sd = lora_sd.copy() + block_down_name = None + block_up_name = None + + with torch.no_grad(): + for key, value in tqdm(lora_sd.items()): + key_parts = key.split(".") + block_down_name = None + for _format in LORA_DOWN_UP_FORMATS: + # Currently we only match lora_down_name in the last two parts of key + # because ("down", "up") are general words and may appear in block_down_name + if len(key_parts) >= 2 and _format[0] == key_parts[-2]: + block_down_name = ".".join(key_parts[:-2]) + lora_down_name = "." + _format[0] + lora_up_name = "." + _format[1] + weight_name = "." + key_parts[-1] + break + if len(key_parts) >= 1 and _format[0] == key_parts[-1]: + block_down_name = ".".join(key_parts[:-1]) + lora_down_name = "." + _format[0] + lora_up_name = "." + _format[1] + weight_name = "" + break + + if block_down_name is None: + # This parameter is not lora_down + continue + + # Now weight_name can be ".weight" or "" + # Find corresponding lora_up and alpha + block_up_name = block_down_name + lora_down_weight = value + lora_up_weight = lora_sd.get(block_up_name + lora_up_name + weight_name, None) + lora_alpha = lora_sd.get(block_down_name + ".alpha", None) + + weights_loaded = lora_down_weight is not None and lora_up_weight is not None + + if weights_loaded: + + conv2d = len(lora_down_weight.size()) == 4 + old_rank = lora_down_weight.size()[0] + max_old_rank = max(max_old_rank or 0, old_rank) + + if lora_alpha is None: + scale = 1.0 + else: + scale = lora_alpha / old_rank + + if conv2d: + full_weight_matrix = merge_conv(lora_down_weight, lora_up_weight, device) + param_dict = extract_conv(full_weight_matrix, new_conv_rank, dynamic_method, dynamic_param, device, scale) + else: + full_weight_matrix = merge_linear(lora_down_weight, lora_up_weight, device) + param_dict = extract_linear(full_weight_matrix, new_rank, dynamic_method, dynamic_param, device, scale) + + if verbose: + max_ratio = param_dict["max_ratio"] + sum_retained = param_dict["sum_retained"] + fro_retained = param_dict["fro_retained"] + if not np.isnan(fro_retained): + fro_list.append(float(fro_retained)) + + verbose_str += f"{block_down_name:75} | " + verbose_str += ( + f"sum(S) retained: {sum_retained:.1%}, fro retained: {fro_retained:.1%}, max(S) ratio: {max_ratio:0.1f}" + ) + + if verbose and dynamic_method: + verbose_str += f", dynamic | dim: {param_dict['new_rank']}, alpha: {param_dict['new_alpha']}\n" + else: + verbose_str += "\n" + + new_alpha = param_dict["new_alpha"] + o_lora_sd[block_down_name + lora_down_name + weight_name] = param_dict["lora_down"].to(save_dtype).contiguous() + o_lora_sd[block_up_name + lora_up_name + weight_name] = param_dict["lora_up"].to(save_dtype).contiguous() + o_lora_sd[block_down_name + ".alpha"] = torch.tensor(param_dict["new_alpha"]).to(save_dtype) + + block_down_name = None + block_up_name = None + lora_down_weight = None + lora_up_weight = None + weights_loaded = False + del param_dict + + if verbose: + print(verbose_str) + print(f"Average Frobenius norm retention: {np.mean(fro_list):.2%} | std: {np.std(fro_list):0.3f}") + logger.info("resizing complete") + return o_lora_sd, max_old_rank, new_alpha + + +def resize(args): + if args.save_to is None or not ( + args.save_to.endswith(".ckpt") + or args.save_to.endswith(".pt") + or args.save_to.endswith(".pth") + or args.save_to.endswith(".safetensors") + ): + raise Exception("The --save_to argument must be specified and must be a .ckpt , .pt, .pth or .safetensors file.") + + args.new_conv_rank = args.new_conv_rank if args.new_conv_rank is not None else args.new_rank + + def str_to_dtype(p): + if p == "float": + return torch.float + if p == "fp16": + return torch.float16 + if p == "bf16": + return torch.bfloat16 + return None + + if args.dynamic_method and not args.dynamic_param: + raise Exception("If using dynamic_method, then dynamic_param is required") + + merge_dtype = str_to_dtype("float") # matmul method above only seems to work in float32 + save_dtype = str_to_dtype(args.save_precision) + if save_dtype is None: + save_dtype = merge_dtype + + logger.info("loading Model...") + lora_sd, metadata = load_state_dict(args.model, merge_dtype) + + logger.info("Resizing Lora...") + state_dict, old_dim, new_alpha = resize_lora_model( + lora_sd, args.new_rank, args.new_conv_rank, save_dtype, args.device, args.dynamic_method, args.dynamic_param, args.verbose + ) + + # update metadata + if metadata is None: + metadata = {} + + comment = metadata.get("ss_training_comment", "") + + if not args.dynamic_method: + conv_desc = "" if args.new_rank == args.new_conv_rank else f" (conv: {args.new_conv_rank})" + metadata["ss_training_comment"] = f"dimension is resized from {old_dim} to {args.new_rank}{conv_desc}; {comment}" + metadata["ss_network_dim"] = str(args.new_rank) + metadata["ss_network_alpha"] = str(new_alpha) + else: + metadata["ss_training_comment"] = ( + f"Dynamic resize with {args.dynamic_method}: {args.dynamic_param} from {old_dim}; {comment}" + ) + metadata["ss_network_dim"] = "Dynamic" + metadata["ss_network_alpha"] = "Dynamic" + + # cast to save_dtype before calculating hashes + for key in list(state_dict.keys()): + value = state_dict[key] + if type(value) == torch.Tensor and value.dtype.is_floating_point and value.dtype != save_dtype: + state_dict[key] = value.to(save_dtype) + + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + logger.info(f"saving model to: {args.save_to}") + save_to_file(args.save_to, state_dict, metadata) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + + parser.add_argument( + "--save_precision", + type=str, + default=None, + choices=[None, "float", "fp16", "bf16"], + help="precision in saving, float if omitted / 保存時の精度、未指定時はfloat", + ) + parser.add_argument("--new_rank", type=int, default=4, help="Specify rank of output LoRA / 出力するLoRAのrank (dim)") + parser.add_argument( + "--new_conv_rank", + type=int, + default=None, + help="Specify rank of output LoRA for Conv2d 3x3, None for same as new_rank / 出力するConv2D 3x3 LoRAのrank (dim)、Noneでnew_rankと同じ", + ) + parser.add_argument( + "--save_to", + type=str, + default=None, + help="destination file name: ckpt or safetensors file / 保存先のファイル名、ckptまたはsafetensors", + ) + parser.add_argument( + "--model", + type=str, + default=None, + help="LoRA model to resize at to new rank: ckpt or safetensors file / 読み込むLoRAモデル、ckptまたはsafetensors", + ) + parser.add_argument( + "--device", type=str, default=None, help="device to use, cuda for GPU / 計算を行うデバイス、cuda でGPUを使う" + ) + parser.add_argument( + "--verbose", action="store_true", help="Display verbose resizing information / rank変更時の詳細情報を出力する" + ) + parser.add_argument( + "--dynamic_method", + type=str, + default=None, + choices=[None, "sv_ratio", "sv_fro", "sv_cumulative"], + help="Specify dynamic resizing method, --new_rank is used as a hard limit for max rank", + ) + parser.add_argument("--dynamic_param", type=float, default=None, help="Specify target for dynamic reduction") + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + resize(args) diff --git a/sai_model_spec.py b/sai_model_spec.py new file mode 100644 index 0000000000000000000000000000000000000000..a63bd82ecf91f9bc13f28e4811d94252464d4a5b --- /dev/null +++ b/sai_model_spec.py @@ -0,0 +1,309 @@ +# based on https://github.com/Stability-AI/ModelSpec +import datetime +import hashlib +from io import BytesIO +import os +from typing import List, Optional, Tuple, Union +import safetensors +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +r""" +# Metadata Example +metadata = { + # === Must === + "modelspec.sai_model_spec": "1.0.0", # Required version ID for the spec + "modelspec.architecture": "stable-diffusion-xl-v1-base", # Architecture, reference the ID of the original model of the arch to match the ID + "modelspec.implementation": "sgm", + "modelspec.title": "Example Model Version 1.0", # Clean, human-readable title. May use your own phrasing/language/etc + # === Should === + "modelspec.author": "Example Corp", # Your name or company name + "modelspec.description": "This is my example model to show you how to do it!", # Describe the model in your own words/language/etc. Focus on what users need to know + "modelspec.date": "2023-07-20", # ISO-8601 compliant date of when the model was created + # === Can === + "modelspec.license": "ExampleLicense-1.0", # eg CreativeML Open RAIL, etc. + "modelspec.usage_hint": "Use keyword 'example'" # In your own language, very short hints about how the user should use the model +} +""" + +BASE_METADATA = { + # === Must === + "modelspec.sai_model_spec": "1.0.0", # Required version ID for the spec + "modelspec.architecture": None, + "modelspec.implementation": None, + "modelspec.title": None, + "modelspec.resolution": None, + # === Should === + "modelspec.description": None, + "modelspec.author": None, + "modelspec.date": None, + # === Can === + "modelspec.license": None, + "modelspec.tags": None, + "modelspec.merged_from": None, + "modelspec.prediction_type": None, + "modelspec.timestep_range": None, + "modelspec.encoder_layer": None, +} + +# 別に使うやつだけ定義 +MODELSPEC_TITLE = "modelspec.title" + +ARCH_SD_V1 = "stable-diffusion-v1" +ARCH_SD_V2_512 = "stable-diffusion-v2-512" +ARCH_SD_V2_768_V = "stable-diffusion-v2-768-v" +ARCH_SD_XL_V1_BASE = "stable-diffusion-xl-v1-base" + +ADAPTER_LORA = "lora" +ADAPTER_TEXTUAL_INVERSION = "textual-inversion" + +IMPL_STABILITY_AI = "https://github.com/Stability-AI/generative-models" +IMPL_DIFFUSERS = "diffusers" + +PRED_TYPE_EPSILON = "epsilon" +PRED_TYPE_V = "v" + + +def load_bytes_in_safetensors(tensors): + bytes = safetensors.torch.save(tensors) + b = BytesIO(bytes) + + b.seek(0) + header = b.read(8) + n = int.from_bytes(header, "little") + + offset = n + 8 + b.seek(offset) + + return b.read() + + +def precalculate_safetensors_hashes(state_dict): + # calculate each tensor one by one to reduce memory usage + hash_sha256 = hashlib.sha256() + for tensor in state_dict.values(): + single_tensor_sd = {"tensor": tensor} + bytes_for_tensor = load_bytes_in_safetensors(single_tensor_sd) + hash_sha256.update(bytes_for_tensor) + + return f"0x{hash_sha256.hexdigest()}" + + +def update_hash_sha256(metadata: dict, state_dict: dict): + raise NotImplementedError + + +def build_metadata( + state_dict: Optional[dict], + v2: bool, + v_parameterization: bool, + sdxl: bool, + lora: bool, + textual_inversion: bool, + timestamp: float, + title: Optional[str] = None, + reso: Optional[Union[int, Tuple[int, int]]] = None, + is_stable_diffusion_ckpt: Optional[bool] = None, + author: Optional[str] = None, + description: Optional[str] = None, + license: Optional[str] = None, + tags: Optional[str] = None, + merged_from: Optional[str] = None, + timesteps: Optional[Tuple[int, int]] = None, + clip_skip: Optional[int] = None, +): + # if state_dict is None, hash is not calculated + + metadata = {} + metadata.update(BASE_METADATA) + + # TODO メモリを消費せずかつ正しいハッシュ計算の方法がわかったら実装する + # if state_dict is not None: + # hash = precalculate_safetensors_hashes(state_dict) + # metadata["modelspec.hash_sha256"] = hash + + if sdxl: + arch = ARCH_SD_XL_V1_BASE + elif v2: + if v_parameterization: + arch = ARCH_SD_V2_768_V + else: + arch = ARCH_SD_V2_512 + else: + arch = ARCH_SD_V1 + + if lora: + arch += f"/{ADAPTER_LORA}" + elif textual_inversion: + arch += f"/{ADAPTER_TEXTUAL_INVERSION}" + + metadata["modelspec.architecture"] = arch + + if not lora and not textual_inversion and is_stable_diffusion_ckpt is None: + is_stable_diffusion_ckpt = True # default is stable diffusion ckpt if not lora and not textual_inversion + + if (lora and sdxl) or textual_inversion or is_stable_diffusion_ckpt: + # Stable Diffusion ckpt, TI, SDXL LoRA + impl = IMPL_STABILITY_AI + else: + # v1/v2 LoRA or Diffusers + impl = IMPL_DIFFUSERS + metadata["modelspec.implementation"] = impl + + if title is None: + if lora: + title = "LoRA" + elif textual_inversion: + title = "TextualInversion" + else: + title = "Checkpoint" + title += f"@{timestamp}" + metadata[MODELSPEC_TITLE] = title + + if author is not None: + metadata["modelspec.author"] = author + else: + del metadata["modelspec.author"] + + if description is not None: + metadata["modelspec.description"] = description + else: + del metadata["modelspec.description"] + + if merged_from is not None: + metadata["modelspec.merged_from"] = merged_from + else: + del metadata["modelspec.merged_from"] + + if license is not None: + metadata["modelspec.license"] = license + else: + del metadata["modelspec.license"] + + if tags is not None: + metadata["modelspec.tags"] = tags + else: + del metadata["modelspec.tags"] + + # remove microsecond from time + int_ts = int(timestamp) + + # time to iso-8601 compliant date + date = datetime.datetime.fromtimestamp(int_ts).isoformat() + metadata["modelspec.date"] = date + + if reso is not None: + # comma separated to tuple + if isinstance(reso, str): + reso = tuple(map(int, reso.split(","))) + if len(reso) == 1: + reso = (reso[0], reso[0]) + else: + # resolution is defined in dataset, so use default + if sdxl: + reso = 1024 + elif v2 and v_parameterization: + reso = 768 + else: + reso = 512 + if isinstance(reso, int): + reso = (reso, reso) + + metadata["modelspec.resolution"] = f"{reso[0]}x{reso[1]}" + + if v_parameterization: + metadata["modelspec.prediction_type"] = PRED_TYPE_V + else: + metadata["modelspec.prediction_type"] = PRED_TYPE_EPSILON + + if timesteps is not None: + if isinstance(timesteps, str) or isinstance(timesteps, int): + timesteps = (timesteps, timesteps) + if len(timesteps) == 1: + timesteps = (timesteps[0], timesteps[0]) + metadata["modelspec.timestep_range"] = f"{timesteps[0]},{timesteps[1]}" + else: + del metadata["modelspec.timestep_range"] + + if clip_skip is not None: + metadata["modelspec.encoder_layer"] = f"{clip_skip}" + else: + del metadata["modelspec.encoder_layer"] + + # # assert all values are filled + # assert all([v is not None for v in metadata.values()]), metadata + if not all([v is not None for v in metadata.values()]): + logger.error(f"Internal error: some metadata values are None: {metadata}") + + return metadata + + +# region utils + + +def get_title(metadata: dict) -> Optional[str]: + return metadata.get(MODELSPEC_TITLE, None) + + +def load_metadata_from_safetensors(model: str) -> dict: + if not model.endswith(".safetensors"): + return {} + + with safetensors.safe_open(model, framework="pt") as f: + metadata = f.metadata() + if metadata is None: + metadata = {} + return metadata + + +def build_merged_from(models: List[str]) -> str: + def get_title(model: str): + metadata = load_metadata_from_safetensors(model) + title = metadata.get(MODELSPEC_TITLE, None) + if title is None: + title = os.path.splitext(os.path.basename(model))[0] # use filename + return title + + titles = [get_title(model) for model in models] + return ", ".join(titles) + + +# endregion + + +r""" +if __name__ == "__main__": + import argparse + import torch + from safetensors.torch import load_file + from library import train_util + + parser = argparse.ArgumentParser() + parser.add_argument("--ckpt", type=str, required=True) + args = parser.parse_args() + + print(f"Loading {args.ckpt}") + state_dict = load_file(args.ckpt) + + print(f"Calculating metadata") + metadata = get(state_dict, False, False, False, False, "sgm", False, False, "title", "date", 256, 1000, 0) + print(metadata) + del state_dict + + # by reference implementation + with open(args.ckpt, mode="rb") as file_data: + file_hash = hashlib.sha256() + head_len = struct.unpack("Q", file_data.read(8)) # int64 header length prefix + header = json.loads(file_data.read(head_len[0])) # header itself, json string + content = ( + file_data.read() + ) # All other content is tightly packed tensors. Copy to RAM for simplicity, but you can avoid this read with a more careful FS-dependent impl. + file_hash.update(content) + # ===== Update the hash for modelspec ===== + by_ref = f"0x{file_hash.hexdigest()}" + print(by_ref) + print("is same?", by_ref == metadata["modelspec.hash_sha256"]) + +""" diff --git a/sdxl_lpw_stable_diffusion.py b/sdxl_lpw_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..03b182566c9ad0be52bccf69b311f1e2e0d3be70 --- /dev/null +++ b/sdxl_lpw_stable_diffusion.py @@ -0,0 +1,1347 @@ +# copy from https://github.com/huggingface/diffusers/blob/main/examples/community/lpw_stable_diffusion.py +# and modify to support SD2.x + +import inspect +import re +from typing import Callable, List, Optional, Union + +import numpy as np +import PIL.Image +import torch +from packaging import version +from tqdm import tqdm +from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer + +from diffusers import SchedulerMixin, StableDiffusionPipeline +from diffusers.models import AutoencoderKL, UNet2DConditionModel +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker +from diffusers.utils import logging +from PIL import Image + +from library import sdxl_model_util, sdxl_train_util, train_util + + +try: + from diffusers.utils import PIL_INTERPOLATION +except ImportError: + if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"): + PIL_INTERPOLATION = { + "linear": PIL.Image.Resampling.BILINEAR, + "bilinear": PIL.Image.Resampling.BILINEAR, + "bicubic": PIL.Image.Resampling.BICUBIC, + "lanczos": PIL.Image.Resampling.LANCZOS, + "nearest": PIL.Image.Resampling.NEAREST, + } + else: + PIL_INTERPOLATION = { + "linear": PIL.Image.LINEAR, + "bilinear": PIL.Image.BILINEAR, + "bicubic": PIL.Image.BICUBIC, + "lanczos": PIL.Image.LANCZOS, + "nearest": PIL.Image.NEAREST, + } +# ------------------------------------------------------------------------------ + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +re_attention = re.compile( + r""" +\\\(| +\\\)| +\\\[| +\\]| +\\\\| +\\| +\(| +\[| +:([+-]?[.\d]+)\)| +\)| +]| +[^\\()\[\]:]+| +: +""", + re.X, +) + + +def parse_prompt_attention(text): + """ + Parses a string with attention tokens and returns a list of pairs: text and its associated weight. + Accepted tokens are: + (abc) - increases attention to abc by a multiplier of 1.1 + (abc:3.12) - increases attention to abc by a multiplier of 3.12 + [abc] - decreases attention to abc by a multiplier of 1.1 + \( - literal character '(' + \[ - literal character '[' + \) - literal character ')' + \] - literal character ']' + \\ - literal character '\' + anything else - just text + >>> parse_prompt_attention('normal text') + [['normal text', 1.0]] + >>> parse_prompt_attention('an (important) word') + [['an ', 1.0], ['important', 1.1], [' word', 1.0]] + >>> parse_prompt_attention('(unbalanced') + [['unbalanced', 1.1]] + >>> parse_prompt_attention('\(literal\]') + [['(literal]', 1.0]] + >>> parse_prompt_attention('(unnecessary)(parens)') + [['unnecessaryparens', 1.1]] + >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).') + [['a ', 1.0], + ['house', 1.5730000000000004], + [' ', 1.1], + ['on', 1.0], + [' a ', 1.1], + ['hill', 0.55], + [', sun, ', 1.1], + ['sky', 1.4641000000000006], + ['.', 1.1]] + """ + + res = [] + round_brackets = [] + square_brackets = [] + + round_bracket_multiplier = 1.1 + square_bracket_multiplier = 1 / 1.1 + + def multiply_range(start_position, multiplier): + for p in range(start_position, len(res)): + res[p][1] *= multiplier + + for m in re_attention.finditer(text): + text = m.group(0) + weight = m.group(1) + + if text.startswith("\\"): + res.append([text[1:], 1.0]) + elif text == "(": + round_brackets.append(len(res)) + elif text == "[": + square_brackets.append(len(res)) + elif weight is not None and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), float(weight)) + elif text == ")" and len(round_brackets) > 0: + multiply_range(round_brackets.pop(), round_bracket_multiplier) + elif text == "]" and len(square_brackets) > 0: + multiply_range(square_brackets.pop(), square_bracket_multiplier) + else: + res.append([text, 1.0]) + + for pos in round_brackets: + multiply_range(pos, round_bracket_multiplier) + + for pos in square_brackets: + multiply_range(pos, square_bracket_multiplier) + + if len(res) == 0: + res = [["", 1.0]] + + # merge runs of identical weights + i = 0 + while i + 1 < len(res): + if res[i][1] == res[i + 1][1]: + res[i][0] += res[i + 1][0] + res.pop(i + 1) + else: + i += 1 + + return res + + +def get_prompts_with_weights(pipe: StableDiffusionPipeline, prompt: List[str], max_length: int): + r""" + Tokenize a list of prompts and return its tokens with weights of each token. + + No padding, starting or ending token is included. + """ + tokens = [] + weights = [] + truncated = False + for text in prompt: + texts_and_weights = parse_prompt_attention(text) + text_token = [] + text_weight = [] + for word, weight in texts_and_weights: + # tokenize and discard the starting and the ending token + token = pipe.tokenizer(word).input_ids[1:-1] + text_token += token + # copy the weight by length of token + text_weight += [weight] * len(token) + # stop if the text is too long (longer than truncation limit) + if len(text_token) > max_length: + truncated = True + break + # truncate + if len(text_token) > max_length: + truncated = True + text_token = text_token[:max_length] + text_weight = text_weight[:max_length] + tokens.append(text_token) + weights.append(text_weight) + if truncated: + logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples") + return tokens, weights + + +def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, pad, no_boseos_middle=True, chunk_length=77): + r""" + Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length. + """ + max_embeddings_multiples = (max_length - 2) // (chunk_length - 2) + weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length + for i in range(len(tokens)): + tokens[i] = [bos] + tokens[i] + [eos] + [pad] * (max_length - 2 - len(tokens[i])) + if no_boseos_middle: + weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i])) + else: + w = [] + if len(weights[i]) == 0: + w = [1.0] * weights_length + else: + for j in range(max_embeddings_multiples): + w.append(1.0) # weight for starting token in this chunk + w += weights[i][j * (chunk_length - 2) : min(len(weights[i]), (j + 1) * (chunk_length - 2))] + w.append(1.0) # weight for ending token in this chunk + w += [1.0] * (weights_length - len(w)) + weights[i] = w[:] + + return tokens, weights + + +def get_hidden_states(text_encoder, input_ids, is_sdxl_text_encoder2: bool, eos_token_id, device): + if not is_sdxl_text_encoder2: + # text_encoder1: same as SD1/2 + enc_out = text_encoder(input_ids.to(text_encoder.device), output_hidden_states=True, return_dict=True) + hidden_states = enc_out["hidden_states"][11] + pool = None + else: + # text_encoder2 + enc_out = text_encoder(input_ids.to(text_encoder.device), output_hidden_states=True, return_dict=True) + hidden_states = enc_out["hidden_states"][-2] # penuultimate layer + # pool = enc_out["text_embeds"] + pool = train_util.pool_workaround(text_encoder, enc_out["last_hidden_state"], input_ids, eos_token_id) + hidden_states = hidden_states.to(device) + if pool is not None: + pool = pool.to(device) + return hidden_states, pool + + +def get_unweighted_text_embeddings( + pipe: StableDiffusionPipeline, + text_input: torch.Tensor, + chunk_length: int, + clip_skip: int, + eos: int, + pad: int, + is_sdxl_text_encoder2: bool, + no_boseos_middle: Optional[bool] = True, +): + """ + When the length of tokens is a multiple of the capacity of the text encoder, + it should be split into chunks and sent to the text encoder individually. + """ + max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2) + text_pool = None + if max_embeddings_multiples > 1: + text_embeddings = [] + for i in range(max_embeddings_multiples): + # extract the i-th chunk + text_input_chunk = text_input[:, i * (chunk_length - 2) : (i + 1) * (chunk_length - 2) + 2].clone() + + # cover the head and the tail by the starting and the ending tokens + text_input_chunk[:, 0] = text_input[0, 0] + if pad == eos: # v1 + text_input_chunk[:, -1] = text_input[0, -1] + else: # v2 + for j in range(len(text_input_chunk)): + if text_input_chunk[j, -1] != eos and text_input_chunk[j, -1] != pad: # 最後に普通の文字がある + text_input_chunk[j, -1] = eos + if text_input_chunk[j, 1] == pad: # BOSだけであとはPAD + text_input_chunk[j, 1] = eos + + text_embedding, current_text_pool = get_hidden_states( + pipe.text_encoder, text_input_chunk, is_sdxl_text_encoder2, eos, pipe.device + ) + if text_pool is None: + text_pool = current_text_pool + + if no_boseos_middle: + if i == 0: + # discard the ending token + text_embedding = text_embedding[:, :-1] + elif i == max_embeddings_multiples - 1: + # discard the starting token + text_embedding = text_embedding[:, 1:] + else: + # discard both starting and ending tokens + text_embedding = text_embedding[:, 1:-1] + + text_embeddings.append(text_embedding) + text_embeddings = torch.concat(text_embeddings, axis=1) + else: + text_embeddings, text_pool = get_hidden_states(pipe.text_encoder, text_input, is_sdxl_text_encoder2, eos, pipe.device) + return text_embeddings, text_pool + + +def get_weighted_text_embeddings( + pipe, # : SdxlStableDiffusionLongPromptWeightingPipeline, + prompt: Union[str, List[str]], + uncond_prompt: Optional[Union[str, List[str]]] = None, + max_embeddings_multiples: Optional[int] = 3, + no_boseos_middle: Optional[bool] = False, + skip_parsing: Optional[bool] = False, + skip_weighting: Optional[bool] = False, + clip_skip=None, + is_sdxl_text_encoder2=False, +): + r""" + Prompts can be assigned with local weights using brackets. For example, + prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful', + and the embedding tokens corresponding to the words get multiplied by a constant, 1.1. + + Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean. + + Args: + pipe (`StableDiffusionPipeline`): + Pipe to provide access to the tokenizer and the text encoder. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + uncond_prompt (`str` or `List[str]`): + The unconditional prompt or prompts for guide the image generation. If unconditional prompt + is provided, the embeddings of prompt and uncond_prompt are concatenated. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + no_boseos_middle (`bool`, *optional*, defaults to `False`): + If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and + ending token in each of the chunk in the middle. + skip_parsing (`bool`, *optional*, defaults to `False`): + Skip the parsing of brackets. + skip_weighting (`bool`, *optional*, defaults to `False`): + Skip the weighting. When the parsing is skipped, it is forced True. + """ + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + if isinstance(prompt, str): + prompt = [prompt] + + if not skip_parsing: + prompt_tokens, prompt_weights = get_prompts_with_weights(pipe, prompt, max_length - 2) + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens, uncond_weights = get_prompts_with_weights(pipe, uncond_prompt, max_length - 2) + else: + prompt_tokens = [token[1:-1] for token in pipe.tokenizer(prompt, max_length=max_length, truncation=True).input_ids] + prompt_weights = [[1.0] * len(token) for token in prompt_tokens] + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens = [ + token[1:-1] for token in pipe.tokenizer(uncond_prompt, max_length=max_length, truncation=True).input_ids + ] + uncond_weights = [[1.0] * len(token) for token in uncond_tokens] + + # round up the longest length of tokens to a multiple of (model_max_length - 2) + max_length = max([len(token) for token in prompt_tokens]) + if uncond_prompt is not None: + max_length = max(max_length, max([len(token) for token in uncond_tokens])) + + max_embeddings_multiples = min( + max_embeddings_multiples, + (max_length - 1) // (pipe.tokenizer.model_max_length - 2) + 1, + ) + max_embeddings_multiples = max(1, max_embeddings_multiples) + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + + # pad the length of tokens and weights + bos = pipe.tokenizer.bos_token_id + eos = pipe.tokenizer.eos_token_id + pad = pipe.tokenizer.pad_token_id + prompt_tokens, prompt_weights = pad_tokens_and_weights( + prompt_tokens, + prompt_weights, + max_length, + bos, + eos, + pad, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, + ) + prompt_tokens = torch.tensor(prompt_tokens, dtype=torch.long, device=pipe.device) + if uncond_prompt is not None: + uncond_tokens, uncond_weights = pad_tokens_and_weights( + uncond_tokens, + uncond_weights, + max_length, + bos, + eos, + pad, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, + ) + uncond_tokens = torch.tensor(uncond_tokens, dtype=torch.long, device=pipe.device) + + # get the embeddings + text_embeddings, text_pool = get_unweighted_text_embeddings( + pipe, + prompt_tokens, + pipe.tokenizer.model_max_length, + clip_skip, + eos, + pad, + is_sdxl_text_encoder2, + no_boseos_middle=no_boseos_middle, + ) + prompt_weights = torch.tensor(prompt_weights, dtype=text_embeddings.dtype, device=pipe.device) + + if uncond_prompt is not None: + uncond_embeddings, uncond_pool = get_unweighted_text_embeddings( + pipe, + uncond_tokens, + pipe.tokenizer.model_max_length, + clip_skip, + eos, + pad, + is_sdxl_text_encoder2, + no_boseos_middle=no_boseos_middle, + ) + uncond_weights = torch.tensor(uncond_weights, dtype=uncond_embeddings.dtype, device=pipe.device) + + # assign weights to the prompts and normalize in the sense of mean + # TODO: should we normalize by chunk or in a whole (current implementation)? + if (not skip_parsing) and (not skip_weighting): + previous_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= prompt_weights.unsqueeze(-1) + current_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + if uncond_prompt is not None: + previous_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= uncond_weights.unsqueeze(-1) + current_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + + if uncond_prompt is not None: + return text_embeddings, text_pool, uncond_embeddings, uncond_pool + return text_embeddings, text_pool, None, None + + +def preprocess_image(image): + w, h = image.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask, scale_factor=8): + mask = mask.convert("L") + w, h = mask.size + w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 + mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"]) + mask = np.array(mask).astype(np.float32) / 255.0 + mask = np.tile(mask, (4, 1, 1)) + mask = mask[None].transpose(0, 1, 2, 3) # what does this step do? + mask = 1 - mask # repaint white, keep black + mask = torch.from_numpy(mask) + return mask + + +def prepare_controlnet_image( + image: PIL.Image.Image, + width: int, + height: int, + batch_size: int, + num_images_per_prompt: int, + device: torch.device, + dtype: torch.dtype, + do_classifier_free_guidance: bool = False, + guess_mode: bool = False, +): + if not isinstance(image, torch.Tensor): + if isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + images = [] + + for image_ in image: + image_ = image_.convert("RGB") + image_ = image_.resize((width, height), resample=PIL_INTERPOLATION["lanczos"]) + image_ = np.array(image_) + image_ = image_[None, :] + images.append(image_) + + image = images + + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + +class SdxlStableDiffusionLongPromptWeightingPipeline: + r""" + Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing + weighting in prompt. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details. + feature_extractor ([`CLIPFeatureExtractor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + # if version.parse(version.parse(diffusers.__version__).base_version) >= version.parse("0.9.0"): + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: List[CLIPTextModel], + tokenizer: List[CLIPTokenizer], + unet: UNet2DConditionModel, + scheduler: SchedulerMixin, + # clip_skip: int, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPFeatureExtractor, + requires_safety_checker: bool = True, + clip_skip: int = 1, + ): + # clip skip is ignored currently + self.tokenizer = tokenizer[0] + self.text_encoder = text_encoder[0] + self.unet = unet + self.scheduler = scheduler + self.safety_checker = safety_checker + self.feature_extractor = feature_extractor + self.requires_safety_checker = requires_safety_checker + self.vae = vae + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.progress_bar = lambda x: tqdm(x, leave=False) + + self.clip_skip = clip_skip + self.tokenizers = tokenizer + self.text_encoders = text_encoder + + # self.__init__additional__() + + # def __init__additional__(self): + # if not hasattr(self, "vae_scale_factor"): + # setattr(self, "vae_scale_factor", 2 ** (len(self.vae.config.block_out_channels) - 1)) + + def to(self, device=None, dtype=None): + if device is not None: + self.device = device + # self.vae.to(device=self.device) + if dtype is not None: + self.dtype = dtype + + # do not move Text Encoders to device, because Text Encoder should be on CPU + + @property + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"): + return self.device + for module in self.unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + max_embeddings_multiples, + is_sdxl_text_encoder2, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + """ + batch_size = len(prompt) if isinstance(prompt, list) else 1 + + if negative_prompt is None: + negative_prompt = [""] * batch_size + elif isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + if batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + text_embeddings, text_pool, uncond_embeddings, uncond_pool = get_weighted_text_embeddings( + pipe=self, + prompt=prompt, + uncond_prompt=negative_prompt if do_classifier_free_guidance else None, + max_embeddings_multiples=max_embeddings_multiples, + clip_skip=self.clip_skip, + is_sdxl_text_encoder2=is_sdxl_text_encoder2, + ) + bs_embed, seq_len, _ = text_embeddings.shape + text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1) # ?? + text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + if text_pool is not None: + text_pool = text_pool.repeat(1, num_images_per_prompt) + text_pool = text_pool.view(bs_embed * num_images_per_prompt, -1) + + if do_classifier_free_guidance: + bs_embed, seq_len, _ = uncond_embeddings.shape + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1) + uncond_embeddings = uncond_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + if uncond_pool is not None: + uncond_pool = uncond_pool.repeat(1, num_images_per_prompt) + uncond_pool = uncond_pool.view(bs_embed * num_images_per_prompt, -1) + + return text_embeddings, text_pool, uncond_embeddings, uncond_pool + + return text_embeddings, text_pool, None, None + + def check_inputs(self, prompt, height, width, strength, callback_steps): + if not isinstance(prompt, str) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" f" {type(callback_steps)}." + ) + + def get_timesteps(self, num_inference_steps, strength, device, is_text2img): + if is_text2img: + return self.scheduler.timesteps.to(device), num_inference_steps + else: + # get the original timestep using init_timestep + offset = self.scheduler.config.get("steps_offset", 0) + init_timestep = int(num_inference_steps * strength) + offset + init_timestep = min(init_timestep, num_inference_steps) + + t_start = max(num_inference_steps - init_timestep + offset, 0) + timesteps = self.scheduler.timesteps[t_start:].to(device) + return timesteps, num_inference_steps - t_start + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker(images=image, clip_input=safety_checker_input.pixel_values.to(dtype)) + else: + has_nsfw_concept = None + return image, has_nsfw_concept + + def decode_latents(self, latents): + with torch.no_grad(): + latents = 1 / sdxl_model_util.VAE_SCALE_FACTOR * latents + + # print("post_quant_conv dtype:", self.vae.post_quant_conv.weight.dtype) # torch.float32 + # x = torch.nn.functional.conv2d(latents, self.vae.post_quant_conv.weight.detach(), stride=1, padding=0) + # print("latents dtype:", latents.dtype, "x dtype:", x.dtype) # torch.float32, torch.float16 + # self.vae.to("cpu") + # self.vae.set_use_memory_efficient_attention_xformers(False) + # image = self.vae.decode(latents.to("cpu")).sample + + image = self.vae.decode(latents.to(self.vae.dtype)).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def prepare_latents(self, image, timestep, batch_size, height, width, dtype, device, generator, latents=None): + if image is None: + shape = ( + batch_size, + self.unet.in_channels, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + + if latents is None: + if device.type == "mps": + # randn does not work reproducibly on mps + latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + latents = torch.randn(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents, None, None + else: + init_latent_dist = self.vae.encode(image).latent_dist + init_latents = init_latent_dist.sample(generator=generator) + init_latents = sdxl_model_util.VAE_SCALE_FACTOR * init_latents + init_latents = torch.cat([init_latents] * batch_size, dim=0) + init_latents_orig = init_latents + shape = init_latents.shape + + # add noise to latents using the timesteps + if device.type == "mps": + noise = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device) + else: + noise = torch.randn(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.add_noise(init_latents, noise, timestep) + return latents, init_latents_orig, noise + + @torch.no_grad() + def __call__( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + image: Union[torch.FloatTensor, PIL.Image.Image] = None, + mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + strength: float = 0.8, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + controlnet=None, + controlnet_image=None, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. + `image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + controlnet (`diffusers.ControlNetModel`, *optional*): + A controlnet model to be used for the inference. If not provided, controlnet will be disabled. + controlnet_image (`torch.FloatTensor` or `PIL.Image.Image`, *optional*): + `Image`, or tensor representing an image batch, to be used as the starting point for the controlnet + inference. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + `None` if cancelled by `is_cancelled_callback`, + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + if controlnet is not None and controlnet_image is None: + raise ValueError("controlnet_image must be provided if controlnet is not None.") + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, height, width, strength, callback_steps) + + # 2. Define call parameters + batch_size = 1 if isinstance(prompt, str) else len(prompt) + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + # 実装を簡単にするためにtokenzer/text encoderを切り替えて二回呼び出す + # To simplify the implementation, switch the tokenzer/text encoder and call it twice + text_embeddings_list = [] + text_pool = None + uncond_embeddings_list = [] + uncond_pool = None + for i in range(len(self.tokenizers)): + self.tokenizer = self.tokenizers[i] + self.text_encoder = self.text_encoders[i] + + text_embeddings, tp1, uncond_embeddings, up1 = self._encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + max_embeddings_multiples, + is_sdxl_text_encoder2=i == 1, + ) + text_embeddings_list.append(text_embeddings) + uncond_embeddings_list.append(uncond_embeddings) + + if tp1 is not None: + text_pool = tp1 + if up1 is not None: + uncond_pool = up1 + + unet_dtype = self.unet.dtype + dtype = unet_dtype + if hasattr(dtype, "itemsize") and dtype.itemsize == 1: # fp8 + dtype = torch.float16 + self.unet.to(dtype) + + # 4. Preprocess image and mask + if isinstance(image, PIL.Image.Image): + image = preprocess_image(image) + if image is not None: + image = image.to(device=self.device, dtype=dtype) + if isinstance(mask_image, PIL.Image.Image): + mask_image = preprocess_mask(mask_image, self.vae_scale_factor) + if mask_image is not None: + mask = mask_image.to(device=self.device, dtype=dtype) + mask = torch.cat([mask] * batch_size * num_images_per_prompt) + else: + mask = None + + # ControlNet is not working yet in SDXL, but keep the code here for future use + if controlnet_image is not None: + controlnet_image = prepare_controlnet_image( + controlnet_image, width, height, batch_size, 1, self.device, controlnet.dtype, do_classifier_free_guidance, False + ) + + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device, image is None) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents, init_latents_orig, noise = self.prepare_latents( + image, + latent_timestep, + batch_size * num_images_per_prompt, + height, + width, + dtype, + device, + generator, + latents, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # create size embs and concat embeddings for SDXL + orig_size = torch.tensor([height, width]).repeat(batch_size * num_images_per_prompt, 1).to(dtype) + crop_size = torch.zeros_like(orig_size) + target_size = orig_size + embs = sdxl_train_util.get_size_embeddings(orig_size, crop_size, target_size, device).to(dtype) + + # make conditionings + if do_classifier_free_guidance: + text_embeddings = torch.cat(text_embeddings_list, dim=2) + uncond_embeddings = torch.cat(uncond_embeddings_list, dim=2) + text_embedding = torch.cat([uncond_embeddings, text_embeddings]).to(dtype) + + cond_vector = torch.cat([text_pool, embs], dim=1) + uncond_vector = torch.cat([uncond_pool, embs], dim=1) + vector_embedding = torch.cat([uncond_vector, cond_vector]).to(dtype) + else: + text_embedding = torch.cat(text_embeddings_list, dim=2).to(dtype) + vector_embedding = torch.cat([text_pool, embs], dim=1).to(dtype) + + # 8. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + unet_additional_args = {} + if controlnet is not None: + down_block_res_samples, mid_block_res_sample = controlnet( + latent_model_input, + t, + encoder_hidden_states=text_embeddings, + controlnet_cond=controlnet_image, + conditioning_scale=1.0, + guess_mode=False, + return_dict=False, + ) + unet_additional_args["down_block_additional_residuals"] = down_block_res_samples + unet_additional_args["mid_block_additional_residual"] = mid_block_res_sample + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, text_embedding, vector_embedding) + noise_pred = noise_pred.to(dtype) # U-Net changes dtype in LoRA training + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + if mask is not None: + # masking + init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t])) + latents = (init_latents_proper * mask) + (latents * (1 - mask)) + + # call the callback, if provided + if i % callback_steps == 0: + if callback is not None: + callback(i, t, latents) + if is_cancelled_callback is not None and is_cancelled_callback(): + return None + + self.unet.to(unet_dtype) + return latents + + def latents_to_image(self, latents): + # 9. Post-processing + image = self.decode_latents(latents.to(self.vae.dtype)) + image = self.numpy_to_pil(image) + return image + + # copy from pil_utils.py + def numpy_to_pil(self, images: np.ndarray) -> Image.Image: + """ + Convert a numpy image or a batch of images to a PIL image. + """ + if images.ndim == 3: + images = images[None, ...] + images = (images * 255).round().astype("uint8") + if images.shape[-1] == 1: + # special case for grayscale (single channel) images + pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images] + else: + pil_images = [Image.fromarray(image) for image in images] + + return pil_images + + def text2img( + self, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + height: int = 512, + width: int = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[torch.Generator] = None, + latents: Optional[torch.FloatTensor] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function for text-to-image generation. + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + height=height, + width=width, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + latents=latents, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, + ) + + def img2img( + self, + image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[torch.Generator] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function for image-to-image generation. + Args: + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. + `image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + image=image, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + strength=strength, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, + ) + + def inpaint( + self, + image: Union[torch.FloatTensor, PIL.Image.Image], + mask_image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + strength: float = 0.8, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, + num_images_per_prompt: Optional[int] = 1, + eta: Optional[float] = 0.0, + generator: Optional[torch.Generator] = None, + max_embeddings_multiples: Optional[int] = 3, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, + ): + r""" + Function for inpaint. + Args: + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. This is the image whose masked region will be inpainted. + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength` + is 1, the denoising process will be run on the masked area for the full number of iterations specified + in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more + noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur. + num_inference_steps (`int`, *optional*, defaults to 50): + The reference number of denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. This parameter will be modulated by `strength`, as explained above. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation + deterministic. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + return self.__call__( + prompt=prompt, + negative_prompt=negative_prompt, + image=image, + mask_image=mask_image, + num_inference_steps=num_inference_steps, + guidance_scale=guidance_scale, + strength=strength, + num_images_per_prompt=num_images_per_prompt, + eta=eta, + generator=generator, + max_embeddings_multiples=max_embeddings_multiples, + output_type=output_type, + return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, + ) diff --git a/sdxl_merge_lora.py b/sdxl_merge_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..b147eb4462cb9e97a2b14b7289148fe698972cce --- /dev/null +++ b/sdxl_merge_lora.py @@ -0,0 +1,513 @@ +import itertools +import math +import argparse +import os +import time +import concurrent.futures +import torch +from safetensors.torch import load_file, save_file +from tqdm import tqdm +from library import sai_model_spec, sdxl_model_util, train_util +import library.model_util as model_util +import lora +import oft +from svd_merge_lora import format_lbws, get_lbw_block_index, LAYER26 +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + + +def load_state_dict(file_name, dtype): + if os.path.splitext(file_name)[1] == ".safetensors": + sd = load_file(file_name) + metadata = train_util.load_metadata_from_safetensors(file_name) + else: + sd = torch.load(file_name, map_location="cpu") + metadata = {} + + for key in list(sd.keys()): + if type(sd[key]) == torch.Tensor: + sd[key] = sd[key].to(dtype) + + return sd, metadata + + +def save_to_file(file_name, model, metadata): + if os.path.splitext(file_name)[1] == ".safetensors": + save_file(model, file_name, metadata=metadata) + else: + torch.save(model, file_name) + + +def detect_method_from_training_model(models, dtype): + for model in models: + # TODO It is better to use key names to detect the method + lora_sd, _ = load_state_dict(model, dtype) + for key in tqdm(lora_sd.keys()): + if "lora_up" in key or "lora_down" in key: + return "LoRA" + elif "oft_blocks" in key: + return "OFT" + + +def merge_to_sd_model(text_encoder1, text_encoder2, unet, models, ratios, lbws, merge_dtype): + text_encoder1.to(merge_dtype) + text_encoder2.to(merge_dtype) + unet.to(merge_dtype) + + # detect the method: OFT or LoRA_module + method = detect_method_from_training_model(models, merge_dtype) + logger.info(f"method:{method}") + + if lbws: + lbws, _, LBW_TARGET_IDX = format_lbws(lbws) + else: + LBW_TARGET_IDX = [] + + # create module map + name_to_module = {} + for i, root_module in enumerate([text_encoder1, text_encoder2, unet]): + if method == "LoRA": + if i <= 1: + if i == 0: + prefix = lora.LoRANetwork.LORA_PREFIX_TEXT_ENCODER1 + else: + prefix = lora.LoRANetwork.LORA_PREFIX_TEXT_ENCODER2 + target_replace_modules = lora.LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE + else: + prefix = lora.LoRANetwork.LORA_PREFIX_UNET + target_replace_modules = ( + lora.LoRANetwork.UNET_TARGET_REPLACE_MODULE + lora.LoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + ) + elif method == "OFT": + prefix = oft.OFTNetwork.OFT_PREFIX_UNET + # ALL_LINEAR includes ATTN_ONLY, so we don't need to specify ATTN_ONLY + target_replace_modules = ( + oft.OFTNetwork.UNET_TARGET_REPLACE_MODULE_ALL_LINEAR + oft.OFTNetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 + ) + + for name, module in root_module.named_modules(): + if module.__class__.__name__ in target_replace_modules: + for child_name, child_module in module.named_modules(): + if child_module.__class__.__name__ == "Linear" or child_module.__class__.__name__ == "Conv2d": + lora_name = prefix + "." + name + "." + child_name + lora_name = lora_name.replace(".", "_") + name_to_module[lora_name] = child_module + + for model, ratio, lbw in itertools.zip_longest(models, ratios, lbws): + logger.info(f"loading: {model}") + lora_sd, _ = load_state_dict(model, merge_dtype) + + logger.info(f"merging...") + + if lbw: + lbw_weights = [1] * 26 + for index, value in zip(LBW_TARGET_IDX, lbw): + lbw_weights[index] = value + logger.info(f"lbw: {dict(zip(LAYER26.keys(), lbw_weights))}") + + if method == "LoRA": + for key in tqdm(lora_sd.keys()): + if "lora_down" in key: + up_key = key.replace("lora_down", "lora_up") + alpha_key = key[: key.index("lora_down")] + "alpha" + + # find original module for this lora + module_name = ".".join(key.split(".")[:-2]) # remove trailing ".lora_down.weight" + if module_name not in name_to_module: + logger.info(f"no module found for LoRA weight: {key}") + continue + module = name_to_module[module_name] + # logger.info(f"apply {key} to {module}") + + down_weight = lora_sd[key] + up_weight = lora_sd[up_key] + + dim = down_weight.size()[0] + alpha = lora_sd.get(alpha_key, dim) + scale = alpha / dim + + if lbw: + index = get_lbw_block_index(key, True) + is_lbw_target = index in LBW_TARGET_IDX + if is_lbw_target: + scale *= lbw_weights[index] # keyがlbwの対象であれば、lbwの重みを掛ける + + # W <- W + U * D + weight = module.weight + # logger.info(module_name, down_weight.size(), up_weight.size()) + if len(weight.size()) == 2: + # linear + weight = weight + ratio * (up_weight @ down_weight) * scale + elif down_weight.size()[2:4] == (1, 1): + # conv2d 1x1 + weight = ( + weight + + ratio + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * scale + ) + else: + # conv2d 3x3 + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + # logger.info(conved.size(), weight.size(), module.stride, module.padding) + weight = weight + ratio * conved * scale + + module.weight = torch.nn.Parameter(weight) + + elif method == "OFT": + + device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + for key in tqdm(lora_sd.keys()): + if "oft_blocks" in key: + oft_blocks = lora_sd[key] + dim = oft_blocks.shape[0] + break + for key in tqdm(lora_sd.keys()): + if "alpha" in key: + oft_blocks = lora_sd[key] + alpha = oft_blocks.item() + break + + def merge_to(key): + if "alpha" in key: + return + + # find original module for this OFT + module_name = ".".join(key.split(".")[:-1]) + if module_name not in name_to_module: + logger.info(f"no module found for OFT weight: {key}") + return + module = name_to_module[module_name] + + # logger.info(f"apply {key} to {module}") + + oft_blocks = lora_sd[key] + + if isinstance(module, torch.nn.Linear): + out_dim = module.out_features + elif isinstance(module, torch.nn.Conv2d): + out_dim = module.out_channels + + num_blocks = dim + block_size = out_dim // dim + constraint = (0 if alpha is None else alpha) * out_dim + + multiplier = 1 + if lbw: + index = get_lbw_block_index(key, False) + is_lbw_target = index in LBW_TARGET_IDX + if is_lbw_target: + multiplier *= lbw_weights[index] + + block_Q = oft_blocks - oft_blocks.transpose(1, 2) + norm_Q = torch.norm(block_Q.flatten()) + new_norm_Q = torch.clamp(norm_Q, max=constraint) + block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8)) + I = torch.eye(block_size, device=oft_blocks.device).unsqueeze(0).repeat(num_blocks, 1, 1) + block_R = torch.matmul(I + block_Q, (I - block_Q).inverse()) + block_R_weighted = multiplier * block_R + (1 - multiplier) * I + R = torch.block_diag(*block_R_weighted) + + # get org weight + org_sd = module.state_dict() + org_weight = org_sd["weight"].to(device) + + R = R.to(org_weight.device, dtype=org_weight.dtype) + + if org_weight.dim() == 4: + weight = torch.einsum("oihw, op -> pihw", org_weight, R) + else: + weight = torch.einsum("oi, op -> pi", org_weight, R) + + weight = weight.contiguous() # Make Tensor contiguous; required due to ThreadPoolExecutor + + module.weight = torch.nn.Parameter(weight) + + # TODO multi-threading may cause OOM on CPU if cpu_count is too high and RAM is not enough + max_workers = 1 if device.type != "cpu" else None # avoid OOM on GPU + with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor: + list(tqdm(executor.map(merge_to, lora_sd.keys()), total=len(lora_sd.keys()))) + + +def merge_lora_models(models, ratios, lbws, merge_dtype, concat=False, shuffle=False): + base_alphas = {} # alpha for merged model + base_dims = {} + + # detect the method: OFT or LoRA_module + method = detect_method_from_training_model(models, merge_dtype) + if method == "OFT": + raise ValueError( + "OFT model is not supported for merging OFT models. / OFTモデルはOFTモデル同士のマージには対応していません" + ) + + if lbws: + lbws, _, LBW_TARGET_IDX = format_lbws(lbws) + else: + LBW_TARGET_IDX = [] + + merged_sd = {} + v2 = None + base_model = None + for model, ratio, lbw in itertools.zip_longest(models, ratios, lbws): + logger.info(f"loading: {model}") + lora_sd, lora_metadata = load_state_dict(model, merge_dtype) + + if lbw: + lbw_weights = [1] * 26 + for index, value in zip(LBW_TARGET_IDX, lbw): + lbw_weights[index] = value + logger.info(f"lbw: {dict(zip(LAYER26.keys(), lbw_weights))}") + + if lora_metadata is not None: + if v2 is None: + v2 = lora_metadata.get(train_util.SS_METADATA_KEY_V2, None) # returns string, SDXLはv2がないのでFalseのはず + if base_model is None: + base_model = lora_metadata.get(train_util.SS_METADATA_KEY_BASE_MODEL_VERSION, None) + + # get alpha and dim + alphas = {} # alpha for current model + dims = {} # dims for current model + for key in lora_sd.keys(): + if "alpha" in key: + lora_module_name = key[: key.rfind(".alpha")] + alpha = float(lora_sd[key].detach().numpy()) + alphas[lora_module_name] = alpha + if lora_module_name not in base_alphas: + base_alphas[lora_module_name] = alpha + elif "lora_down" in key: + lora_module_name = key[: key.rfind(".lora_down")] + dim = lora_sd[key].size()[0] + dims[lora_module_name] = dim + if lora_module_name not in base_dims: + base_dims[lora_module_name] = dim + + for lora_module_name in dims.keys(): + if lora_module_name not in alphas: + alpha = dims[lora_module_name] + alphas[lora_module_name] = alpha + if lora_module_name not in base_alphas: + base_alphas[lora_module_name] = alpha + + logger.info(f"dim: {list(set(dims.values()))}, alpha: {list(set(alphas.values()))}") + + # merge + logger.info(f"merging...") + for key in tqdm(lora_sd.keys()): + if "alpha" in key: + continue + + if "lora_up" in key and concat: + concat_dim = 1 + elif "lora_down" in key and concat: + concat_dim = 0 + else: + concat_dim = None + + lora_module_name = key[: key.rfind(".lora_")] + + base_alpha = base_alphas[lora_module_name] + alpha = alphas[lora_module_name] + + scale = math.sqrt(alpha / base_alpha) * ratio + scale = abs(scale) if "lora_up" in key else scale # マイナスの重みに対応する。 + + if lbw: + index = get_lbw_block_index(key, True) + is_lbw_target = index in LBW_TARGET_IDX + if is_lbw_target: + scale *= lbw_weights[index] # keyがlbwの対象であれば、lbwの重みを掛ける + + if key in merged_sd: + assert ( + merged_sd[key].size() == lora_sd[key].size() or concat_dim is not None + ), f"weights shape mismatch merging v1 and v2, different dims? / 重みのサイズが合いません。v1とv2、または次元数の異なるモデルはマージできません" + if concat_dim is not None: + merged_sd[key] = torch.cat([merged_sd[key], lora_sd[key] * scale], dim=concat_dim) + else: + merged_sd[key] = merged_sd[key] + lora_sd[key] * scale + else: + merged_sd[key] = lora_sd[key] * scale + + # set alpha to sd + for lora_module_name, alpha in base_alphas.items(): + key = lora_module_name + ".alpha" + merged_sd[key] = torch.tensor(alpha) + if shuffle: + key_down = lora_module_name + ".lora_down.weight" + key_up = lora_module_name + ".lora_up.weight" + dim = merged_sd[key_down].shape[0] + perm = torch.randperm(dim) + merged_sd[key_down] = merged_sd[key_down][perm] + merged_sd[key_up] = merged_sd[key_up][:, perm] + + logger.info("merged model") + logger.info(f"dim: {list(set(base_dims.values()))}, alpha: {list(set(base_alphas.values()))}") + + # check all dims are same + dims_list = list(set(base_dims.values())) + alphas_list = list(set(base_alphas.values())) + all_same_dims = True + all_same_alphas = True + for dims in dims_list: + if dims != dims_list[0]: + all_same_dims = False + break + for alphas in alphas_list: + if alphas != alphas_list[0]: + all_same_alphas = False + break + + # build minimum metadata + dims = f"{dims_list[0]}" if all_same_dims else "Dynamic" + alphas = f"{alphas_list[0]}" if all_same_alphas else "Dynamic" + metadata = train_util.build_minimum_network_metadata(v2, base_model, "networks.lora", dims, alphas, None) + + return merged_sd, metadata + + +def merge(args): + assert len(args.models) == len( + args.ratios + ), f"number of models must be equal to number of ratios / モデルの数と重みの数は合わせてください" + if args.lbws: + assert len(args.models) == len( + args.lbws + ), f"number of models must be equal to number of ratios / モデルの数と層別適用率の数は合わせてください" + else: + args.lbws = [] # zip_longestで扱えるようにlbws未使用時には空のリストにしておく + + def str_to_dtype(p): + if p == "float": + return torch.float + if p == "fp16": + return torch.float16 + if p == "bf16": + return torch.bfloat16 + return None + + merge_dtype = str_to_dtype(args.precision) + save_dtype = str_to_dtype(args.save_precision) + if save_dtype is None: + save_dtype = merge_dtype + + if args.sd_model is not None: + logger.info(f"loading SD model: {args.sd_model}") + + ( + text_model1, + text_model2, + vae, + unet, + logit_scale, + ckpt_info, + ) = sdxl_model_util.load_models_from_sdxl_checkpoint(sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, args.sd_model, "cpu") + + merge_to_sd_model(text_model1, text_model2, unet, args.models, args.ratios, args.lbws, merge_dtype) + + if args.no_metadata: + sai_metadata = None + else: + merged_from = sai_model_spec.build_merged_from([args.sd_model] + args.models) + title = os.path.splitext(os.path.basename(args.save_to))[0] + sai_metadata = sai_model_spec.build_metadata( + None, False, False, True, False, False, time.time(), title=title, merged_from=merged_from + ) + + logger.info(f"saving SD model to: {args.save_to}") + sdxl_model_util.save_stable_diffusion_checkpoint( + args.save_to, text_model1, text_model2, unet, 0, 0, ckpt_info, vae, logit_scale, sai_metadata, save_dtype + ) + else: + state_dict, metadata = merge_lora_models(args.models, args.ratios, args.lbws, merge_dtype, args.concat, args.shuffle) + + # cast to save_dtype before calculating hashes + for key in list(state_dict.keys()): + value = state_dict[key] + if type(value) == torch.Tensor and value.dtype.is_floating_point and value.dtype != save_dtype: + state_dict[key] = value.to(save_dtype) + + logger.info(f"calculating hashes and creating metadata...") + + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + if not args.no_metadata: + merged_from = sai_model_spec.build_merged_from(args.models) + title = os.path.splitext(os.path.basename(args.save_to))[0] + sai_metadata = sai_model_spec.build_metadata( + state_dict, False, False, True, True, False, time.time(), title=title, merged_from=merged_from + ) + metadata.update(sai_metadata) + + logger.info(f"saving model to: {args.save_to}") + save_to_file(args.save_to, state_dict, metadata) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument( + "--save_precision", + type=str, + default=None, + choices=[None, "float", "fp16", "bf16"], + help="precision in saving, same to merging if omitted / 保存時に精度を変更して保存する、省略時はマージ時の精度と同じ", + ) + parser.add_argument( + "--precision", + type=str, + default="float", + choices=["float", "fp16", "bf16"], + help="precision in merging (float is recommended) / マージの計算時の精度（floatを推奨）", + ) + parser.add_argument( + "--sd_model", + type=str, + default=None, + help="Stable Diffusion model to load: ckpt or safetensors file, merge LoRA models if omitted / 読み込むモデル、ckptまたはsafetensors。省略時はLoRAモデル同士をマージする", + ) + parser.add_argument( + "--save_to", + type=str, + default=None, + help="destination file name: ckpt or safetensors file / 保存先のファイル名、ckptまたはsafetensors", + ) + parser.add_argument( + "--models", + type=str, + nargs="*", + help="LoRA models to merge: ckpt or safetensors file / マージするLoRAモデル、ckptまたはsafetensors", + ) + parser.add_argument("--ratios", type=float, nargs="*", help="ratios for each model / それぞれのLoRAモデルの比率") + parser.add_argument("--lbws", type=str, nargs="*", help="lbw for each model / それぞれのLoRAモデルの層別適用率") + parser.add_argument( + "--no_metadata", + action="store_true", + help="do not save sai modelspec metadata (minimum ss_metadata for LoRA is saved) / " + + "sai modelspecのメタデータを保存しない（LoRAの最低限のss_metadataは保存される）", + ) + parser.add_argument( + "--concat", + action="store_true", + help="concat lora instead of merge (The dim(rank) of the output LoRA is the sum of the input dims) / " + + "マージの代わりに結合する（LoRAのdim(rank)は入力dimの合計になる）", + ) + parser.add_argument( + "--shuffle", + action="store_true", + help="shuffle lora weight./ " + "LoRAの重みをシャッフルする", + ) + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + merge(args) diff --git a/sdxl_model_util.py b/sdxl_model_util.py new file mode 100644 index 0000000000000000000000000000000000000000..4fad78a1c81aedcd27f980f453bf8127c10dd184 --- /dev/null +++ b/sdxl_model_util.py @@ -0,0 +1,583 @@ +import torch +import safetensors +from accelerate import init_empty_weights +from accelerate.utils.modeling import set_module_tensor_to_device +from safetensors.torch import load_file, save_file +from transformers import CLIPTextModel, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer +from typing import List +from diffusers import AutoencoderKL, EulerDiscreteScheduler, UNet2DConditionModel +from library import model_util +from library import sdxl_original_unet +from .utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +VAE_SCALE_FACTOR = 0.13025 +MODEL_VERSION_SDXL_BASE_V1_0 = "sdxl_base_v1-0" + +# Diffusersの設定を読み込むための参照モデル +DIFFUSERS_REF_MODEL_ID_SDXL = "stabilityai/stable-diffusion-xl-base-1.0" + +DIFFUSERS_SDXL_UNET_CONFIG = { + "act_fn": "silu", + "addition_embed_type": "text_time", + "addition_embed_type_num_heads": 64, + "addition_time_embed_dim": 256, + "attention_head_dim": [5, 10, 20], + "block_out_channels": [320, 640, 1280], + "center_input_sample": False, + "class_embed_type": None, + "class_embeddings_concat": False, + "conv_in_kernel": 3, + "conv_out_kernel": 3, + "cross_attention_dim": 2048, + "cross_attention_norm": None, + "down_block_types": ["DownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D"], + "downsample_padding": 1, + "dual_cross_attention": False, + "encoder_hid_dim": None, + "encoder_hid_dim_type": None, + "flip_sin_to_cos": True, + "freq_shift": 0, + "in_channels": 4, + "layers_per_block": 2, + "mid_block_only_cross_attention": None, + "mid_block_scale_factor": 1, + "mid_block_type": "UNetMidBlock2DCrossAttn", + "norm_eps": 1e-05, + "norm_num_groups": 32, + "num_attention_heads": None, + "num_class_embeds": None, + "only_cross_attention": False, + "out_channels": 4, + "projection_class_embeddings_input_dim": 2816, + "resnet_out_scale_factor": 1.0, + "resnet_skip_time_act": False, + "resnet_time_scale_shift": "default", + "sample_size": 128, + "time_cond_proj_dim": None, + "time_embedding_act_fn": None, + "time_embedding_dim": None, + "time_embedding_type": "positional", + "timestep_post_act": None, + "transformer_layers_per_block": [1, 2, 10], + "up_block_types": ["CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "UpBlock2D"], + "upcast_attention": False, + "use_linear_projection": True, +} + + +def convert_sdxl_text_encoder_2_checkpoint(checkpoint, max_length): + SDXL_KEY_PREFIX = "conditioner.embedders.1.model." + + # SD2のと、基本的には同じ。logit_scaleを後で使うので、それを追加で返す + # logit_scaleはcheckpointの保存時に使用する + def convert_key(key): + # common conversion + key = key.replace(SDXL_KEY_PREFIX + "transformer.", "text_model.encoder.") + key = key.replace(SDXL_KEY_PREFIX, "text_model.") + + if "resblocks" in key: + # resblocks conversion + key = key.replace(".resblocks.", ".layers.") + if ".ln_" in key: + key = key.replace(".ln_", ".layer_norm") + elif ".mlp." in key: + key = key.replace(".c_fc.", ".fc1.") + key = key.replace(".c_proj.", ".fc2.") + elif ".attn.out_proj" in key: + key = key.replace(".attn.out_proj.", ".self_attn.out_proj.") + elif ".attn.in_proj" in key: + key = None # 特殊なので後で処理する + else: + raise ValueError(f"unexpected key in SD: {key}") + elif ".positional_embedding" in key: + key = key.replace(".positional_embedding", ".embeddings.position_embedding.weight") + elif ".text_projection" in key: + key = key.replace("text_model.text_projection", "text_projection.weight") + elif ".logit_scale" in key: + key = None # 後で処理する + elif ".token_embedding" in key: + key = key.replace(".token_embedding.weight", ".embeddings.token_embedding.weight") + elif ".ln_final" in key: + key = key.replace(".ln_final", ".final_layer_norm") + # ckpt from comfy has this key: text_model.encoder.text_model.embeddings.position_ids + elif ".embeddings.position_ids" in key: + key = None # remove this key: position_ids is not used in newer transformers + return key + + keys = list(checkpoint.keys()) + new_sd = {} + for key in keys: + new_key = convert_key(key) + if new_key is None: + continue + new_sd[new_key] = checkpoint[key] + + # attnの変換 + for key in keys: + if ".resblocks" in key and ".attn.in_proj_" in key: + # 三つに分割 + values = torch.chunk(checkpoint[key], 3) + + key_suffix = ".weight" if "weight" in key else ".bias" + key_pfx = key.replace(SDXL_KEY_PREFIX + "transformer.resblocks.", "text_model.encoder.layers.") + key_pfx = key_pfx.replace("_weight", "") + key_pfx = key_pfx.replace("_bias", "") + key_pfx = key_pfx.replace(".attn.in_proj", ".self_attn.") + new_sd[key_pfx + "q_proj" + key_suffix] = values[0] + new_sd[key_pfx + "k_proj" + key_suffix] = values[1] + new_sd[key_pfx + "v_proj" + key_suffix] = values[2] + + # logit_scale はDiffusersには含まれないが、保存時に戻したいので別途返す + logit_scale = checkpoint.get(SDXL_KEY_PREFIX + "logit_scale", None) + + # temporary workaround for text_projection.weight.weight for Playground-v2 + if "text_projection.weight.weight" in new_sd: + logger.info("convert_sdxl_text_encoder_2_checkpoint: convert text_projection.weight.weight to text_projection.weight") + new_sd["text_projection.weight"] = new_sd["text_projection.weight.weight"] + del new_sd["text_projection.weight.weight"] + + return new_sd, logit_scale + + +# load state_dict without allocating new tensors +def _load_state_dict_on_device(model, state_dict, device, dtype=None): + # dtype will use fp32 as default + missing_keys = list(model.state_dict().keys() - state_dict.keys()) + unexpected_keys = list(state_dict.keys() - model.state_dict().keys()) + + # similar to model.load_state_dict() + if not missing_keys and not unexpected_keys: + for k in list(state_dict.keys()): + set_module_tensor_to_device(model, k, device, value=state_dict.pop(k), dtype=dtype) + return "" + + # error_msgs + error_msgs: List[str] = [] + if missing_keys: + error_msgs.insert(0, "Missing key(s) in state_dict: {}. ".format(", ".join('"{}"'.format(k) for k in missing_keys))) + if unexpected_keys: + error_msgs.insert(0, "Unexpected key(s) in state_dict: {}. ".format(", ".join('"{}"'.format(k) for k in unexpected_keys))) + + raise RuntimeError("Error(s) in loading state_dict for {}:\n\t{}".format(model.__class__.__name__, "\n\t".join(error_msgs))) + + +def load_models_from_sdxl_checkpoint(model_version, ckpt_path, map_location, dtype=None, disable_mmap=False): + # model_version is reserved for future use + # dtype is used for full_fp16/bf16 integration. Text Encoder will remain fp32, because it runs on CPU when caching + + # Load the state dict + if model_util.is_safetensors(ckpt_path): + checkpoint = None + if disable_mmap: + state_dict = safetensors.torch.load(open(ckpt_path, "rb").read()) + else: + try: + state_dict = load_file(ckpt_path, device=map_location) + except: + state_dict = load_file(ckpt_path) # prevent device invalid Error + epoch = None + global_step = None + else: + checkpoint = torch.load(ckpt_path, map_location=map_location) + if "state_dict" in checkpoint: + state_dict = checkpoint["state_dict"] + epoch = checkpoint.get("epoch", 0) + global_step = checkpoint.get("global_step", 0) + else: + state_dict = checkpoint + epoch = 0 + global_step = 0 + checkpoint = None + + # U-Net + logger.info("building U-Net") + with init_empty_weights(): + unet = sdxl_original_unet.SdxlUNet2DConditionModel() + + logger.info("loading U-Net from checkpoint") + unet_sd = {} + for k in list(state_dict.keys()): + if k.startswith("model.diffusion_model."): + unet_sd[k.replace("model.diffusion_model.", "")] = state_dict.pop(k) + info = _load_state_dict_on_device(unet, unet_sd, device=map_location, dtype=dtype) + logger.info(f"U-Net: {info}") + + # Text Encoders + logger.info("building text encoders") + + # Text Encoder 1 is same to Stability AI's SDXL + text_model1_cfg = CLIPTextConfig( + vocab_size=49408, + hidden_size=768, + intermediate_size=3072, + num_hidden_layers=12, + num_attention_heads=12, + max_position_embeddings=77, + hidden_act="quick_gelu", + layer_norm_eps=1e-05, + dropout=0.0, + attention_dropout=0.0, + initializer_range=0.02, + initializer_factor=1.0, + pad_token_id=1, + bos_token_id=0, + eos_token_id=2, + model_type="clip_text_model", + projection_dim=768, + # torch_dtype="float32", + # transformers_version="4.25.0.dev0", + ) + with init_empty_weights(): + text_model1 = CLIPTextModel._from_config(text_model1_cfg) + + # Text Encoder 2 is different from Stability AI's SDXL. SDXL uses open clip, but we use the model from HuggingFace. + # Note: Tokenizer from HuggingFace is different from SDXL. We must use open clip's tokenizer. + text_model2_cfg = CLIPTextConfig( + vocab_size=49408, + hidden_size=1280, + intermediate_size=5120, + num_hidden_layers=32, + num_attention_heads=20, + max_position_embeddings=77, + hidden_act="gelu", + layer_norm_eps=1e-05, + dropout=0.0, + attention_dropout=0.0, + initializer_range=0.02, + initializer_factor=1.0, + pad_token_id=1, + bos_token_id=0, + eos_token_id=2, + model_type="clip_text_model", + projection_dim=1280, + # torch_dtype="float32", + # transformers_version="4.25.0.dev0", + ) + with init_empty_weights(): + text_model2 = CLIPTextModelWithProjection(text_model2_cfg) + + logger.info("loading text encoders from checkpoint") + te1_sd = {} + te2_sd = {} + for k in list(state_dict.keys()): + if k.startswith("conditioner.embedders.0.transformer."): + te1_sd[k.replace("conditioner.embedders.0.transformer.", "")] = state_dict.pop(k) + elif k.startswith("conditioner.embedders.1.model."): + te2_sd[k] = state_dict.pop(k) + + # 最新の transformers では position_ids を含むとエラーになるので削除 / remove position_ids for latest transformers + if "text_model.embeddings.position_ids" in te1_sd: + te1_sd.pop("text_model.embeddings.position_ids") + + info1 = _load_state_dict_on_device(text_model1, te1_sd, device=map_location) # remain fp32 + logger.info(f"text encoder 1: {info1}") + + converted_sd, logit_scale = convert_sdxl_text_encoder_2_checkpoint(te2_sd, max_length=77) + info2 = _load_state_dict_on_device(text_model2, converted_sd, device=map_location) # remain fp32 + logger.info(f"text encoder 2: {info2}") + + # prepare vae + logger.info("building VAE") + vae_config = model_util.create_vae_diffusers_config() + with init_empty_weights(): + vae = AutoencoderKL(**vae_config) + + logger.info("loading VAE from checkpoint") + converted_vae_checkpoint = model_util.convert_ldm_vae_checkpoint(state_dict, vae_config) + info = _load_state_dict_on_device(vae, converted_vae_checkpoint, device=map_location, dtype=dtype) + logger.info(f"VAE: {info}") + + ckpt_info = (epoch, global_step) if epoch is not None else None + return text_model1, text_model2, vae, unet, logit_scale, ckpt_info + + +def make_unet_conversion_map(): + unet_conversion_map_layer = [] + + for i in range(3): # num_blocks is 3 in sdxl + # loop over downblocks/upblocks + for j in range(2): + # loop over resnets/attentions for downblocks + hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}." + sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0." + unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix)) + + if i < 3: + # no attention layers in down_blocks.3 + hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}." + sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1." + unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix)) + + for j in range(3): + # loop over resnets/attentions for upblocks + hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}." + sd_up_res_prefix = f"output_blocks.{3*i + j}.0." + unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix)) + + # if i > 0: commentout for sdxl + # no attention layers in up_blocks.0 + hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}." + sd_up_atn_prefix = f"output_blocks.{3*i + j}.1." + unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix)) + + if i < 3: + # no downsample in down_blocks.3 + hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv." + sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op." + unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix)) + + # no upsample in up_blocks.3 + hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0." + sd_upsample_prefix = f"output_blocks.{3*i + 2}.{2}." # change for sdxl + unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix)) + + hf_mid_atn_prefix = "mid_block.attentions.0." + sd_mid_atn_prefix = "middle_block.1." + unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix)) + + for j in range(2): + hf_mid_res_prefix = f"mid_block.resnets.{j}." + sd_mid_res_prefix = f"middle_block.{2*j}." + unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix)) + + unet_conversion_map_resnet = [ + # (stable-diffusion, HF Diffusers) + ("in_layers.0.", "norm1."), + ("in_layers.2.", "conv1."), + ("out_layers.0.", "norm2."), + ("out_layers.3.", "conv2."), + ("emb_layers.1.", "time_emb_proj."), + ("skip_connection.", "conv_shortcut."), + ] + + unet_conversion_map = [] + for sd, hf in unet_conversion_map_layer: + if "resnets" in hf: + for sd_res, hf_res in unet_conversion_map_resnet: + unet_conversion_map.append((sd + sd_res, hf + hf_res)) + else: + unet_conversion_map.append((sd, hf)) + + for j in range(2): + hf_time_embed_prefix = f"time_embedding.linear_{j+1}." + sd_time_embed_prefix = f"time_embed.{j*2}." + unet_conversion_map.append((sd_time_embed_prefix, hf_time_embed_prefix)) + + for j in range(2): + hf_label_embed_prefix = f"add_embedding.linear_{j+1}." + sd_label_embed_prefix = f"label_emb.0.{j*2}." + unet_conversion_map.append((sd_label_embed_prefix, hf_label_embed_prefix)) + + unet_conversion_map.append(("input_blocks.0.0.", "conv_in.")) + unet_conversion_map.append(("out.0.", "conv_norm_out.")) + unet_conversion_map.append(("out.2.", "conv_out.")) + + return unet_conversion_map + + +def convert_diffusers_unet_state_dict_to_sdxl(du_sd): + unet_conversion_map = make_unet_conversion_map() + + conversion_map = {hf: sd for sd, hf in unet_conversion_map} + return convert_unet_state_dict(du_sd, conversion_map) + + +def convert_unet_state_dict(src_sd, conversion_map): + converted_sd = {} + for src_key, value in src_sd.items(): + # さすがに全部回すのは時間がかかるので右から要素を削りつつprefixを探す + src_key_fragments = src_key.split(".")[:-1] # remove weight/bias + while len(src_key_fragments) > 0: + src_key_prefix = ".".join(src_key_fragments) + "." + if src_key_prefix in conversion_map: + converted_prefix = conversion_map[src_key_prefix] + converted_key = converted_prefix + src_key[len(src_key_prefix) :] + converted_sd[converted_key] = value + break + src_key_fragments.pop(-1) + assert len(src_key_fragments) > 0, f"key {src_key} not found in conversion map" + + return converted_sd + + +def convert_sdxl_unet_state_dict_to_diffusers(sd): + unet_conversion_map = make_unet_conversion_map() + + conversion_dict = {sd: hf for sd, hf in unet_conversion_map} + return convert_unet_state_dict(sd, conversion_dict) + + +def convert_text_encoder_2_state_dict_to_sdxl(checkpoint, logit_scale): + def convert_key(key): + # position_idsの除去 + if ".position_ids" in key: + return None + + # common + key = key.replace("text_model.encoder.", "transformer.") + key = key.replace("text_model.", "") + if "layers" in key: + # resblocks conversion + key = key.replace(".layers.", ".resblocks.") + if ".layer_norm" in key: + key = key.replace(".layer_norm", ".ln_") + elif ".mlp." in key: + key = key.replace(".fc1.", ".c_fc.") + key = key.replace(".fc2.", ".c_proj.") + elif ".self_attn.out_proj" in key: + key = key.replace(".self_attn.out_proj.", ".attn.out_proj.") + elif ".self_attn." in key: + key = None # 特殊なので後で処理する + else: + raise ValueError(f"unexpected key in DiffUsers model: {key}") + elif ".position_embedding" in key: + key = key.replace("embeddings.position_embedding.weight", "positional_embedding") + elif ".token_embedding" in key: + key = key.replace("embeddings.token_embedding.weight", "token_embedding.weight") + elif "text_projection" in key: # no dot in key + key = key.replace("text_projection.weight", "text_projection") + elif "final_layer_norm" in key: + key = key.replace("final_layer_norm", "ln_final") + return key + + keys = list(checkpoint.keys()) + new_sd = {} + for key in keys: + new_key = convert_key(key) + if new_key is None: + continue + new_sd[new_key] = checkpoint[key] + + # attnの変換 + for key in keys: + if "layers" in key and "q_proj" in key: + # 三つを結合 + key_q = key + key_k = key.replace("q_proj", "k_proj") + key_v = key.replace("q_proj", "v_proj") + + value_q = checkpoint[key_q] + value_k = checkpoint[key_k] + value_v = checkpoint[key_v] + value = torch.cat([value_q, value_k, value_v]) + + new_key = key.replace("text_model.encoder.layers.", "transformer.resblocks.") + new_key = new_key.replace(".self_attn.q_proj.", ".attn.in_proj_") + new_sd[new_key] = value + + if logit_scale is not None: + new_sd["logit_scale"] = logit_scale + + return new_sd + + +def save_stable_diffusion_checkpoint( + output_file, + text_encoder1, + text_encoder2, + unet, + epochs, + steps, + ckpt_info, + vae, + logit_scale, + metadata, + save_dtype=None, +): + state_dict = {} + + def update_sd(prefix, sd): + for k, v in sd.items(): + key = prefix + k + if save_dtype is not None: + v = v.detach().clone().to("cpu").to(save_dtype) + state_dict[key] = v + + # Convert the UNet model + update_sd("model.diffusion_model.", unet.state_dict()) + + # Convert the text encoders + update_sd("conditioner.embedders.0.transformer.", text_encoder1.state_dict()) + + text_enc2_dict = convert_text_encoder_2_state_dict_to_sdxl(text_encoder2.state_dict(), logit_scale) + update_sd("conditioner.embedders.1.model.", text_enc2_dict) + + # Convert the VAE + vae_dict = model_util.convert_vae_state_dict(vae.state_dict()) + update_sd("first_stage_model.", vae_dict) + + # Put together new checkpoint + key_count = len(state_dict.keys()) + new_ckpt = {"state_dict": state_dict} + + # epoch and global_step are sometimes not int + if ckpt_info is not None: + epochs += ckpt_info[0] + steps += ckpt_info[1] + + new_ckpt["epoch"] = epochs + new_ckpt["global_step"] = steps + + if model_util.is_safetensors(output_file): + save_file(state_dict, output_file, metadata) + else: + torch.save(new_ckpt, output_file) + + return key_count + + +def save_diffusers_checkpoint( + output_dir, text_encoder1, text_encoder2, unet, pretrained_model_name_or_path, vae=None, use_safetensors=False, save_dtype=None +): + from diffusers import StableDiffusionXLPipeline + + # convert U-Net + unet_sd = unet.state_dict() + du_unet_sd = convert_sdxl_unet_state_dict_to_diffusers(unet_sd) + + diffusers_unet = UNet2DConditionModel(**DIFFUSERS_SDXL_UNET_CONFIG) + if save_dtype is not None: + diffusers_unet.to(save_dtype) + diffusers_unet.load_state_dict(du_unet_sd) + + # create pipeline to save + if pretrained_model_name_or_path is None: + pretrained_model_name_or_path = DIFFUSERS_REF_MODEL_ID_SDXL + + scheduler = EulerDiscreteScheduler.from_pretrained(pretrained_model_name_or_path, subfolder="scheduler") + tokenizer1 = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer") + tokenizer2 = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer_2") + if vae is None: + vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae") + + # prevent local path from being saved + def remove_name_or_path(model): + if hasattr(model, "config"): + model.config._name_or_path = None + model.config._name_or_path = None + + remove_name_or_path(diffusers_unet) + remove_name_or_path(text_encoder1) + remove_name_or_path(text_encoder2) + remove_name_or_path(scheduler) + remove_name_or_path(tokenizer1) + remove_name_or_path(tokenizer2) + remove_name_or_path(vae) + + pipeline = StableDiffusionXLPipeline( + unet=diffusers_unet, + text_encoder=text_encoder1, + text_encoder_2=text_encoder2, + vae=vae, + scheduler=scheduler, + tokenizer=tokenizer1, + tokenizer_2=tokenizer2, + ) + if save_dtype is not None: + pipeline.to(None, save_dtype) + pipeline.save_pretrained(output_dir, safe_serialization=use_safetensors) diff --git a/sdxl_original_unet.py b/sdxl_original_unet.py new file mode 100644 index 0000000000000000000000000000000000000000..17c345a89ac7c178dedc2a14eddb5a14b0f41e6b --- /dev/null +++ b/sdxl_original_unet.py @@ -0,0 +1,1286 @@ +# Diffusersのコードをベースとした sd_xl_baseのU-Net +# state dictの形式をSDXLに合わせてある + +""" + target: sgm.modules.diffusionmodules.openaimodel.UNetModel + params: + adm_in_channels: 2816 + num_classes: sequential + use_checkpoint: True + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [4, 2] + num_res_blocks: 2 + channel_mult: [1, 2, 4] + num_head_channels: 64 + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: [1, 2, 10] # note: the first is unused (due to attn_res starting at 2) 32, 16, 8 --> 64, 32, 16 + context_dim: 2048 + spatial_transformer_attn_type: softmax-xformers + legacy: False +""" + +import math +from types import SimpleNamespace +from typing import Any, Optional +import torch +import torch.utils.checkpoint +from torch import nn +from torch.nn import functional as F +from einops import rearrange +from .utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +IN_CHANNELS: int = 4 +OUT_CHANNELS: int = 4 +ADM_IN_CHANNELS: int = 2816 +CONTEXT_DIM: int = 2048 +MODEL_CHANNELS: int = 320 +TIME_EMBED_DIM = 320 * 4 + +USE_REENTRANT = True + +# region memory efficient attention + +# FlashAttentionを使うCrossAttention +# based on https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/memory_efficient_attention_pytorch/flash_attention.py +# LICENSE MIT https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/LICENSE + +# constants + +EPSILON = 1e-6 + +# helper functions + + +def exists(val): + return val is not None + + +def default(val, d): + return val if exists(val) else d + + +# flash attention forwards and backwards + +# https://arxiv.org/abs/2205.14135 + + +class FlashAttentionFunction(torch.autograd.Function): + @staticmethod + @torch.no_grad() + def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size): + """Algorithm 2 in the paper""" + + device = q.device + dtype = q.dtype + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + o = torch.zeros_like(q) + all_row_sums = torch.zeros((*q.shape[:-1], 1), dtype=dtype, device=device) + all_row_maxes = torch.full((*q.shape[:-1], 1), max_neg_value, dtype=dtype, device=device) + + scale = q.shape[-1] ** -0.5 + + if not exists(mask): + mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size) + else: + mask = rearrange(mask, "b n -> b 1 1 n") + mask = mask.split(q_bucket_size, dim=-1) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + mask, + all_row_sums.split(q_bucket_size, dim=-2), + all_row_maxes.split(q_bucket_size, dim=-2), + ) + + for ind, (qc, oc, row_mask, row_sums, row_maxes) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale + + if exists(row_mask): + attn_weights.masked_fill_(~row_mask, max_neg_value) + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device).triu( + q_start_index - k_start_index + 1 + ) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + block_row_maxes = attn_weights.amax(dim=-1, keepdims=True) + attn_weights -= block_row_maxes + exp_weights = torch.exp(attn_weights) + + if exists(row_mask): + exp_weights.masked_fill_(~row_mask, 0.0) + + block_row_sums = exp_weights.sum(dim=-1, keepdims=True).clamp(min=EPSILON) + + new_row_maxes = torch.maximum(block_row_maxes, row_maxes) + + exp_values = torch.einsum("... i j, ... j d -> ... i d", exp_weights, vc) + + exp_row_max_diff = torch.exp(row_maxes - new_row_maxes) + exp_block_row_max_diff = torch.exp(block_row_maxes - new_row_maxes) + + new_row_sums = exp_row_max_diff * row_sums + exp_block_row_max_diff * block_row_sums + + oc.mul_((row_sums / new_row_sums) * exp_row_max_diff).add_((exp_block_row_max_diff / new_row_sums) * exp_values) + + row_maxes.copy_(new_row_maxes) + row_sums.copy_(new_row_sums) + + ctx.args = (causal, scale, mask, q_bucket_size, k_bucket_size) + ctx.save_for_backward(q, k, v, o, all_row_sums, all_row_maxes) + + return o + + @staticmethod + @torch.no_grad() + def backward(ctx, do): + """Algorithm 4 in the paper""" + + causal, scale, mask, q_bucket_size, k_bucket_size = ctx.args + q, k, v, o, l, m = ctx.saved_tensors + + device = q.device + + max_neg_value = -torch.finfo(q.dtype).max + qk_len_diff = max(k.shape[-2] - q.shape[-2], 0) + + dq = torch.zeros_like(q) + dk = torch.zeros_like(k) + dv = torch.zeros_like(v) + + row_splits = zip( + q.split(q_bucket_size, dim=-2), + o.split(q_bucket_size, dim=-2), + do.split(q_bucket_size, dim=-2), + mask, + l.split(q_bucket_size, dim=-2), + m.split(q_bucket_size, dim=-2), + dq.split(q_bucket_size, dim=-2), + ) + + for ind, (qc, oc, doc, row_mask, lc, mc, dqc) in enumerate(row_splits): + q_start_index = ind * q_bucket_size - qk_len_diff + + col_splits = zip( + k.split(k_bucket_size, dim=-2), + v.split(k_bucket_size, dim=-2), + dk.split(k_bucket_size, dim=-2), + dv.split(k_bucket_size, dim=-2), + ) + + for k_ind, (kc, vc, dkc, dvc) in enumerate(col_splits): + k_start_index = k_ind * k_bucket_size + + attn_weights = torch.einsum("... i d, ... j d -> ... i j", qc, kc) * scale + + if causal and q_start_index < (k_start_index + k_bucket_size - 1): + causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype=torch.bool, device=device).triu( + q_start_index - k_start_index + 1 + ) + attn_weights.masked_fill_(causal_mask, max_neg_value) + + exp_attn_weights = torch.exp(attn_weights - mc) + + if exists(row_mask): + exp_attn_weights.masked_fill_(~row_mask, 0.0) + + p = exp_attn_weights / lc + + dv_chunk = torch.einsum("... i j, ... i d -> ... j d", p, doc) + dp = torch.einsum("... i d, ... j d -> ... i j", doc, vc) + + D = (doc * oc).sum(dim=-1, keepdims=True) + ds = p * scale * (dp - D) + + dq_chunk = torch.einsum("... i j, ... j d -> ... i d", ds, kc) + dk_chunk = torch.einsum("... i j, ... i d -> ... j d", ds, qc) + + dqc.add_(dq_chunk) + dkc.add_(dk_chunk) + dvc.add_(dv_chunk) + + return dq, dk, dv, None, None, None, None + + +# endregion + + +def get_parameter_dtype(parameter: torch.nn.Module): + return next(parameter.parameters()).dtype + + +def get_parameter_device(parameter: torch.nn.Module): + return next(parameter.parameters()).device + + +def get_timestep_embedding( + timesteps: torch.Tensor, + embedding_dim: int, + downscale_freq_shift: float = 1, + scale: float = 1, + max_period: int = 10000, +): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings. + + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the + embeddings. :return: an [N x dim] Tensor of positional embeddings. + """ + assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" + + half_dim = embedding_dim // 2 + exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device) + exponent = exponent / (half_dim - downscale_freq_shift) + + emb = torch.exp(exponent) + emb = timesteps[:, None].float() * emb[None, :] + + # scale embeddings + emb = scale * emb + + # concat sine and cosine embeddings: flipped from Diffusers original ver because always flip_sin_to_cos=True + emb = torch.cat([torch.cos(emb), torch.sin(emb)], dim=-1) + + # zero pad + if embedding_dim % 2 == 1: + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +# Deep Shrink: We do not common this function, because minimize dependencies. +def resize_like(x, target, mode="bicubic", align_corners=False): + org_dtype = x.dtype + if org_dtype == torch.bfloat16: + x = x.to(torch.float32) + + if x.shape[-2:] != target.shape[-2:]: + if mode == "nearest": + x = F.interpolate(x, size=target.shape[-2:], mode=mode) + else: + x = F.interpolate(x, size=target.shape[-2:], mode=mode, align_corners=align_corners) + + if org_dtype == torch.bfloat16: + x = x.to(org_dtype) + return x + + +class GroupNorm32(nn.GroupNorm): + def forward(self, x): + if self.weight.dtype != torch.float32: + return super().forward(x) + return super().forward(x.float()).type(x.dtype) + + +class ResnetBlock2D(nn.Module): + def __init__( + self, + in_channels, + out_channels, + ): + super().__init__() + self.in_channels = in_channels + self.out_channels = out_channels + + self.in_layers = nn.Sequential( + GroupNorm32(32, in_channels), + nn.SiLU(), + nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1), + ) + + self.emb_layers = nn.Sequential(nn.SiLU(), nn.Linear(TIME_EMBED_DIM, out_channels)) + + self.out_layers = nn.Sequential( + GroupNorm32(32, out_channels), + nn.SiLU(), + nn.Identity(), # to make state_dict compatible with original model + nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1), + ) + + if in_channels != out_channels: + self.skip_connection = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0) + else: + self.skip_connection = nn.Identity() + + self.gradient_checkpointing = False + + def forward_body(self, x, emb): + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + h = h + emb_out[:, :, None, None] + h = self.out_layers(h) + x = self.skip_connection(x) + return x + h + + def forward(self, x, emb): + if self.training and self.gradient_checkpointing: + # logger.info("ResnetBlock2D: gradient_checkpointing") + + def create_custom_forward(func): + def custom_forward(*inputs): + return func(*inputs) + + return custom_forward + + x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.forward_body), x, emb, use_reentrant=USE_REENTRANT) + else: + x = self.forward_body(x, emb) + + return x + + +class Downsample2D(nn.Module): + def __init__(self, channels, out_channels): + super().__init__() + + self.channels = channels + self.out_channels = out_channels + + self.op = nn.Conv2d(self.channels, self.out_channels, 3, stride=2, padding=1) + + self.gradient_checkpointing = False + + def forward_body(self, hidden_states): + assert hidden_states.shape[1] == self.channels + hidden_states = self.op(hidden_states) + + return hidden_states + + def forward(self, hidden_states): + if self.training and self.gradient_checkpointing: + # logger.info("Downsample2D: gradient_checkpointing") + + def create_custom_forward(func): + def custom_forward(*inputs): + return func(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(self.forward_body), hidden_states, use_reentrant=USE_REENTRANT + ) + else: + hidden_states = self.forward_body(hidden_states) + + return hidden_states + + +class CrossAttention(nn.Module): + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + dim_head: int = 64, + upcast_attention: bool = False, + ): + super().__init__() + inner_dim = dim_head * heads + cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.upcast_attention = upcast_attention + + self.scale = dim_head**-0.5 + self.heads = heads + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=False) + self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=False) + + self.to_out = nn.ModuleList([]) + self.to_out.append(nn.Linear(inner_dim, query_dim)) + # no dropout here + + self.use_memory_efficient_attention_xformers = False + self.use_memory_efficient_attention_mem_eff = False + self.use_sdpa = False + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + self.use_memory_efficient_attention_xformers = xformers + self.use_memory_efficient_attention_mem_eff = mem_eff + + def set_use_sdpa(self, sdpa): + self.use_sdpa = sdpa + + def reshape_heads_to_batch_dim(self, tensor): + batch_size, seq_len, dim = tensor.shape + head_size = self.heads + tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size) + tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size) + return tensor + + def reshape_batch_dim_to_heads(self, tensor): + batch_size, seq_len, dim = tensor.shape + head_size = self.heads + tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim) + tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size) + return tensor + + def forward(self, hidden_states, context=None, mask=None): + if self.use_memory_efficient_attention_xformers: + return self.forward_memory_efficient_xformers(hidden_states, context, mask) + if self.use_memory_efficient_attention_mem_eff: + return self.forward_memory_efficient_mem_eff(hidden_states, context, mask) + if self.use_sdpa: + return self.forward_sdpa(hidden_states, context, mask) + + query = self.to_q(hidden_states) + context = context if context is not None else hidden_states + key = self.to_k(context) + value = self.to_v(context) + + query = self.reshape_heads_to_batch_dim(query) + key = self.reshape_heads_to_batch_dim(key) + value = self.reshape_heads_to_batch_dim(value) + + hidden_states = self._attention(query, key, value) + + # linear proj + hidden_states = self.to_out[0](hidden_states) + # hidden_states = self.to_out[1](hidden_states) # no dropout + return hidden_states + + def _attention(self, query, key, value): + if self.upcast_attention: + query = query.float() + key = key.float() + + attention_scores = torch.baddbmm( + torch.empty(query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device), + query, + key.transpose(-1, -2), + beta=0, + alpha=self.scale, + ) + attention_probs = attention_scores.softmax(dim=-1) + + # cast back to the original dtype + attention_probs = attention_probs.to(value.dtype) + + # compute attention output + hidden_states = torch.bmm(attention_probs, value) + + # reshape hidden_states + hidden_states = self.reshape_batch_dim_to_heads(hidden_states) + return hidden_states + + # TODO support Hypernetworks + def forward_memory_efficient_xformers(self, x, context=None, mask=None): + import xformers.ops + + h = self.heads + q_in = self.to_q(x) + context = context if context is not None else x + context = context.to(x.dtype) + k_in = self.to_k(context) + v_in = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b n h d", h=h), (q_in, k_in, v_in)) + del q_in, k_in, v_in + + q = q.contiguous() + k = k.contiguous() + v = v.contiguous() + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None) # 最適なのを選んでくれる + del q, k, v + + out = rearrange(out, "b n h d -> b n (h d)", h=h) + + out = self.to_out[0](out) + return out + + def forward_memory_efficient_mem_eff(self, x, context=None, mask=None): + flash_func = FlashAttentionFunction + + q_bucket_size = 512 + k_bucket_size = 1024 + + h = self.heads + q = self.to_q(x) + context = context if context is not None else x + context = context.to(x.dtype) + k = self.to_k(context) + v = self.to_v(context) + del context, x + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v)) + + out = flash_func.apply(q, k, v, mask, False, q_bucket_size, k_bucket_size) + + out = rearrange(out, "b h n d -> b n (h d)") + + out = self.to_out[0](out) + return out + + def forward_sdpa(self, x, context=None, mask=None): + h = self.heads + q_in = self.to_q(x) + context = context if context is not None else x + context = context.to(x.dtype) + k_in = self.to_k(context) + v_in = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q_in, k_in, v_in)) + del q_in, k_in, v_in + + out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False) + + out = rearrange(out, "b h n d -> b n (h d)", h=h) + + out = self.to_out[0](out) + return out + + +# feedforward +class GEGLU(nn.Module): + r""" + A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + """ + + def __init__(self, dim_in: int, dim_out: int): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def gelu(self, gate): + if gate.device.type != "mps": + return F.gelu(gate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype) + + def forward(self, hidden_states): + hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1) + return hidden_states * self.gelu(gate) + + +class FeedForward(nn.Module): + def __init__( + self, + dim: int, + ): + super().__init__() + inner_dim = int(dim * 4) # mult is always 4 + + self.net = nn.ModuleList([]) + # project in + self.net.append(GEGLU(dim, inner_dim)) + # project dropout + self.net.append(nn.Identity()) # nn.Dropout(0)) # dummy for dropout with 0 + # project out + self.net.append(nn.Linear(inner_dim, dim)) + + def forward(self, hidden_states): + for module in self.net: + hidden_states = module(hidden_states) + return hidden_states + + +class BasicTransformerBlock(nn.Module): + def __init__( + self, dim: int, num_attention_heads: int, attention_head_dim: int, cross_attention_dim: int, upcast_attention: bool = False + ): + super().__init__() + + self.gradient_checkpointing = False + + # 1. Self-Attn + self.attn1 = CrossAttention( + query_dim=dim, + cross_attention_dim=None, + heads=num_attention_heads, + dim_head=attention_head_dim, + upcast_attention=upcast_attention, + ) + self.ff = FeedForward(dim) + + # 2. Cross-Attn + self.attn2 = CrossAttention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + upcast_attention=upcast_attention, + ) + + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(dim) + + def set_use_memory_efficient_attention(self, xformers: bool, mem_eff: bool): + self.attn1.set_use_memory_efficient_attention(xformers, mem_eff) + self.attn2.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa: bool): + self.attn1.set_use_sdpa(sdpa) + self.attn2.set_use_sdpa(sdpa) + + def forward_body(self, hidden_states, context=None, timestep=None): + # 1. Self-Attention + norm_hidden_states = self.norm1(hidden_states) + + hidden_states = self.attn1(norm_hidden_states) + hidden_states + + # 2. Cross-Attention + norm_hidden_states = self.norm2(hidden_states) + hidden_states = self.attn2(norm_hidden_states, context=context) + hidden_states + + # 3. Feed-forward + hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states + + return hidden_states + + def forward(self, hidden_states, context=None, timestep=None): + if self.training and self.gradient_checkpointing: + # logger.info("BasicTransformerBlock: checkpointing") + + def create_custom_forward(func): + def custom_forward(*inputs): + return func(*inputs) + + return custom_forward + + output = torch.utils.checkpoint.checkpoint( + create_custom_forward(self.forward_body), hidden_states, context, timestep, use_reentrant=USE_REENTRANT + ) + else: + output = self.forward_body(hidden_states, context, timestep) + + return output + + +class Transformer2DModel(nn.Module): + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + use_linear_projection: bool = False, + upcast_attention: bool = False, + num_transformer_layers: int = 1, + ): + super().__init__() + self.in_channels = in_channels + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + self.use_linear_projection = use_linear_projection + + self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + # self.norm = GroupNorm32(32, in_channels, eps=1e-6, affine=True) + + if use_linear_projection: + self.proj_in = nn.Linear(in_channels, inner_dim) + else: + self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + + blocks = [] + for _ in range(num_transformer_layers): + blocks.append( + BasicTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + upcast_attention=upcast_attention, + ) + ) + + self.transformer_blocks = nn.ModuleList(blocks) + + if use_linear_projection: + self.proj_out = nn.Linear(in_channels, inner_dim) + else: + self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + + self.gradient_checkpointing = False + + def set_use_memory_efficient_attention(self, xformers, mem_eff): + for transformer in self.transformer_blocks: + transformer.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa): + for transformer in self.transformer_blocks: + transformer.set_use_sdpa(sdpa) + + def forward(self, hidden_states, encoder_hidden_states=None, timestep=None): + # 1. Input + batch, _, height, weight = hidden_states.shape + residual = hidden_states + + hidden_states = self.norm(hidden_states) + if not self.use_linear_projection: + hidden_states = self.proj_in(hidden_states) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) + hidden_states = self.proj_in(hidden_states) + + # 2. Blocks + for block in self.transformer_blocks: + hidden_states = block(hidden_states, context=encoder_hidden_states, timestep=timestep) + + # 3. Output + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous() + hidden_states = self.proj_out(hidden_states) + else: + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous() + + output = hidden_states + residual + + return output + + +class Upsample2D(nn.Module): + def __init__(self, channels, out_channels): + super().__init__() + self.channels = channels + self.out_channels = out_channels + self.conv = nn.Conv2d(self.channels, self.out_channels, 3, padding=1) + + self.gradient_checkpointing = False + + def forward_body(self, hidden_states, output_size=None): + assert hidden_states.shape[1] == self.channels + + # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16 + # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch + # https://github.com/pytorch/pytorch/issues/86679 + dtype = hidden_states.dtype + if dtype == torch.bfloat16: + hidden_states = hidden_states.to(torch.float32) + + # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984 + if hidden_states.shape[0] >= 64: + hidden_states = hidden_states.contiguous() + + # if `output_size` is passed we force the interpolation output size and do not make use of `scale_factor=2` + if output_size is None: + hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest") + else: + hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest") + + # If the input is bfloat16, we cast back to bfloat16 + if dtype == torch.bfloat16: + hidden_states = hidden_states.to(dtype) + + hidden_states = self.conv(hidden_states) + + return hidden_states + + def forward(self, hidden_states, output_size=None): + if self.training and self.gradient_checkpointing: + # logger.info("Upsample2D: gradient_checkpointing") + + def create_custom_forward(func): + def custom_forward(*inputs): + return func(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(self.forward_body), hidden_states, output_size, use_reentrant=USE_REENTRANT + ) + else: + hidden_states = self.forward_body(hidden_states, output_size) + + return hidden_states + + +class SdxlUNet2DConditionModel(nn.Module): + _supports_gradient_checkpointing = True + + def __init__( + self, + **kwargs, + ): + super().__init__() + + self.in_channels = IN_CHANNELS + self.out_channels = OUT_CHANNELS + self.model_channels = MODEL_CHANNELS + self.time_embed_dim = TIME_EMBED_DIM + self.adm_in_channels = ADM_IN_CHANNELS + + self.gradient_checkpointing = False + # self.sample_size = sample_size + + # time embedding + self.time_embed = nn.Sequential( + nn.Linear(self.model_channels, self.time_embed_dim), + nn.SiLU(), + nn.Linear(self.time_embed_dim, self.time_embed_dim), + ) + + # label embedding + self.label_emb = nn.Sequential( + nn.Sequential( + nn.Linear(self.adm_in_channels, self.time_embed_dim), + nn.SiLU(), + nn.Linear(self.time_embed_dim, self.time_embed_dim), + ) + ) + + # input + self.input_blocks = nn.ModuleList( + [ + nn.Sequential( + nn.Conv2d(self.in_channels, self.model_channels, kernel_size=3, padding=(1, 1)), + ) + ] + ) + + # level 0 + for i in range(2): + layers = [ + ResnetBlock2D( + in_channels=1 * self.model_channels, + out_channels=1 * self.model_channels, + ), + ] + self.input_blocks.append(nn.ModuleList(layers)) + + self.input_blocks.append( + nn.Sequential( + Downsample2D( + channels=1 * self.model_channels, + out_channels=1 * self.model_channels, + ), + ) + ) + + # level 1 + for i in range(2): + layers = [ + ResnetBlock2D( + in_channels=(1 if i == 0 else 2) * self.model_channels, + out_channels=2 * self.model_channels, + ), + Transformer2DModel( + num_attention_heads=2 * self.model_channels // 64, + attention_head_dim=64, + in_channels=2 * self.model_channels, + num_transformer_layers=2, + use_linear_projection=True, + cross_attention_dim=2048, + ), + ] + self.input_blocks.append(nn.ModuleList(layers)) + + self.input_blocks.append( + nn.Sequential( + Downsample2D( + channels=2 * self.model_channels, + out_channels=2 * self.model_channels, + ), + ) + ) + + # level 2 + for i in range(2): + layers = [ + ResnetBlock2D( + in_channels=(2 if i == 0 else 4) * self.model_channels, + out_channels=4 * self.model_channels, + ), + Transformer2DModel( + num_attention_heads=4 * self.model_channels // 64, + attention_head_dim=64, + in_channels=4 * self.model_channels, + num_transformer_layers=10, + use_linear_projection=True, + cross_attention_dim=2048, + ), + ] + self.input_blocks.append(nn.ModuleList(layers)) + + # mid + self.middle_block = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=4 * self.model_channels, + out_channels=4 * self.model_channels, + ), + Transformer2DModel( + num_attention_heads=4 * self.model_channels // 64, + attention_head_dim=64, + in_channels=4 * self.model_channels, + num_transformer_layers=10, + use_linear_projection=True, + cross_attention_dim=2048, + ), + ResnetBlock2D( + in_channels=4 * self.model_channels, + out_channels=4 * self.model_channels, + ), + ] + ) + + # output + self.output_blocks = nn.ModuleList([]) + + # level 2 + for i in range(3): + layers = [ + ResnetBlock2D( + in_channels=4 * self.model_channels + (4 if i <= 1 else 2) * self.model_channels, + out_channels=4 * self.model_channels, + ), + Transformer2DModel( + num_attention_heads=4 * self.model_channels // 64, + attention_head_dim=64, + in_channels=4 * self.model_channels, + num_transformer_layers=10, + use_linear_projection=True, + cross_attention_dim=2048, + ), + ] + if i == 2: + layers.append( + Upsample2D( + channels=4 * self.model_channels, + out_channels=4 * self.model_channels, + ) + ) + + self.output_blocks.append(nn.ModuleList(layers)) + + # level 1 + for i in range(3): + layers = [ + ResnetBlock2D( + in_channels=2 * self.model_channels + (4 if i == 0 else (2 if i == 1 else 1)) * self.model_channels, + out_channels=2 * self.model_channels, + ), + Transformer2DModel( + num_attention_heads=2 * self.model_channels // 64, + attention_head_dim=64, + in_channels=2 * self.model_channels, + num_transformer_layers=2, + use_linear_projection=True, + cross_attention_dim=2048, + ), + ] + if i == 2: + layers.append( + Upsample2D( + channels=2 * self.model_channels, + out_channels=2 * self.model_channels, + ) + ) + + self.output_blocks.append(nn.ModuleList(layers)) + + # level 0 + for i in range(3): + layers = [ + ResnetBlock2D( + in_channels=1 * self.model_channels + (2 if i == 0 else 1) * self.model_channels, + out_channels=1 * self.model_channels, + ), + ] + + self.output_blocks.append(nn.ModuleList(layers)) + + # output + self.out = nn.ModuleList( + [GroupNorm32(32, self.model_channels), nn.SiLU(), nn.Conv2d(self.model_channels, self.out_channels, 3, padding=1)] + ) + + # region diffusers compatibility + def prepare_config(self): + self.config = SimpleNamespace() + + @property + def dtype(self) -> torch.dtype: + # `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype). + return get_parameter_dtype(self) + + @property + def device(self) -> torch.device: + # `torch.device`: The device on which the module is (assuming that all the module parameters are on the same device). + return get_parameter_device(self) + + def set_attention_slice(self, slice_size): + raise NotImplementedError("Attention slicing is not supported for this model.") + + def is_gradient_checkpointing(self) -> bool: + return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for m in self.modules()) + + def enable_gradient_checkpointing(self): + self.gradient_checkpointing = True + self.set_gradient_checkpointing(value=True) + + def disable_gradient_checkpointing(self): + self.gradient_checkpointing = False + self.set_gradient_checkpointing(value=False) + + def set_use_memory_efficient_attention(self, xformers: bool, mem_eff: bool) -> None: + blocks = self.input_blocks + [self.middle_block] + self.output_blocks + for block in blocks: + for module in block: + if hasattr(module, "set_use_memory_efficient_attention"): + # logger.info(module.__class__.__name__) + module.set_use_memory_efficient_attention(xformers, mem_eff) + + def set_use_sdpa(self, sdpa: bool) -> None: + blocks = self.input_blocks + [self.middle_block] + self.output_blocks + for block in blocks: + for module in block: + if hasattr(module, "set_use_sdpa"): + module.set_use_sdpa(sdpa) + + def set_gradient_checkpointing(self, value=False): + blocks = self.input_blocks + [self.middle_block] + self.output_blocks + for block in blocks: + for module in block.modules(): + if hasattr(module, "gradient_checkpointing"): + # logger.info(f{module.__class__.__name__} {module.gradient_checkpointing} -> {value}") + module.gradient_checkpointing = value + + # endregion + + def forward(self, x, timesteps=None, context=None, y=None, **kwargs): + # broadcast timesteps to batch dimension + timesteps = timesteps.expand(x.shape[0]) + + hs = [] + t_emb = get_timestep_embedding(timesteps, self.model_channels, downscale_freq_shift=0) # , repeat_only=False) + t_emb = t_emb.to(x.dtype) + emb = self.time_embed(t_emb) + + assert x.shape[0] == y.shape[0], f"batch size mismatch: {x.shape[0]} != {y.shape[0]}" + assert x.dtype == y.dtype, f"dtype mismatch: {x.dtype} != {y.dtype}" + # assert x.dtype == self.dtype + emb = emb + self.label_emb(y) + + def call_module(module, h, emb, context): + x = h + for layer in module: + # logger.info(layer.__class__.__name__, x.dtype, emb.dtype, context.dtype if context is not None else None) + if isinstance(layer, ResnetBlock2D): + x = layer(x, emb) + elif isinstance(layer, Transformer2DModel): + x = layer(x, context) + else: + x = layer(x) + return x + + # h = x.type(self.dtype) + h = x + + for module in self.input_blocks: + h = call_module(module, h, emb, context) + hs.append(h) + + h = call_module(self.middle_block, h, emb, context) + + for module in self.output_blocks: + h = torch.cat([h, hs.pop()], dim=1) + h = call_module(module, h, emb, context) + + h = h.type(x.dtype) + h = call_module(self.out, h, emb, context) + + return h + + +class InferSdxlUNet2DConditionModel: + def __init__(self, original_unet: SdxlUNet2DConditionModel, **kwargs): + self.delegate = original_unet + + # override original model's forward method: because forward is not called by `__call__` + # overriding `__call__` is not enough, because nn.Module.forward has a special handling + self.delegate.forward = self.forward + + # Deep Shrink + self.ds_depth_1 = None + self.ds_depth_2 = None + self.ds_timesteps_1 = None + self.ds_timesteps_2 = None + self.ds_ratio = None + + # call original model's methods + def __getattr__(self, name): + return getattr(self.delegate, name) + + def __call__(self, *args, **kwargs): + return self.delegate(*args, **kwargs) + + def set_deep_shrink(self, ds_depth_1, ds_timesteps_1=650, ds_depth_2=None, ds_timesteps_2=None, ds_ratio=0.5): + if ds_depth_1 is None: + logger.info("Deep Shrink is disabled.") + self.ds_depth_1 = None + self.ds_timesteps_1 = None + self.ds_depth_2 = None + self.ds_timesteps_2 = None + self.ds_ratio = None + else: + logger.info( + f"Deep Shrink is enabled: [depth={ds_depth_1}/{ds_depth_2}, timesteps={ds_timesteps_1}/{ds_timesteps_2}, ratio={ds_ratio}]" + ) + self.ds_depth_1 = ds_depth_1 + self.ds_timesteps_1 = ds_timesteps_1 + self.ds_depth_2 = ds_depth_2 if ds_depth_2 is not None else -1 + self.ds_timesteps_2 = ds_timesteps_2 if ds_timesteps_2 is not None else 1000 + self.ds_ratio = ds_ratio + + def forward(self, x, timesteps=None, context=None, y=None, **kwargs): + r""" + current implementation is a copy of `SdxlUNet2DConditionModel.forward()` with Deep Shrink. + """ + _self = self.delegate + + # broadcast timesteps to batch dimension + timesteps = timesteps.expand(x.shape[0]) + + hs = [] + t_emb = get_timestep_embedding(timesteps, _self.model_channels, downscale_freq_shift=0) # , repeat_only=False) + t_emb = t_emb.to(x.dtype) + emb = _self.time_embed(t_emb) + + assert x.shape[0] == y.shape[0], f"batch size mismatch: {x.shape[0]} != {y.shape[0]}" + assert x.dtype == y.dtype, f"dtype mismatch: {x.dtype} != {y.dtype}" + # assert x.dtype == _self.dtype + emb = emb + _self.label_emb(y) + + def call_module(module, h, emb, context): + x = h + for layer in module: + # print(layer.__class__.__name__, x.dtype, emb.dtype, context.dtype if context is not None else None) + if isinstance(layer, ResnetBlock2D): + x = layer(x, emb) + elif isinstance(layer, Transformer2DModel): + x = layer(x, context) + else: + x = layer(x) + return x + + # h = x.type(self.dtype) + h = x + + for depth, module in enumerate(_self.input_blocks): + # Deep Shrink + if self.ds_depth_1 is not None: + if (depth == self.ds_depth_1 and timesteps[0] >= self.ds_timesteps_1) or ( + self.ds_depth_2 is not None + and depth == self.ds_depth_2 + and timesteps[0] < self.ds_timesteps_1 + and timesteps[0] >= self.ds_timesteps_2 + ): + # print("downsample", h.shape, self.ds_ratio) + org_dtype = h.dtype + if org_dtype == torch.bfloat16: + h = h.to(torch.float32) + h = F.interpolate(h, scale_factor=self.ds_ratio, mode="bicubic", align_corners=False).to(org_dtype) + + h = call_module(module, h, emb, context) + hs.append(h) + + h = call_module(_self.middle_block, h, emb, context) + + for module in _self.output_blocks: + # Deep Shrink + if self.ds_depth_1 is not None: + if hs[-1].shape[-2:] != h.shape[-2:]: + # print("upsample", h.shape, hs[-1].shape) + h = resize_like(h, hs[-1]) + + h = torch.cat([h, hs.pop()], dim=1) + h = call_module(module, h, emb, context) + + # Deep Shrink: in case of depth 0 + if self.ds_depth_1 == 0 and h.shape[-2:] != x.shape[-2:]: + # print("upsample", h.shape, x.shape) + h = resize_like(h, x) + + h = h.type(x.dtype) + h = call_module(_self.out, h, emb, context) + + return h + + +if __name__ == "__main__": + import time + + logger.info("create unet") + unet = SdxlUNet2DConditionModel() + + unet.to("cuda") + unet.set_use_memory_efficient_attention(True, False) + unet.set_gradient_checkpointing(True) + unet.train() + + # 使用メモリ量確認用の疑似学習ループ + logger.info("preparing optimizer") + + # optimizer = torch.optim.SGD(unet.parameters(), lr=1e-3, nesterov=True, momentum=0.9) # not working + + # import bitsandbytes + # optimizer = bitsandbytes.adam.Adam8bit(unet.parameters(), lr=1e-3) # not working + # optimizer = bitsandbytes.optim.RMSprop8bit(unet.parameters(), lr=1e-3) # working at 23.5 GB with torch2 + # optimizer=bitsandbytes.optim.Adagrad8bit(unet.parameters(), lr=1e-3) # working at 23.5 GB with torch2 + + import transformers + + optimizer = transformers.optimization.Adafactor(unet.parameters(), relative_step=True) # working at 22.2GB with torch2 + + scaler = torch.cuda.amp.GradScaler(enabled=True) + + logger.info("start training") + steps = 10 + batch_size = 1 + + for step in range(steps): + logger.info(f"step {step}") + if step == 1: + time_start = time.perf_counter() + + x = torch.randn(batch_size, 4, 128, 128).cuda() # 1024x1024 + t = torch.randint(low=0, high=10, size=(batch_size,), device="cuda") + ctx = torch.randn(batch_size, 77, 2048).cuda() + y = torch.randn(batch_size, ADM_IN_CHANNELS).cuda() + + with torch.cuda.amp.autocast(enabled=True): + output = unet(x, t, ctx, y) + target = torch.randn_like(output) + loss = torch.nn.functional.mse_loss(output, target) + + scaler.scale(loss).backward() + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad(set_to_none=True) + + time_end = time.perf_counter() + logger.info(f"elapsed time: {time_end - time_start} [sec] for last {steps - 1} steps") diff --git a/sdxl_train_util.py b/sdxl_train_util.py new file mode 100644 index 0000000000000000000000000000000000000000..f78d942447ef186b3a1b2da67c031f2ecb7eb1c5 --- /dev/null +++ b/sdxl_train_util.py @@ -0,0 +1,379 @@ +import argparse +import math +import os +from typing import Optional + +import torch +from library.device_utils import init_ipex, clean_memory_on_device + +init_ipex() + +from accelerate import init_empty_weights +from tqdm import tqdm +from transformers import CLIPTokenizer +from library import model_util, sdxl_model_util, train_util, sdxl_original_unet +from library.sdxl_lpw_stable_diffusion import SdxlStableDiffusionLongPromptWeightingPipeline +from .utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +TOKENIZER1_PATH = "openai/clip-vit-large-patch14" +TOKENIZER2_PATH = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k" + +# DEFAULT_NOISE_OFFSET = 0.0357 + + +def load_target_model(args, accelerator, model_version: str, weight_dtype): + model_dtype = match_mixed_precision(args, weight_dtype) # prepare fp16/bf16 + for pi in range(accelerator.state.num_processes): + if pi == accelerator.state.local_process_index: + logger.info(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}") + + ( + load_stable_diffusion_format, + text_encoder1, + text_encoder2, + vae, + unet, + logit_scale, + ckpt_info, + ) = _load_target_model( + args.pretrained_model_name_or_path, + args.vae, + model_version, + weight_dtype, + accelerator.device if args.lowram else "cpu", + model_dtype, + args.disable_mmap_load_safetensors, + ) + + # work on low-ram device + if args.lowram: + text_encoder1.to(accelerator.device) + text_encoder2.to(accelerator.device) + unet.to(accelerator.device) + vae.to(accelerator.device) + + clean_memory_on_device(accelerator.device) + accelerator.wait_for_everyone() + + return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info + + +def _load_target_model( + name_or_path: str, vae_path: Optional[str], model_version: str, weight_dtype, device="cpu", model_dtype=None, disable_mmap=False +): + # model_dtype only work with full fp16/bf16 + name_or_path = os.readlink(name_or_path) if os.path.islink(name_or_path) else name_or_path + load_stable_diffusion_format = os.path.isfile(name_or_path) # determine SD or Diffusers + + if load_stable_diffusion_format: + logger.info(f"load StableDiffusion checkpoint: {name_or_path}") + ( + text_encoder1, + text_encoder2, + vae, + unet, + logit_scale, + ckpt_info, + ) = sdxl_model_util.load_models_from_sdxl_checkpoint(model_version, name_or_path, device, model_dtype, disable_mmap) + else: + # Diffusers model is loaded to CPU + from diffusers import StableDiffusionXLPipeline + + variant = "fp16" if weight_dtype == torch.float16 else None + logger.info(f"load Diffusers pretrained models: {name_or_path}, variant={variant}") + try: + try: + pipe = StableDiffusionXLPipeline.from_pretrained( + name_or_path, torch_dtype=model_dtype, variant=variant, tokenizer=None + ) + except EnvironmentError as ex: + if variant is not None: + logger.info("try to load fp32 model") + pipe = StableDiffusionXLPipeline.from_pretrained(name_or_path, variant=None, tokenizer=None) + else: + raise ex + except EnvironmentError as ex: + logger.error( + f"model is not found as a file or in Hugging Face, perhaps file name is wrong? / 指定したモデル名のファイル、またはHugging Faceのモデルが見つかりません。ファイル名が誤っているかもしれません: {name_or_path}" + ) + raise ex + + text_encoder1 = pipe.text_encoder + text_encoder2 = pipe.text_encoder_2 + + # convert to fp32 for cache text_encoders outputs + if text_encoder1.dtype != torch.float32: + text_encoder1 = text_encoder1.to(dtype=torch.float32) + if text_encoder2.dtype != torch.float32: + text_encoder2 = text_encoder2.to(dtype=torch.float32) + + vae = pipe.vae + unet = pipe.unet + del pipe + + # Diffusers U-Net to original U-Net + state_dict = sdxl_model_util.convert_diffusers_unet_state_dict_to_sdxl(unet.state_dict()) + with init_empty_weights(): + unet = sdxl_original_unet.SdxlUNet2DConditionModel() # overwrite unet + sdxl_model_util._load_state_dict_on_device(unet, state_dict, device=device, dtype=model_dtype) + logger.info("U-Net converted to original U-Net") + + logit_scale = None + ckpt_info = None + + # VAEを読み込む + if vae_path is not None: + vae = model_util.load_vae(vae_path, weight_dtype) + logger.info("additional VAE loaded") + + return load_stable_diffusion_format, text_encoder1, text_encoder2, vae, unet, logit_scale, ckpt_info + + +def load_tokenizers(args: argparse.Namespace): + logger.info("prepare tokenizers") + + original_paths = [TOKENIZER1_PATH, TOKENIZER2_PATH] + tokeniers = [] + for i, original_path in enumerate(original_paths): + tokenizer: CLIPTokenizer = None + if args.tokenizer_cache_dir: + local_tokenizer_path = os.path.join(args.tokenizer_cache_dir, original_path.replace("/", "_")) + if os.path.exists(local_tokenizer_path): + logger.info(f"load tokenizer from cache: {local_tokenizer_path}") + tokenizer = CLIPTokenizer.from_pretrained(local_tokenizer_path) + + if tokenizer is None: + tokenizer = CLIPTokenizer.from_pretrained(original_path) + + if args.tokenizer_cache_dir and not os.path.exists(local_tokenizer_path): + logger.info(f"save Tokenizer to cache: {local_tokenizer_path}") + tokenizer.save_pretrained(local_tokenizer_path) + + if i == 1: + tokenizer.pad_token_id = 0 # fix pad token id to make same as open clip tokenizer + + tokeniers.append(tokenizer) + + if hasattr(args, "max_token_length") and args.max_token_length is not None: + logger.info(f"update token length: {args.max_token_length}") + + return tokeniers + + +def match_mixed_precision(args, weight_dtype): + if args.full_fp16: + assert ( + weight_dtype == torch.float16 + ), "full_fp16 requires mixed precision='fp16' / full_fp16を使う場合はmixed_precision='fp16'を指定してください。" + return weight_dtype + elif args.full_bf16: + assert ( + weight_dtype == torch.bfloat16 + ), "full_bf16 requires mixed precision='bf16' / full_bf16を使う場合はmixed_precision='bf16'を指定してください。" + return weight_dtype + else: + return None + + +def timestep_embedding(timesteps, dim, max_period=10000): + """ + Create sinusoidal timestep embeddings. + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an [N x dim] Tensor of positional embeddings. + """ + half = dim // 2 + freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to( + device=timesteps.device + ) + args = timesteps[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + return embedding + + +def get_timestep_embedding(x, outdim): + assert len(x.shape) == 2 + b, dims = x.shape[0], x.shape[1] + x = torch.flatten(x) + emb = timestep_embedding(x, outdim) + emb = torch.reshape(emb, (b, dims * outdim)) + return emb + + +def get_size_embeddings(orig_size, crop_size, target_size, device): + emb1 = get_timestep_embedding(orig_size, 256) + emb2 = get_timestep_embedding(crop_size, 256) + emb3 = get_timestep_embedding(target_size, 256) + vector = torch.cat([emb1, emb2, emb3], dim=1).to(device) + return vector + + +def save_sd_model_on_train_end( + args: argparse.Namespace, + src_path: str, + save_stable_diffusion_format: bool, + use_safetensors: bool, + save_dtype: torch.dtype, + epoch: int, + global_step: int, + text_encoder1, + text_encoder2, + unet, + vae, + logit_scale, + ckpt_info, +): + def sd_saver(ckpt_file, epoch_no, global_step): + sai_metadata = train_util.get_sai_model_spec(None, args, True, False, False, is_stable_diffusion_ckpt=True) + sdxl_model_util.save_stable_diffusion_checkpoint( + ckpt_file, + text_encoder1, + text_encoder2, + unet, + epoch_no, + global_step, + ckpt_info, + vae, + logit_scale, + sai_metadata, + save_dtype, + ) + + def diffusers_saver(out_dir): + sdxl_model_util.save_diffusers_checkpoint( + out_dir, + text_encoder1, + text_encoder2, + unet, + src_path, + vae, + use_safetensors=use_safetensors, + save_dtype=save_dtype, + ) + + train_util.save_sd_model_on_train_end_common( + args, save_stable_diffusion_format, use_safetensors, epoch, global_step, sd_saver, diffusers_saver + ) + + +# epochとstepの保存、メタデータにepoch/stepが含まれ引数が同じになるため、統合している +# on_epoch_end: Trueならepoch終了時、Falseならstep経過時 +def save_sd_model_on_epoch_end_or_stepwise( + args: argparse.Namespace, + on_epoch_end: bool, + accelerator, + src_path, + save_stable_diffusion_format: bool, + use_safetensors: bool, + save_dtype: torch.dtype, + epoch: int, + num_train_epochs: int, + global_step: int, + text_encoder1, + text_encoder2, + unet, + vae, + logit_scale, + ckpt_info, +): + def sd_saver(ckpt_file, epoch_no, global_step): + sai_metadata = train_util.get_sai_model_spec(None, args, True, False, False, is_stable_diffusion_ckpt=True) + sdxl_model_util.save_stable_diffusion_checkpoint( + ckpt_file, + text_encoder1, + text_encoder2, + unet, + epoch_no, + global_step, + ckpt_info, + vae, + logit_scale, + sai_metadata, + save_dtype, + ) + + def diffusers_saver(out_dir): + sdxl_model_util.save_diffusers_checkpoint( + out_dir, + text_encoder1, + text_encoder2, + unet, + src_path, + vae, + use_safetensors=use_safetensors, + save_dtype=save_dtype, + ) + + train_util.save_sd_model_on_epoch_end_or_stepwise_common( + args, + on_epoch_end, + accelerator, + save_stable_diffusion_format, + use_safetensors, + epoch, + num_train_epochs, + global_step, + sd_saver, + diffusers_saver, + ) + + +def add_sdxl_training_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--cache_text_encoder_outputs", action="store_true", help="cache text encoder outputs / text encoderの出力をキャッシュする" + ) + parser.add_argument( + "--cache_text_encoder_outputs_to_disk", + action="store_true", + help="cache text encoder outputs to disk / text encoderの出力をディスクにキャッシュする", + ) + parser.add_argument( + "--disable_mmap_load_safetensors", + action="store_true", + help="disable mmap load for safetensors. Speed up model loading in WSL environment / safetensorsのmmapロードを無効にする。WSL環境等でモデル読み込みを高速化できる", + ) + + +def verify_sdxl_training_args(args: argparse.Namespace, supportTextEncoderCaching: bool = True): + assert not args.v2, "v2 cannot be enabled in SDXL training / SDXL学習ではv2を有効にすることはできません" + + if args.clip_skip is not None: + logger.warning("clip_skip will be unexpected / SDXL学習ではclip_skipは動作しません") + + # if args.multires_noise_iterations: + # logger.info( + # f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET}, but noise_offset is disabled due to multires_noise_iterations / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されていますが、multires_noise_iterationsが有効になっているためnoise_offsetは無効になります" + # ) + # else: + # if args.noise_offset is None: + # args.noise_offset = DEFAULT_NOISE_OFFSET + # elif args.noise_offset != DEFAULT_NOISE_OFFSET: + # logger.info( + # f"Warning: SDXL has been trained with noise_offset={DEFAULT_NOISE_OFFSET} / SDXLはnoise_offset={DEFAULT_NOISE_OFFSET}で学習されています" + # ) + # logger.info(f"noise_offset is set to {args.noise_offset} / noise_offsetが{args.noise_offset}に設定されました") + + assert ( + not hasattr(args, "weighted_captions") or not args.weighted_captions + ), "weighted_captions cannot be enabled in SDXL training currently / SDXL学習では今のところweighted_captionsを有効にすることはできません" + + if supportTextEncoderCaching: + if args.cache_text_encoder_outputs_to_disk and not args.cache_text_encoder_outputs: + args.cache_text_encoder_outputs = True + logger.warning( + "cache_text_encoder_outputs is enabled because cache_text_encoder_outputs_to_disk is enabled / " + + "cache_text_encoder_outputs_to_diskが有効になっているためcache_text_encoder_outputsが有効になりました" + ) + + +def sample_images(*args, **kwargs): + return train_util.sample_images_common(SdxlStableDiffusionLongPromptWeightingPipeline, *args, **kwargs) diff --git a/show_metadata.py b/show_metadata.py new file mode 100644 index 0000000000000000000000000000000000000000..05bfbe0a4a155ffe8da380af4cb55817ebef3505 --- /dev/null +++ b/show_metadata.py @@ -0,0 +1,23 @@ +import json +import argparse +from safetensors import safe_open +from library.utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +parser = argparse.ArgumentParser() +parser.add_argument("--model", type=str, required=True) +args = parser.parse_args() + +with safe_open(args.model, framework="pt") as f: + metadata = f.metadata() + +if metadata is None: + logger.error("No metadata found") +else: + # metadata is json dict, but not pretty printed + # sort by key and pretty print + print(json.dumps(metadata, indent=4, sort_keys=True)) + + diff --git a/slicing_vae.py b/slicing_vae.py new file mode 100644 index 0000000000000000000000000000000000000000..ea765342961ab0c55721f5b8a81cbca2ec01feb5 --- /dev/null +++ b/slicing_vae.py @@ -0,0 +1,682 @@ +# Modified from Diffusers to reduce VRAM usage + +# Copyright 2022 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn as nn + + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.models.modeling_utils import ModelMixin +from diffusers.models.unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block +from diffusers.models.vae import DecoderOutput, DiagonalGaussianDistribution +from diffusers.models.autoencoder_kl import AutoencoderKLOutput +from .utils import setup_logging +setup_logging() +import logging +logger = logging.getLogger(__name__) + +def slice_h(x, num_slices): + # slice with pad 1 both sides: to eliminate side effect of padding of conv2d + # Conv2dのpaddingの副作用を排除するために、両側にpad 1しながらHをスライスする + # NCHWでもNHWCでもどちらでも動く + size = (x.shape[2] + num_slices - 1) // num_slices + sliced = [] + for i in range(num_slices): + if i == 0: + sliced.append(x[:, :, : size + 1, :]) + else: + end = size * (i + 1) + 1 + if x.shape[2] - end < 3: # if the last slice is too small, use the rest of the tensor 最後が細すぎるとconv2dできないので全部使う + end = x.shape[2] + sliced.append(x[:, :, size * i - 1 : end, :]) + if end >= x.shape[2]: + break + return sliced + + +def cat_h(sliced): + # padding分を除いて結合する + cat = [] + for i, x in enumerate(sliced): + if i == 0: + cat.append(x[:, :, :-1, :]) + elif i == len(sliced) - 1: + cat.append(x[:, :, 1:, :]) + else: + cat.append(x[:, :, 1:-1, :]) + del x + x = torch.cat(cat, dim=2) + return x + + +def resblock_forward(_self, num_slices, input_tensor, temb, **kwargs): + assert _self.upsample is None and _self.downsample is None + assert _self.norm1.num_groups == _self.norm2.num_groups + assert temb is None + + # make sure norms are on cpu + org_device = input_tensor.device + cpu_device = torch.device("cpu") + _self.norm1.to(cpu_device) + _self.norm2.to(cpu_device) + + # GroupNormがCPUでfp16で動かない対策 + org_dtype = input_tensor.dtype + if org_dtype == torch.float16: + _self.norm1.to(torch.float32) + _self.norm2.to(torch.float32) + + # すべてのテンソルをCPUに移動する + input_tensor = input_tensor.to(cpu_device) + hidden_states = input_tensor + + # どうもこれは結果が異なるようだ…… + # def sliced_norm1(norm, x): + # num_div = 4 if up_block_idx <= 2 else x.shape[1] // norm.num_groups + # sliced_tensor = torch.chunk(x, num_div, dim=1) + # sliced_weight = torch.chunk(norm.weight, num_div, dim=0) + # sliced_bias = torch.chunk(norm.bias, num_div, dim=0) + # logger.info(sliced_tensor[0].shape, num_div, sliced_weight[0].shape, sliced_bias[0].shape) + # normed_tensor = [] + # for i in range(num_div): + # n = torch.group_norm(sliced_tensor[i], norm.num_groups, sliced_weight[i], sliced_bias[i], norm.eps) + # normed_tensor.append(n) + # del n + # x = torch.cat(normed_tensor, dim=1) + # return num_div, x + + # normを分割すると結果が変わるので、ここだけは分割しない。GPUで計算するとVRAMが足りなくなるので、CPUで計算する。幸いCPUでもそこまで遅くない + if org_dtype == torch.float16: + hidden_states = hidden_states.to(torch.float32) + hidden_states = _self.norm1(hidden_states) # run on cpu + if org_dtype == torch.float16: + hidden_states = hidden_states.to(torch.float16) + + sliced = slice_h(hidden_states, num_slices) + del hidden_states + + for i in range(len(sliced)): + x = sliced[i] + sliced[i] = None + + # 計算する部分だけGPUに移動する、以下同様 + x = x.to(org_device) + x = _self.nonlinearity(x) + x = _self.conv1(x) + x = x.to(cpu_device) + sliced[i] = x + del x + + hidden_states = cat_h(sliced) + del sliced + + if org_dtype == torch.float16: + hidden_states = hidden_states.to(torch.float32) + hidden_states = _self.norm2(hidden_states) # run on cpu + if org_dtype == torch.float16: + hidden_states = hidden_states.to(torch.float16) + + sliced = slice_h(hidden_states, num_slices) + del hidden_states + + for i in range(len(sliced)): + x = sliced[i] + sliced[i] = None + + x = x.to(org_device) + x = _self.nonlinearity(x) + x = _self.dropout(x) + x = _self.conv2(x) + x = x.to(cpu_device) + sliced[i] = x + del x + + hidden_states = cat_h(sliced) + del sliced + + # make shortcut + if _self.conv_shortcut is not None: + sliced = list(torch.chunk(input_tensor, num_slices, dim=2)) # no padding in conv_shortcut パディングがないので普通にスライスする + del input_tensor + + for i in range(len(sliced)): + x = sliced[i] + sliced[i] = None + + x = x.to(org_device) + x = _self.conv_shortcut(x) + x = x.to(cpu_device) + sliced[i] = x + del x + + input_tensor = torch.cat(sliced, dim=2) + del sliced + + output_tensor = (input_tensor + hidden_states) / _self.output_scale_factor + + output_tensor = output_tensor.to(org_device) # 次のレイヤーがGPUで計算する + return output_tensor + + +class SlicingEncoder(nn.Module): + def __init__( + self, + in_channels=3, + out_channels=3, + down_block_types=("DownEncoderBlock2D",), + block_out_channels=(64,), + layers_per_block=2, + norm_num_groups=32, + act_fn="silu", + double_z=True, + num_slices=2, + ): + super().__init__() + self.layers_per_block = layers_per_block + + self.conv_in = torch.nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, stride=1, padding=1) + + self.mid_block = None + self.down_blocks = nn.ModuleList([]) + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=self.layers_per_block, + in_channels=input_channel, + out_channels=output_channel, + add_downsample=not is_final_block, + resnet_eps=1e-6, + downsample_padding=0, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + attention_head_dim=output_channel, + temb_channels=None, + ) + self.down_blocks.append(down_block) + + # mid + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + resnet_eps=1e-6, + resnet_act_fn=act_fn, + output_scale_factor=1, + resnet_time_scale_shift="default", + attention_head_dim=block_out_channels[-1], + resnet_groups=norm_num_groups, + temb_channels=None, + ) + self.mid_block.attentions[0].set_use_memory_efficient_attention_xformers(True) # とりあえずDiffusersのxformersを使う + + # out + self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6) + self.conv_act = nn.SiLU() + + conv_out_channels = 2 * out_channels if double_z else out_channels + self.conv_out = nn.Conv2d(block_out_channels[-1], conv_out_channels, 3, padding=1) + + # replace forward of ResBlocks + def wrapper(func, module, num_slices): + def forward(*args, **kwargs): + return func(module, num_slices, *args, **kwargs) + + return forward + + self.num_slices = num_slices + div = num_slices / (2 ** (len(self.down_blocks) - 1)) # 深い層はそこまで分割しなくていいので適宜減らす + # logger.info(f"initial divisor: {div}") + if div >= 2: + div = int(div) + for resnet in self.mid_block.resnets: + resnet.forward = wrapper(resblock_forward, resnet, div) + # midblock doesn't have downsample + + for i, down_block in enumerate(self.down_blocks[::-1]): + if div >= 2: + div = int(div) + # logger.info(f"down block: {i} divisor: {div}") + for resnet in down_block.resnets: + resnet.forward = wrapper(resblock_forward, resnet, div) + if down_block.downsamplers is not None: + # logger.info("has downsample") + for downsample in down_block.downsamplers: + downsample.forward = wrapper(self.downsample_forward, downsample, div * 2) + div *= 2 + + def forward(self, x): + sample = x + del x + + org_device = sample.device + cpu_device = torch.device("cpu") + + # sample = self.conv_in(sample) + sample = sample.to(cpu_device) + sliced = slice_h(sample, self.num_slices) + del sample + + for i in range(len(sliced)): + x = sliced[i] + sliced[i] = None + + x = x.to(org_device) + x = self.conv_in(x) + x = x.to(cpu_device) + sliced[i] = x + del x + + sample = cat_h(sliced) + del sliced + + sample = sample.to(org_device) + + # down + for down_block in self.down_blocks: + sample = down_block(sample) + + # middle + sample = self.mid_block(sample) + + # post-process + # ここも省メモリ化したいが、恐らくそこまでメモリを食わないので省略 + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + return sample + + def downsample_forward(self, _self, num_slices, hidden_states): + assert hidden_states.shape[1] == _self.channels + assert _self.use_conv and _self.padding == 0 + logger.info(f"downsample forward {num_slices} {hidden_states.shape}") + + org_device = hidden_states.device + cpu_device = torch.device("cpu") + + hidden_states = hidden_states.to(cpu_device) + pad = (0, 1, 0, 1) + hidden_states = torch.nn.functional.pad(hidden_states, pad, mode="constant", value=0) + + # slice with even number because of stride 2 + # strideが2なので偶数でスライスする + # slice with pad 1 both sides: to eliminate side effect of padding of conv2d + size = (hidden_states.shape[2] + num_slices - 1) // num_slices + size = size + 1 if size % 2 == 1 else size + + sliced = [] + for i in range(num_slices): + if i == 0: + sliced.append(hidden_states[:, :, : size + 1, :]) + else: + end = size * (i + 1) + 1 + if hidden_states.shape[2] - end < 4: # if the last slice is too small, use the rest of the tensor + end = hidden_states.shape[2] + sliced.append(hidden_states[:, :, size * i - 1 : end, :]) + if end >= hidden_states.shape[2]: + break + del hidden_states + + for i in range(len(sliced)): + x = sliced[i] + sliced[i] = None + + x = x.to(org_device) + x = _self.conv(x) + x = x.to(cpu_device) + + # ここだけ雰囲気が違うのはCopilotのせい + if i == 0: + hidden_states = x + else: + hidden_states = torch.cat([hidden_states, x], dim=2) + + hidden_states = hidden_states.to(org_device) + # logger.info(f"downsample forward done {hidden_states.shape}") + return hidden_states + + +class SlicingDecoder(nn.Module): + def __init__( + self, + in_channels=3, + out_channels=3, + up_block_types=("UpDecoderBlock2D",), + block_out_channels=(64,), + layers_per_block=2, + norm_num_groups=32, + act_fn="silu", + num_slices=2, + ): + super().__init__() + self.layers_per_block = layers_per_block + + self.conv_in = nn.Conv2d(in_channels, block_out_channels[-1], kernel_size=3, stride=1, padding=1) + + self.mid_block = None + self.up_blocks = nn.ModuleList([]) + + # mid + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + resnet_eps=1e-6, + resnet_act_fn=act_fn, + output_scale_factor=1, + resnet_time_scale_shift="default", + attention_head_dim=block_out_channels[-1], + resnet_groups=norm_num_groups, + temb_channels=None, + ) + self.mid_block.attentions[0].set_use_memory_efficient_attention_xformers(True) # とりあえずDiffusersのxformersを使う + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + + is_final_block = i == len(block_out_channels) - 1 + + up_block = get_up_block( + up_block_type, + num_layers=self.layers_per_block + 1, + in_channels=prev_output_channel, + out_channels=output_channel, + prev_output_channel=None, + add_upsample=not is_final_block, + resnet_eps=1e-6, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + attention_head_dim=output_channel, + temb_channels=None, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6) + self.conv_act = nn.SiLU() + self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1) + + # replace forward of ResBlocks + def wrapper(func, module, num_slices): + def forward(*args, **kwargs): + return func(module, num_slices, *args, **kwargs) + + return forward + + self.num_slices = num_slices + div = num_slices / (2 ** (len(self.up_blocks) - 1)) + logger.info(f"initial divisor: {div}") + if div >= 2: + div = int(div) + for resnet in self.mid_block.resnets: + resnet.forward = wrapper(resblock_forward, resnet, div) + # midblock doesn't have upsample + + for i, up_block in enumerate(self.up_blocks): + if div >= 2: + div = int(div) + # logger.info(f"up block: {i} divisor: {div}") + for resnet in up_block.resnets: + resnet.forward = wrapper(resblock_forward, resnet, div) + if up_block.upsamplers is not None: + # logger.info("has upsample") + for upsample in up_block.upsamplers: + upsample.forward = wrapper(self.upsample_forward, upsample, div * 2) + div *= 2 + + def forward(self, z): + sample = z + del z + sample = self.conv_in(sample) + + # middle + sample = self.mid_block(sample) + + # up + for i, up_block in enumerate(self.up_blocks): + sample = up_block(sample) + + # post-process + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + + # conv_out with slicing because of VRAM usage + # conv_outはとてもVRAM使うのでスライスして対応 + org_device = sample.device + cpu_device = torch.device("cpu") + sample = sample.to(cpu_device) + + sliced = slice_h(sample, self.num_slices) + del sample + for i in range(len(sliced)): + x = sliced[i] + sliced[i] = None + + x = x.to(org_device) + x = self.conv_out(x) + x = x.to(cpu_device) + sliced[i] = x + sample = cat_h(sliced) + del sliced + + sample = sample.to(org_device) + return sample + + def upsample_forward(self, _self, num_slices, hidden_states, output_size=None): + assert hidden_states.shape[1] == _self.channels + assert _self.use_conv_transpose == False and _self.use_conv + + org_dtype = hidden_states.dtype + org_device = hidden_states.device + cpu_device = torch.device("cpu") + + hidden_states = hidden_states.to(cpu_device) + sliced = slice_h(hidden_states, num_slices) + del hidden_states + + for i in range(len(sliced)): + x = sliced[i] + sliced[i] = None + + x = x.to(org_device) + + # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16 + # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch + # https://github.com/pytorch/pytorch/issues/86679 + # PyTorch 2で直らないかね…… + if org_dtype == torch.bfloat16: + x = x.to(torch.float32) + + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + + if org_dtype == torch.bfloat16: + x = x.to(org_dtype) + + x = _self.conv(x) + + # upsampleされてるのでpadは2になる + if i == 0: + x = x[:, :, :-2, :] + elif i == num_slices - 1: + x = x[:, :, 2:, :] + else: + x = x[:, :, 2:-2, :] + + x = x.to(cpu_device) + sliced[i] = x + del x + + hidden_states = torch.cat(sliced, dim=2) + # logger.info(f"us hidden_states {hidden_states.shape}") + del sliced + + hidden_states = hidden_states.to(org_device) + return hidden_states + + +class SlicingAutoencoderKL(ModelMixin, ConfigMixin): + r"""Variational Autoencoder (VAE) model with KL loss from the paper Auto-Encoding Variational Bayes by Diederik P. Kingma + and Max Welling. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library + implements for all the model (such as downloading or saving, etc.) + + Parameters: + in_channels (int, *optional*, defaults to 3): Number of channels in the input image. + out_channels (int, *optional*, defaults to 3): Number of channels in the output. + down_block_types (`Tuple[str]`, *optional*, defaults to : + obj:`("DownEncoderBlock2D",)`): Tuple of downsample block types. + up_block_types (`Tuple[str]`, *optional*, defaults to : + obj:`("UpDecoderBlock2D",)`): Tuple of upsample block types. + block_out_channels (`Tuple[int]`, *optional*, defaults to : + obj:`(64,)`): Tuple of block output channels. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + latent_channels (`int`, *optional*, defaults to `4`): Number of channels in the latent space. + sample_size (`int`, *optional*, defaults to `32`): TODO + """ + + @register_to_config + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + down_block_types: Tuple[str] = ("DownEncoderBlock2D",), + up_block_types: Tuple[str] = ("UpDecoderBlock2D",), + block_out_channels: Tuple[int] = (64,), + layers_per_block: int = 1, + act_fn: str = "silu", + latent_channels: int = 4, + norm_num_groups: int = 32, + sample_size: int = 32, + num_slices: int = 16, + ): + super().__init__() + + # pass init params to Encoder + self.encoder = SlicingEncoder( + in_channels=in_channels, + out_channels=latent_channels, + down_block_types=down_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + act_fn=act_fn, + norm_num_groups=norm_num_groups, + double_z=True, + num_slices=num_slices, + ) + + # pass init params to Decoder + self.decoder = SlicingDecoder( + in_channels=latent_channels, + out_channels=out_channels, + up_block_types=up_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + norm_num_groups=norm_num_groups, + act_fn=act_fn, + num_slices=num_slices, + ) + + self.quant_conv = torch.nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) + self.post_quant_conv = torch.nn.Conv2d(latent_channels, latent_channels, 1) + self.use_slicing = False + + def encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput: + h = self.encoder(x) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + z = self.post_quant_conv(z) + dec = self.decoder(z) + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + # これはバッチ方向のスライシング　紛らわしい + def enable_slicing(self): + r""" + Enable sliced VAE decoding. + + When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several + steps. This is useful to save some memory and allow larger batch sizes. + """ + self.use_slicing = True + + def disable_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_slicing` was previously invoked, this method will go back to computing + decoding in one step. + """ + self.use_slicing = False + + def decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + if self.use_slicing and z.shape[0] > 1: + decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)] + decoded = torch.cat(decoded_slices) + else: + decoded = self._decode(z).sample + + if not return_dict: + return (decoded,) + + return DecoderOutput(sample=decoded) + + def forward( + self, + sample: torch.FloatTensor, + sample_posterior: bool = False, + return_dict: bool = True, + generator: Optional[torch.Generator] = None, + ) -> Union[DecoderOutput, torch.FloatTensor]: + r""" + Args: + sample (`torch.FloatTensor`): Input sample. + sample_posterior (`bool`, *optional*, defaults to `False`): + Whether to sample from the posterior. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`DecoderOutput`] instead of a plain tuple. + """ + x = sample + posterior = self.encode(x).latent_dist + if sample_posterior: + z = posterior.sample(generator=generator) + else: + z = posterior.mode() + dec = self.decode(z).sample + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) diff --git a/svd_merge_lora.py b/svd_merge_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..c79b45acfd15a47ec4fed31faad0db1c38107729 --- /dev/null +++ b/svd_merge_lora.py @@ -0,0 +1,515 @@ +import argparse +import itertools +import json +import os +import re +import time +import torch +from safetensors.torch import load_file, save_file +from tqdm import tqdm +from library import sai_model_spec, train_util +import library.model_util as model_util +import lora +from library.utils import setup_logging + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +CLAMP_QUANTILE = 0.99 + +ACCEPTABLE = [12, 17, 20, 26] +SDXL_LAYER_NUM = [12, 20] + +LAYER12 = { + "BASE": True, + "IN00": False, + "IN01": False, + "IN02": False, + "IN03": False, + "IN04": True, + "IN05": True, + "IN06": False, + "IN07": True, + "IN08": True, + "IN09": False, + "IN10": False, + "IN11": False, + "MID": True, + "OUT00": True, + "OUT01": True, + "OUT02": True, + "OUT03": True, + "OUT04": True, + "OUT05": True, + "OUT06": False, + "OUT07": False, + "OUT08": False, + "OUT09": False, + "OUT10": False, + "OUT11": False, +} + +LAYER17 = { + "BASE": True, + "IN00": False, + "IN01": True, + "IN02": True, + "IN03": False, + "IN04": True, + "IN05": True, + "IN06": False, + "IN07": True, + "IN08": True, + "IN09": False, + "IN10": False, + "IN11": False, + "MID": True, + "OUT00": False, + "OUT01": False, + "OUT02": False, + "OUT03": True, + "OUT04": True, + "OUT05": True, + "OUT06": True, + "OUT07": True, + "OUT08": True, + "OUT09": True, + "OUT10": True, + "OUT11": True, +} + +LAYER20 = { + "BASE": True, + "IN00": True, + "IN01": True, + "IN02": True, + "IN03": True, + "IN04": True, + "IN05": True, + "IN06": True, + "IN07": True, + "IN08": True, + "IN09": False, + "IN10": False, + "IN11": False, + "MID": True, + "OUT00": True, + "OUT01": True, + "OUT02": True, + "OUT03": True, + "OUT04": True, + "OUT05": True, + "OUT06": True, + "OUT07": True, + "OUT08": True, + "OUT09": False, + "OUT10": False, + "OUT11": False, +} + +LAYER26 = { + "BASE": True, + "IN00": True, + "IN01": True, + "IN02": True, + "IN03": True, + "IN04": True, + "IN05": True, + "IN06": True, + "IN07": True, + "IN08": True, + "IN09": True, + "IN10": True, + "IN11": True, + "MID": True, + "OUT00": True, + "OUT01": True, + "OUT02": True, + "OUT03": True, + "OUT04": True, + "OUT05": True, + "OUT06": True, + "OUT07": True, + "OUT08": True, + "OUT09": True, + "OUT10": True, + "OUT11": True, +} + +assert len([v for v in LAYER12.values() if v]) == 12 +assert len([v for v in LAYER17.values() if v]) == 17 +assert len([v for v in LAYER20.values() if v]) == 20 +assert len([v for v in LAYER26.values() if v]) == 26 + +RE_UPDOWN = re.compile(r"(up|down)_blocks_(\d+)_(resnets|upsamplers|downsamplers|attentions)_(\d+)_") + + +def get_lbw_block_index(lora_name: str, is_sdxl: bool = False) -> int: + # lbw block index is 0-based, but 0 for text encoder, so we return 0 for text encoder + if "text_model_encoder_" in lora_name: # LoRA for text encoder + return 0 + + # lbw block index is 1-based for U-Net, and no "input_blocks.0" in CompVis SD, so "input_blocks.1" have index 2 + block_idx = -1 # invalid lora name + if not is_sdxl: + NUM_OF_BLOCKS = 12 # up/down blocks + m = RE_UPDOWN.search(lora_name) + if m: + g = m.groups() + up_down = g[0] + i = int(g[1]) + j = int(g[3]) + if up_down == "down": + if g[2] == "resnets" or g[2] == "attentions": + idx = 3 * i + j + 1 + elif g[2] == "downsamplers": + idx = 3 * (i + 1) + else: + return block_idx # invalid lora name + elif up_down == "up": + if g[2] == "resnets" or g[2] == "attentions": + idx = 3 * i + j + elif g[2] == "upsamplers": + idx = 3 * i + 2 + else: + return block_idx # invalid lora name + + if g[0] == "down": + block_idx = 1 + idx # 1-based index, down block index + elif g[0] == "up": + block_idx = 1 + NUM_OF_BLOCKS + 1 + idx # 1-based index, num blocks, mid block, up block index + + elif "mid_block_" in lora_name: + block_idx = 1 + NUM_OF_BLOCKS # 1-based index, num blocks, mid block + else: + # SDXL: some numbers are skipped + if lora_name.startswith("lora_unet_"): + name = lora_name[len("lora_unet_") :] + if name.startswith("time_embed_") or name.startswith("label_emb_"): # 1, No LoRA in sd-scripts + block_idx = 1 + elif name.startswith("input_blocks_"): # 1-8 to 2-9 + block_idx = 1 + int(name.split("_")[2]) + elif name.startswith("middle_block_"): # 13 + block_idx = 13 + elif name.startswith("output_blocks_"): # 0-8 to 14-22 + block_idx = 14 + int(name.split("_")[2]) + elif name.startswith("out_"): # 23, No LoRA in sd-scripts + block_idx = 23 + + return block_idx + + +def load_state_dict(file_name, dtype): + if os.path.splitext(file_name)[1] == ".safetensors": + sd = load_file(file_name) + metadata = train_util.load_metadata_from_safetensors(file_name) + else: + sd = torch.load(file_name, map_location="cpu") + metadata = {} + + for key in list(sd.keys()): + if type(sd[key]) == torch.Tensor: + sd[key] = sd[key].to(dtype) + + return sd, metadata + + +def save_to_file(file_name, state_dict, metadata): + if os.path.splitext(file_name)[1] == ".safetensors": + save_file(state_dict, file_name, metadata=metadata) + else: + torch.save(state_dict, file_name) + + +def format_lbws(lbws): + try: + # lbwは"[1,1,1,1,1,1,1,1,1,1,1,1]"のような文字列で与えられることを期待している + lbws = [json.loads(lbw) for lbw in lbws] + except Exception: + raise ValueError(f"format of lbws are must be json / 層別適用率はJSON形式で書いてください") + assert all(isinstance(lbw, list) for lbw in lbws), f"lbws are must be list / 層別適用率はリストにしてください" + assert len(set(len(lbw) for lbw in lbws)) == 1, "all lbws should have the same length / 層別適用率は同じ長さにしてください" + assert all( + len(lbw) in ACCEPTABLE for lbw in lbws + ), f"length of lbw are must be in {ACCEPTABLE} / 層別適用率の長さは{ACCEPTABLE}のいずれかにしてください" + assert all( + all(isinstance(weight, (int, float)) for weight in lbw) for lbw in lbws + ), f"values of lbs are must be numbers / 層別適用率の値はすべて数値にしてください" + + layer_num = len(lbws[0]) + is_sdxl = True if layer_num in SDXL_LAYER_NUM else False + FLAGS = { + "12": LAYER12.values(), + "17": LAYER17.values(), + "20": LAYER20.values(), + "26": LAYER26.values(), + }[str(layer_num)] + LBW_TARGET_IDX = [i for i, flag in enumerate(FLAGS) if flag] + return lbws, is_sdxl, LBW_TARGET_IDX + + +def merge_lora_models(models, ratios, lbws, new_rank, new_conv_rank, device, merge_dtype): + logger.info(f"new rank: {new_rank}, new conv rank: {new_conv_rank}") + merged_sd = {} + v2 = None # This is meaning LoRA Metadata v2, Not meaning SD2 + base_model = None + + if lbws: + lbws, is_sdxl, LBW_TARGET_IDX = format_lbws(lbws) + else: + is_sdxl = False + LBW_TARGET_IDX = [] + + for model, ratio, lbw in itertools.zip_longest(models, ratios, lbws): + logger.info(f"loading: {model}") + lora_sd, lora_metadata = load_state_dict(model, merge_dtype) + + if lora_metadata is not None: + if v2 is None: + v2 = lora_metadata.get(train_util.SS_METADATA_KEY_V2, None) # return string + if base_model is None: + base_model = lora_metadata.get(train_util.SS_METADATA_KEY_BASE_MODEL_VERSION, None) + + if lbw: + lbw_weights = [1] * 26 + for index, value in zip(LBW_TARGET_IDX, lbw): + lbw_weights[index] = value + logger.info(f"lbw: {dict(zip(LAYER26.keys(), lbw_weights))}") + + # merge + logger.info(f"merging...") + for key in tqdm(list(lora_sd.keys())): + if "lora_down" not in key: + continue + + lora_module_name = key[: key.rfind(".lora_down")] + + down_weight = lora_sd[key] + network_dim = down_weight.size()[0] + + up_weight = lora_sd[lora_module_name + ".lora_up.weight"] + alpha = lora_sd.get(lora_module_name + ".alpha", network_dim) + + in_dim = down_weight.size()[1] + out_dim = up_weight.size()[0] + conv2d = len(down_weight.size()) == 4 + kernel_size = None if not conv2d else down_weight.size()[2:4] + # logger.info(lora_module_name, network_dim, alpha, in_dim, out_dim, kernel_size) + + # make original weight if not exist + if lora_module_name not in merged_sd: + weight = torch.zeros((out_dim, in_dim, *kernel_size) if conv2d else (out_dim, in_dim), dtype=merge_dtype) + else: + weight = merged_sd[lora_module_name] + if device: + weight = weight.to(device) + + # merge to weight + if device: + up_weight = up_weight.to(device) + down_weight = down_weight.to(device) + + # W <- W + U * D + scale = alpha / network_dim + + if lbw: + index = get_lbw_block_index(key, is_sdxl) + is_lbw_target = index in LBW_TARGET_IDX + if is_lbw_target: + scale *= lbw_weights[index] # keyがlbwの対象であれば、lbwの重みを掛ける + + if device: # and isinstance(scale, torch.Tensor): + scale = scale.to(device) + + if not conv2d: # linear + weight = weight + ratio * (up_weight @ down_weight) * scale + elif kernel_size == (1, 1): + weight = ( + weight + + ratio + * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) + * scale + ) + else: + conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3) + weight = weight + ratio * conved * scale + + merged_sd[lora_module_name] = weight.to("cpu") + + # extract from merged weights + logger.info("extract new lora...") + merged_lora_sd = {} + with torch.no_grad(): + for lora_module_name, mat in tqdm(list(merged_sd.items())): + if device: + mat = mat.to(device) + + conv2d = len(mat.size()) == 4 + kernel_size = None if not conv2d else mat.size()[2:4] + conv2d_3x3 = conv2d and kernel_size != (1, 1) + out_dim, in_dim = mat.size()[0:2] + + if conv2d: + if conv2d_3x3: + mat = mat.flatten(start_dim=1) + else: + mat = mat.squeeze() + + module_new_rank = new_conv_rank if conv2d_3x3 else new_rank + module_new_rank = min(module_new_rank, in_dim, out_dim) # LoRA rank cannot exceed the original dim + + U, S, Vh = torch.linalg.svd(mat) + + U = U[:, :module_new_rank] + S = S[:module_new_rank] + U = U @ torch.diag(S) + + Vh = Vh[:module_new_rank, :] + + dist = torch.cat([U.flatten(), Vh.flatten()]) + hi_val = torch.quantile(dist, CLAMP_QUANTILE) + low_val = -hi_val + + U = U.clamp(low_val, hi_val) + Vh = Vh.clamp(low_val, hi_val) + + if conv2d: + U = U.reshape(out_dim, module_new_rank, 1, 1) + Vh = Vh.reshape(module_new_rank, in_dim, kernel_size[0], kernel_size[1]) + + up_weight = U + down_weight = Vh + + merged_lora_sd[lora_module_name + ".lora_up.weight"] = up_weight.to("cpu").contiguous() + merged_lora_sd[lora_module_name + ".lora_down.weight"] = down_weight.to("cpu").contiguous() + merged_lora_sd[lora_module_name + ".alpha"] = torch.tensor(module_new_rank, device="cpu") + + # build minimum metadata + dims = f"{new_rank}" + alphas = f"{new_rank}" + if new_conv_rank is not None: + network_args = {"conv_dim": new_conv_rank, "conv_alpha": new_conv_rank} + else: + network_args = None + metadata = train_util.build_minimum_network_metadata(v2, base_model, "networks.lora", dims, alphas, network_args) + + return merged_lora_sd, metadata, v2 == "True", base_model + + +def merge(args): + assert len(args.models) == len( + args.ratios + ), f"number of models must be equal to number of ratios / モデルの数と重みの数は合わせてください" + if args.lbws: + assert len(args.models) == len( + args.lbws + ), f"number of models must be equal to number of ratios / モデルの数と層別適用率の数は合わせてください" + else: + args.lbws = [] # zip_longestで扱えるようにlbws未使用時には空のリストにしておく + + def str_to_dtype(p): + if p == "float": + return torch.float + if p == "fp16": + return torch.float16 + if p == "bf16": + return torch.bfloat16 + return None + + merge_dtype = str_to_dtype(args.precision) + save_dtype = str_to_dtype(args.save_precision) + if save_dtype is None: + save_dtype = merge_dtype + + new_conv_rank = args.new_conv_rank if args.new_conv_rank is not None else args.new_rank + state_dict, metadata, v2, base_model = merge_lora_models( + args.models, args.ratios, args.lbws, args.new_rank, new_conv_rank, args.device, merge_dtype + ) + + # cast to save_dtype before calculating hashes + for key in list(state_dict.keys()): + value = state_dict[key] + if type(value) == torch.Tensor and value.dtype.is_floating_point and value.dtype != save_dtype: + state_dict[key] = value.to(save_dtype) + + logger.info(f"calculating hashes and creating metadata...") + + model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata) + metadata["sshs_model_hash"] = model_hash + metadata["sshs_legacy_hash"] = legacy_hash + + if not args.no_metadata: + is_sdxl = base_model is not None and base_model.lower().startswith("sdxl") + merged_from = sai_model_spec.build_merged_from(args.models) + title = os.path.splitext(os.path.basename(args.save_to))[0] + sai_metadata = sai_model_spec.build_metadata( + state_dict, v2, v2, is_sdxl, True, False, time.time(), title=title, merged_from=merged_from + ) + if v2: + # TODO read sai modelspec + logger.warning( + "Cannot determine if LoRA is for v-prediction, so save metadata as v-prediction / LoRAがv-prediction用か否か不明なため、仮にv-prediction用としてmetadataを保存します" + ) + metadata.update(sai_metadata) + + logger.info(f"saving model to: {args.save_to}") + save_to_file(args.save_to, state_dict, metadata) + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument( + "--save_precision", + type=str, + default=None, + choices=[None, "float", "fp16", "bf16"], + help="precision in saving, same to merging if omitted / 保存時に精度を変更して保存する、省略時はマージ時の精度と同じ", + ) + parser.add_argument( + "--precision", + type=str, + default="float", + choices=["float", "fp16", "bf16"], + help="precision in merging (float is recommended) / マージの計算時の精度（floatを推奨）", + ) + parser.add_argument( + "--save_to", + type=str, + default=None, + help="destination file name: ckpt or safetensors file / 保存先のファイル名、ckptまたはsafetensors", + ) + parser.add_argument( + "--models", + type=str, + nargs="*", + help="LoRA models to merge: ckpt or safetensors file / マージするLoRAモデル、ckptまたはsafetensors", + ) + parser.add_argument("--ratios", type=float, nargs="*", help="ratios for each model / それぞれのLoRAモデルの比率") + parser.add_argument("--lbws", type=str, nargs="*", help="lbw for each model / それぞれのLoRAモデルの層別適用率") + parser.add_argument("--new_rank", type=int, default=4, help="Specify rank of output LoRA / 出力するLoRAのrank (dim)") + parser.add_argument( + "--new_conv_rank", + type=int, + default=None, + help="Specify rank of output LoRA for Conv2d 3x3, None for same as new_rank / 出力するConv2D 3x3 LoRAのrank (dim)、Noneでnew_rankと同じ", + ) + parser.add_argument( + "--device", type=str, default=None, help="device to use, cuda for GPU / 計算を行うデバイス、cuda でGPUを使う" + ) + parser.add_argument( + "--no_metadata", + action="store_true", + help="do not save sai modelspec metadata (minimum ss_metadata for LoRA is saved) / " + + "sai modelspecのメタデータを保存しない（LoRAの最低限のss_metadataは保存される）", + ) + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + merge(args) diff --git a/tag_images_by_wd14_tagger.py b/tag_images_by_wd14_tagger.py new file mode 100644 index 0000000000000000000000000000000000000000..cbc3d2d6bae70b71e6985c2f682e5250cec5b53e --- /dev/null +++ b/tag_images_by_wd14_tagger.py @@ -0,0 +1,515 @@ +import argparse +import csv +import os +from pathlib import Path + +import cv2 +import numpy as np +import torch +from huggingface_hub import hf_hub_download +from PIL import Image +from tqdm import tqdm + +import library.train_util as train_util +from library.utils import setup_logging, pil_resize + +setup_logging() +import logging + +logger = logging.getLogger(__name__) + +# from wd14 tagger +IMAGE_SIZE = 448 + +# wd-v1-4-swinv2-tagger-v2 / wd-v1-4-vit-tagger / wd-v1-4-vit-tagger-v2/ wd-v1-4-convnext-tagger / wd-v1-4-convnext-tagger-v2 +DEFAULT_WD14_TAGGER_REPO = "SmilingWolf/wd-v1-4-convnext-tagger-v2" +FILES = ["keras_metadata.pb", "saved_model.pb", "selected_tags.csv"] +FILES_ONNX = ["model.onnx"] +SUB_DIR = "variables" +SUB_DIR_FILES = ["variables.data-00000-of-00001", "variables.index"] +CSV_FILE = FILES[-1] + + +def preprocess_image(image): + image = np.array(image) + image = image[:, :, ::-1] # RGB->BGR + + # pad to square + size = max(image.shape[0:2]) + pad_x = size - image.shape[1] + pad_y = size - image.shape[0] + pad_l = pad_x // 2 + pad_t = pad_y // 2 + image = np.pad(image, ((pad_t, pad_y - pad_t), (pad_l, pad_x - pad_l), (0, 0)), mode="constant", constant_values=255) + + if size > IMAGE_SIZE: + image = cv2.resize(image, (IMAGE_SIZE, IMAGE_SIZE), cv2.INTER_AREA) + else: + image = pil_resize(image, (IMAGE_SIZE, IMAGE_SIZE)) + + image = image.astype(np.float32) + return image + + +class ImageLoadingPrepDataset(torch.utils.data.Dataset): + def __init__(self, image_paths): + self.images = image_paths + + def __len__(self): + return len(self.images) + + def __getitem__(self, idx): + img_path = str(self.images[idx]) + + try: + image = Image.open(img_path).convert("RGB") + image = preprocess_image(image) + # tensor = torch.tensor(image) # これ Tensor に変換する必要ないな……(;･∀･) + except Exception as e: + logger.error(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}") + return None + + return (image, img_path) + + +def collate_fn_remove_corrupted(batch): + """Collate function that allows to remove corrupted examples in the + dataloader. It expects that the dataloader returns 'None' when that occurs. + The 'None's in the batch are removed. + """ + # Filter out all the Nones (corrupted examples) + batch = list(filter(lambda x: x is not None, batch)) + return batch + + +def main(args): + # model location is model_dir + repo_id + # repo id may be like "user/repo" or "user/repo/branch", so we need to remove slash + model_location = os.path.join(args.model_dir, args.repo_id.replace("/", "_")) + + # hf_hub_downloadをそのまま使うとsymlink関係で問題があるらしいので、キャッシュディレクトリとforce_filenameを指定してなんとかする + # depreacatedの警告が出るけどなくなったらその時 + # https://github.com/toriato/stable-diffusion-webui-wd14-tagger/issues/22 + if not os.path.exists(model_location) or args.force_download: + os.makedirs(args.model_dir, exist_ok=True) + logger.info(f"downloading wd14 tagger model from hf_hub. id: {args.repo_id}") + files = FILES + if args.onnx: + files = ["selected_tags.csv"] + files += FILES_ONNX + else: + for file in SUB_DIR_FILES: + hf_hub_download( + args.repo_id, + file, + subfolder=SUB_DIR, + cache_dir=os.path.join(model_location, SUB_DIR), + force_download=True, + force_filename=file, + ) + for file in files: + hf_hub_download(args.repo_id, file, cache_dir=model_location, force_download=True, force_filename=file) + else: + logger.info("using existing wd14 tagger model") + + # モデルを読み込む + if args.onnx: + import onnx + import onnxruntime as ort + + onnx_path = f"{model_location}/model.onnx" + logger.info("Running wd14 tagger with onnx") + logger.info(f"loading onnx model: {onnx_path}") + + if not os.path.exists(onnx_path): + raise Exception( + f"onnx model not found: {onnx_path}, please redownload the model with --force_download" + + " / onnxモデルが見つかりませんでした。--force_downloadで再ダウンロードしてください" + ) + + model = onnx.load(onnx_path) + input_name = model.graph.input[0].name + try: + batch_size = model.graph.input[0].type.tensor_type.shape.dim[0].dim_value + except Exception: + batch_size = model.graph.input[0].type.tensor_type.shape.dim[0].dim_param + + if args.batch_size != batch_size and not isinstance(batch_size, str) and batch_size > 0: + # some rebatch model may use 'N' as dynamic axes + logger.warning( + f"Batch size {args.batch_size} doesn't match onnx model batch size {batch_size}, use model batch size {batch_size}" + ) + args.batch_size = batch_size + + del model + + if "OpenVINOExecutionProvider" in ort.get_available_providers(): + # requires provider options for gpu support + # fp16 causes nonsense outputs + ort_sess = ort.InferenceSession( + onnx_path, + providers=(["OpenVINOExecutionProvider"]), + provider_options=[{'device_type' : "GPU_FP32"}], + ) + else: + ort_sess = ort.InferenceSession( + onnx_path, + providers=( + ["CUDAExecutionProvider"] if "CUDAExecutionProvider" in ort.get_available_providers() else + ["ROCMExecutionProvider"] if "ROCMExecutionProvider" in ort.get_available_providers() else + ["CPUExecutionProvider"] + ), + ) + else: + from tensorflow.keras.models import load_model + + model = load_model(f"{model_location}") + + # label_names = pd.read_csv("2022_0000_0899_6549/selected_tags.csv") + # 依存ライブラリを増やしたくないので自力で読むよ + + with open(os.path.join(model_location, CSV_FILE), "r", encoding="utf-8") as f: + reader = csv.reader(f) + line = [row for row in reader] + header = line[0] # tag_id,name,category,count + rows = line[1:] + assert header[0] == "tag_id" and header[1] == "name" and header[2] == "category", f"unexpected csv format: {header}" + + rating_tags = [row[1] for row in rows[0:] if row[2] == "9"] + general_tags = [row[1] for row in rows[0:] if row[2] == "0"] + character_tags = [row[1] for row in rows[0:] if row[2] == "4"] + + # preprocess tags in advance + if args.character_tag_expand: + for i, tag in enumerate(character_tags): + if tag.endswith(")"): + # chara_name_(series) -> chara_name, series + # chara_name_(costume)_(series) -> chara_name_(costume), series + tags = tag.split("(") + character_tag = "(".join(tags[:-1]) + if character_tag.endswith("_"): + character_tag = character_tag[:-1] + series_tag = tags[-1].replace(")", "") + character_tags[i] = character_tag + args.caption_separator + series_tag + + if args.remove_underscore: + rating_tags = [tag.replace("_", " ") if len(tag) > 3 else tag for tag in rating_tags] + general_tags = [tag.replace("_", " ") if len(tag) > 3 else tag for tag in general_tags] + character_tags = [tag.replace("_", " ") if len(tag) > 3 else tag for tag in character_tags] + + if args.tag_replacement is not None: + # escape , and ; in tag_replacement: wd14 tag names may contain , and ; + escaped_tag_replacements = args.tag_replacement.replace("\\,", "@@@@").replace("\\;", "####") + tag_replacements = escaped_tag_replacements.split(";") + for tag_replacement in tag_replacements: + tags = tag_replacement.split(",") # source, target + assert len(tags) == 2, f"tag replacement must be in the format of `source,target` / タグの置換は `置換元,置換先` の形式で指定してください: {args.tag_replacement}" + + source, target = [tag.replace("@@@@", ",").replace("####", ";") for tag in tags] + logger.info(f"replacing tag: {source} -> {target}") + + if source in general_tags: + general_tags[general_tags.index(source)] = target + elif source in character_tags: + character_tags[character_tags.index(source)] = target + elif source in rating_tags: + rating_tags[rating_tags.index(source)] = target + + # 画像を読み込む + train_data_dir_path = Path(args.train_data_dir) + image_paths = train_util.glob_images_pathlib(train_data_dir_path, args.recursive) + logger.info(f"found {len(image_paths)} images.") + + tag_freq = {} + + caption_separator = args.caption_separator + stripped_caption_separator = caption_separator.strip() + undesired_tags = args.undesired_tags.split(stripped_caption_separator) + undesired_tags = set([tag.strip() for tag in undesired_tags if tag.strip() != ""]) + + always_first_tags = None + if args.always_first_tags is not None: + always_first_tags = [tag for tag in args.always_first_tags.split(stripped_caption_separator) if tag.strip() != ""] + + def run_batch(path_imgs): + imgs = np.array([im for _, im in path_imgs]) + + if args.onnx: + # if len(imgs) < args.batch_size: + # imgs = np.concatenate([imgs, np.zeros((args.batch_size - len(imgs), IMAGE_SIZE, IMAGE_SIZE, 3))], axis=0) + probs = ort_sess.run(None, {input_name: imgs})[0] # onnx output numpy + probs = probs[: len(path_imgs)] + else: + probs = model(imgs, training=False) + probs = probs.numpy() + + for (image_path, _), prob in zip(path_imgs, probs): + combined_tags = [] + rating_tag_text = "" + character_tag_text = "" + general_tag_text = "" + + # 最初の4つ以降はタグなのでconfidenceがthreshold以上のものを追加する + # First 4 labels are ratings, the rest are tags: pick any where prediction confidence >= threshold + for i, p in enumerate(prob[4:]): + if i < len(general_tags) and p >= args.general_threshold: + tag_name = general_tags[i] + + if tag_name not in undesired_tags: + tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1 + general_tag_text += caption_separator + tag_name + combined_tags.append(tag_name) + elif i >= len(general_tags) and p >= args.character_threshold: + tag_name = character_tags[i - len(general_tags)] + + if tag_name not in undesired_tags: + tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1 + character_tag_text += caption_separator + tag_name + if args.character_tags_first: # insert to the beginning + combined_tags.insert(0, tag_name) + else: + combined_tags.append(tag_name) + + # 最初の4つはratingなのでargmaxで選ぶ + # First 4 labels are actually ratings: pick one with argmax + if args.use_rating_tags or args.use_rating_tags_as_last_tag: + ratings_probs = prob[:4] + rating_index = ratings_probs.argmax() + found_rating = rating_tags[rating_index] + + if found_rating not in undesired_tags: + tag_freq[found_rating] = tag_freq.get(found_rating, 0) + 1 + rating_tag_text = found_rating + if args.use_rating_tags: + combined_tags.insert(0, found_rating) # insert to the beginning + else: + combined_tags.append(found_rating) + + # 一番最初に置くタグを指定する + # Always put some tags at the beginning + if always_first_tags is not None: + for tag in always_first_tags: + if tag in combined_tags: + combined_tags.remove(tag) + combined_tags.insert(0, tag) + + # 先頭のカンマを取る + if len(general_tag_text) > 0: + general_tag_text = general_tag_text[len(caption_separator) :] + if len(character_tag_text) > 0: + character_tag_text = character_tag_text[len(caption_separator) :] + + caption_file = os.path.splitext(image_path)[0] + args.caption_extension + + tag_text = caption_separator.join(combined_tags) + + if args.append_tags: + # Check if file exists + if os.path.exists(caption_file): + with open(caption_file, "rt", encoding="utf-8") as f: + # Read file and remove new lines + existing_content = f.read().strip("\n") # Remove newlines + + # Split the content into tags and store them in a list + existing_tags = [tag.strip() for tag in existing_content.split(stripped_caption_separator) if tag.strip()] + + # Check and remove repeating tags in tag_text + new_tags = [tag for tag in combined_tags if tag not in existing_tags] + + # Create new tag_text + tag_text = caption_separator.join(existing_tags + new_tags) + + with open(caption_file, "wt", encoding="utf-8") as f: + f.write(tag_text + "\n") + if args.debug: + logger.info("") + logger.info(f"{image_path}:") + logger.info(f"\tRating tags: {rating_tag_text}") + logger.info(f"\tCharacter tags: {character_tag_text}") + logger.info(f"\tGeneral tags: {general_tag_text}") + + # 読み込みの高速化のためにDataLoaderを使うオプション + if args.max_data_loader_n_workers is not None: + dataset = ImageLoadingPrepDataset(image_paths) + data = torch.utils.data.DataLoader( + dataset, + batch_size=args.batch_size, + shuffle=False, + num_workers=args.max_data_loader_n_workers, + collate_fn=collate_fn_remove_corrupted, + drop_last=False, + ) + else: + data = [[(None, ip)] for ip in image_paths] + + b_imgs = [] + for data_entry in tqdm(data, smoothing=0.0): + for data in data_entry: + if data is None: + continue + + image, image_path = data + if image is None: + try: + image = Image.open(image_path) + if image.mode != "RGB": + image = image.convert("RGB") + image = preprocess_image(image) + except Exception as e: + logger.error(f"Could not load image path / 画像を読み込めません: {image_path}, error: {e}") + continue + b_imgs.append((image_path, image)) + + if len(b_imgs) >= args.batch_size: + b_imgs = [(str(image_path), image) for image_path, image in b_imgs] # Convert image_path to string + run_batch(b_imgs) + b_imgs.clear() + + if len(b_imgs) > 0: + b_imgs = [(str(image_path), image) for image_path, image in b_imgs] # Convert image_path to string + run_batch(b_imgs) + + if args.frequency_tags: + sorted_tags = sorted(tag_freq.items(), key=lambda x: x[1], reverse=True) + print("Tag frequencies:") + for tag, freq in sorted_tags: + print(f"{tag}: {freq}") + + logger.info("done!") + + +def setup_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser() + parser.add_argument( + "train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ" + ) + parser.add_argument( + "--repo_id", + type=str, + default=DEFAULT_WD14_TAGGER_REPO, + help="repo id for wd14 tagger on Hugging Face / Hugging Faceのwd14 taggerのリポジトリID", + ) + parser.add_argument( + "--model_dir", + type=str, + default="wd14_tagger_model", + help="directory to store wd14 tagger model / wd14 taggerのモデルを格納するディレクトリ", + ) + parser.add_argument( + "--force_download", + action="store_true", + help="force downloading wd14 tagger models / wd14 taggerのモデルを再ダウンロードします", + ) + parser.add_argument( + "--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ" + ) + parser.add_argument( + "--max_data_loader_n_workers", + type=int, + default=None, + help="enable image reading by DataLoader with this number of workers (faster) / DataLoaderによる画像読み込みを有効にしてこのワーカー数を適用する（読み込みを高速化）", + ) + parser.add_argument( + "--caption_extention", + type=str, + default=None, + help="extension of caption file (for backward compatibility) / 出力されるキャプションファイルの拡張子（スペルミスしていたのを残してあります）", + ) + parser.add_argument( + "--caption_extension", type=str, default=".txt", help="extension of caption file / 出力されるキャプションファイルの拡張子" + ) + parser.add_argument( + "--thresh", type=float, default=0.35, help="threshold of confidence to add a tag / タグを追加するか判定する閾値" + ) + parser.add_argument( + "--general_threshold", + type=float, + default=None, + help="threshold of confidence to add a tag for general category, same as --thresh if omitted / generalカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ", + ) + parser.add_argument( + "--character_threshold", + type=float, + default=None, + help="threshold of confidence to add a tag for character category, same as --thres if omitted / characterカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ", + ) + parser.add_argument( + "--recursive", action="store_true", help="search for images in subfolders recursively / サブフォルダを再帰的に検索する" + ) + parser.add_argument( + "--remove_underscore", + action="store_true", + help="replace underscores with spaces in the output tags / 出力されるタグのアンダースコアをスペースに置き換える", + ) + parser.add_argument( + "--debug", action="store_true", help="debug mode" + ) + parser.add_argument( + "--undesired_tags", + type=str, + default="", + help="comma-separated list of undesired tags to remove from the output / 出力から除外したいタグのカンマ区切りのリスト", + ) + parser.add_argument( + "--frequency_tags", action="store_true", help="Show frequency of tags for images / タグの出現頻度を表示する" + ) + parser.add_argument( + "--onnx", action="store_true", help="use onnx model for inference / onnxモデルを推論に使用する" + ) + parser.add_argument( + "--append_tags", action="store_true", help="Append captions instead of overwriting / 上書きではなくキャプションを追記する" + ) + parser.add_argument( + "--use_rating_tags", action="store_true", help="Adds rating tags as the first tag / レーティングタグを最初のタグとして追加する", + ) + parser.add_argument( + "--use_rating_tags_as_last_tag", action="store_true", help="Adds rating tags as the last tag / レーティングタグを最後のタグとして追加する", + ) + parser.add_argument( + "--character_tags_first", action="store_true", help="Always inserts character tags before the general tags / characterタグを常にgeneralタグの前に出力する", + ) + parser.add_argument( + "--always_first_tags", + type=str, + default=None, + help="comma-separated list of tags to always put at the beginning, e.g. `1girl,1boy`" + + " / 必ず先頭に置くタグのカンマ区切りリスト、例 : `1girl,1boy`", + ) + parser.add_argument( + "--caption_separator", + type=str, + default=", ", + help="Separator for captions, include space if needed / キャプションの区切り文字、必要ならスペースを含めてください", + ) + parser.add_argument( + "--tag_replacement", + type=str, + default=None, + help="tag replacement in the format of `source1,target1;source2,target2; ...`. Escape `,` and `;` with `\`. e.g. `tag1,tag2;tag3,tag4`" + + " / タグの置換を `置換元1,置換先1;置換元2,置換先2; ...`で指定する。`\` で `,` と `;` をエスケープできる。例: `tag1,tag2;tag3,tag4`", + ) + parser.add_argument( + "--character_tag_expand", + action="store_true", + help="expand tag tail parenthesis to another tag for character tags. `chara_name_(series)` becomes `chara_name, series`" + + " / キャラクタタグの末尾の括弧を別のタグに展開する。`chara_name_(series)` は `chara_name, series` になる", + ) + + return parser + + +if __name__ == "__main__": + parser = setup_parser() + + args = parser.parse_args() + + # スペルミスしていたオプションを復元する + if args.caption_extention is not None: + args.caption_extension = args.caption_extention + + if args.general_threshold is None: + args.general_threshold = args.thresh + if args.character_threshold is None: + args.character_threshold = args.thresh + + main(args) diff --git a/train_README-ja.md b/train_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..cfa5a7d1c0c1f63b9ede89b5d81f969f710f2ddc --- /dev/null +++ b/train_README-ja.md @@ -0,0 +1,1008 @@ +__ドキュメント更新中のため記述に誤りがあるかもしれません。__ + +# 学習について、共通編 + +当リポジトリではモデルのfine tuning、DreamBooth、およびLoRAとTextual Inversion（[XTI:P+](https://github.com/kohya-ss/sd-scripts/pull/327)を含む）の学習をサポートします。この文書ではそれらに共通する、学習データの準備方法やオプション等について説明します。 + +# 概要 + +あらかじめこのリポジトリのREADMEを参照し、環境整備を行ってください。 + + +以下について説明します。 + +1. 学習データの準備について（設定ファイルを用いる新形式） +1. 学習で使われる用語のごく簡単な解説 +1. 以前の指定形式（設定ファイルを用いずコマンドラインから指定） +1. 学習途中のサンプル画像生成 +1. 各スクリプトで共通の、よく使われるオプション +1. fine tuning 方式のメタデータ準備：キャプションニングなど + +1.だけ実行すればとりあえず学習は可能です（学習については各スクリプトのドキュメントを参照）。2.以降は必要に応じて参照してください。 + + +# 学習データの準備について + +任意のフォルダ（複数でも可）に学習データの画像ファイルを用意しておきます。`.png`, `.jpg`, `.jpeg`, `.webp`, `.bmp` をサポートします。リサイズなどの前処理は基本的に必要ありません。 + +ただし学習解像度（後述）よりも極端に小さい画像は使わないか、あらかじめ超解像AIなどで拡大しておくことをお勧めします。また極端に大きな画像（3000x3000ピクセル程度？）よりも大きな画像はエラーになる場合があるようですので事前に縮小してください。 + +学習時には、モデルに学ばせる画像データを整理し、スクリプトに対して指定する必要があります。学習データの数、学習対象、キャプション（画像の説明）が用意できるか否かなどにより、いくつかの方法で学習データを指定できます。以下の方式があります（それぞれの名前は一般的なものではなく、当リポジトリ独自の定義です）。正則化画像については後述します。 + +1. DreamBooth、class+identifier方式（正則化画像使用可） + + 特定の単語 (identifier) に学習対象を紐づけるように学習します。キャプションを用意する必要はありません。たとえば特定のキャラを学ばせる場合に使うとキャプションを用意する必要がない分、手軽ですが、髪型や服装、背景など学習データの全要素が identifier に紐づけられて学習されるため、生成時のプロンプトで服が変えられない、といった事態も起こりえます。 + +1. DreamBooth、キャプション方式（正則化画像使用可） + + 画像ごとにキャプションが記録されたテキストファイルを用意して学習します。たとえば特定のキャラを学ばせると、画像の詳細をキャプションに記述することで（白い服を着たキャラA、赤い服を着たキャラA、など）キャラとそれ以外の要素が分離され、より厳密にモデルがキャラだけを学ぶことが期待できます。 + +1. fine tuning方式（正則化画像使用不可） + + あらかじめキャプションをメタデータファイルにまとめます。タグとキャプションを分けて管理したり、学習を高速化するためlatentsを事前キャッシュしたりなどの機能をサポートします（いずれも別文書で説明しています）。（fine tuning方式という名前ですが fine tuning 以外でも使えます。） + +学習したいものと使用できる指定方法の組み合わせは以下の通りです。 + +| 学習対象または方法 | スクリプト | DB / class+identifier | DB / キャプション | fine tuning | +| ----- | ----- | ----- | ----- | ----- | +| モデルをfine tuning | `fine_tune.py`| x | x | o | +| モデルをDreamBooth | `train_db.py`| o | o | x | +| LoRA | `train_network.py`| o | o | o | +| Textual Invesion | `train_textual_inversion.py`| o | o | o | + +## どれを選ぶか + +LoRA、Textual Inversionについては、手軽にキャプションファイルを用意せずに学習したい場合はDreamBooth class+identifier、用意できるならDreamBooth キャプション方式がよいでしょう。学習データの枚数が多く、かつ正則化画像を使用しない場合はfine tuning方式も検討してください。 + +DreamBoothについても同様ですが、fine tuning方式は使えません。fine tuningの場合はfine tuning方式のみです。 + +# 各方式の指定方法について + +ここではそれぞれの指定方法で典型的なパターンについてだけ説明します。より詳細な指定方法については [データセット設定](./config_README-ja.md) をご覧ください。 + +# DreamBooth、class+identifier方式（正則化画像使用可） + +この方式では、各画像は `class identifier` というキャプションで学習されたのと同じことになります（`shs dog` など）。 + +## step 1. identifierとclassを決める + +学ばせたい対象を結びつける単語identifierと、対象の属するclassを決めます。 + +（instanceなどいろいろな呼び方がありますが、とりあえず元の論文に合わせます。） + +以下ごく簡単に説明します（詳しくは調べてください）。 + +classは学習対象の一般的な種別です。たとえば特定の犬種を学ばせる場合には、classはdogになります。アニメキャラならモデルによりboyやgirl、1boyや1girlになるでしょう。 + +identifierは学習対象を識別して学習するためのものです。任意の単語で構いませんが、元論文によると「tokinizerで1トークンになる3文字以下でレアな単語」が良いとのことです。 + +identifierとclassを使い、たとえば「shs dog」などでモデルを学習することで、学習させたい対象をclassから識別して学習できます。 + +画像生成時には「shs dog」とすれば学ばせた犬種の画像が生成されます。 + +（identifierとして私が最近使っているものを参考までに挙げると、``shs sts scs cpc coc cic msm usu ici lvl cic dii muk ori hru rik koo yos wny`` などです。本当は Danbooru Tag に含まれないやつがより望ましいです。） + +## step 2. 正則化画像を使うか否かを決め、使う場合には正則化画像を生成する + +正則化画像とは、前述のclass全体が、学習対象に引っ張られることを防ぐための画像です（language drift）。正則化画像を使わないと、たとえば `shs 1girl` で特定のキャラクタを学ばせると、単なる `1girl` というプロンプトで生成してもそのキャラに似てきます。これは `1girl` が学習時のキャプションに含まれているためです。 + +学習対象の画像と正則化画像を同時に学ばせることで、class は class のままで留まり、identifier をプロンプトにつけた時だけ学習対象が生成されるようになります。 + +LoRAやDreamBoothで特定のキャラだけ出てくればよい場合は、正則化画像を用いなくても良いといえます。 + +Textual Inversionでは用いなくてよいでしょう（学ばせる token string がキャプションに含まれない場合はなにも学習されないため）。 + +正則化画像としては、学習対象のモデルで、class 名だけで生成した画像を用いるのが一般的です（たとえば `1girl`）。ただし生成画像の品質が悪い場合には、プロンプトを工夫したり、ネットから別途ダウンロードした画像を用いることもできます。 + +（正則化画像も学習されるため、その品質はモデルに影響します。） + +一般的には数百枚程度、用意するのが望ましいようです（枚数が少ないと class 画像が一般化されずそれらの特徴を学んでしまいます）。 + +生成画像を使う場合、通常、生成画像のサイズは学習解像度（より正確にはbucketの解像度、後述）にあわせてください。 + +## step 2. 設定ファイルの記述 + +テキストファイルを作成し、拡張子を `.toml` にします。たとえば以下のように記述します。 + +（`#` で始まっている部分はコメントですので、このままコピペしてそのままでもよいですし、削除しても問題ありません。） + +```toml +[general] +enable_bucket = true # Aspect Ratio Bucketingを使うか否か + +[[datasets]] +resolution = 512 # 学習解像度 +batch_size = 4 # バッチサイズ + + [[datasets.subsets]] + image_dir = 'C:\hoge' # 学習用画像を入れたフォルダを指定 + class_tokens = 'hoge girl' # identifier class を指定 + num_repeats = 10 # 学習用画像の繰り返し回数 + + # 以下は正則化画像を用いる場合のみ記述する。用いない場合は削除する + [[datasets.subsets]] + is_reg = true + image_dir = 'C:\reg' # 正則化画像を入れたフォルダを指定 + class_tokens = 'girl' # class を指定 + num_repeats = 1 # 正則化画像の繰り返し回数、基本的には1でよい +``` + +基本的には以下の場所のみ書き換えれば学習できます。 + +1. 学習解像度 + + 数値1つを指定すると正方形（`512`なら512x512）、鍵カッコカンマ区切りで2つ指定すると横×縦（`[512,768]`なら512x768）になります。SD1.x系ではもともとの学習解像度は512です。`[512,768]` 等の大きめの解像度を指定すると縦長、横長画像生成時の破綻を小さくできるかもしれません。SD2.x 768系では `768` です。 + +1. バッチサイズ + + 同時に何件のデータを学習するかを指定します。GPUのVRAMサイズ、学習解像度によって変わってきます。詳しくは後述します。またfine tuning/DreamBooth/LoRA等でも変わってきますので各スクリプトの説明もご覧ください。 + +1. フォルダ指定 + + 学習用画像、正則化画像（使用する場合のみ）のフォルダを指定します。画像データが含まれているフォルダそのものを指定します。 + +1. identifier と class の指定 + + 前述のサンプルの通りです。 + +1. 繰り返し回数 + + 後述します。 + +### 繰り返し回数について + +繰り返し回数は、正則化画像の枚数と学習用画像の枚数を調整するために用いられます。正則化画像の枚数は学習用画像よりも多いため、学習用画像を繰り返して枚数を合わせ、1対1の比率で学習できるようにします。 + +繰り返し回数は「 __学習用画像の繰り返し回数×学習用画像の枚数≧正則化画像の繰り返し回数×正則化画像の枚数__ 」となるように指定してください。 + +（1 epoch（データが一周すると1 epoch）のデータ数が「学習用画像の繰り返し回数×学習用画像の枚数」となります。正則化画像の枚数がそれより多いと、余った部分の正則化画像は使用されません。） + +## step 3. 学習 + +それぞれのドキュメントを参考に学習を行ってください。 + +# DreamBooth、キャプション方式（正則化画像使用可） + +この方式では各画像はキャプションで学習されます。 + +## step 1. キャプションファイルを準備する + +学習用画像のフォルダに、画像と同じファイル名で、拡張子 `.caption`（設定で変えられます）のファイルを置いてください。それぞれのファイルは1行のみとしてください。エンコーディングは `UTF-8` です。 + +## step 2. 正則化画像を使うか否かを決め、使う場合には正則化画像を生成する + +class+identifier形式と同様です。なお正則化画像にもキャプションを付けることができますが、通常は不要でしょう。 + +## step 2. 設定ファイルの記述 + +テキストファイルを作成し、拡張子を `.toml` にします。たとえば以下のように記述します。 + +```toml +[general] +enable_bucket = true # Aspect Ratio Bucketingを使うか否か + +[[datasets]] +resolution = 512 # 学習解像度 +batch_size = 4 # バッチサイズ + + [[datasets.subsets]] + image_dir = 'C:\hoge' # 学習用画像を入れたフォルダを指定 + caption_extension = '.caption' # キャプションファイルの拡張子　.txt を使う場合には書き換える + num_repeats = 10 # 学習用画像の繰り返し回数 + + # 以下は正則化画像を用いる場合のみ記述する。用いない場合は削除する + [[datasets.subsets]] + is_reg = true + image_dir = 'C:\reg' # 正則化画像を入れたフォルダを指定 + class_tokens = 'girl' # class を指定 + num_repeats = 1 # 正則化画像の繰り返し回数、基本的には1でよい +``` + +基本的には以下を場所のみ書き換えれば学習できます。特に記述がない部分は class+identifier 方式と同じです。 + +1. 学習解像度 +1. バッチサイズ +1. フォルダ指定 +1. キャプションファイルの拡張子 + + 任意の拡張子を指定できます。 +1. 繰り返し回数 + +## step 3. 学習 + +それぞれのドキュメントを参考に学習を行ってください。 + +# fine tuning 方式 + +## step 1. メタデータを準備する + +キャプションやタグをまとめた管理用ファイルをメタデータと呼びます。json形式で拡張子は `.json` + です。作成方法は長くなりますのでこの文書の末尾に書きました。 + +## step 2. 設定ファイルの記述 + +テキストファイルを作成し、拡張子を `.toml` にします。たとえば以下のように記述します。 + +```toml +[general] +shuffle_caption = true +keep_tokens = 1 + +[[datasets]] +resolution = 512 # 学習解像度 +batch_size = 4 # バッチサイズ + + [[datasets.subsets]] + image_dir = 'C:\piyo' # 学習用画像を入れたフォルダを指定 + metadata_file = 'C:\piyo\piyo_md.json' # メタデータファイル名 +``` + +基本的には以下を場所のみ書き換えれば学習できます。特に記述がない部分は DreamBooth, class+identifier 方式と同じです。 + +1. 学習解像度 +1. バッチサイズ +1. フォルダ指定 +1. メタデータファイル名 + + 後述の方法で作成したメタデータファイルを指定します。 + + +## step 3. 学習 + +それぞれのドキュメントを参考に学習を行ってください。 + +# 学習で使われる用語のごく簡単な解説 + +細かいことは省略していますし私も完全には理解していないため、詳しくは各自お調べください。 + +## fine tuning（ファインチューニング） + +モデルを学習して微調整することを指します。使われ方によって意味が異なってきますが、狭義のfine tuningはStable Diffusionの場合、モデルを画像とキャプションで学習することです。DreamBoothは狭義のfine tuningのひとつの特殊なやり方と言えます。広義のfine tuningは、LoRAやTextual Inversion、Hypernetworksなどを含み、モデルを学習することすべてを含みます。 + +## ステップ + +ざっくりいうと学習データで1回計算すると1ステップです。「学習データのキャプションを今のモデルに流してみて、出てくる画像を学習データの画像と比較し、学習データに近づくようにモデルをわずかに変更する」のが1ステップです。 + +## バッチサイズ + +バッチサイズは1ステップで何件のデータをまとめて計算するかを指定する値です。まとめて計算するため速度は相対的に向上します。また一般的には精度も高くなるといわれています。 + +`バッチサイズ×ステップ数` が学習に使われるデータの件数になります。そのため、バッチサイズを増やした分だけステップ数を減らすとよいでしょう。 + +（ただし、たとえば「バッチサイズ1で1600ステップ」と「バッチサイズ4で400ステップ」は同じ結果にはなりません。同じ学習率の場合、一般的には後者のほうが学習不足になります。学習率を多少大きくするか（たとえば `2e-6` など）、ステップ数をたとえば500ステップにするなどして工夫してください。） + +バッチサイズを大きくするとその分だけGPUメモリを消費します。メモリが足りなくなるとエラーになりますし、エラーにならないギリギリでは学習速度が低下します。タスクマネージャーや `nvidia-smi` コマンドで使用メモリ量を確認しながら調整するとよいでしょう。 + +なお、バッチは「一塊のデータ」位の意味です。 + +## 学習率 + +ざっくりいうと1ステップごとにどのくらい変化させるかを表します。大きな値を指定するとそれだけ速く学習が進みますが、変化しすぎてモデルが壊れたり、最適な状態にまで至れない場合があります。小さい値を指定すると学習速度は遅くなり、また最適な状態にやはり至れない場合があります。 + +fine tuning、DreamBoooth、LoRAそれぞれで大きく異なり、また学習データや学習させたいモデル、バッチサイズやステップ数によっても変わってきます。一般的な値から初めて学習状態を見ながら増減してください。 + +デフォルトでは学習全体を通して学習率は固定です。スケジューラの指定で学習率をどう変化させるか決められますので、それらによっても結果は変わってきます。 + +## エポック（epoch） + +学習データが一通り学習されると（データが一周すると）1 epochです。繰り返し回数を指定した場合は、その繰り返し後のデータが一周すると1 epochです。 + +1 epochのステップ数は、基本的には `データ件数÷バッチサイズ` ですが、Aspect Ratio Bucketing を使うと微妙に増えます（異なるbucketのデータは同じバッチにできないため、ステップ数が増えます）。 + +## Aspect Ratio Bucketing + +Stable Diffusion のv1は512\*512で学習されていますが、それに加えて256\*1024や384\*640といった解像度でも学習します。これによりトリミングされる部分が減り、より正しくキャプションと画像の関係が学習されることが期待されます。 + +また任意の解像度で学習するため、事前に画像データの縦横比を統一しておく必要がなくなります。 + +設定で有効、無効が切り替えられますが、ここまでの設定ファイルの記述例では有効になっています（`true` が設定されています）。 + +学習解像度はパラメータとして与えられた解像度の面積（＝メモリ使用量）を超えない範囲で、64ピクセル単位（デフォルト、変更可）で縦横に調整、作成されます。 + +機械学習では入力サイズをすべて統一するのが一般的ですが、特に制約があるわけではなく、実際は同一のバッチ内で統一されていれば大丈夫です。NovelAIの言うbucketingは、あらかじめ教師データを、アスペクト比に応じた学習解像度ごとに分類しておくことを指しているようです。そしてバッチを各bucket内の画像で作成することで、バッチの画像サイズを統一します。 + +# 以前の指定形式（設定ファイルを用いずコマンドラインから指定） + +`.toml` ファイルを指定せずコマンドラインオプションで指定する方法です。DreamBooth class+identifier方式、DreamBooth キャプション方式、fine tuning方式があります。 + +## DreamBooth、class+identifier方式 + +フォルダ名で繰り返し回数を指定します。また `train_data_dir` オプションと `reg_data_dir` オプションを用います。 + +### step 1. 学習用画像の準備 + +学習用画像を格納するフォルダを作成します。 __さらにその中に__ 、以下の名前でディレクトリを作成します。 + +``` +<繰り返し回数>_ +``` + +間の``_``を忘れないでください。 + +たとえば「sls frog」というプロンプトで、データを20回繰り返す場合、「20_sls frog」となります。以下のようになります。 + +![image](https://user-images.githubusercontent.com/52813779/210770636-1c851377-5936-4c15-90b7-8ac8ad6c2074.png) + +### 複数class、複数対象（identifier）の学習 + +方法は単純で、学習用画像のフォルダ内に ``繰り返し回数_ `` のフォルダを複数、正則化画像フォルダにも同様に ``繰り返し回数_`` のフォルダを複数、用意してください。 + +たとえば「sls frog」と「cpc rabbit」を同時に学習する場合、以下のようになります。 + +![image](https://user-images.githubusercontent.com/52813779/210777933-a22229db-b219-4cd8-83ca-e87320fc4192.png) + +classがひとつで対象が複数の場合、正則化画像フォルダはひとつで構いません。たとえば1girlにキャラAとキャラBがいる場合は次のようにします。 + +- train_girls + - 10_sls 1girl + - 10_cpc 1girl +- reg_girls + - 1_1girl + +### step 2. 正則化画像の準備 + +正則化画像を使う場合の手順です。 + +正則化画像を格納するフォルダを作成します。 __さらにその中に__ ``<繰り返し回数>_`` という名前でディレクトリを作成します。 + +たとえば「frog」というプロンプトで、データを繰り返さない（1回だけ）場合、以下のようになります。 + +![image](https://user-images.githubusercontent.com/52813779/210770897-329758e5-3675-49f1-b345-c135f1725832.png) + + +### step 3. 学習の実行 + +各学習スクリプトを実行します。 `--train_data_dir` オプションで前述の学習用データのフォルダを（__画像を含むフォルダではなく、その親フォルダ__）、`--reg_data_dir` オプションで正則化画像のフォルダ（__画像を含むフォルダではなく、その親フォルダ__）を指定してください。 + +## DreamBooth、キャプション方式 + +学習用画像、正則化画像のフォルダに、画像と同じファイル名で、拡張子.caption（オプションで変えられます）のファイルを置くと、そのファイルからキャプションを読み込みプロンプトとして学習します。 + +※それらの画像の学習に、フォルダ名（identifier class）は使用されなくなります。 + +キャプションファイルの拡張子はデフォルトで.captionです。学習スクリプトの `--caption_extension` オプションで変更できます。`--shuffle_caption` オプションで学習時のキャプションについて、カンマ区切りの各部分をシャッフルしながら学習します。 + +## fine tuning 方式 + +メタデータを作るところまでは設定ファイルを使う場合と同様です。`in_json` オプションでメタデータファイルを指定します。 + +# 学習途中でのサンプル出力 + +学習中のモデルで試しに画像生成することで学習の進み方を確認できます。学習スクリプトに以下のオプションを指定します。 + +- `--sample_every_n_steps` / `--sample_every_n_epochs` + + サンプル出力するステップ数またはエポック数を指定します。この数ごとにサンプル出力します。両方指定するとエポック数が優先されます。 + +- `--sample_at_first` + + 学習開始前にサンプル出力します。学習前との比較ができます。 + +- `--sample_prompts` + + サンプル出力用プロンプトのファイルを指定します。 + +- `--sample_sampler` + + サンプル出力に使うサンプラーを指定します。 + `'ddim', 'pndm', 'heun', 'dpmsolver', 'dpmsolver++', 'dpmsingle', 'k_lms', 'k_euler', 'k_euler_a', 'k_dpm_2', 'k_dpm_2_a'`が選べます。 + +サンプル出力を行うにはあらかじめプロンプトを記述したテキストファイルを用意しておく必要があります。1行につき1プロンプトで記述します。 + +たとえば以下のようになります。 + +```txt +# prompt 1 +masterpiece, best quality, 1girl, in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28 + +# prompt 2 +masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40 +``` + +先頭が `#` の行はコメントになります。`--n` のように 「`--` + 英小文字」で生成画像へのオプションを指定できます。以下が使えます。 + +- `--n` 次のオプションまでをネガティブプロンプトとします。 +- `--w` 生成画像の横幅を指定します。 +- `--h` 生成画像の高さを指定します。 +- `--d` 生成画像のseedを指定します。 +- `--l` 生成画像のCFG scaleを指定します。 +- `--s` 生成時のステップ数を指定します。 + + +# 各スクリプトで共通の、よく使われるオプション + +スクリプトの更新後、ドキュメントの更新が追い付いていない場合があります。その場合は `--help` オプションで使用できるオプションを確認してください。 + +## 学習に使うモデル指定 + +- `--v2` / `--v_parameterization` + + 学習対象モデルとしてHugging Faceのstable-diffusion-2-base、またはそこからのfine tuningモデルを使う場合（推論時に `v2-inference.yaml` を使うように指示されているモデルの場合）は `--v2` オプションを、stable-diffusion-2や768-v-ema.ckpt、およびそれらのfine tuningモデルを使う場合（推論時に `v2-inference-v.yaml` を使うモデルの場合）は `--v2` と `--v_parameterization` の両方のオプションを指定してください。 + + Stable Diffusion 2.0では大きく以下の点が変わっています。 + + 1. 使用するTokenizer + 2. 使用するText Encoderおよび使用する出力層（2.0は最後から二番目の層を使う） + 3. Text Encoderの出力次元数（768->1024） + 4. U-Netの構造（CrossAttentionのhead数など） + 5. v-parameterization（サンプリング方法が変更されているらしい） + + このうちbaseでは1～4が、baseのつかない方（768-v）では1～5が採用されています。1～4を有効にするのがv2オプション、5を有効にするのがv_parameterizationオプションです。 + +- `--pretrained_model_name_or_path` + + 追加学習を行う元となるモデルを指定します。Stable Diffusionのcheckpointファイル（.ckptまたは.safetensors）、Diffusersのローカルディスクにあるモデルディレクトリ、DiffusersのモデルID（"stabilityai/stable-diffusion-2"など）が指定できます。 + +## 学習に関する設定 + +- `--output_dir` + + 学習後のモデルを保存するフォルダを指定します。 + +- `--output_name` + + モデルのファイル名を拡張子を除いて指定します。 + +- `--dataset_config` + + データセットの設定を記述した `.toml` ファイルを指定します。 + +- `--max_train_steps` / `--max_train_epochs` + + 学習するステップ数やエポック数を指定します。両方指定するとエポック数のほうが優先されます。 + +- `--mixed_precision` + + 省メモリ化のため mixed precision （混合精度）で学習します。`--mixed_precision="fp16"` のように指定します。mixed precision なし（デフォルト）と比べて精度が低くなる可能性がありますが、学習に必要なGPUメモリ量が大きく減ります。 + + （RTX30 シリーズ以降では `bf16` も指定できます。環境整備時にaccelerateに行った設定と合わせてください）。 + +- `--gradient_checkpointing` + + 学習時の重みの計算をまとめて行うのではなく少しずつ行うことで、学習に必要なGPUメモリ量を減らします。オンオフは精度には影響しませんが、オンにするとバッチサイズを大きくできるため、そちらでの影響はあります。 + + また一般的にはオンにすると速度は低下しますが、バッチサイズを大きくできるので、トータルでの学習時間はむしろ速くなるかもしれません。 + +- `--xformers` / `--mem_eff_attn` + + xformersオプションを指定するとxformersのCrossAttentionを用います。xformersをインストールしていない場合やエラーとなる場合（環境にもよりますが `mixed_precision="no"` の場合など）、代わりに `mem_eff_attn` オプションを指定すると省メモリ版CrossAttentionを使用します（xformersよりも速度は遅くなります）。 + +- `--clip_skip` + + `2` を指定すると、Text Encoder (CLIP) の後ろから二番目の層の出力を用います。1またはオプション省略時は最後の層を用います。 + + ※SD2.0はデフォルトで後ろから二番目の層を使うため、SD2.0の学習では指定しないでください。 + + 学習対象のモデルがもともと二番目の層を使うように学習されている場合は、2を指定するとよいでしょう。 + + そうではなく最後の層を使用していた場合はモデル全体がそれを前提に学習されています。そのため改めて二番目の層を使用して学習すると、望ましい学習結果を得るにはある程度の枚数の教師データ、長めの学習が必要になるかもしれません。 + +- `--max_token_length` + + デフォルトは75です。`150` または `225` を指定することでトークン長を拡張して学習できます。長いキャプションで学習する場合に指定してください。 + + ただし学習時のトークン拡張の仕様は Automatic1111 氏のWeb UIとは微妙に異なるため（分割の仕様など）、必要なければ75で学習することをお勧めします。 + + clip_skipと同様に、モデルの学習状態と異なる長さで学習するには、ある程度の教師データ枚数、長めの学習時間が必要になると思われます。 + +- `--weighted_captions` + + 指定するとAutomatic1111氏のWeb UIと同様の重み付きキャプションが有効になります。「Textual Inversion と XTI」以外の学習に使用できます。キャプションだけでなく DreamBooth 手法の token string でも有効です。 + + 重みづけキャプションの記法はWeb UIとほぼ同じで、(abc)や[abc]、(abc:1.23)などが使用できます。入れ子も可能です。括弧内にカンマを含めるとプロンプトのshuffle/dropoutで括弧の対応付けがおかしくなるため、括弧内にはカンマを含めないでください。 + +- `--persistent_data_loader_workers` + + Windows環境で指定するとエポック間の待ち時間が大幅に短縮されます。 + +- `--max_data_loader_n_workers` + + データ読み込みのプロセス数を指定します。プロセス数が多いとデータ読み込みが速くなりGPUを効率的に利用できますが、メインメモリを消費します。デフォルトは「`8` または `CPU同時実行スレッド数-1` の小さいほう」なので、メインメモリに余裕がない場合や、GPU使用率が90%程度以上なら、それらの数値を見ながら `2` または `1` 程度まで下げてください。 + +- `--logging_dir` / `--log_prefix` + + 学習ログの保存に関するオプションです。logging_dirオプションにログ保存先フォルダを指定してください。TensorBoard形式のログが保存されます。 + + たとえば--logging_dir=logsと指定すると、作業フォルダにlogsフォルダが作成され、その中の日時フォルダにログが保存されます。 + また--log_prefixオプションを指定すると、日時の前に指定した文字列が追加されます。「--logging_dir=logs --log_prefix=db_style1_」などとして識別用にお使いください。 + + TensorBoardでログを確認するには、別のコマンドプロンプトを開き、作業フォルダで以下のように入力します。 + + ``` + tensorboard --logdir=logs + ``` + + （tensorboardは環境整備時にあわせてインストールされると思いますが、もし入っていないなら `pip install tensorboard` で入れてください。） + + その後ブラウザを開き、http://localhost:6006/ へアクセスすると表示されます。 + +- `--log_with` / `--log_tracker_name` + + 学習ログの保存に関するオプションです。`tensorboard` だけでなく `wandb`への保存が可能です。詳細は [PR#428](https://github.com/kohya-ss/sd-scripts/pull/428)をご覧ください。 + +- `--noise_offset` + + こちらの記事の実装になります: https://www.crosslabs.org//blog/diffusion-with-offset-noise + + 全体的に暗い、明るい画像の生成結果が良くなる可能性があるようです。LoRA学習でも有効なようです。`0.1` 程度の値を指定するとよいようです。 + +- `--adaptive_noise_scale` （実験的オプション） + + Noise offsetの値を、latentsの各チャネルの平均値の絶対値に応じて自動調整するオプションです。`--noise_offset` と同時に指定することで有効になります。Noise offsetの値は `noise_offset + abs(mean(latents, dim=(2,3))) * adaptive_noise_scale` で計算されます。latentは正規分布に近いためnoise_offsetの1/10～同程度の値を指定するとよいかもしれません。 + + 負の値も指定でき、その場合はnoise offsetは0以上にclipされます。 + +- `--multires_noise_iterations` / `--multires_noise_discount` + + Multi resolution noise (pyramid noise)の設定です。詳細は [PR#471](https://github.com/kohya-ss/sd-scripts/pull/471) およびこちらのページ [Multi-Resolution Noise for Diffusion Model Training](https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2) を参照してください。 + + `--multires_noise_iterations` に数値を指定すると有効になります。6~10程度の値が良いようです。`--multires_noise_discount` に0.1~0.3 程度の値（LoRA学習等比較的データセットが小さい場合のPR作者の推奨）、ないしは0.8程度の値（元記事の推奨）を指定してください（デフォルトは 0.3）。 + +- `--debug_dataset` + + このオプションを付けることで学習を行う前に事前にどのような画像データ、キャプションで学習されるかを確認できます。Escキーを押すと終了してコマンドラインに戻ります。`S`キーで次のステップ（バッチ）、`E`キーで次のエポックに進みます。 + + ※Linux環境（Colabを含む）では画像は表示されません。 + +- `--vae` + + vaeオプションにStable Diffusionのcheckpoint、VAEのcheckpointファイル、DiffusesのモデルまたはVAE（ともにローカルまたはHugging FaceのモデルIDが指定できます）のいずれかを指定すると、そのVAEを使って学習します（latentsのキャッシュ時または学習中のlatents取得時）。 + + DreamBoothおよびfine tuningでは、保存されるモデルはこのVAEを組み込んだものになります。 + +- `--cache_latents` / `--cache_latents_to_disk` + + 使用VRAMを減らすためVAEの出力をメインメモリにキャッシュします。`flip_aug` 以外のaugmentationは使えなくなります。また全体の学習速度が若干速くなります。 + + cache_latents_to_diskを指定するとキャッシュをディスクに保存します。スクリプトを終了し、再度起動した場合もキャッシュが有効になります。 + +- `--min_snr_gamma` + + Min-SNR Weighting strategyを指定します。詳細は[こちら](https://github.com/kohya-ss/sd-scripts/pull/308)を参照してください。論文では`5`が推奨されています。 + +## モデルの保存に関する設定 + +- `--save_precision` + + 保存時のデータ精度を指定します。save_precisionオプションにfloat、fp16、bf16のいずれかを指定すると、その形式でモデルを保存します（DreamBooth、fine tuningでDiffusers形式でモデルを保存する場合は無効です）。モデルのサイズを削減したい場合などにお使いください。 + +- `--save_every_n_epochs` / `--save_state` / `--resume` + + save_every_n_epochsオプションに数値を指定すると、そのエポックごとに学習途中のモデルを保存します。 + + save_stateオプションを同時に指定すると、optimizer等の状態も含めた学習状態を合わせて保存します（保存したモデルからも学習再開できますが、それに比べると精度の向上、学習時間の短縮が期待できます）。保存先はフォルダになります。 + + 学習状態は保存先フォルダに `-??????-state`（??????はエポック数）という名前のフォルダで出力されます。長時間にわたる学習時にご利用ください。 + + 保存された学習状態から学習を再開するにはresumeオプションを使います。学習状態のフォルダ（`output_dir` ではなくその中のstateのフォルダ）を指定してください。 + + なおAcceleratorの仕様により、エポック数、global stepは保存されておらず、resumeしたときにも1からになりますがご容赦ください。 + +- `--save_every_n_steps` + + save_every_n_stepsオプションに数値を指定すると、そのステップごとに学習途中のモデルを保存します。save_every_n_epochsと同時に指定できます。 + +- `--save_model_as` （DreamBooth, fine tuning のみ） + + モデルの保存形式を`ckpt, safetensors, diffusers, diffusers_safetensors` から選べます。 + + `--save_model_as=safetensors` のように指定します。Stable Diffusion形式（ckptまたはsafetensors）を読み込み、Diffusers形式で保存する場合、不足する情報はHugging Faceからv1.5またはv2.1の情報を落としてきて補完します。 + +- `--huggingface_repo_id` 等 + + huggingface_repo_idが指定されているとモデル保存時に同時にHuggingFaceにアップロードします。アクセストークンの取り扱いに注意してください（HuggingFaceのドキュメントを参照してください）。 + + 他の引数をたとえば以下のように指定してください。 + + - `--huggingface_repo_id "your-hf-name/your-model" --huggingface_path_in_repo "path" --huggingface_repo_type model --huggingface_repo_visibility private --huggingface_token hf_YourAccessTokenHere` + + huggingface_repo_visibilityに`public`を指定するとリポジトリが公開されます。省略時または`private`（などpublic以外）を指定すると非公開になります。 + + `--save_state`オプション指定時に`--save_state_to_huggingface`を指定するとstateもアップロードします。 + + `--resume`オプション指定時に`--resume_from_huggingface`を指定するとHuggingFaceからstateをダウンロードして再開します。その時の --resumeオプションは `--resume {repo_id}/{path_in_repo}:{revision}:{repo_type}`になります。 + + 例: `--resume_from_huggingface --resume your-hf-name/your-model/path/test-000002-state:main:model` + + `--async_upload`オプションを指定するとアップロードを非同期で行います。 + +## オプティマイザ関係 + +- `--optimizer_type` + --オプティマイザの種類を指定します。以下が指定できます。 + - AdamW : [torch.optim.AdamW](https://pytorch.org/docs/stable/generated/torch.optim.AdamW.html) + - 過去のバージョンのオプション未指定時と同じ + - AdamW8bit : 引数は同上 + - PagedAdamW8bit : 引数は同上 + - 過去のバージョンの--use_8bit_adam指定時と同じ + - Lion : https://github.com/lucidrains/lion-pytorch + - 過去のバージョンの--use_lion_optimizer指定時と同じ + - Lion8bit : 引数は同上 + - PagedLion8bit : 引数は同上 + - SGDNesterov : [torch.optim.SGD](https://pytorch.org/docs/stable/generated/torch.optim.SGD.html), nesterov=True + - SGDNesterov8bit : 引数は同上 + - DAdaptation(DAdaptAdamPreprint) : https://github.com/facebookresearch/dadaptation + - DAdaptAdam : 引数は同上 + - DAdaptAdaGrad : 引数は同上 + - DAdaptAdan : 引数は同上 + - DAdaptAdanIP : 引数は同上 + - DAdaptLion : 引数は同上 + - DAdaptSGD : 引数は同上 + - Prodigy : https://github.com/konstmish/prodigy + - AdaFactor : [Transformers AdaFactor](https://huggingface.co/docs/transformers/main_classes/optimizer_schedules) + - 任意のオプティマイザ + +- `--learning_rate` + + 学習率を指定します。適切な学習率は学習スクリプトにより異なりますので、それぞれの説明を参照してください。 + +- `--lr_scheduler` / `--lr_warmup_steps` / `--lr_scheduler_num_cycles` / `--lr_scheduler_power` + + 学習率のスケジューラ関連の指定です。 + + lr_schedulerオプションで学習率のスケジューラをlinear, cosine, cosine_with_restarts, polynomial, constant, constant_with_warmup, 任意のスケジューラから選べます。デフォルトはconstantです。 + + lr_warmup_stepsでスケジューラのウォームアップ（だんだん学習率を変えていく）ステップ数を指定できます。 + + lr_scheduler_num_cycles は cosine with restartsスケジューラでのリスタート回数、lr_scheduler_power は polynomialスケジューラでのpolynomial power です。 + + 詳細については各自お調べください。 + + 任意のスケジューラを使う場合、任意のオプティマイザと同様に、`--lr_scheduler_args`でオプション引数を指定してください。 + +### オプティマイザの指定について + +オプティマイザのオプション引数は--optimizer_argsオプションで指定してください。key=valueの形式で、複数の値が指定できます。また、valueはカンマ区切りで複数の値が指定できます。たとえばAdamWオプティマイザに引数を指定する場合は、``--optimizer_args weight_decay=0.01 betas=.9,.999``のようになります。 + +オプション引数を指定する場合は、それぞれのオプティマイザの仕様をご確認ください。 + +一部のオプティマイザでは必須の引数があり、省略すると自動的に追加されます（SGDNesterovのmomentumなど）。コンソールの出力を確認してください。 + +D-Adaptationオプティマイザは学習率を自動調整します。学習率のオプションに指定した値は学習率そのものではなくD-Adaptationが決定した学習率の適用率になりますので、通常は1.0を指定してください。Text EncoderにU-Netの半分の学習率を指定したい場合は、``--text_encoder_lr=0.5 --unet_lr=1.0``と指定します。 + +AdaFactorオプティマイザはrelative_step=Trueを指定すると学習率を自動調整できます（省略時はデフォルトで追加されます）。自動調整する場合は学習率のスケジューラにはadafactor_schedulerが強制的に使用されます。またscale_parameterとwarmup_initを指定するとよいようです。 + +自動調整する場合のオプション指定はたとえば ``--optimizer_args "relative_step=True" "scale_parameter=True" "warmup_init=True"`` のようになります。 + +学習率を自動調整しない場合はオプション引数 ``relative_step=False`` を追加してください。その場合、学習率のスケジューラにはconstant_with_warmupが、また勾配のclip normをしないことが推奨されているようです。そのため引数は ``--optimizer_type=adafactor --optimizer_args "relative_step=False" --lr_scheduler="constant_with_warmup" --max_grad_norm=0.0`` のようになります。 + +### 任意のオプティマイザを使う + +``torch.optim`` のオプティマイザを使う場合にはクラス名のみを（``--optimizer_type=RMSprop``など）、他のモジュールのオプティマイザを使う時は「モジュール名.クラス名」を指定してください（``--optimizer_type=bitsandbytes.optim.lamb.LAMB``など）。 + +（内部でimportlibしているだけで動作は未確認です。必要ならパッケージをインストールしてください。） + + + + +# メタデータファイルの作成 + +## 教師データの用意 + +前述のように学習させたい画像データを用意し、任意のフォルダに入れてください。 + +たとえば以下のように画像を格納します。 + +![教師データフォルダのスクショ](https://user-images.githubusercontent.com/52813779/208907739-8e89d5fa-6ca8-4b60-8927-f484d2a9ae04.png) + +## 自動キャプショニング + +キャプションを使わずタグだけで学習する場合はスキップしてください。 + +また手動でキャプションを用意する場合、キャプションは教師データ画像と同じディレクトリに、同じファイル名、拡張子.caption等で用意してください。各ファイルは1行のみのテキストファイルとします。 + +### BLIPによるキャプショニング + +最新版ではBLIPのダウンロード、重みのダウンロード、仮想環境の追加は不要になりました。そのままで動作します。 + +finetuneフォルダ内のmake_captions.pyを実行します。 + +``` +python finetune\make_captions.py --batch_size <バッチサイズ> <教師データフォルダ> +``` + +バッチサイズ8、教師データを親フォルダのtrain_dataに置いた場合、以下のようになります。 + +``` +python finetune\make_captions.py --batch_size 8 ..\train_data +``` + +キャプションファイルが教師データ画像と同じディレクトリに、同じファイル名、拡張子.captionで作成されます。 + +batch_sizeはGPUのVRAM容量に応じて増減してください。大きいほうが速くなります（VRAM 12GBでももう少し増やせると思います）。 +max_lengthオプションでキャプションの最大長を指定できます。デフォルトは75です。モデルをトークン長225で学習する場合には長くしても良いかもしれません。 +caption_extensionオプションでキャプションの拡張子を変更できます。デフォルトは.captionです（.txtにすると後述のDeepDanbooruと競合します）。 + +複数の教師データフォルダがある場合には、それぞれのフォルダに対して実行してください。 + +なお、推論にランダム性があるため、実行するたびに結果が変わります。固定する場合には--seedオプションで `--seed 42` のように乱数seedを指定してください。 + +その他のオプションは `--help` でヘルプをご参照ください（パラメータの意味についてはドキュメントがまとまっていないようで、ソースを見るしかないようです）。 + +デフォルトでは拡張子.captionでキャプションファイルが生成されます。 + +![captionが生成されたフォルダ](https://user-images.githubusercontent.com/52813779/208908845-48a9d36c-f6ee-4dae-af71-9ab462d1459e.png) + +たとえば以下のようなキャプションが付きます。 + +![キャプションと画像](https://user-images.githubusercontent.com/52813779/208908947-af936957-5d73-4339-b6c8-945a52857373.png) + +## DeepDanbooruによるタグ付け + +danbooruタグのタグ付け自体を行わない場合は「キャプションとタグ情報の前処理」に進んでください。 + +タグ付けはDeepDanbooruまたはWD14Taggerで行います。WD14Taggerのほうが精度が良いようです。WD14Taggerでタグ付けする場合は、次の章へ進んでください。 + +### 環境整備 + +DeepDanbooru https://github.com/KichangKim/DeepDanbooru を作業フォルダにcloneしてくるか、zipをダウンロードして展開します。私はzipで展開しました。 +またDeepDanbooruのReleasesのページ https://github.com/KichangKim/DeepDanbooru/releases の「DeepDanbooru Pretrained Model v3-20211112-sgd-e28」のAssetsから、deepdanbooru-v3-20211112-sgd-e28.zipをダウンロードしてきてDeepDanbooruのフォルダに展開します。 + +以下からダウンロードします。Assetsをクリックして開き、そこからダウンロードします。 + +![DeepDanbooruダウンロードページ](https://user-images.githubusercontent.com/52813779/208909417-10e597df-7085-41ee-bd06-3e856a1339df.png) + +以下のようなこういうディレクトリ構造にしてください + +![DeepDanbooruのディレクトリ構造](https://user-images.githubusercontent.com/52813779/208909486-38935d8b-8dc6-43f1-84d3-fef99bc471aa.png) + +Diffusersの環境に必要なライブラリをインストールします。DeepDanbooruのフォルダに移動してインストールします（実質的にはtensorflow-ioが追加されるだけだと思います）。 + +``` +pip install -r requirements.txt +``` + +続いてDeepDanbooru自体をインストールします。 + +``` +pip install . +``` + +以上でタグ付けの環境整備は完了です。 + +### タグ付けの実施 +DeepDanbooruのフォルダに移動し、deepdanbooruを実行してタグ付けを行います。 + +``` +deepdanbooru evaluate <教師データフォルダ> --project-path deepdanbooru-v3-20211112-sgd-e28 --allow-folder --save-txt +``` + +教師データを親フォルダのtrain_dataに置いた場合、以下のようになります。 + +``` +deepdanbooru evaluate ../train_data --project-path deepdanbooru-v3-20211112-sgd-e28 --allow-folder --save-txt +``` + +タグファイルが教師データ画像と同じディレクトリに、同じファイル名、拡張子.txtで作成されます。1件ずつ処理されるためわりと遅いです。 + +複数の教師データフォルダがある場合には、それぞれのフォルダに対して実行してください。 + +以下のように生成されます。 + +![DeepDanbooruの生成ファイル](https://user-images.githubusercontent.com/52813779/208909855-d21b9c98-f2d3-4283-8238-5b0e5aad6691.png) + +こんな感じにタグが付きます（すごい情報量……）。 + +![DeepDanbooruタグと画像](https://user-images.githubusercontent.com/52813779/208909908-a7920174-266e-48d5-aaef-940aba709519.png) + +## WD14Taggerによるタグ付け + +DeepDanbooruの代わりにWD14Taggerを用いる手順です。 + +Automatic1111氏のWebUIで使用しているtaggerを利用します。こちらのgithubページ（https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger ）の情報を参考にさせていただきました。 + +最初の環境整備で必要なモジュールはインストール済みです。また重みはHugging Faceから自動的にダウンロードしてきます。 + +### タグ付けの実施 + +スクリプトを実行してタグ付けを行います。 +``` +python tag_images_by_wd14_tagger.py --batch_size <バッチサイズ> <教師データフォルダ> +``` + +教師データを親フォルダのtrain_dataに置いた場合、以下のようになります。 +``` +python tag_images_by_wd14_tagger.py --batch_size 4 ..\train_data +``` + +初回起動時にはモデルファイルがwd14_tagger_modelフォルダに自動的にダウンロードされます（フォルダはオプションで変えられます）。以下のようになります。 + +![ダウンロードされたファイル](https://user-images.githubusercontent.com/52813779/208910447-f7eb0582-90d6-49d3-a666-2b508c7d1842.png) + +タグファイルが教師データ画像と同じディレクトリに、同じファイル名、拡張子.txtで作成されます。 + +![生成されたタグファイル](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png) + +![タグと画像](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png) + +threshオプションで、判定されたタグのconfidence（確信度）がいくつ以上でタグをつけるかが指定できます。デフォルトはWD14Taggerのサンプルと同じ0.35です。値を下げるとより多くのタグが付与されますが、精度は下がります。 + +batch_sizeはGPUのVRAM容量に応じて増減してください。大きいほうが速くなります（VRAM 12GBでももう少し増やせると思います）。caption_extensionオプションでタグファイルの拡張子を変更できます。デフォルトは.txtです。 + +model_dirオプションでモデルの保存先フォルダを指定できます。 + +またforce_downloadオプションを指定すると保存先フォルダがあってもモデルを再ダウンロードします。 + +複数の教師データフォルダがある場合には、それぞれのフォルダに対して実行してください。 + +## キャプションとタグ情報の前処理 + +スクリプトから処理しやすいようにキャプションとタグをメタデータとしてひとつのファイルにまとめます。 + +### キャプションの前処理 + +キャプションをメタデータに入れるには、作業フォルダ内で以下を実行してください（キャプションを学習に使わない場合は実行不要です）（実際は1行で記述します、以下同様）。`--full_path` オプションを指定してメタデータに画像ファイルの場所をフルパスで格納します。このオプションを省略すると相対パスで記録されますが、フォルダ指定が `.toml` ファイル内で別途必要になります。 + +``` +python merge_captions_to_metadata.py --full_path <教師データフォルダ> +　 --in_json <読み込むメタデータファイル名> <メタデータファイル名> +``` + +メタデータファイル名は任意の名前です。 +教師データがtrain_data、読み込むメタデータファイルなし、メタデータファイルがmeta_cap.jsonの場合、以下のようになります。 + +``` +python merge_captions_to_metadata.py --full_path train_data meta_cap.json +``` + +caption_extensionオプションでキャプションの拡張子を指定できます。 + +複数の教師データフォルダがある場合には、full_path引数を指定しつつ、それぞれのフォルダに対して実行してください。 + +``` +python merge_captions_to_metadata.py --full_path + train_data1 meta_cap1.json +python merge_captions_to_metadata.py --full_path --in_json meta_cap1.json + train_data2 meta_cap2.json +``` + +in_jsonを省略すると書き込み先メタデータファイルがあるとそこから読み込み、そこに上書きします。 + +__※in_jsonオプションと書き込み先を都度書き換えて、別のメタデータファイルへ書き出すようにすると安全です。__ + +### タグの前処理 + +同様にタグもメタデータにまとめます（タグを学習に使わない場合は実行不要です）。 +``` +python merge_dd_tags_to_metadata.py --full_path <教師データフォルダ> + --in_json <読み込むメタデータファイル名> <書き込むメタデータファイル名> +``` + +先と同じディレクトリ構成で、meta_cap.jsonを読み、meta_cap_dd.jsonに書きだす場合、以下となります。 +``` +python merge_dd_tags_to_metadata.py --full_path train_data --in_json meta_cap.json meta_cap_dd.json +``` + +複数の教師データフォルダがある場合には、full_path引数を指定しつつ、それぞれのフォルダに対して実行してください。 + +``` +python merge_dd_tags_to_metadata.py --full_path --in_json meta_cap2.json + train_data1 meta_cap_dd1.json +python merge_dd_tags_to_metadata.py --full_path --in_json meta_cap_dd1.json + train_data2 meta_cap_dd2.json +``` + +in_jsonを省略すると書き込み先メタデータファイルがあるとそこから読み込み、そこに上書きします。 + +__※in_jsonオプションと書き込み先を都度書き換えて、別のメタデータファイルへ書き出すようにすると安全です。__ + +### キャプションとタグのクリーニング + +ここまででメタデータファイルにキャプションとDeepDanbooruのタグがまとめられています。ただ自動キャプショニングにしたキャプションは表記ゆれなどがあり微妙（※）ですし、タグにはアンダースコアが含まれていたりratingが付いていたりしますので（DeepDanbooruの場合）、エディタの置換機能などを用いてキャプションとタグのクリーニングをしたほうがいいでしょう。 + +※たとえばアニメ絵の少女を学習する場合、キャプションにはgirl/girls/woman/womenなどのばらつきがあります。また「anime girl」なども単に「girl」としたほうが適切かもしれません。 + +クリーニング用のスクリプトが用意してありますので、スクリプトの内容を状況に応じて編集してお使いください。 + +（教師データフォルダの指定は不要になりました。メタデータ内の全データをクリーニングします。） + +``` +python clean_captions_and_tags.py <読み込むメタデータファイル名> <書き込むメタデータファイル名> +``` + +--in_jsonは付きませんのでご注意ください。たとえば次のようになります。 + +``` +python clean_captions_and_tags.py meta_cap_dd.json meta_clean.json +``` + +以上でキャプションとタグの前処理は完了です。 + +## latentsの事前取得 + +※ このステップは必須ではありません。省略しても学習時にlatentsを取得しながら学習できます。 +また学習時に `random_crop` や `color_aug` などを行う場合にはlatentsの事前取得はできません（画像を毎回変えながら学習するため）。事前取得をしない場合、ここまでのメタデータで学習できます。 + +あらかじめ画像の潜在表現を取得しディスクに保存しておきます。それにより、学習を高速に進めることができます。あわせてbucketing（教師データをアスペクト比に応じて分類する）を行います。 + +作業フォルダで以下のように入力してください。 +``` +python prepare_buckets_latents.py --full_path <教師データフォルダ> + <読み込むメタデータファイル名> <書き込むメタデータファイル名> + + --batch_size <バッチサイズ> + --max_resolution <解像度 幅,高さ> + --mixed_precision <精度> +``` + +モデルがmodel.ckpt、バッチサイズ4、学習解像度は512\*512、精度no（float32）で、meta_clean.jsonからメタデータを読み込み、meta_lat.jsonに書き込む場合、以下のようになります。 + +``` +python prepare_buckets_latents.py --full_path + train_data meta_clean.json meta_lat.json model.ckpt + --batch_size 4 --max_resolution 512,512 --mixed_precision no +``` + +教師データフォルダにnumpyのnpz形式でlatentsが保存されます。 + +解像度の最小サイズを--min_bucket_resoオプションで、最大サイズを--max_bucket_resoで指定できます。デフォルトはそれぞれ256、1024です。たとえば最小サイズに384を指定すると、256\*1024や320\*768などの解像度は使わなくなります。 +解像度を768\*768のように大きくした場合、最大サイズに1280などを指定すると良いでしょう。 + +--flip_augオプションを指定すると左右反転のaugmentation（データ拡張）を行います。疑似的にデータ量を二倍に増やすことができますが、データが左右対称でない場合に指定すると（例えばキャラクタの外見、髪型など）学習がうまく行かなくなります。 + + +（反転した画像についてもlatentsを取得し、\*\_flip.npzファイルを保存する単純な実装です。fline_tune.pyには特にオプション指定は必要ありません。\_flip付きのファイルがある場合、flip付き・なしのファイルを、ランダムに読み込みます。） + +バッチサイズはVRAM 12GBでももう少し増やせるかもしれません。 +解像度は64で割り切れる数字で、"幅,高さ"で指定します。解像度はfine tuning時のメモリサイズに直結します。VRAM 12GBでは512,512が限界と思われます（※）。16GBなら512,704や512,768まで上げられるかもしれません。なお256,256等にしてもVRAM 8GBでは厳しいようです（パラメータやoptimizerなどは解像度に関係せず一定のメモリが必要なため）。 + +※batch size 1の学習で12GB VRAM、640,640で動いたとの報告もありました。 + +以下のようにbucketingの結果が表示されます。 + +![bucketingの結果](https://user-images.githubusercontent.com/52813779/208911419-71c00fbb-2ce6-49d5-89b5-b78d7715e441.png) + +複数の教師データフォルダがある場合には、full_path引数を指定しつつ、それぞれのフォルダに対して実行してください。 +``` +python prepare_buckets_latents.py --full_path + train_data1 meta_clean.json meta_lat1.json model.ckpt + --batch_size 4 --max_resolution 512,512 --mixed_precision no + +python prepare_buckets_latents.py --full_path + train_data2 meta_lat1.json meta_lat2.json model.ckpt + --batch_size 4 --max_resolution 512,512 --mixed_precision no + +``` +読み込み元と書き込み先を同じにすることも可能ですが別々の方が安全です。 + +__※引数を都度書き換えて、別のメタデータファイルに書き込むと安全です。__ + diff --git a/train_README-zh.md b/train_README-zh.md new file mode 100644 index 0000000000000000000000000000000000000000..1bc47e0f5c7027a36cc5ea27d71389b04e2674bb --- /dev/null +++ b/train_README-zh.md @@ -0,0 +1,912 @@ +__由于文档正在更新中，描述可能有错误。__ + +# 关于训练，通用描述 +本库支持模型微调(fine tuning)、DreamBooth、训练LoRA和文本反转(Textual Inversion)（包括[XTI:P+](https://github.com/kohya-ss/sd-scripts/pull/327) +） +本文档将说明它们通用的训练数据准备方法和选项等。 + +# 概要 + +请提前参考本仓库的README，准备好环境。 + + +以下本节说明。 + +1. 准备训练数据（使用设置文件的新格式） +1. 训练中使用的术语的简要解释 +1. 先前的指定格式（不使用设置文件，而是从命令行指定） +1. 生成训练过程中的示例图像 +1. 各脚本中常用的共同选项 +1. 准备 fine tuning 方法的元数据：如说明文字(打标签)等 + + +1. 如果只执行一次，训练就可以进行（相关内容，请参阅各个脚本的文档）。如果需要，以后可以随时参考。 + + + +# 关于准备训练数据 + +在任意文件夹（也可以是多个文件夹）中准备好训练数据的图像文件。支持 `.png`, `.jpg`, `.jpeg`, `.webp`, `.bmp` 格式的文件。通常不需要进行任何预处理，如调整大小等。 + +但是请勿使用极小的图像，若其尺寸比训练分辨率（稍后将提到）还小，建议事先使用超分辨率AI等进行放大。另外，请注意不要使用过大的图像（约为3000 x 3000像素以上），因为这可能会导致错误，建议事先缩小。 + +在训练时，需要整理要用于训练模型的图像数据，并将其指定给脚本。根据训练数据的数量、训练目标和说明（图像描述）是否可用等因素，可以使用几种方法指定训练数据。以下是其中的一些方法（每个名称都不是通用的，而是该存储库自定义的定义）。有关正则化图像的信息将在稍后提供。 + +1. DreamBooth、class + identifier方式（可使用正则化图像） + + 将训练目标与特定单词（identifier）相关联进行训练。无需准备说明。例如，当要学习特定角色时，由于无需准备说明，因此比较方便，但由于训练数据的所有元素都与identifier相关联，例如发型、服装、背景等，因此在生成时可能会出现无法更换服装的情况。 + +2. DreamBooth、说明方式（可使用正则化图像） + + 事先给每个图片写说明（caption），存放到文本文件中，然后进行训练。例如，通过将图像详细信息（如穿着白色衣服的角色A、穿着红色衣服的角色A等）记录在caption中，可以将角色和其他元素分离，并期望模型更准确地学习角色。 + +3. 微调方式（不可使用正则化图像） + + 先将说明收集到元数据文件中。支持分离标签和说明以及预先缓存latents等功能，以加速训练（这些将在另一篇文档中介绍）。（虽然名为fine tuning方式，但不仅限于fine tuning。） + +训练对象和你可以使用的规范方法的组合如下。 + +| 训练对象或方法 | 脚本 | DB/class+identifier | DB/caption | fine tuning | +|----------------| ----- | ----- | ----- | ----- | +| fine tuning微调模型 | `fine_tune.py`| x | x | o | +| DreamBooth训练模型 | `train_db.py`| o | o | x | +| LoRA | `train_network.py`| o | o | o | +| Textual Invesion | `train_textual_inversion.py`| o | o | o | + +## 选择哪一个 + +如果您想要训练LoRA、Textual Inversion而不需要准备说明（caption）文件，则建议使用DreamBooth class+identifier。如果您能够准备caption文件，则DreamBooth Captions方法更好。如果您有大量的训练数据并且不使用正则化图像，则请考虑使用fine-tuning方法。 + +对于DreamBooth也是一样的，但不能使用fine-tuning方法。若要进行微调，只能使用fine-tuning方式。 + +# 每种方法的指定方式 + +在这里，我们只介绍每种指定方法的典型模式。有关更详细的指定方法，请参见[数据集设置](./config_README-ja.md)。 + +# DreamBooth，class+identifier方法（可使用正则化图像） + +在该方法中，每个图像将被视为使用与 `class identifier` 相同的标题进行训练（例如 `shs dog`）。 + +这样一来，每张图片都相当于使用标题“分类标识”（例如“shs dog”）进行训练。 + +## step 1.确定identifier和class + +要将训练的目标与identifier和属于该目标的class相关联。 + +（虽然有很多称呼，但暂时按照原始论文的说法。） + +以下是简要说明（请查阅详细信息）。 + +class是训练目标的一般类别。例如，如果要学习特定品种的狗，则class将是“dog”。对于动漫角色，根据模型不同，可能是“boy”或“girl”，也可能是“1boy”或“1girl”。 + +identifier是用于识别训练目标并进行学习的单词。可以使用任何单词，但是根据原始论文，“Tokenizer生成的3个或更少字符的罕见单词”是最好的选择。 + +使用identifier和class，例如，“shs dog”可以将模型训练为从class中识别并学习所需的目标。 + +在图像生成时，使用“shs dog”将生成所学习狗种的图像。 + +（作为identifier，我最近使用的一些参考是“shs sts scs cpc coc cic msm usu ici lvl cic dii muk ori hru rik koo yos wny”等。最好是不包含在Danbooru标签中的单词。） + +## step 2. 决定是否使用正则化图像，并在使用时生成正则化图像 + +正则化图像是为防止前面提到的语言漂移，即整个类别被拉扯成为训练目标而生成的图像。如果不使用正则化图像，例如在 `shs 1girl` 中学习特定角色时，即使在简单的 `1girl` 提示下生成，也会越来越像该角色。这是因为 `1girl` 在训练时的标题中包含了该角色的信息。 + +通过同时学习目标图像和正则化图像，类别仍然保持不变，仅在将标识符附加到提示中时才生成目标图像。 + +如果您只想在LoRA或DreamBooth中使用特定的角色，则可以不使用正则化图像。 + +在Textual Inversion中也不需要使用（如果要学习的token string不包含在标题中，则不会学习任何内容）。 + +一般情况下，使用在训练目标模型时只使用类别名称生成的图像作为正则化图像是常见的做法（例如 `1girl`）。但是，如果生成的图像质量不佳，可以尝试修改提示或使用从网络上另外下载的图像。 + +（由于正则化图像也被训练，因此其质量会影响模型。） + +通常，准备数百张图像是理想的（图像数量太少会导致类别图像无法被归纳，特征也不会被学习）。 + +如果要使用生成的图像，生成图像的大小通常应与训练分辨率（更准确地说，是bucket的分辨率，见下文）相匹配。 + + + +## step 2. 设置文件的描述 + +创建一个文本文件，并将其扩展名更改为`.toml`。例如，您可以按以下方式进行描述： + +（以`＃`开头的部分是注释，因此您可以直接复制粘贴，或者将其删除。） + +```toml +[general] +enable_bucket = true # 是否使用Aspect Ratio Bucketing + +[[datasets]] +resolution = 512 # 训练分辨率 +batch_size = 4 # 批次大小 + + [[datasets.subsets]] + image_dir = 'C:\hoge' # 指定包含训练图像的文件夹 + class_tokens = 'hoge girl' # 指定标识符类 + num_repeats = 10 # 训练图像的重复次数 + + # 以下仅在使用正则化图像时进行描述。不使用则删除 + [[datasets.subsets]] + is_reg = true + image_dir = 'C:\reg' # 指定包含正则化图像的文件夹 + class_tokens = 'girl' # 指定class + num_repeats = 1 # 正则化图像的重复次数，基本上1就可以了 +``` + +基本上只需更改以下几个地方即可进行训练。 + +1. 训练分辨率 + + 指定一个数字表示正方形（如果是 `512`，则为 512x512），如果使用方括号和逗号分隔的两个数字，则表示横向×纵向（如果是`[512,768]`，则为 512x768）。在SD1.x系列中，原始训练分辨率为512。指定较大的分辨率，如 `[512,768]` 可能会减少纵向和横向图像生成时的错误。在SD2.x 768系列中，分辨率为 `768`。 + +1. 批次大小 + + 指定同时训练多少个数据。这取决于GPU的VRAM大小和训练分辨率。详细信息将在后面说明。此外，fine tuning/DreamBooth/LoRA等也会影响批次大小，请查看各个脚本的说明。 + +1. 文件夹指定 + + 指定用于学习的图像和正则化图像（仅在使用时）的文件夹。指定包含图像数据的文件夹。 + +1. identifier 和 class 的指定 + + 如前所述，与示例相同。 + +1. 重复次数 + + 将在后面说明。 + +### 关于重复次数 + +重复次数用于调整正则化图像和训练用图像的数量。由于正则化图像的数量多于训练用图像，因此需要重复使用训练用图像来达到一对一的比例，从而实现训练。 + +请将重复次数指定为“ __训练用图像的重复次数×训练用图像的数量≥正则化图像的重复次数×正则化图像的数量__ ”。 + +（1个epoch（指训练数据过完一遍）的数据量为“训练用图像的重复次数×训练用图像的数量”。如果正则化图像的数量多于这个值，则剩余的正则化图像将不会被使用。） + +## 步骤 3. 训练 + +详情请参考相关文档进行训练。 + +# DreamBooth，文本说明（caption）方式（可使用正则化图像） + +在此方式中，每个图像都将通过caption进行训练。 + +## 步骤 1. 准备文本说明文件 + +请将与图像具有相同文件名且扩展名为 `.caption`（可以在设置中更改）的文件放置在用于训练图像的文件夹中。每个文件应该只有一行。编码为 `UTF-8`。 + +## 步骤 2. 决定是否使用正则化图像，并在使用时生成正则化图像 + +与class+identifier格式相同。可以在规范化图像上附加caption，但通常不需要。 + +## 步骤 2. 编写设置文件 + +创建一个文本文件并将扩展名更改为 `.toml`。例如，您可以按以下方式进行描述： + +```toml +[general] +enable_bucket = true # 是否使用Aspect Ratio Bucketing + +[[datasets]] +resolution = 512 # 训练分辨率 +batch_size = 4 # 批次大小 + + [[datasets.subsets]] + image_dir = 'C:\hoge' # 指定包含训练图像的文件夹 + caption_extension = '.caption' # 若使用txt文件,更改此项 + num_repeats = 10 # 训练图像的重复次数 + + # 以下仅在使用正则化图像时进行描述。不使用则删除 + [[datasets.subsets]] + is_reg = true + image_dir = 'C:\reg' # 指定包含正则化图像的文件夹 + class_tokens = 'girl' # 指定class + num_repeats = 1 # 正则化图像的重复次数，基本上1就可以了 +``` + +基本上只需更改以下几个地方来训练。除非另有说明，否则与class+identifier方法相同。 + +1. 训练分辨率 +2. 批次大小 +3. 文件夹指定 +4. caption文件的扩展名 + + 可以指定任意的扩展名。 +5. 重复次数 + +## 步骤 3. 训练 + +详情请参考相关文档进行训练。 + +# 微调方法(fine tuning) + +## 步骤 1. 准备元数据 + +将caption和标签整合到管理文件中称为元数据。它的扩展名为 `.json`，格式为json。由于创建方法较长，因此在本文档的末尾进行描述。 + +## 步骤 2. 编写设置文件 + +创建一个文本文件，将扩展名设置为 `.toml`。例如，可以按以下方式编写： +```toml +[general] +shuffle_caption = true +keep_tokens = 1 + +[[datasets]] +resolution = 512 # 图像分辨率 +batch_size = 4 # 批次大小 + + [[datasets.subsets]] + image_dir = 'C:\piyo' # 指定包含训练图像的文件夹 + metadata_file = 'C:\piyo\piyo_md.json' # 元数据文件名 +``` + +基本上只需更改以下几个地方来训练。除非另有说明，否则与DreamBooth, class+identifier方法相同。 + +1. 训练分辨率 +2. 批次大小 +3. 指定文件夹 +4. 元数据文件名 + + 指定使用后面所述方法创建的元数据文件。 + + +## 第三步：训练 + +详情请参考相关文档进行训练。 + +# 训练中使用的术语简单解释 + +由于省略了细节并且我自己也没有完全理解，因此请自行查阅详细信息。 + +## 微调（fine tuning） + +指训练模型并微调其性能。具体含义因用法而异，但在 Stable Diffusion 中，狭义的微调是指使用图像和caption进行训练模型。DreamBooth 可视为狭义微调的一种特殊方法。广义的微调包括 LoRA、Textual Inversion、Hypernetworks 等，包括训练模型的所有内容。 + +## 步骤（step） + +粗略地说，每次在训练数据上进行一次计算即为一步。具体来说，“将训练数据的caption传递给当前模型，将生成的图像与训练数据的图像进行比较，稍微更改模型，以使其更接近训练数据”即为一步。 + +## 批次大小（batch size） + +批次大小指定每个步骤要计算多少数据。批次计算可以提高速度。一般来说，批次大小越大，精度也越高。 + +“批次大小×步数”是用于训练的数据数量。因此，建议减少步数以增加批次大小。 + +（但是，例如，“批次大小为 1，步数为 1600”和“批次大小为 4，步数为 400”将不会产生相同的结果。如果使用相同的学习速率，通常后者会导致模型欠拟合。请尝试增加学习率（例如 `2e-6`），将步数设置为 500 等。） + +批次大小越大，GPU 内存消耗就越大。如果内存不足，将导致错误，或者在边缘时将导致训练速度降低。建议在任务管理器或 `nvidia-smi` 命令中检查使用的内存量进行调整。 + +注意，一个批次是指“一个数据单位”。 + +## 学习率 + + 学习率指的是每个步骤中改变的程度。如果指定一个大的值，学习速度就会加快，但是可能会出现变化太大导致模型崩溃或无法达到最佳状态的情况。如果指定一个小的值，学习速度会变慢，同时可能无法达到最佳状态。 + +在fine tuning、DreamBooth、LoRA等过程中，学习率会有很大的差异，并且也会受到训练数据、所需训练的模型、批次大小和步骤数等因素的影响。建议从通常值开始，观察训练状态并逐渐调整。 + +默认情况下，整个训练过程中学习率是固定的。但是可以通过调度程序指定学习率如何变化，因此结果也会有所不同。 + +## Epoch + +Epoch指的是训练数据被完整训练一遍（即数据已经迭代一轮）。如果指定了重复次数，则在重复后的数据迭代一轮后，为1个epoch。 + +1个epoch的步骤数通常为“数据量÷批次大小”，但如果使用Aspect Ratio Bucketing，则略微增加（由于不同bucket的数据不能在同一个批次中，因此步骤数会增加）。 + +## 长宽比分桶（Aspect Ratio Bucketing） + +Stable Diffusion 的 v1 是以 512\*512 的分辨率进行训练的，但同时也可以在其他分辨率下进行训练，例如 256\*1024 和 384\*640。这样可以减少裁剪的部分，希望更准确地学习图像和标题之间的关系。 + +此外，由于可以在任意分辨率下进行训练，因此不再需要事先统一图像数据的长宽比。 + +此值可以被设定，其在此之前的配置文件示例中已被启用（设置为 `true`）。 + +只要不超过作为参数给出的分辨率区域（= 内存使用量），就可以按 64 像素的增量（默认值，可更改）在垂直和水平方向上调整和创建训练分辨率。 + +在机器学习中，通常需要将所有输入大小统一，但实际上只要在同一批次中统一即可。 NovelAI 所说的分桶(bucketing) 指的是，预先将训练数据按照长宽比分类到每个学习分辨率下，并通过使用每个 bucket 内的图像创建批次来统一批次图像大小。 + +# 以前的指定格式（不使用 .toml 文件，而是使用命令行选项指定） + +这是一种通过命令行选项而不是指定 .toml 文件的方法。有 DreamBooth 类+标识符方法、DreamBooth caption方法、微调方法三种方式。 + +## DreamBooth、类+标识符方式 + +指定文件夹名称以指定迭代次数。还要使用 `train_data_dir` 和 `reg_data_dir` 选项。 + +### 第1步。准备用于训练的图像 + +创建一个用于存储训练图像的文件夹。__此外__，按以下名称创建目录。 + +``` +<迭代次数>_<标识符> <类别> +``` + +不要忘记下划线``_``。 + +例如，如果在名为“sls frog”的提示下重复数据 20 次，则为“20_sls frog”。如下所示： + +![image](https://user-images.githubusercontent.com/52813779/210770636-1c851377-5936-4c15-90b7-8ac8ad6c2074.png) + +### 多个类别、多个标识符的训练 + +该方法很简单，在用于训练的图像文件夹中，需要准备多个文件夹，每个文件夹都是以“重复次数_<标识符> <类别>”命名的，同样，在正则化图像文件夹中，也需要准备多个文件夹，每个文件夹都是以“重复次数_<类别>”命名的。 + +例如，如果要同时训练“sls青蛙”和“cpc兔子”，则应按以下方式准备文件夹。 + +![image](https://user-images.githubusercontent.com/52813779/210777933-a22229db-b219-4cd8-83ca-e87320fc4192.png) + +如果一个类别包含多个对象，可以只使用一个正则化图像文件夹。例如，如果在1girl类别中有角色A和角色B，则可以按照以下方式处理： + +- train_girls + - 10_sls 1girl + - 10_cpc 1girl +- reg_girls + - 1_1girl + +### step 2. 准备正规化图像 + +这是使用正则化图像时的过程。 + +创建一个文件夹来存储正则化的图像。 __此外，__ 创建一个名为``_`` 的目录。 + +例如，使用提示“frog”并且不重复数据（仅一次）： +![image](https://user-images.githubusercontent.com/52813779/210770897-329758e5-3675-49f1-b345-c135f1725832.png) + + +步骤3. 执行训练 + +执行每个训练脚本。使用 `--train_data_dir` 选项指定包含训练数据文件夹的父文件夹（不是包含图像的文件夹），使用 `--reg_data_dir` 选项指定包含正则化图像的父文件夹（不是包含图像的文件夹）。 + +## DreamBooth，带文本说明（caption）的方式 + +在包含训练图像和正则化图像的文件夹中，将与图像具有相同文件名的文件.caption（可以使用选项进行更改）放置在该文件夹中，然后从该文件中加载caption所作为提示进行训练。 + +※文件夹名称（标识符类）不再用于这些图像的训练。 + +默认的caption文件扩展名为.caption。可以使用训练脚本的 `--caption_extension` 选项进行更改。 使用 `--shuffle_caption` 选项，同时对每个逗号分隔的部分进行训练时会对训练时的caption进行混洗。 + +## 微调方式 + +创建元数据的方式与使用配置文件相同。 使用 `in_json` 选项指定元数据文件。 + +# 训练过程中的样本输出 + +通过在训练中使用模型生成图像，可以检查训练进度。将以下选项指定为训练脚本。 + +- `--sample_every_n_steps` / `--sample_every_n_epochs` + + 指定要采样的步数或epoch数。为这些数字中的每一个输出样本。如果两者都指定，则 epoch 数优先。 +- `--sample_prompts` + + 指定示例输出的提示文件。 + +- `--sample_sampler` + + 指定用于采样输出的采样器。 + `'ddim', 'pndm', 'heun', 'dpmsolver', 'dpmsolver++', 'dpmsingle', 'k_lms', 'k_euler', 'k_euler_a', 'k_dpm_2', 'k_dpm_2_a'`が選べます。 + +要输出样本，您需要提前准备一个包含提示的文本文件。每行输入一个提示。 + +```txt +# prompt 1 +masterpiece, best quality, 1girl, in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28 + +# prompt 2 +masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40 +``` + +以“#”开头的行是注释。您可以使用“`--` + 小写字母”为生成的图像指定选项，例如 `--n`。您可以使用： + +- `--n` 否定提示到下一个选项。 +- `--w` 指定生成图像的宽度。 +- `--h` 指定生成图像的高度。 +- `--d` 指定生成图像的种子。 +- `--l` 指定生成图像的 CFG 比例。 +- `--s` 指定生成过程中的步骤数。 + + +# 每个脚本通用的常用选项 + +文档更新可能跟不上脚本更新。在这种情况下，请使用 `--help` 选项检查可用选项。 +## 学习模型规范 + +- `--v2` / `--v_parameterization` + + 如果使用 Hugging Face 的 stable-diffusion-2-base 或来自它的微调模型作为学习目标模型（对于在推理时指示使用 `v2-inference.yaml` 的模型），`- 当使用-v2` 选项与 stable-diffusion-2、768-v-ema.ckpt 及其微调模型（对于在推理过程中使用 `v2-inference-v.yaml` 的模型），`- 指定两个 -v2`和 `--v_parameterization` 选项。 + + 以下几点在 Stable Diffusion 2.0 中发生了显着变化。 + + 1. 使用分词器 + 2. 使用哪个Text Encoder，使用哪个输出层（2.0使用倒数第二层） + 3. Text Encoder的输出维度(768->1024) + 4. U-Net的结构（CrossAttention的头数等） + 5. v-parameterization（采样方式好像变了） + + 其中base使用1-4，非base使用1-5（768-v）。使用 1-4 进行 v2 选择，使用 5 进行 v_parameterization 选择。 +- `--pretrained_model_name_or_path` + + 指定要从中执行额外训练的模型。您可以指定Stable Diffusion检查点文件（.ckpt 或 .safetensors）、diffusers本地磁盘上的模型目录或diffusers模型 ID（例如“stabilityai/stable-diffusion-2”）。 +## 训练设置 + +- `--output_dir` + + 指定训练后保存模型的文件夹。 + +- `--output_name` + + 指定不带扩展名的模型文件名。 + +- `--dataset_config` + + 指定描述数据集配置的 .toml 文件。 + +- `--max_train_steps` / `--max_train_epochs` + + 指定要训练的步数或epoch数。如果两者都指定，则 epoch 数优先。 +- +- `--mixed_precision` + + 训练混合精度以节省内存。指定像`--mixed_precision = "fp16"`。与无混合精度（默认）相比，精度可能较低，但训练所需的 GPU 内存明显较少。 + + （在RTX30系列以后也可以指定`bf16`，请配合您在搭建环境时做的加速设置）。 +- `--gradient_checkpointing` + + 通过逐步计算权重而不是在训练期间一次计算所有权重来减少训练所需的 GPU 内存量。关闭它不会影响准确性，但打开它允许更大的批次大小，所以那里有影响。 + + 另外，打开它通常会减慢速度，但可以增加批次大小，因此总的训练时间实际上可能会更快。 + +- `--xformers` / `--mem_eff_attn` + + 当指定 xformers 选项时，使用 xformers 的 CrossAttention。如果未安装 xformers 或发生错误（取决于环境，例如 `mixed_precision="no"`），请指定 `mem_eff_attn` 选项而不是使用 CrossAttention 的内存节省版本（xformers 比 慢）。 +- `--save_precision` + + 指定保存时的数据精度。为 save_precision 选项指定 float、fp16 或 bf16 将以该格式保存模型（在 DreamBooth 中保存 Diffusers 格式时无效，微调）。当您想缩小模型的尺寸时请使用它。 +- `--save_every_n_epochs` / `--save_state` / `--resume` + 为 save_every_n_epochs 选项指定一个数字可以在每个时期的训练期间保存模型。 + + 如果同时指定save_state选项，训练状态包括优化器的状态等都会一起保存。。保存目的地将是一个文件夹。 + + 训练状态输出到目标文件夹中名为“-??????-state”（??????是epoch数）的文件夹中。长时间训练时请使用。 + + 使用 resume 选项从保存的训练状态恢复训练。指定训练状态文件夹（其中的状态文件夹，而不是 `output_dir`）。 + + 请注意，由于 Accelerator 规范，epoch 数和全局步数不会保存，即使恢复时它们也从 1 开始。 +- `--save_model_as` （DreamBooth, fine tuning 仅有的） + + 您可以从 `ckpt, safetensors, diffusers, diffusers_safetensors` 中选择模型保存格式。 + +- `--save_model_as=safetensors` 指定喜欢当读取Stable Diffusion格式（ckpt 或safetensors）并以diffusers格式保存时，缺少的信息通过从 Hugging Face 中删除 v1.5 或 v2.1 信息来补充。 + +- `--clip_skip` + + `2` 如果指定，则使用文本编码器 (CLIP) 的倒数第二层的输出。如果省略 1 或选项，则使用最后一层。 + + *SD2.0默认使用倒数第二层，训练SD2.0时请不要指定。 + + 如果被训练的模型最初被训练为使用第二层，则 2 是一个很好的值。 + + 如果您使用的是最后一层，那么整个模型都会根据该假设进行训练。因此，如果再次使用第二层进行训练，可能需要一定数量的teacher数据和更长时间的训练才能得到想要的训练结果。 +- `--max_token_length` + + 默认值为 75。您可以通过指定“150”或“225”来扩展令牌长度来训练。使用长字幕训练时指定。 + + 但由于训练时token展开的规范与Automatic1111的web UI（除法等规范）略有不同，如非必要建议用75训练。 + + 与clip_skip一样，训练与模型训练状态不同的长度可能需要一定量的teacher数据和更长的学习时间。 + +- `--persistent_data_loader_workers` + + 在 Windows 环境中指定它可以显着减少时期之间的延迟。 + +- `--max_data_loader_n_workers` + + 指定数据加载的进程数。大量的进程会更快地加载数据并更有效地使用 GPU，但会消耗更多的主内存。默认是"`8`或者`CPU并发执行线程数 - 1`，取小者"，所以如果主存没有空间或者GPU使用率大概在90%以上，就看那些数字和 `2` 或将其降低到大约 `1`。 +- `--logging_dir` / `--log_prefix` + + 保存训练日志的选项。在 logging_dir 选项中指定日志保存目标文件夹。以 TensorBoard 格式保存日志。 + + 例如，如果您指定 --logging_dir=logs，将在您的工作文件夹中创建一个日志文件夹，并将日志保存在日期/时间文件夹中。 + 此外，如果您指定 --log_prefix 选项，则指定的字符串将添加到日期和时间之前。使用“--logging_dir=logs --log_prefix=db_style1_”进行识别。 + + 要检查 TensorBoard 中的日志，请打开另一个命令提示符并在您的工作文件夹中键入： + ``` + tensorboard --logdir=logs + ``` + + 我觉得tensorboard会在环境搭建的时候安装，如果没有安装，请用`pip install tensorboard`安装。） + + 然后打开浏览器到http://localhost:6006/就可以看到了。 +- `--noise_offset` +本文的实现：https://www.crosslabs.org//blog/diffusion-with-offset-noise + + 看起来它可能会为整体更暗和更亮的图像产生更好的结果。它似乎对 LoRA 训练也有效。指定一个大约 0.1 的值似乎很好。 + +- `--debug_dataset` + + 通过添加此选项，您可以在训练之前检查将训练什么样的图像数据和标题。按 Esc 退出并返回命令行。按 `S` 进入下一步（批次），按 `E` 进入下一个epoch。 + + *图片在 Linux 环境（包括 Colab）下不显示。 + +- `--vae` + + 如果您在 vae 选项中指定Stable Diffusion检查点、VAE 检查点文件、扩散模型或 VAE（两者都可以指定本地或拥抱面模型 ID），则该 VAE 用于训练（缓存时的潜伏）或在训练过程中获得潜伏）。 + + 对于 DreamBooth 和微调，保存的模型将包含此 VAE + +- `--cache_latents` + + 在主内存中缓存 VAE 输出以减少 VRAM 使用。除 flip_aug 之外的任何增强都将不可用。此外，整体训练速度略快。 +- `--min_snr_gamma` + + 指定最小 SNR 加权策略。细节是[这里](https://github.com/kohya-ss/sd-scripts/pull/308)请参阅。论文中推荐`5`。 + +## 优化器相关 + +- `--optimizer_type` + -- 指定优化器类型。您可以指定 + - AdamW : [torch.optim.AdamW](https://pytorch.org/docs/stable/generated/torch.optim.AdamW.html) + - 与过去版本中未指定选项时相同 + - AdamW8bit : 参数同上 + - PagedAdamW8bit : 参数同上 + - 与过去版本中指定的 --use_8bit_adam 相同 + - Lion : https://github.com/lucidrains/lion-pytorch + - Lion8bit : 参数同上 + - PagedLion8bit : 参数同上 + - 与过去版本中指定的 --use_lion_optimizer 相同 + - SGDNesterov : [torch.optim.SGD](https://pytorch.org/docs/stable/generated/torch.optim.SGD.html), nesterov=True + - SGDNesterov8bit : 参数同上 + - DAdaptation(DAdaptAdamPreprint) : https://github.com/facebookresearch/dadaptation + - DAdaptAdam : 参数同上 + - DAdaptAdaGrad : 参数同上 + - DAdaptAdan : 参数同上 + - DAdaptAdanIP : 参数同上 + - DAdaptLion : 参数同上 + - DAdaptSGD : 参数同上 + - Prodigy : https://github.com/konstmish/prodigy + - AdaFactor : [Transformers AdaFactor](https://huggingface.co/docs/transformers/main_classes/optimizer_schedules) + - 任何优化器 + +- `--learning_rate` + + 指定学习率。合适的学习率取决于训练脚本，所以请参考每个解释。 +- `--lr_scheduler` / `--lr_warmup_steps` / `--lr_scheduler_num_cycles` / `--lr_scheduler_power` + + 学习率的调度程序相关规范。 + + 使用 lr_scheduler 选项，您可以从线性、余弦、cosine_with_restarts、多项式、常数、constant_with_warmup 或任何调度程序中选择学习率调度程序。默认值是常量。 + + 使用 lr_warmup_steps，您可以指定预热调度程序的步数（逐渐改变学习率）。 + + lr_scheduler_num_cycles 是 cosine with restarts 调度器中的重启次数，lr_scheduler_power 是多项式调度器中的多项式幂。 + + 有关详细信息，请自行研究。 + + 要使用任何调度程序，请像使用任何优化器一样使用“--lr_scheduler_args”指定可选参数。 +### 关于指定优化器 + +使用 --optimizer_args 选项指定优化器选项参数。可以以key=value的格式指定多个值。此外，您可以指定多个值，以逗号分隔。例如，要指定 AdamW 优化器的参数，``--optimizer_args weight_decay=0.01 betas=.9,.999``。 + +指定可选参数时，请检查每个优化器的规格。 +一些优化器有一个必需的参数，如果省略它会自动添加（例如 SGDNesterov 的动量）。检查控制台输出。 + +D-Adaptation 优化器自动调整学习率。学习率选项指定的值不是学习率本身，而是D-Adaptation决定的学习率的应用率，所以通常指定1.0。如果您希望 Text Encoder 的学习率是 U-Net 的一半，请指定 ``--text_encoder_lr=0.5 --unet_lr=1.0``。 +如果指定 relative_step=True，AdaFactor 优化器可以自动调整学习率（如果省略，将默认添加）。自动调整时，学习率调度器被迫使用 adafactor_scheduler。此外，指定 scale_parameter 和 warmup_init 似乎也不错。 + +自动调整的选项类似于``--optimizer_args "relative_step=True" "scale_parameter=True" "warmup_init=True"``。 + +如果您不想自动调整学习率，请添加可选参数 ``relative_step=False``。在那种情况下，似乎建议将 constant_with_warmup 用于学习率调度程序，而不要为梯度剪裁范数。所以参数就像``--optimizer_type=adafactor --optimizer_args "relative_step=False" --lr_scheduler="constant_with_warmup" --max_grad_norm=0.0``。 + +### 使用任何优化器 + +使用 ``torch.optim`` 优化器时，仅指定类名（例如 ``--optimizer_type=RMSprop``），使用其他模块的优化器时，指定“模块名.类名”。（例如``--optimizer_type=bitsandbytes.optim.lamb.LAMB``）。 + +（内部仅通过 importlib 未确认操作。如果需要，请安装包。） + + +# 创建元数据文件 + +## 准备训练数据 + +如上所述准备好你要训练的图像数据，放在任意文件夹中。 + +例如，存储这样的图像： + +![教师数据文件夹的屏幕截图](https://user-images.githubusercontent.com/52813779/208907739-8e89d5fa-6ca8-4b60-8927-f484d2a9ae04.png) + +## 自动captioning + +如果您只想训练没有标题的标签，请跳过。 + +另外，手动准备caption时，请准备在与教师数据图像相同的目录下，文件名相同，扩展名.caption等。每个文件应该是只有一行的文本文件。 +### 使用 BLIP 添加caption + +最新版本不再需要 BLIP 下载、权重下载和额外的虚拟环境。按原样工作。 + +运行 finetune 文件夹中的 make_captions.py。 + +``` +python finetune\make_captions.py --batch_size <バッチサイズ> <教師データフォルダ> +``` + +如果batch size为8，训练数据放在父文件夹train_data中，则会如下所示 +``` +python finetune\make_captions.py --batch_size 8 ..\train_data +``` + +caption文件创建在与教师数据图像相同的目录中，具有相同的文件名和扩展名.caption。 + +根据 GPU 的 VRAM 容量增加或减少 batch_size。越大越快（我认为 12GB 的 VRAM 可以多一点）。 +您可以使用 max_length 选项指定caption的最大长度。默认值为 75。如果使用 225 的令牌长度训练模型，它可能会更长。 +您可以使用 caption_extension 选项更改caption扩展名。默认为 .caption（.txt 与稍后描述的 DeepDanbooru 冲突）。 +如果有多个教师数据文件夹，则对每个文件夹执行。 + +请注意，推理是随机的，因此每次运行时结果都会发生变化。如果要修复它，请使用 --seed 选项指定一个随机数种子，例如 `--seed 42`。 + +其他的选项，请参考help with `--help`（好像没有文档说明参数的含义，得看源码）。 + +默认情况下，会生成扩展名为 .caption 的caption文件。 + +![caption生成的文件夹](https://user-images.githubusercontent.com/52813779/208908845-48a9d36c-f6ee-4dae-af71-9ab462d1459e.png) + +例如，标题如下： + +![caption和图像](https://user-images.githubusercontent.com/52813779/208908947-af936957-5d73-4339-b6c8-945a52857373.png) + +## 由 DeepDanbooru 标记 + +如果不想给danbooru标签本身打标签，请继续“标题和标签信息的预处理”。 + +标记是使用 DeepDanbooru 或 WD14Tagger 完成的。 WD14Tagger 似乎更准确。如果您想使用 WD14Tagger 进行标记，请跳至下一章。 +### 环境布置 + +将 DeepDanbooru https://github.com/KichangKim/DeepDanbooru 克隆到您的工作文件夹中，或下载并展开 zip。我解压缩了它。 +另外，从 DeepDanbooru 发布页面 https://github.com/KichangKim/DeepDanbooru/releases 上的“DeepDanbooru 预训练模型 v3-20211112-sgd-e28”的资产下载 deepdanbooru-v3-20211112-sgd-e28.zip 并解压到 DeepDanbooru 文件夹。 + +从下面下载。单击以打开资产并从那里下载。 + +![DeepDanbooru下载页面](https://user-images.githubusercontent.com/52813779/208909417-10e597df-7085-41ee-bd06-3e856a1339df.png) + +做一个这样的目录结构 + +![DeepDanbooru的目录结构](https://user-images.githubusercontent.com/52813779/208909486-38935d8b-8dc6-43f1-84d3-fef99bc471aa.png) +为diffusers环境安装必要的库。进入 DeepDanbooru 文件夹并安装它（我认为它实际上只是添加了 tensorflow-io）。 +``` +pip install -r requirements.txt +``` + +接下来，安装 DeepDanbooru 本身。 + +``` +pip install . +``` + +这样就完成了标注环境的准备工作。 + +### 实施标记 +转到 DeepDanbooru 的文件夹并运行 deepdanbooru 进行标记。 +``` +deepdanbooru evaluate <教师资料夹> --project-path deepdanbooru-v3-20211112-sgd-e28 --allow-folder --save-txt +``` + +如果将训练数据放在父文件夹train_data中，则如下所示。 +``` +deepdanbooru evaluate ../train_data --project-path deepdanbooru-v3-20211112-sgd-e28 --allow-folder --save-txt +``` + +在与教师数据图像相同的目录中创建具有相同文件名和扩展名.txt 的标记文件。它很慢，因为它是一个接一个地处理的。 + +如果有多个教师数据文件夹，则对每个文件夹执行。 + +它生成如下。 + +![DeepDanbooru生成的文件](https://user-images.githubusercontent.com/52813779/208909855-d21b9c98-f2d3-4283-8238-5b0e5aad6691.png) + +它会被这样标记（信息量很大...）。 + +![DeepDanbooru标签和图片](https://user-images.githubusercontent.com/52813779/208909908-a7920174-266e-48d5-aaef-940aba709519.png) + +## WD14Tagger标记为 + +此过程使用 WD14Tagger 而不是 DeepDanbooru。 + +使用 Mr. Automatic1111 的 WebUI 中使用的标记器。我参考了这个 github 页面上的信息 (https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger)。 + +初始环境维护所需的模块已经安装。权重自动从 Hugging Face 下载。 +### 实施标记 + +运行脚本以进行标记。 +``` +python tag_images_by_wd14_tagger.py --batch_size <バッチサイズ> <教師データフォルダ> +``` + +如果将训练数据放在父文件夹train_data中，则如下所示 +``` +python tag_images_by_wd14_tagger.py --batch_size 4 ..\train_data +``` + +模型文件将在首次启动时自动下载到 wd14_tagger_model 文件夹（文件夹可以在选项中更改）。它将如下所示。 +![下载文件](https://user-images.githubusercontent.com/52813779/208910447-f7eb0582-90d6-49d3-a666-2b508c7d1842.png) + +在与教师数据图像相同的目录中创建具有相同文件名和扩展名.txt 的标记文件。 +![生成的标签文件](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png) + +![标签和图片](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png) + +使用 thresh 选项，您可以指定确定的标签的置信度数以附加标签。默认值为 0.35，与 WD14Tagger 示例相同。较低的值给出更多的标签，但准确性较低。 + +根据 GPU 的 VRAM 容量增加或减少 batch_size。越大越快（我认为 12GB 的 VRAM 可以多一点）。您可以使用 caption_extension 选项更改标记文件扩展名。默认为 .txt。 + +您可以使用 model_dir 选项指定保存模型的文件夹。 + +此外，如果指定 force_download 选项，即使有保存目标文件夹，也会重新下载模型。 + +如果有多个教师数据文件夹，则对每个文件夹执行。 + +## 预处理caption和标签信息 + +将caption和标签作为元数据合并到一个文件中，以便从脚本中轻松处理。 +### caption预处理 + +要将caption放入元数据，请在您的工作文件夹中运行以下命令（如果您不使用caption进行训练，则不需要运行它）（它实际上是一行，依此类推）。指定 `--full_path` 选项以将图像文件的完整路径存储在元数据中。如果省略此选项，则会记录相对路径，但 .toml 文件中需要单独的文件夹规范。 +``` +python merge_captions_to_metadata.py --full_path <教师资料夹> +　 --in_json <要读取的元数据文件名> <元数据文件名> +``` + +元数据文件名是任意名称。 +如果训练数据为train_data，没有读取元数据文件，元数据文件为meta_cap.json，则会如下。 +``` +python merge_captions_to_metadata.py --full_path train_data meta_cap.json +``` + +您可以使用 caption_extension 选项指定标题扩展。 + +如果有多个教师数据文件夹，请指定 full_path 参数并为每个文件夹执行。 +``` +python merge_captions_to_metadata.py --full_path + train_data1 meta_cap1.json +python merge_captions_to_metadata.py --full_path --in_json meta_cap1.json + train_data2 meta_cap2.json +``` +如果省略in_json，如果有写入目标元数据文件，将从那里读取并覆盖。 + +__* 每次重写 in_json 选项和写入目标并写入单独的元数据文件是安全的。 __ +### 标签预处理 + +同样，标签也收集在元数据中（如果标签不用于训练，则无需这样做）。 +``` +python merge_dd_tags_to_metadata.py --full_path <教师资料夹> + --in_json <要读取的元数据文件名> <要写入的元数据文件名> +``` + +同样的目录结构，读取meta_cap.json和写入meta_cap_dd.json时，会是这样的。 +``` +python merge_dd_tags_to_metadata.py --full_path train_data --in_json meta_cap.json meta_cap_dd.json +``` + +如果有多个教师数据文件夹，请指定 full_path 参数并为每个文件夹执行。 + +``` +python merge_dd_tags_to_metadata.py --full_path --in_json meta_cap2.json + train_data1 meta_cap_dd1.json +python merge_dd_tags_to_metadata.py --full_path --in_json meta_cap_dd1.json + train_data2 meta_cap_dd2.json +``` + +如果省略in_json，如果有写入目标元数据文件，将从那里读取并覆盖。 +__※ 通过每次重写 in_json 选项和写入目标，写入单独的元数据文件是安全的。 __ +### 标题和标签清理 + +到目前为止，标题和DeepDanbooru标签已经被整理到元数据文件中。然而，自动标题生成的标题存在表达差异等微妙问题（※），而标签中可能包含下划线和评级（DeepDanbooru的情况下）。因此，最好使用编辑器的替换功能清理标题和标签。 + +※例如，如果要学习动漫中的女孩，标题可能会包含girl/girls/woman/women等不同的表达方式。另外，将"anime girl"简单地替换为"girl"可能更合适。 + +我们提供了用于清理的脚本，请根据情况编辑脚本并使用它。 + +（不需要指定教师数据文件夹。将清理元数据中的所有数据。） + +``` +python clean_captions_and_tags.py <要读取的元数据文件名> <要写入的元数据文件名> +``` + +--in_json 请注意，不包括在内。例如： + +``` +python clean_captions_and_tags.py meta_cap_dd.json meta_clean.json +``` + +标题和标签的预处理现已完成。 + +## 预先获取 latents + +※ 这一步骤并非必须。即使省略此步骤，也可以在训练过程中获取 latents。但是，如果在训练时执行 `random_crop` 或 `color_aug` 等操作，则无法预先获取 latents（因为每次图像都会改变）。如果不进行预先获取，则可以使用到目前为止的元数据进行训练。 + +提前获取图像的潜在表达并保存到磁盘上。这样可以加速训练过程。同时进行 bucketing（根据宽高比对训练数据进行分类）。 + +请在工作文件夹中输入以下内容。 + +``` +python prepare_buckets_latents.py --full_path <教师资料夹> + <要读取的元数据文件名> <要写入的元数据文件名> + <要微调的模型名称或检查点> + --batch_size <批次大小> + --max_resolution <分辨率宽、高> + --mixed_precision <准确性> +``` + +如果要从meta_clean.json中读取元数据，并将其写入meta_lat.json，使用模型model.ckpt，批处理大小为4，训练分辨率为512*512，精度为no（float32），则应如下所示。 +``` +python prepare_buckets_latents.py --full_path + train_data meta_clean.json meta_lat.json model.ckpt + --batch_size 4 --max_resolution 512,512 --mixed_precision no +``` + +教师数据文件夹中，latents以numpy的npz格式保存。 + +您可以使用--min_bucket_reso选项指定最小分辨率大小，--max_bucket_reso指定最大大小。默认值分别为256和1024。例如，如果指定最小大小为384，则将不再使用分辨率为256 * 1024或320 * 768等。如果将分辨率增加到768 * 768等较大的值，则最好将最大大小指定为1280等。 + +如果指定--flip_aug选项，则进行左右翻转的数据增强。虽然这可以使数据量伪造一倍，但如果数据不是左右对称的（例如角色外观、发型等），则可能会导致训练不成功。 + +对于翻转的图像，也会获取latents，并保存名为\ *_flip.npz的文件，这是一个简单的实现。在fline_tune.py中不需要特定的选项。如果有带有\_flip的文件，则会随机加载带有和不带有flip的文件。 + +即使VRAM为12GB，批次大小也可以稍微增加。分辨率以“宽度，高度”的形式指定，必须是64的倍数。分辨率直接影响fine tuning时的内存大小。在12GB VRAM中，512,512似乎是极限（*）。如果有16GB，则可以将其提高到512,704或512,768。即使分辨率为256,256等，VRAM 8GB也很难承受（因为参数、优化器等与分辨率无关，需要一定的内存）。 + +*有报道称，在batch size为1的训练中，使用12GB VRAM和640,640的分辨率。 + +以下是bucketing结果的显示方式。 + +![bucketing的結果](https://user-images.githubusercontent.com/52813779/208911419-71c00fbb-2ce6-49d5-89b5-b78d7715e441.png) + +如果有多个教师数据文件夹，请指定 full_path 参数并为每个文件夹执行 + +``` +python prepare_buckets_latents.py --full_path + train_data1 meta_clean.json meta_lat1.json model.ckpt + --batch_size 4 --max_resolution 512,512 --mixed_precision no + +python prepare_buckets_latents.py --full_path + train_data2 meta_lat1.json meta_lat2.json model.ckpt + --batch_size 4 --max_resolution 512,512 --mixed_precision no + +``` +可以将读取源和写入目标设为相同，但分开设定更为安全。 + +__※建议每次更改参数并将其写入另一个元数据文件，以确保安全性。__ diff --git a/train_SDXL-en.md b/train_SDXL-en.md new file mode 100644 index 0000000000000000000000000000000000000000..a4c55b3fd78aa1e5a6332c9b7d4b976a92cd0a37 --- /dev/null +++ b/train_SDXL-en.md @@ -0,0 +1,84 @@ +## SDXL training + +The documentation will be moved to the training documentation in the future. The following is a brief explanation of the training scripts for SDXL. + +### Training scripts for SDXL + +- `sdxl_train.py` is a script for SDXL fine-tuning. The usage is almost the same as `fine_tune.py`, but it also supports DreamBooth dataset. + - `--full_bf16` option is added. Thanks to KohakuBlueleaf! + - This option enables the full bfloat16 training (includes gradients). This option is useful to reduce the GPU memory usage. + - The full bfloat16 training might be unstable. Please use it at your own risk. + - The different learning rates for each U-Net block are now supported in sdxl_train.py. Specify with `--block_lr` option. Specify 23 values separated by commas like `--block_lr 1e-3,1e-3 ... 1e-3`. + - 23 values correspond to `0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out`. +- `prepare_buckets_latents.py` now supports SDXL fine-tuning. + +- `sdxl_train_network.py` is a script for LoRA training for SDXL. The usage is almost the same as `train_network.py`. + +- Both scripts has following additional options: + - `--cache_text_encoder_outputs` and `--cache_text_encoder_outputs_to_disk`: Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions. + - `--no_half_vae`: Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs. + +- `--weighted_captions` option is not supported yet for both scripts. + +- `sdxl_train_textual_inversion.py` is a script for Textual Inversion training for SDXL. The usage is almost the same as `train_textual_inversion.py`. + - `--cache_text_encoder_outputs` is not supported. + - There are two options for captions: + 1. Training with captions. All captions must include the token string. The token string is replaced with multiple tokens. + 2. Use `--use_object_template` or `--use_style_template` option. The captions are generated from the template. The existing captions are ignored. + - See below for the format of the embeddings. + +- `--min_timestep` and `--max_timestep` options are added to each training script. These options can be used to train U-Net with different timesteps. The default values are 0 and 1000. + +### Utility scripts for SDXL + +- `tools/cache_latents.py` is added. This script can be used to cache the latents to disk in advance. + - The options are almost the same as `sdxl_train.py'. See the help message for the usage. + - Please launch the script as follows: + `accelerate launch --num_cpu_threads_per_process 1 tools/cache_latents.py ...` + - This script should work with multi-GPU, but it is not tested in my environment. + +- `tools/cache_text_encoder_outputs.py` is added. This script can be used to cache the text encoder outputs to disk in advance. + - The options are almost the same as `cache_latents.py` and `sdxl_train.py`. See the help message for the usage. + +- `sdxl_gen_img.py` is added. This script can be used to generate images with SDXL, including LoRA, Textual Inversion and ControlNet-LLLite. See the help message for the usage. + +### Tips for SDXL training + +- The default resolution of SDXL is 1024x1024. +- The fine-tuning can be done with 24GB GPU memory with the batch size of 1. For 24GB GPU, the following options are recommended __for the fine-tuning with 24GB GPU memory__: + - Train U-Net only. + - Use gradient checkpointing. + - Use `--cache_text_encoder_outputs` option and caching latents. + - Use Adafactor optimizer. RMSprop 8bit or Adagrad 8bit may work. AdamW 8bit doesn't seem to work. +- The LoRA training can be done with 8GB GPU memory (10GB recommended). For reducing the GPU memory usage, the following options are recommended: + - Train U-Net only. + - Use gradient checkpointing. + - Use `--cache_text_encoder_outputs` option and caching latents. + - Use one of 8bit optimizers or Adafactor optimizer. + - Use lower dim (4 to 8 for 8GB GPU). +- `--network_train_unet_only` option is highly recommended for SDXL LoRA. Because SDXL has two text encoders, the result of the training will be unexpected. +- PyTorch 2 seems to use slightly less GPU memory than PyTorch 1. +- `--bucket_reso_steps` can be set to 32 instead of the default value 64. Smaller values than 32 will not work for SDXL training. + +Example of the optimizer settings for Adafactor with the fixed learning rate: +```toml +optimizer_type = "adafactor" +optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ] +lr_scheduler = "constant_with_warmup" +lr_warmup_steps = 100 +learning_rate = 4e-7 # SDXL original learning rate +``` + +### Format of Textual Inversion embeddings for SDXL + +```python +from safetensors.torch import save_file + +state_dict = {"clip_g": embs_for_text_encoder_1280, "clip_l": embs_for_text_encoder_768} +save_file(state_dict, file) +``` + +### ControlNet-LLLite + +ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [documentation](./docs/train_lllite_README.md) for details. + diff --git a/train_db_README-ja.md b/train_db_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..0d0747bb41223a52a4f609f58eb1314639924913 --- /dev/null +++ b/train_db_README-ja.md @@ -0,0 +1,167 @@ +DreamBoothのガイドです。 + +[学習についての共通ドキュメント](./train_README-ja.md) もあわせてご覧ください。 + +# 概要 + +DreamBoothとは、画像生成モデルに特定の主題を追加学習し、それを特定の識別子で生成する技術です。[論文はこちら](https://arxiv.org/abs/2208.12242)。 + +具体的には、Stable Diffusionのモデルにキャラや画風などを学ばせ、それを `shs` のような特定の単語で呼び出せる（生成画像に出現させる）ことができます。 + +スクリプトは[DiffusersのDreamBooth](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth)を元にしていますが、以下のような機能追加を行っています（いくつかの機能は元のスクリプト側もその後対応しています）。 + +スクリプトの主な機能は以下の通りです。 + +- 8bit Adam optimizerおよびlatentのキャッシュによる省メモリ化（[Shivam Shrirao氏版](https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth)と同様）。 +- xformersによる省メモリ化。 +- 512x512だけではなく任意サイズでの学習。 +- augmentationによる品質の向上。 +- DreamBoothだけではなくText Encoder+U-Netのfine tuningに対応。 +- Stable Diffusion形式でのモデルの読み書き。 +- Aspect Ratio Bucketing。 +- Stable Diffusion v2.0対応。 + +# 学習の手順 + +あらかじめこのリポジトリのREADMEを参照し、環境整備を行ってください。 + +## データの準備 + +[学習データの準備について](./train_README-ja.md) を参照してください。 + +## 学習の実行 + +スクリプトを実行します。最大限、メモリを節約したコマンドは以下のようになります（実際には1行で入力します）。それぞれの行を必要に応じて書き換えてください。12GB程度のVRAMで動作するようです。 + +``` +accelerate launch --num_cpu_threads_per_process 1 train_db.py + --pretrained_model_name_or_path=<.ckptまたは.safetensordまたはDiffusers版モデルのディレクトリ> + --dataset_config=<データ準備で作成した.tomlファイル> + --output_dir=<学習したモデルの出力先フォルダ> + --output_name=<学習したモデル出力時のファイル名> + --save_model_as=safetensors + --prior_loss_weight=1.0 + --max_train_steps=1600 + --learning_rate=1e-6 + --optimizer_type="AdamW8bit" + --xformers + --mixed_precision="fp16" + --cache_latents + --gradient_checkpointing +``` + +`num_cpu_threads_per_process` には通常は1を指定するとよいようです。 + +`pretrained_model_name_or_path` に追加学習を行う元となるモデルを指定します。Stable Diffusionのcheckpointファイル（.ckptまたは.safetensors）、Diffusersのローカルディスクにあるモデルディレクトリ、DiffusersのモデルID（"stabilityai/stable-diffusion-2"など）が指定できます。 + +`output_dir` に学習後のモデルを保存するフォルダを指定します。`output_name` にモデルのファイル名を拡張子を除いて指定します。`save_model_as` でsafetensors形式での保存を指定しています。 + +`dataset_config` に `.toml` ファイルを指定します。ファイル内でのバッチサイズ指定は、当初はメモリ消費を抑えるために `1` としてください。 + +`prior_loss_weight` は正則化画像のlossの重みです。通常は1.0を指定します。 + +学習させるステップ数 `max_train_steps` を1600とします。学習率 `learning_rate` はここでは1e-6を指定しています。 + +省メモリ化のため `mixed_precision="fp16"` を指定します（RTX30 シリーズ以降では `bf16` も指定できます。環境整備時にaccelerateに行った設定と合わせてください）。また `gradient_checkpointing` を指定します。 + +オプティマイザ（モデルを学習データにあうように最適化＝学習させるクラス）にメモリ消費の少ない 8bit AdamW を使うため、 `optimizer_type="AdamW8bit"` を指定します。 + +`xformers` オプションを指定し、xformersのCrossAttentionを用います。xformersをインストールしていない場合やエラーとなる場合（環境にもよりますが `mixed_precision="no"` の場合など）、代わりに `mem_eff_attn` オプションを指定すると省メモリ版CrossAttentionを使用します（速度は遅くなります）。 + +省メモリ化のため `cache_latents` オプションを指定してVAEの出力をキャッシュします。 + +ある程度メモリがある場合は、`.toml` ファイルを編集してバッチサイズをたとえば `4` くらいに増やしてください（高速化と精度向上の可能性があります）。また `cache_latents` を外すことで augmentation が可能になります。 + +### よく使われるオプションについて + +以下の場合には [学習の共通ドキュメント](./train_README-ja.md) の「よく使われるオプション」を参照してください。 + +- Stable Diffusion 2.xまたはそこからの派生モデルを学習する +- clip skipを2以上を前提としたモデルを学習する +- 75トークンを超えたキャプションで学習する + +### DreamBoothでのステップ数について + +当スクリプトでは省メモリ化のため、ステップ当たりの学習回数が元のスクリプトの半分になっています（対象の画像と正則化画像を同一のバッチではなく別のバッチに分割して学習するため）。 + +元のDiffusers版やXavierXiao氏のStable Diffusion版とほぼ同じ学習を行うには、ステップ数を倍にしてください。 + +（学習画像と正則化画像をまとめてから shuffle するため厳密にはデータの順番が変わってしまいますが、学習には大きな影響はないと思います。） + +### DreamBoothでのバッチサイズについて + +モデル全体を学習するためLoRA等の学習に比べるとメモリ消費量は多くなります（fine tuningと同じ）。 + +### 学習率について + +Diffusers版では5e-6ですがStable Diffusion版は1e-6ですので、上のサンプルでは1e-6を指定しています。 + +### 以前の形式のデータセット指定をした場合のコマンドライン + +解像度やバッチサイズをオプションで指定します。コマンドラインの例は以下の通りです。 + +``` +accelerate launch --num_cpu_threads_per_process 1 train_db.py + --pretrained_model_name_or_path=<.ckptまたは.safetensordまたはDiffusers版モデルのディレクトリ> + --train_data_dir=<学習用データのディレクトリ> + --reg_data_dir=<正則化画像のディレクトリ> + --output_dir=<学習したモデルの出力先ディレクトリ> + --output_name=<学習したモデル出力時のファイル名> + --prior_loss_weight=1.0 + --resolution=512 + --train_batch_size=1 + --learning_rate=1e-6 + --max_train_steps=1600 + --use_8bit_adam + --xformers + --mixed_precision="bf16" + --cache_latents + --gradient_checkpointing +``` + +## 学習したモデルで画像生成する + +学習が終わると指定したフォルダに指定した名前でsafetensorsファイルが出力されます。 + +v1.4/1.5およびその他の派生モデルの場合、このモデルでAutomatic1111氏のWebUIなどで推論できます。models\Stable-diffusionフォルダに置いてください。 + +v2.xモデルでWebUIで画像生成する場合、モデルの仕様が記述された.yamlファイルが別途必要になります。v2.x baseの場合はv2-inference.yamlを、768/vの場合はv2-inference-v.yamlを、同じフォルダに置き、拡張子の前の部分をモデルと同じ名前にしてください。 + +![image](https://user-images.githubusercontent.com/52813779/210776915-061d79c3-6582-42c2-8884-8b91d2f07313.png) + +各yamlファイルは[Stability AIのSD2.0のリポジトリ](https://github.com/Stability-AI/stablediffusion/tree/main/configs/stable-diffusion)にあります。 + +# DreamBooth特有のその他の主なオプション + +すべてのオプションについては別文書を参照してください。 + +## Text Encoderの学習を途中から行わない --stop_text_encoder_training + +stop_text_encoder_trainingオプションに数値を指定すると、そのステップ数以降はText Encoderの学習を行わずU-Netだけ学習します。場合によっては精度の向上が期待できるかもしれません。 + +（恐らくText Encoderだけ先に過学習することがあり、それを防げるのではないかと推測していますが、詳細な影響は不明です。） + +## Tokenizerのパディングをしない --no_token_padding +no_token_paddingオプションを指定するとTokenizerの出力をpaddingしません（Diffusers版の旧DreamBoothと同じ動きになります）。 + + + diff --git a/train_db_README-zh.md b/train_db_README-zh.md new file mode 100644 index 0000000000000000000000000000000000000000..d8ea5f3edae856248c334d7c2a779dc44a86c0f5 --- /dev/null +++ b/train_db_README-zh.md @@ -0,0 +1,162 @@ +这是DreamBooth的指南。 + +请同时查看[关于学习的通用文档](./train_README-zh.md)。 + +# 概要 + +DreamBooth是一种将特定主题添加到图像生成模型中进行学习，并使用特定识别子生成它的技术。论文链接。 + +具体来说，它可以将角色和绘画风格等添加到Stable Diffusion模型中进行学习，并使用特定的单词（例如`shs`）来调用（呈现在生成的图像中）。 + +脚本基于Diffusers的DreamBooth，但添加了以下功能（一些功能已在原始脚本中得到支持）。 + +脚本的主要功能如下： + +- 使用8位Adam优化器和潜在变量的缓存来节省内存（与Shivam Shrirao版相似）。 +- 使用xformers来节省内存。 +- 不仅支持512x512，还支持任意尺寸的训练。 +- 通过数据增强来提高质量。 +- 支持DreamBooth和Text Encoder + U-Net的微调。 +- 支持以Stable Diffusion格式读写模型。 +- 支持Aspect Ratio Bucketing。 +- 支持Stable Diffusion v2.0。 + +# 训练步骤 + +请先参阅此存储库的README以进行环境设置。 + +## 准备数据 + +请参阅[有关准备训练数据的说明](./train_README-zh.md)。 + +## 运行训练 + +运行脚本。以下是最大程度地节省内存的命令（实际上，这将在一行中输入）。请根据需要修改每行。它似乎需要约12GB的VRAM才能运行。 +``` +accelerate launch --num_cpu_threads_per_process 1 train_db.py + --pretrained_model_name_or_path=<.ckpt或.safetensord或Diffusers版模型的目录> + --dataset_config=<数据准备时创建的.toml文件> + --output_dir=<训练模型的输出目录> + --output_name=<训练模型输出时的文件名> + --save_model_as=safetensors + --prior_loss_weight=1.0 + --max_train_steps=1600 + --learning_rate=1e-6 + --optimizer_type="AdamW8bit" + --xformers + --mixed_precision="fp16" + --cache_latents + --gradient_checkpointing +``` +`num_cpu_threads_per_process` 通常应该设置为1。 + +`pretrained_model_name_or_path` 指定要进行追加训练的基础模型。可以指定 Stable Diffusion 的 checkpoint 文件（.ckpt 或 .safetensors）、Diffusers 的本地模型目录或模型 ID（如 "stabilityai/stable-diffusion-2"）。 + +`output_dir` 指定保存训练后模型的文件夹。在 `output_name` 中指定模型文件名，不包括扩展名。使用 `save_model_as` 指定以 safetensors 格式保存。 + +在 `dataset_config` 中指定 `.toml` 文件。初始批处理大小应为 `1`，以减少内存消耗。 + +`prior_loss_weight` 是正则化图像损失的权重。通常设为1.0。 + +将要训练的步数 `max_train_steps` 设置为1600。在这里，学习率 `learning_rate` 被设置为1e-6。 + +为了节省内存，设置 `mixed_precision="fp16"`（在 RTX30 系列及更高版本中也可以设置为 `bf16`）。同时指定 `gradient_checkpointing`。 + +为了使用内存消耗较少的 8bit AdamW 优化器（将模型优化为适合于训练数据的状态），指定 `optimizer_type="AdamW8bit"`。 + +指定 `xformers` 选项，并使用 xformers 的 CrossAttention。如果未安装 xformers 或出现错误（具体情况取决于环境，例如使用 `mixed_precision="no"`），则可以指定 `mem_eff_attn` 选项以使用省内存版的 CrossAttention（速度会变慢）。 + +为了节省内存，指定 `cache_latents` 选项以缓存 VAE 的输出。 + +如果有足够的内存，请编辑 `.toml` 文件将批处理大小增加到大约 `4`（可能会提高速度和精度）。此外，取消 `cache_latents` 选项可以进行数据增强。 + +### 常用选项 + +对于以下情况，请参阅“常用选项”部分。 + +- 学习 Stable Diffusion 2.x 或其衍生模型。 +- 学习基于 clip skip 大于等于2的模型。 +- 学习超过75个令牌的标题。 + +### 关于DreamBooth中的步数 + +为了实现省内存化，该脚本中每个步骤的学习次数减半（因为学习和正则化的图像在训练时被分为不同的批次）。 + +要进行与原始Diffusers版或XavierXiao的Stable Diffusion版几乎相同的学习，请将步骤数加倍。 + +（虽然在将学习图像和正则化图像整合后再打乱顺序，但我认为对学习没有太大影响。） + +关于DreamBooth的批量大小 + +与像LoRA这样的学习相比，为了训练整个模型，内存消耗量会更大（与微调相同）。 + +关于学习率 + +在Diffusers版中，学习率为5e-6，而在Stable Diffusion版中为1e-6，因此在上面的示例中指定了1e-6。 + +当使用旧格式的数据集指定命令行时 + +使用选项指定分辨率和批量大小。命令行示例如下。 +``` +accelerate launch --num_cpu_threads_per_process 1 train_db.py + --pretrained_model_name_or_path=<.ckpt或.safetensord或Diffusers版模型的目录> + --train_data_dir=<训练数据的目录> + --reg_data_dir=<正则化图像的目录> + --output_dir=<训练后模型的输出目录> + --output_name=<训练后模型输出文件的名称> + --prior_loss_weight=1.0 + --resolution=512 + --train_batch_size=1 + --learning_rate=1e-6 + --max_train_steps=1600 + --use_8bit_adam + --xformers + --mixed_precision="bf16" + --cache_latents + --gradient_checkpointing +``` + +## 使用训练好的模型生成图像 + +训练完成后，将在指定的文件夹中以指定的名称输出safetensors文件。 + +对于v1.4/1.5和其他派生模型，可以在此模型中使用Automatic1111先生的WebUI进行推断。请将其放置在models\Stable-diffusion文件夹中。 + +对于使用v2.x模型在WebUI中生成图像的情况，需要单独的.yaml文件来描述模型的规格。对于v2.x base，需要v2-inference.yaml，对于768/v，则需要v2-inference-v.yaml。请将它们放置在相同的文件夹中，并将文件扩展名之前的部分命名为与模型相同的名称。 +![image](https://user-images.githubusercontent.com/52813779/210776915-061d79c3-6582-42c2-8884-8b91d2f07313.png) + +每个yaml文件都在[Stability AI的SD2.0存储库](https://github.com/Stability-AI/stablediffusion/tree/main/configs/stable-diffusion)……之中。 + +# DreamBooth的其他主要选项 + +有关所有选项的详细信息，请参阅另一份文档。 + +## 不在中途开始对文本编码器进行训练 --stop_text_encoder_training + +如果在stop_text_encoder_training选项中指定一个数字，则在该步骤之后，将不再对文本编码器进行训练，只会对U-Net进行训练。在某些情况下，可能会期望提高精度。 + +（我们推测可能会有时候仅仅文本编码器会过度学习，而这样做可以避免这种情况，但详细影响尚不清楚。） + +## 不进行分词器的填充 --no_token_padding + +如果指定no_token_padding选项，则不会对分词器的输出进行填充（与Diffusers版本的旧DreamBooth相同）。 + + diff --git a/train_lllite_README-ja.md b/train_lllite_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..1f6a78d5cb078e802180a22d561583d04a0fafb7 --- /dev/null +++ b/train_lllite_README-ja.md @@ -0,0 +1,218 @@ +# ControlNet-LLLite について + +__きわめて実験的な実装のため、将来的に大きく変更される可能性があります。__ + +## 概要 +ControlNet-LLLite は、[ControlNet](https://github.com/lllyasviel/ControlNet) の軽量版です。LoRA Like Lite という意味で、LoRAからインスピレーションを得た構造を持つ、軽量なControlNetです。現在はSDXLにのみ対応しています。 + +## サンプルの重みファイルと推論 + +こちらにあります: https://huggingface.co/kohya-ss/controlnet-lllite + +ComfyUIのカスタムノードを用意しています。: https://github.com/kohya-ss/ControlNet-LLLite-ComfyUI + +生成サンプルはこのページの末尾にあります。 + +## モデル構造 +ひとつのLLLiteモジュールは、制御用画像（以下conditioning image）を潜在空間に写像するconditioning image embeddingと、LoRAにちょっと似た構造を持つ小型のネットワークからなります。LLLiteモジュールを、LoRAと同様にU-NetのLinearやConvに追加します。詳しくはソースコードを参照してください。 + +推論環境の制限で、現在はCrossAttentionのみ（attn1のq/k/v、attn2のq）に追加されます。 + +## モデルの学習 + +### データセットの準備 +DreamBooth 方式の dataset で、`conditioning_data_dir` で指定したディレクトリにconditioning imageを格納してください。 + +（finetuning 方式の dataset はサポートしていません。） + +conditioning imageは学習用画像と同じbasenameを持つ必要があります。また、conditioning imageは学習用画像と同じサイズに自動的にリサイズされます。conditioning imageにはキャプションファイルは不要です。 + +たとえば、キャプションにフォルダ名ではなくキャプションファイルを用いる場合の設定ファイルは以下のようになります。 + +```toml +[[datasets.subsets]] +image_dir = "path/to/image/dir" +caption_extension = ".txt" +conditioning_data_dir = "path/to/conditioning/image/dir" +``` + +現時点の制約として、random_cropは使用できません。 + +学習データとしては、元のモデルで生成した画像を学習用画像として、そこから加工した画像をconditioning imageとした、合成によるデータセットを用いるのがもっとも簡単です（データセットの品質的には問題があるかもしれません）。具体的なデータセットの合成方法については後述します。 + +なお、元モデルと異なる画風の画像を学習用画像とすると、制御に加えて、その画風についても学ぶ必要が生じます。ControlNet-LLLiteは容量が少ないため、画風学習には不向きです。このような場合には、後述の次元数を多めにしてください。 + +### 学習 +スクリプトで生成する場合は、`sdxl_train_control_net_lllite.py` を実行してください。`--cond_emb_dim` でconditioning image embeddingの次元数を指定できます。`--network_dim` でLoRA的モジュールのrankを指定できます。その他のオプションは`sdxl_train_network.py`に準じますが、`--network_module`の指定は不要です。 + +学習時にはメモリを大量に使用しますので、キャッシュやgradient checkpointingなどの省メモリ化のオプションを有効にしてください。また`--full_bf16` オプションで、BFloat16を使用するのも有効です（RTX 30シリーズ以降のGPUが必要です）。24GB VRAMで動作確認しています。 + +conditioning image embeddingの次元数は、サンプルのCannyでは32を指定しています。LoRA的モジュールのrankは同じく64です。対象とするconditioning imageの特徴に合わせて調整してください。 + +（サンプルのCannyは恐らくかなり難しいと思われます。depthなどでは半分程度にしてもいいかもしれません。） + +以下は .toml の設定例です。 + +```toml +pretrained_model_name_or_path = "/path/to/model_trained_on.safetensors" +max_train_epochs = 12 +max_data_loader_n_workers = 4 +persistent_data_loader_workers = true +seed = 42 +gradient_checkpointing = true +mixed_precision = "bf16" +save_precision = "bf16" +full_bf16 = true +optimizer_type = "adamw8bit" +learning_rate = 2e-4 +xformers = true +output_dir = "/path/to/output/dir" +output_name = "output_name" +save_every_n_epochs = 1 +save_model_as = "safetensors" +vae_batch_size = 4 +cache_latents = true +cache_latents_to_disk = true +cache_text_encoder_outputs = true +cache_text_encoder_outputs_to_disk = true +network_dim = 64 +cond_emb_dim = 32 +dataset_config = "/path/to/dataset.toml" +``` + +### 推論 + +スクリプトで生成する場合は、`sdxl_gen_img.py` を実行してください。`--control_net_lllite_models` でLLLiteのモデルファイルを指定できます。次元数はモデルファイルから自動取得します。 + +`--guide_image_path`で推論に用いるconditioning imageを指定してください。なおpreprocessは行われないため、たとえばCannyならCanny処理を行った画像を指定してください（背景黒に白線）。`--control_net_preps`, `--control_net_weights`, `--control_net_ratios` には未対応です。 + +## データセットの合成方法 + +### 学習用画像の生成 + +学習のベースとなるモデルで画像生成を行います。Web UIやComfyUIなどで生成してください。画像サイズはモデルのデフォルトサイズで良いと思われます（1024x1024など）。bucketingを用いることもできます。その場合は適宜適切な解像度で生成してください。 + +生成時のキャプション等は、ControlNet-LLLiteの利用時に生成したい画像にあわせるのが良いと思われます。 + +生成した画像を任意のディレクトリに保存してください。このディレクトリをデータセットの設定ファイルで指定します。 + +当リポジトリ内の `sdxl_gen_img.py` でも生成できます。例えば以下のように実行します。 + +```dos +python sdxl_gen_img.py --ckpt path/to/model.safetensors --n_iter 1 --scale 10 --steps 36 --outdir path/to/output/dir --xformers --W 1024 --H 1024 --original_width 2048 --original_height 2048 --bf16 --sampler ddim --batch_size 4 --vae_batch_size 2 --images_per_prompt 512 --max_embeddings_multiples 1 --prompt "{portrait|digital art|anime screen cap|detailed illustration} of 1girl, {standing|sitting|walking|running|dancing} on {classroom|street|town|beach|indoors|outdoors}, {looking at viewer|looking away|looking at another}, {in|wearing} {shirt and skirt|school uniform|casual wear} { |, dynamic pose}, (solo), teen age, {0-1$$smile,|blush,|kind smile,|expression less,|happy,|sadness,} {0-1$$upper body,|full body,|cowboy shot,|face focus,} trending on pixiv, {0-2$$depth of fields,|8k wallpaper,|highly detailed,|pov,} {0-1$$summer, |winter, |spring, |autumn, } beautiful face { |, from below|, from above|, from side|, from behind|, from back} --n nsfw, bad face, lowres, low quality, worst quality, low effort, watermark, signature, ugly, poorly drawn" +``` + +VRAM 24GBの設定です。VRAMサイズにより`--batch_size` `--vae_batch_size`を調整してください。 + +`--prompt`でワイルドカードを利用してランダムに生成しています。適宜調整してください。 + +### 画像の加工 + +外部のプログラムを用いて、生成した画像を加工します。加工した画像を任意のディレクトリに保存してください。これらがconditioning imageになります。 + +加工にはたとえばCannyなら以下のようなスクリプトが使えます。 + +```python +import glob +import os +import random +import cv2 +import numpy as np + +IMAGES_DIR = "path/to/generated/images" +CANNY_DIR = "path/to/canny/images" + +os.makedirs(CANNY_DIR, exist_ok=True) +img_files = glob.glob(IMAGES_DIR + "/*.png") +for img_file in img_files: + can_file = CANNY_DIR + "/" + os.path.basename(img_file) + if os.path.exists(can_file): + print("Skip: " + img_file) + continue + + print(img_file) + + img = cv2.imread(img_file) + + # random threshold + # while True: + # threshold1 = random.randint(0, 127) + # threshold2 = random.randint(128, 255) + # if threshold2 - threshold1 > 80: + # break + + # fixed threshold + threshold1 = 100 + threshold2 = 200 + + img = cv2.Canny(img, threshold1, threshold2) + + cv2.imwrite(can_file, img) +``` + +### キャプションファイルの作成 + +学習用画像のbasenameと同じ名前で、それぞれの画像に対応したキャプションファイルを作成してください。生成時のプロンプトをそのまま利用すれば良いと思われます。 + +`sdxl_gen_img.py` で生成した場合は、画像内のメタデータに生成時のプロンプトが記録されていますので、以下のようなスクリプトで学習用画像と同じディレクトリにキャプションファイルを作成できます（拡張子 `.txt`）。 + +```python +import glob +import os +from PIL import Image + +IMAGES_DIR = "path/to/generated/images" + +img_files = glob.glob(IMAGES_DIR + "/*.png") +for img_file in img_files: + cap_file = img_file.replace(".png", ".txt") + if os.path.exists(cap_file): + print(f"Skip: {img_file}") + continue + print(img_file) + + img = Image.open(img_file) + prompt = img.text["prompt"] if "prompt" in img.text else "" + if prompt == "": + print(f"Prompt not found in {img_file}") + + with open(cap_file, "w") as f: + f.write(prompt + "\n") +``` + +### データセットの設定ファイルの作成 + +コマンドラインオプションからの指定も可能ですが、`.toml`ファイルを作成する場合は `conditioning_data_dir` に加工した画像を保存したディレクトリを指定します。 + +以下は設定ファイルの例です。 + +```toml +[general] +flip_aug = false +color_aug = false +resolution = [1024,1024] + +[[datasets]] +batch_size = 8 +enable_bucket = false + + [[datasets.subsets]] + image_dir = "path/to/generated/image/dir" + caption_extension = ".txt" + conditioning_data_dir = "path/to/canny/image/dir" +``` + +## 謝辞 + +ControlNetの作者である lllyasviel 氏、実装上のアドバイスとトラブル解決へのご尽力をいただいた furusu 氏、ControlNetデータセットを実装していただいた ddPn08 氏に感謝いたします。 + +## サンプル +Canny +![kohya_ss_girl_standing_at_classroom_smiling_to_the_viewer_class_78976b3e-0d4d-4ea0-b8e3-053ae493abbc](https://github.com/kohya-ss/sd-scripts/assets/52813779/37e9a736-649b-4c0f-ab26-880a1bf319b5) + +![im_20230820104253_000_1](https://github.com/kohya-ss/sd-scripts/assets/52813779/c8896900-ab86-4120-932f-6e2ae17b77c0) + +![im_20230820104302_000_1](https://github.com/kohya-ss/sd-scripts/assets/52813779/b12457a0-ee3c-450e-ba9a-b712d0fe86bb) + +![im_20230820104310_000_1](https://github.com/kohya-ss/sd-scripts/assets/52813779/8845b8d9-804a-44ac-9618-113a28eac8a1) + diff --git a/train_lllite_README.md b/train_lllite_README.md new file mode 100644 index 0000000000000000000000000000000000000000..a05f87f5f93e46dba8e2454a7e7833bf31c8f79b --- /dev/null +++ b/train_lllite_README.md @@ -0,0 +1,219 @@ +# About ControlNet-LLLite + +__This is an extremely experimental implementation and may change significantly in the future.__ + +日本語版は[こちら](./train_lllite_README-ja.md) + +## Overview + +ControlNet-LLLite is a lightweight version of [ControlNet](https://github.com/lllyasviel/ControlNet). It is a "LoRA Like Lite" that is inspired by LoRA and has a lightweight structure. Currently, only SDXL is supported. + +## Sample weight file and inference + +Sample weight file is available here: https://huggingface.co/kohya-ss/controlnet-lllite + +A custom node for ComfyUI is available: https://github.com/kohya-ss/ControlNet-LLLite-ComfyUI + +Sample images are at the end of this page. + +## Model structure + +A single LLLite module consists of a conditioning image embedding that maps a conditioning image to a latent space and a small network with a structure similar to LoRA. The LLLite module is added to U-Net's Linear and Conv in the same way as LoRA. Please refer to the source code for details. + +Due to the limitations of the inference environment, only CrossAttention (attn1 q/k/v, attn2 q) is currently added. + +## Model training + +### Preparing the dataset + +In addition to the normal DreamBooth method dataset, please store the conditioning image in the directory specified by `conditioning_data_dir`. The conditioning image must have the same basename as the training image. The conditioning image will be automatically resized to the same size as the training image. The conditioning image does not require a caption file. + +(We do not support the finetuning method dataset.) + +```toml +[[datasets.subsets]] +image_dir = "path/to/image/dir" +caption_extension = ".txt" +conditioning_data_dir = "path/to/conditioning/image/dir" +``` + +At the moment, random_crop cannot be used. + +For training data, it is easiest to use a synthetic dataset with the original model-generated images as training images and processed images as conditioning images (the quality of the dataset may be problematic). See below for specific methods of synthesizing datasets. + +Note that if you use an image with a different art style than the original model as a training image, the model will have to learn not only the control but also the art style. ControlNet-LLLite has a small capacity, so it is not suitable for learning art styles. In such cases, increase the number of dimensions as described below. + +### Training + +Run `sdxl_train_control_net_lllite.py`. You can specify the dimension of the conditioning image embedding with `--cond_emb_dim`. You can specify the rank of the LoRA-like module with `--network_dim`. Other options are the same as `sdxl_train_network.py`, but `--network_module` is not required. + +Since a large amount of memory is used during training, please enable memory-saving options such as cache and gradient checkpointing. It is also effective to use BFloat16 with the `--full_bf16` option (requires RTX 30 series or later GPU). It has been confirmed to work with 24GB VRAM. + +For the sample Canny, the dimension of the conditioning image embedding is 32. The rank of the LoRA-like module is also 64. Adjust according to the features of the conditioning image you are targeting. + +(The sample Canny is probably quite difficult. It may be better to reduce it to about half for depth, etc.) + +The following is an example of a .toml configuration. + +```toml +pretrained_model_name_or_path = "/path/to/model_trained_on.safetensors" +max_train_epochs = 12 +max_data_loader_n_workers = 4 +persistent_data_loader_workers = true +seed = 42 +gradient_checkpointing = true +mixed_precision = "bf16" +save_precision = "bf16" +full_bf16 = true +optimizer_type = "adamw8bit" +learning_rate = 2e-4 +xformers = true +output_dir = "/path/to/output/dir" +output_name = "output_name" +save_every_n_epochs = 1 +save_model_as = "safetensors" +vae_batch_size = 4 +cache_latents = true +cache_latents_to_disk = true +cache_text_encoder_outputs = true +cache_text_encoder_outputs_to_disk = true +network_dim = 64 +cond_emb_dim = 32 +dataset_config = "/path/to/dataset.toml" +``` + +### Inference + +If you want to generate images with a script, run `sdxl_gen_img.py`. You can specify the LLLite model file with `--control_net_lllite_models`. The dimension is automatically obtained from the model file. + +Specify the conditioning image to be used for inference with `--guide_image_path`. Since preprocess is not performed, if it is Canny, specify an image processed with Canny (white line on black background). `--control_net_preps`, `--control_net_weights`, and `--control_net_ratios` are not supported. + +## How to synthesize a dataset + +### Generating training images + +Generate images with the base model for training. Please generate them with Web UI or ComfyUI etc. The image size should be the default size of the model (1024x1024, etc.). You can also use bucketing. In that case, please generate it at an arbitrary resolution. + +The captions and other settings when generating the images should be the same as when generating the images with the trained ControlNet-LLLite model. + +Save the generated images in an arbitrary directory. Specify this directory in the dataset configuration file. + + +You can also generate them with `sdxl_gen_img.py` in this repository. For example, run as follows: + +```dos +python sdxl_gen_img.py --ckpt path/to/model.safetensors --n_iter 1 --scale 10 --steps 36 --outdir path/to/output/dir --xformers --W 1024 --H 1024 --original_width 2048 --original_height 2048 --bf16 --sampler ddim --batch_size 4 --vae_batch_size 2 --images_per_prompt 512 --max_embeddings_multiples 1 --prompt "{portrait|digital art|anime screen cap|detailed illustration} of 1girl, {standing|sitting|walking|running|dancing} on {classroom|street|town|beach|indoors|outdoors}, {looking at viewer|looking away|looking at another}, {in|wearing} {shirt and skirt|school uniform|casual wear} { |, dynamic pose}, (solo), teen age, {0-1$$smile,|blush,|kind smile,|expression less,|happy,|sadness,} {0-1$$upper body,|full body,|cowboy shot,|face focus,} trending on pixiv, {0-2$$depth of fields,|8k wallpaper,|highly detailed,|pov,} {0-1$$summer, |winter, |spring, |autumn, } beautiful face { |, from below|, from above|, from side|, from behind|, from back} --n nsfw, bad face, lowres, low quality, worst quality, low effort, watermark, signature, ugly, poorly drawn" +``` + +This is a setting for VRAM 24GB. Adjust `--batch_size` and `--vae_batch_size` according to the VRAM size. + +The images are generated randomly using wildcards in `--prompt`. Adjust as necessary. + +### Processing images + +Use an external program to process the generated images. Save the processed images in an arbitrary directory. These will be the conditioning images. + +For example, you can use the following script to process the images with Canny. + +```python +import glob +import os +import random +import cv2 +import numpy as np + +IMAGES_DIR = "path/to/generated/images" +CANNY_DIR = "path/to/canny/images" + +os.makedirs(CANNY_DIR, exist_ok=True) +img_files = glob.glob(IMAGES_DIR + "/*.png") +for img_file in img_files: + can_file = CANNY_DIR + "/" + os.path.basename(img_file) + if os.path.exists(can_file): + print("Skip: " + img_file) + continue + + print(img_file) + + img = cv2.imread(img_file) + + # random threshold + # while True: + # threshold1 = random.randint(0, 127) + # threshold2 = random.randint(128, 255) + # if threshold2 - threshold1 > 80: + # break + + # fixed threshold + threshold1 = 100 + threshold2 = 200 + + img = cv2.Canny(img, threshold1, threshold2) + + cv2.imwrite(can_file, img) +``` + +### Creating caption files + +Create a caption file for each image with the same basename as the training image. It is fine to use the same caption as the one used when generating the image. + +If you generated the images with `sdxl_gen_img.py`, you can use the following script to create the caption files (`*.txt`) from the metadata in the generated images. + +```python +import glob +import os +from PIL import Image + +IMAGES_DIR = "path/to/generated/images" + +img_files = glob.glob(IMAGES_DIR + "/*.png") +for img_file in img_files: + cap_file = img_file.replace(".png", ".txt") + if os.path.exists(cap_file): + print(f"Skip: {img_file}") + continue + print(img_file) + + img = Image.open(img_file) + prompt = img.text["prompt"] if "prompt" in img.text else "" + if prompt == "": + print(f"Prompt not found in {img_file}") + + with open(cap_file, "w") as f: + f.write(prompt + "\n") +``` + +### Creating a dataset configuration file + +You can use the command line arguments of `sdxl_train_control_net_lllite.py` to specify the conditioning image directory. However, if you want to use a `.toml` file, specify the conditioning image directory in `conditioning_data_dir`. + +```toml +[general] +flip_aug = false +color_aug = false +resolution = [1024,1024] + +[[datasets]] +batch_size = 8 +enable_bucket = false + + [[datasets.subsets]] + image_dir = "path/to/generated/image/dir" + caption_extension = ".txt" + conditioning_data_dir = "path/to/canny/image/dir" +``` + +## Credit + +I would like to thank lllyasviel, the author of ControlNet, furusu, who provided me with advice on implementation and helped me solve problems, and ddPn08, who implemented the ControlNet dataset. + +## Sample + +Canny +![kohya_ss_girl_standing_at_classroom_smiling_to_the_viewer_class_78976b3e-0d4d-4ea0-b8e3-053ae493abbc](https://github.com/kohya-ss/sd-scripts/assets/52813779/37e9a736-649b-4c0f-ab26-880a1bf319b5) + +![im_20230820104253_000_1](https://github.com/kohya-ss/sd-scripts/assets/52813779/c8896900-ab86-4120-932f-6e2ae17b77c0) + +![im_20230820104302_000_1](https://github.com/kohya-ss/sd-scripts/assets/52813779/b12457a0-ee3c-450e-ba9a-b712d0fe86bb) + +![im_20230820104310_000_1](https://github.com/kohya-ss/sd-scripts/assets/52813779/8845b8d9-804a-44ac-9618-113a28eac8a1) diff --git a/train_network_README-ja.md b/train_network_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..55c80c4b0ae3676b047484190b5bbc461c0b8df2 --- /dev/null +++ b/train_network_README-ja.md @@ -0,0 +1,491 @@ +# LoRAの学習について + +[LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685)（arxiv）、[LoRA](https://github.com/microsoft/LoRA)（github）をStable Diffusionに適用したものです。 + +[cloneofsimo氏のリポジトリ](https://github.com/cloneofsimo/lora)を大いに参考にさせていただきました。ありがとうございます。 + +通常のLoRAは Linear およぴカーネルサイズ 1x1 の Conv2d にのみ適用されますが、カーネルサイズ 3x3 のConv2dに適用を拡大することもできます。 + +Conv2d 3x3への拡大は [cloneofsimo氏](https://github.com/cloneofsimo/lora) が最初にリリースし、KohakuBlueleaf氏が [LoCon](https://github.com/KohakuBlueleaf/LoCon) でその有効性を明らかにしたものです。KohakuBlueleaf氏に深く感謝します。 + +8GB VRAMでもぎりぎり動作するようです。 + +[学習についての共通ドキュメント](./train_README-ja.md) もあわせてご覧ください。 + +# 学習できるLoRAの種類 + +以下の二種類をサポートします。以下は当リポジトリ内の独自の名称です。 + +1. __LoRA-LierLa__ : (LoRA for __Li__ n __e__ a __r__ __La__ yers、リエラと読みます) + + Linear およびカーネルサイズ 1x1 の Conv2d に適用されるLoRA + +2. __LoRA-C3Lier__ : (LoRA for __C__ olutional layers with __3__ x3 Kernel and __Li__ n __e__ a __r__ layers、セリアと読みます) + + 1.に加え、カーネルサイズ 3x3 の Conv2d に適用されるLoRA + +LoRA-LierLaに比べ、LoRA-C3Liarは適用される層が増える分、高い精度が期待できるかもしれません。 + +また学習時は __DyLoRA__ を使用することもできます（後述します）。 + +## 学習したモデルに関する注意 + +LoRA-LierLa は、AUTOMATIC1111氏のWeb UIのLoRA機能で使用することができます。 + +LoRA-C3Liarを使いWeb UIで生成するには、こちらの[WebUI用extension](https://github.com/kohya-ss/sd-webui-additional-networks)を使ってください。 + +いずれも学習したLoRAのモデルを、Stable Diffusionのモデルにこのリポジトリ内のスクリプトであらかじめマージすることもできます。 + +cloneofsimo氏のリポジトリ、およびd8ahazard氏の[Dreambooth Extension for Stable-Diffusion-WebUI](https://github.com/d8ahazard/sd_dreambooth_extension)とは、現時点では互換性がありません。いくつかの機能拡張を行っているためです（後述）。 + +# 学習の手順 + +あらかじめこのリポジトリのREADMEを参照し、環境整備を行ってください。 + +## データの準備 + +[学習データの準備について](./train_README-ja.md) を参照してください。 + + +## 学習の実行 + +`train_network.py`を用います。 + +`train_network.py`では `--network_module` オプションに、学習対象のモジュール名を指定します。LoRAに対応するのは`network.lora`となりますので、それを指定してください。 + +なお学習率は通常のDreamBoothやfine tuningよりも高めの、`1e-4`～`1e-3`程度を指定するとよいようです。 + +以下はコマンドラインの例です。 + +``` +accelerate launch --num_cpu_threads_per_process 1 train_network.py + --pretrained_model_name_or_path=<.ckptまたは.safetensordまたはDiffusers版モデルのディレクトリ> + --dataset_config=<データ準備で作成した.tomlファイル> + --output_dir=<学習したモデルの出力先フォルダ> + --output_name=<学習したモデル出力時のファイル名> + --save_model_as=safetensors + --prior_loss_weight=1.0 + --max_train_steps=400 + --learning_rate=1e-4 + --optimizer_type="AdamW8bit" + --xformers + --mixed_precision="fp16" + --cache_latents + --gradient_checkpointing + --save_every_n_epochs=1 + --network_module=networks.lora +``` + +このコマンドラインでは LoRA-LierLa が学習されます。 + +`--output_dir` オプションで指定したフォルダに、LoRAのモデルが保存されます。他のオプション、オプティマイザ等については [学習の共通ドキュメント](./train_README-ja.md) の「よく使われるオプション」も参照してください。 + +その他、以下のオプションが指定できます。 + +* `--network_dim` + * LoRAのRANKを指定します（``--networkdim=4``など）。省略時は4になります。数が多いほど表現力は増しますが、学習に必要なメモリ、時間は増えます。また闇雲に増やしても良くないようです。 +* `--network_alpha` + * アンダーフローを防ぎ安定して学習するための ``alpha`` 値を指定します。デフォルトは1です。``network_dim``と同じ値を指定すると以前のバージョンと同じ動作になります。 +* `--persistent_data_loader_workers` + * Windows環境で指定するとエポック間の待ち時間が大幅に短縮されます。 +* `--max_data_loader_n_workers` + * データ読み込みのプロセス数を指定します。プロセス数が多いとデータ読み込みが速くなりGPUを効率的に利用できますが、メインメモリを消費します。デフォルトは「`8` または `CPU同時実行スレッド数-1` の小さいほう」なので、メインメモリに余裕がない場合や、GPU使用率が90%程度以上なら、それらの数値を見ながら `2` または `1` 程度まで下げてください。 +* `--network_weights` + * 学習前に学習済みのLoRAの重みを読み込み、そこから追加で学習します。 +* `--network_train_unet_only` + * U-Netに関連するLoRAモジュールのみ有効とします。fine tuning的な学習で指定するとよいかもしれません。 +* `--network_train_text_encoder_only` + * Text Encoderに関連するLoRAモジュールのみ有効とします。Textual Inversion的な効果が期待できるかもしれません。 +* `--unet_lr` + * U-Netに関連するLoRAモジュールに、通常の学習率（--learning_rateオプションで指定）とは異なる学習率を使う時に指定します。 +* `--text_encoder_lr` + * Text Encoderに関連するLoRAモジュールに、通常の学習率（--learning_rateオプションで指定）とは異なる学習率を使う時に指定します。Text Encoderのほうを若干低めの学習率（5e-5など）にしたほうが良い、という話もあるようです。 +* `--network_args` + * 複数の引数を指定できます。後述します。 +* `--alpha_mask` + * 画像のアルファ値をマスクとして使用します。透過画像を学習する際に使用します。[PR #1223](https://github.com/kohya-ss/sd-scripts/pull/1223) + +`--network_train_unet_only` と `--network_train_text_encoder_only` の両方とも未指定時（デフォルト）はText EncoderとU-Netの両方のLoRAモジュールを有効にします。 + +# その他の学習方法 + +## LoRA-C3Lier を学習する + +`--network_args` に以下のように指定してください。`conv_dim` で Conv2d (3x3) の rank を、`conv_alpha` で alpha を指定してください。 + +``` +--network_args "conv_dim=4" "conv_alpha=1" +``` + +以下のように alpha 省略時は1になります。 + +``` +--network_args "conv_dim=4" +``` + +## DyLoRA + +DyLoRAはこちらの論文で提案されたものです。[DyLoRA: Parameter Efficient Tuning of Pre-trained Models using Dynamic Search-Free Low-Rank Adaptation](https://arxiv.org/abs/2210.07558)　公式実装は[こちら](https://github.com/huawei-noah/KD-NLP/tree/main/DyLoRA)です。 + +論文によると、LoRAのrankは必ずしも高いほうが良いわけではなく、対象のモデル、データセット、タスクなどにより適切なrankを探す必要があるようです。DyLoRAを使うと、指定したdim(rank)以下のさまざまなrankで同時にLoRAを学習します。これにより最適なrankをそれぞれ学習して探す手間を省くことができます。 + +当リポジトリの実装は公式実装をベースに独自の拡張を加えています（そのため不具合などあるかもしれません）。 + +### 当リポジトリのDyLoRAの特徴 + +学習後のDyLoRAのモデルファイルはLoRAと互換性があります。また、モデルファイルから指定したdim(rank)以下の複数のdimのLoRAを抽出できます。 + +DyLoRA-LierLa、DyLoRA-C3Lierのどちらも学習できます。 + +### DyLoRAで学習する + +`--network_module=networks.dylora` のように、DyLoRAに対応する`network.dylora`を指定してください。 + +また `--network_args` に、たとえば`--network_args "unit=4"`のように`unit`を指定します。`unit`はrankを分割する単位です。たとえば`--network_dim=16 --network_args "unit=4"` のように指定します。`unit`は`network_dim`を割り切れる値（`network_dim`は`unit`の倍数）としてください。 + +`unit`を指定しない場合は、`unit=1`として扱われます。 + +記述例は以下です。 + +``` +--network_module=networks.dylora --network_dim=16 --network_args "unit=4" + +--network_module=networks.dylora --network_dim=32 --network_alpha=16 --network_args "unit=4" +``` + +DyLoRA-C3Lierの場合は、`--network_args` に`"conv_dim=4"`のように`conv_dim`を指定します。通常のLoRAと異なり、`conv_dim`は`network_dim`と同じ値である必要があります。記述例は以下です。 + +``` +--network_module=networks.dylora --network_dim=16 --network_args "conv_dim=16" "unit=4" + +--network_module=networks.dylora --network_dim=32 --network_alpha=16 --network_args "conv_dim=32" "conv_alpha=16" "unit=8" +``` + +たとえばdim=16、unit=4（後述）で学習すると、4、8、12、16の4つのrankのLoRAを学習、抽出できます。抽出した各モデルで画像を生成し、比較することで、最適なrankのLoRAを選択できます。 + +その他のオプションは通常のLoRAと同じです。 + +※ `unit`は当リポジトリの独自拡張で、DyLoRAでは同dim(rank)の通常LoRAに比べると学習時間が長くなることが予想されるため、分割単位を大きくしたものです。 + +### DyLoRAのモデルからLoRAモデルを抽出する + +`networks`フォルダ内の `extract_lora_from_dylora.py`を使用します。指定した`unit`単位で、DyLoRAのモデルからLoRAのモデルを抽出します。 + +コマンドラインはたとえば以下のようになります。 + +```powershell +python networks\extract_lora_from_dylora.py --model "foldername/dylora-model.safetensors" --save_to "foldername/dylora-model-split.safetensors" --unit 4 +``` + +`--model` にはDyLoRAのモデルファイルを指定します。`--save_to` には抽出したモデルを保存するファイル名を指定します（rankの数値がファイル名に付加されます）。`--unit` にはDyLoRAの学習時の`unit`を指定します。 + +## 階層別学習率 + +詳細は[PR #355](https://github.com/kohya-ss/sd-scripts/pull/355) をご覧ください。 + +フルモデルの25個のブロックの重みを指定できます。最初のブロックに該当するLoRAは存在しませんが、階層別LoRA適用等との互換性のために25個としています。またconv2d3x3に拡張しない場合も一部のブロックにはLoRAが存在しませんが、記述を統一するため常に25個の値を指定してください。 + +SDXL では down/up 9 個、middle 3 個の値を指定してください。 + +`--network_args` で以下の引数を指定してください。 + +- `down_lr_weight` : U-Netのdown blocksの学習率の重みを指定します。以下が指定可能です。 + - ブロックごとの重み : `"down_lr_weight=0,0,0,0,0,0,1,1,1,1,1,1"` のように12個（SDXL では 9 個）の数値を指定します。 + - プリセットからの指定 : `"down_lr_weight=sine"` のように指定します（サインカーブで重みを指定します）。sine, cosine, linear, reverse_linear, zeros が指定可能です。また `"down_lr_weight=cosine+.25"` のように `+数値` を追加すると、指定した数値を加算します（0.25~1.25になります）。 +- `mid_lr_weight` : U-Netのmid blockの学習率の重みを指定します。`"down_lr_weight=0.5"` のように数値を一つだけ指定します（SDXL の場合は 3 個）。 +- `up_lr_weight` : U-Netのup blocksの学習率の重みを指定します。down_lr_weightと同様です。 +- 指定を省略した部分は1.0として扱われます。また重みを0にするとそのブロックのLoRAモジュールは作成されません。 +- `block_lr_zero_threshold` : 重みがこの値以下の場合、LoRAモジュールを作成しません。デフォルトは0です。 + +### 階層別学習率コマンドライン指定例: + +```powershell +--network_args "down_lr_weight=0.5,0.5,0.5,0.5,1.0,1.0,1.0,1.0,1.5,1.5,1.5,1.5" "mid_lr_weight=2.0" "up_lr_weight=1.5,1.5,1.5,1.5,1.0,1.0,1.0,1.0,0.5,0.5,0.5,0.5" + +--network_args "block_lr_zero_threshold=0.1" "down_lr_weight=sine+.5" "mid_lr_weight=1.5" "up_lr_weight=cosine+.5" +``` + +### 階層別学習率tomlファイル指定例: + +```toml +network_args = [ "down_lr_weight=0.5,0.5,0.5,0.5,1.0,1.0,1.0,1.0,1.5,1.5,1.5,1.5", "mid_lr_weight=2.0", "up_lr_weight=1.5,1.5,1.5,1.5,1.0,1.0,1.0,1.0,0.5,0.5,0.5,0.5",] + +network_args = [ "block_lr_zero_threshold=0.1", "down_lr_weight=sine+.5", "mid_lr_weight=1.5", "up_lr_weight=cosine+.5", ] +``` + +## 階層別dim (rank) + +フルモデルの25個のブロックのdim (rank)を指定できます。階層別学習率と同様に一部のブロックにはLoRAが存在しない場合がありますが、常に25個の値を指定してください。 + +SDXL では 23 個の値を指定してください。一部のブロックにはLoRA が存在しませんが、`sdxl_train.py` の[階層別学習率](./train_SDXL-en.md) との互換性のためです。 +対応は、`0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out` です。 + +`--network_args` で以下の引数を指定してください。 + +- `block_dims` : 各ブロックのdim (rank)を指定します。`"block_dims=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2"` のように25個の数値を指定します。 +- `block_alphas` : 各ブロックのalphaを指定します。block_dimsと同様に25個の数値を指定します。省略時はnetwork_alphaの値が使用されます。 +- `conv_block_dims` : LoRAをConv2d 3x3に拡張し、各ブロックのdim (rank)を指定します。 +- `conv_block_alphas` : LoRAをConv2d 3x3に拡張したときの各ブロックのalphaを指定します。省略時はconv_alphaの値が使用されます。 + +### 階層別dim (rank)コマンドライン指定例: + +```powershell +--network_args "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2" + +--network_args "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2" "conv_block_dims=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2" + +--network_args "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2" "block_alphas=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2" +``` + +### 階層別dim (rank)tomlファイル指定例: + +```toml +network_args = [ "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2",] + +network_args = [ "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2", "block_alphas=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2",] +``` + +# その他のスクリプト + +マージ等LoRAに関連するスクリプト群です。 + +## マージスクリプトについて + +merge_lora.pyでStable DiffusionのモデルにLoRAの学習結果をマージしたり、複数のLoRAモデルをマージしたりできます。 + +SDXL向けにはsdxl_merge_lora.pyを用意しています。オプション等は同一ですので、以下のmerge_lora.pyを読み替えてください。 + +### Stable DiffusionのモデルにLoRAのモデルをマージする + +マージ後のモデルは通常のStable Diffusionのckptと同様に扱えます。たとえば以下のようなコマンドラインになります。 + +``` +python networks\merge_lora.py --sd_model ..\model\model.ckpt + --save_to ..\lora_train1\model-char1-merged.safetensors + --models ..\lora_train1\last.safetensors --ratios 0.8 +``` + +Stable Diffusion v2.xのモデルで学習し、それにマージする場合は、--v2オプションを指定してください。 + +--sd_modelオプションにマージの元となるStable Diffusionのモデルファイルを指定します（.ckptまたは.safetensorsのみ対応で、Diffusersは今のところ対応していません）。 + +--save_toオプションにマージ後のモデルの保存先を指定します（.ckptまたは.safetensors、拡張子で自動判定）。 + +--modelsに学習したLoRAのモデルファイルを指定します。複数指定も可能で、その時は順にマージします。 + +--ratiosにそれぞれのモデルの適用率（どのくらい重みを元モデルに反映するか）を0~1.0の数値で指定します。例えば過学習に近いような場合は、適用率を下げるとマシになるかもしれません。モデルの数と同じだけ指定してください。 + +複数指定時は以下のようになります。 + +``` +python networks\merge_lora.py --sd_model ..\model\model.ckpt + --save_to ..\lora_train1\model-char1-merged.safetensors + --models ..\lora_train1\last.safetensors ..\lora_train2\last.safetensors --ratios 0.8 0.5 +``` + +### 複数のLoRAのモデルをマージする + +--concatオプションを指定すると、複数のLoRAを単純に結合して新しいLoRAモデルを作成できます。ファイルサイズ（およびdim/rank）は指定したLoRAの合計サイズになります（マージ時にdim (rank)を変更する場合は `svd_merge_lora.py` を使用してください）。 + +たとえば以下のようなコマンドラインになります。 + +``` +python networks\merge_lora.py --save_precision bf16 + --save_to ..\lora_train1\model-char1-style1-merged.safetensors + --models ..\lora_train1\last.safetensors ..\lora_train2\last.safetensors + --ratios 1.0 -1.0 --concat --shuffle +``` + +--concatオプションを指定します。 + +また--shuffleオプションを追加し、重みをシャッフルします。シャッフルしないとマージ後のLoRAから元のLoRAを取り出せるため、コピー機学習などの場合には学習元データが明らかになります。ご注意ください。 + +--save_toオプションにマージ後のLoRAモデルの保存先を指定します（.ckptまたは.safetensors、拡張子で自動判定）。 + +--modelsに学習したLoRAのモデルファイルを指定します。三つ以上も指定可能です。 + +--ratiosにそれぞれのモデルの比率（どのくらい重みを元モデルに反映するか）を0~1.0の数値で指定します。二つのモデルを一対一でマージする場合は、「0.5 0.5」になります。「1.0 1.0」では合計の重みが大きくなりすぎて、恐らく結果はあまり望ましくないものになると思われます。 + +v1で学習したLoRAとv2で学習したLoRA、rank（次元数）の異なるLoRAはマージできません。U-NetだけのLoRAとU-Net+Text EncoderのLoRAはマージできるはずですが、結果は未知数です。 + +### その他のオプション + +* precision + * マージ計算時の精度をfloat、fp16、bf16から指定できます。省略時は精度を確保するためfloatになります。メモリ使用量を減らしたい場合はfp16/bf16を指定してください。 +* save_precision + * モデル保存時の精度をfloat、fp16、bf16から指定できます。省略時はprecisionと同じ精度になります。 + +他にもいくつかのオプションがありますので、--helpで確認してください。 + +## 複数のrankが異なるLoRAのモデルをマージする + +複数のLoRAをひとつのLoRAで近似します（完全な再現はできません）。`svd_merge_lora.py`を用います。たとえば以下のようなコマンドラインになります。 + +``` +python networks\svd_merge_lora.py + --save_to ..\lora_train1\model-char1-style1-merged.safetensors + --models ..\lora_train1\last.safetensors ..\lora_train2\last.safetensors + --ratios 0.6 0.4 --new_rank 32 --device cuda +``` + +`merge_lora.py` と主なオプションは同一です。以下のオプションが追加されています。 + +- `--new_rank` + - 作成するLoRAのrankを指定します。 +- `--new_conv_rank` + - 作成する Conv2d 3x3 LoRA の rank を指定します。省略時は `new_rank` と同じになります。 +- `--device` + - `--device cuda`としてcudaを指定すると計算をGPU上で行います。処理が速くなります。 + +## 当リポジトリ内の画像生成スクリプトで生成する + +gen_img_diffusers.pyに、--network_module、--network_weightsの各オプションを追加してください。意味は学習時と同様です。 + +--network_mulオプションで0~1.0の数値を指定すると、LoRAの適用率を変えられます。 + +## Diffusersのpipelineで生成する + +以下の例を参考にしてください。必要なファイルはnetworks/lora.pyのみです。Diffusersのバージョンは0.10.2以外では動作しない可能性があります。 + +```python +import torch +from diffusers import StableDiffusionPipeline +from networks.lora import LoRAModule, create_network_from_weights +from safetensors.torch import load_file + +# if the ckpt is CompVis based, convert it to Diffusers beforehand with tools/convert_diffusers20_original_sd.py. See --help for more details. + +model_id_or_dir = r"model_id_on_hugging_face_or_dir" +device = "cuda" + +# create pipe +print(f"creating pipe from {model_id_or_dir}...") +pipe = StableDiffusionPipeline.from_pretrained(model_id_or_dir, revision="fp16", torch_dtype=torch.float16) +pipe = pipe.to(device) +vae = pipe.vae +text_encoder = pipe.text_encoder +unet = pipe.unet + +# load lora networks +print(f"loading lora networks...") + +lora_path1 = r"lora1.safetensors" +sd = load_file(lora_path1) # If the file is .ckpt, use torch.load instead. +network1, sd = create_network_from_weights(0.5, None, vae, text_encoder,unet, sd) +network1.apply_to(text_encoder, unet) +network1.load_state_dict(sd) +network1.to(device, dtype=torch.float16) + +# # You can merge weights instead of apply_to+load_state_dict. network.set_multiplier does not work +# network.merge_to(text_encoder, unet, sd) + +lora_path2 = r"lora2.safetensors" +sd = load_file(lora_path2) +network2, sd = create_network_from_weights(0.7, None, vae, text_encoder,unet, sd) +network2.apply_to(text_encoder, unet) +network2.load_state_dict(sd) +network2.to(device, dtype=torch.float16) + +lora_path3 = r"lora3.safetensors" +sd = load_file(lora_path3) +network3, sd = create_network_from_weights(0.5, None, vae, text_encoder,unet, sd) +network3.apply_to(text_encoder, unet) +network3.load_state_dict(sd) +network3.to(device, dtype=torch.float16) + +# prompts +prompt = "masterpiece, best quality, 1girl, in white shirt, looking at viewer" +negative_prompt = "bad quality, worst quality, bad anatomy, bad hands" + +# exec pipe +print("generating image...") +with torch.autocast("cuda"): + image = pipe(prompt, guidance_scale=7.5, negative_prompt=negative_prompt).images[0] + +# if not merged, you can use set_multiplier +# network1.set_multiplier(0.8) +# and generate image again... + +# save image +image.save(r"by_diffusers..png") +``` + +## 二つのモデルの差分からLoRAモデルを作成する + +[こちらのディスカッション](https://github.com/cloneofsimo/lora/discussions/56)を参考に実装したものです。数式はそのまま使わせていただきました（よく理解していませんが近似には特異値分解を用いるようです）。 + +二つのモデル（たとえばfine tuningの元モデルとfine tuning後のモデル）の差分を、LoRAで近似します。 + +### スクリプトの実行方法 + +以下のように指定してください。 +``` +python networks\extract_lora_from_models.py --model_org base-model.ckpt + --model_tuned fine-tuned-model.ckpt + --save_to lora-weights.safetensors --dim 4 +``` + +--model_orgオプションに元のStable Diffusionモデルを指定します。作成したLoRAモデルを適用する場合は、このモデルを指定して適用することになります。.ckptまたは.safetensorsが指定できます。 + +--model_tunedオプションに差分を抽出する対象のStable Diffusionモデルを指定します。たとえばfine tuningやDreamBooth後のモデルを指定します。.ckptまたは.safetensorsが指定できます。 + +--save_toにLoRAモデルの保存先を指定します。--dimにLoRAの次元数を指定します。 + +生成されたLoRAモデルは、学習したLoRAモデルと同様に使用できます。 + +Text Encoderが二つのモデルで同じ場合にはLoRAはU-NetのみのLoRAとなります。 + +### その他のオプション + +- `--v2` + - v2.xのStable Diffusionモデルを使う場合に指定してください。 +- `--device` + - ``--device cuda``としてcudaを指定すると計算をGPU上で行います。処理が速くなります（CPUでもそこまで遅くないため、せいぜい倍～数倍程度のようです）。 +- `--save_precision` + - LoRAの保存形式を"float", "fp16", "bf16"から指定します。省略時はfloatになります。 +- `--conv_dim` + - 指定するとLoRAの適用範囲を Conv2d 3x3 へ拡大します。Conv2d 3x3 の rank を指定します。 + +## 画像リサイズスクリプト + +（のちほどドキュメントを整理しますがとりあえずここに説明を書いておきます。） + +Aspect Ratio Bucketingの機能拡張で、小さな画像については拡大しないでそのまま教師データとすることが可能になりました。元の教師画像を縮小した画像を、教師データに加えると精度が向上したという報告とともに前処理用のスクリプトをいただきましたので整備して追加しました。bmaltais氏に感謝します。 + +### スクリプトの実行方法 + +以下のように指定してください。元の画像そのまま、およびリサイズ後の画像が変換先フォルダに保存されます。リサイズ後の画像には、ファイル名に ``+512x512`` のようにリサイズ先の解像度が付け加えられます（画像サイズとは異なります）。リサイズ先の解像度より小さい画像は拡大されることはありません。 + +``` +python tools\resize_images_to_resolution.py --max_resolution 512x512,384x384,256x256 --save_as_png + --copy_associated_files 元画像フォルダ 変換先フォルダ +``` + +元画像フォルダ内の画像ファイルが、指定した解像度（複数指定可）と同じ面積になるようにリサイズされ、変換先フォルダに保存されます。画像以外のファイルはそのままコピーされます。 + +``--max_resolution`` オプションにリサイズ先のサイズを例のように指定してください。面積がそのサイズになるようにリサイズします。複数指定すると、それぞれの解像度でリサイズされます。``512x512,384x384,256x256``なら、変換先フォルダの画像は、元サイズとリサイズ後サイズ×3の計4枚になります。 + +``--save_as_png`` オプションを指定するとpng形式で保存します。省略するとjpeg形式（quality=100）で保存されます。 + +``--copy_associated_files`` オプションを指定すると、拡張子を除き画像と同じファイル名（たとえばキャプションなど）のファイルが、リサイズ後の画像のファイル名と同じ名前でコピーされます。 + + +### その他のオプション + +- divisible_by + - リサイズ後の画像のサイズ（縦、横のそれぞれ）がこの値で割り切れるように、画像中心を切り出します。 +- interpolation + - 縮小時の補完方法を指定します。``area, cubic, lanczos4``から選択可能で、デフォルトは``area``です。 + + +# 追加情報 + +## cloneofsimo氏のリポジトリとの違い + +2022/12/25時点では、当リポジトリはLoRAの適用個所をText EncoderのMLP、U-NetのFFN、Transformerのin/out projectionに拡大し、表現力が増しています。ただその代わりメモリ使用量は増え、8GBぎりぎりになりました。 + +またモジュール入れ替え機構は全く異なります。 + +## 将来拡張について + +LoRAだけでなく他の拡張にも対応可能ですので、それらも追加予定です。 diff --git a/train_network_README-zh.md b/train_network_README-zh.md new file mode 100644 index 0000000000000000000000000000000000000000..830014f72d6a1ecb0a8206a18cded308c42d6363 --- /dev/null +++ b/train_network_README-zh.md @@ -0,0 +1,468 @@ +# 关于LoRA的学习。 + +[LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685)（arxiv）、[LoRA](https://github.com/microsoft/LoRA)（github）这是应用于Stable Diffusion“稳定扩散”的内容。 + +[cloneofsimo先生的代码仓库](https://github.com/cloneofsimo/lora) 我们非常感謝您提供的参考。非常感謝。 + +通常情況下，LoRA只适用于Linear和Kernel大小为1x1的Conv2d，但也可以將其擴展到Kernel大小为3x3的Conv2d。 + +Conv2d 3x3的扩展最初是由 [cloneofsimo先生的代码仓库](https://github.com/cloneofsimo/lora) +而KohakuBlueleaf先生在[LoCon](https://github.com/KohakuBlueleaf/LoCon)中揭示了其有效性。我们深深地感谢KohakuBlueleaf先生。 + +看起来即使在8GB VRAM上也可以勉强运行。 + +请同时查看关于[学习的通用文档](./train_README-zh.md)。 +# 可学习的LoRA 类型 + +支持以下两种类型。以下是本仓库中自定义的名称。 + +1. __LoRA-LierLa__：(用于 __Li__ n __e__ a __r__ __La__ yers 的 LoRA，读作 "Liela") + + 适用于 Linear 和卷积层 Conv2d 的 1x1 Kernel 的 LoRA + +2. __LoRA-C3Lier__：(用于具有 3x3 Kernel 的卷积层和 __Li__ n __e__ a __r__ 层的 LoRA，读作 "Seria") + + 除了第一种类型外，还适用于 3x3 Kernel 的 Conv2d 的 LoRA + +与 LoRA-LierLa 相比，LoRA-C3Lier 可能会获得更高的准确性，因为它适用于更多的层。 + +在训练时，也可以使用 __DyLoRA__（将在后面介绍）。 + +## 请注意与所学模型相关的事项。 + +LoRA-LierLa可以用于AUTOMATIC1111先生的Web UI LoRA功能。 + +要使用LoRA-C3Liar并在Web UI中生成，请使用此处的[WebUI用extension](https://github.com/kohya-ss/sd-webui-additional-networks)。 + +在此存储库的脚本中，您还可以预先将经过训练的LoRA模型合并到Stable Diffusion模型中。 + +请注意，与cloneofsimo先生的存储库以及d8ahazard先生的[Stable-Diffusion-WebUI的Dreambooth扩展](https://github.com/d8ahazard/sd_dreambooth_extension)不兼容，因为它们进行了一些功能扩展（如下文所述）。 + +# 学习步骤 + +请先参考此存储库的README文件并进行环境设置。 + +## 准备数据 + +请参考 [关于准备学习数据](./train_README-zh.md)。 + +## 网络训练 + +使用`train_network.py`。 + +在`train_network.py`中，使用`--network_module`选项指定要训练的模块名称。对于LoRA模块，它应该是`network.lora`，请指定它。 + +请注意，学习率应该比通常的DreamBooth或fine tuning要高，建议指定为`1e-4`至`1e-3`左右。 + +以下是命令行示例。 + +``` +accelerate launch --num_cpu_threads_per_process 1 train_network.py + --pretrained_model_name_or_path=<.ckpt或.safetensord或Diffusers版模型目录> + --dataset_config=<数据集配置的.toml文件> + --output_dir=<训练过程中的模型输出文件夹> + --output_name=<训练模型输出时的文件名> + --save_model_as=safetensors + --prior_loss_weight=1.0 + --max_train_steps=400 + --learning_rate=1e-4 + --optimizer_type="AdamW8bit" + --xformers + --mixed_precision="fp16" + --cache_latents + --gradient_checkpointing + --save_every_n_epochs=1 + --network_module=networks.lora +``` + +在这个命令行中，LoRA-LierLa将会被训练。 + +LoRA的模型将会被保存在通过`--output_dir`选项指定的文件夹中。关于其他选项和优化器等，请参阅[学习的通用文档](./train_README-zh.md)中的“常用选项”。 + +此外，还可以指定以下选项： + +* `--network_dim` + * 指定LoRA的RANK（例如：`--network_dim=4`）。默认值为4。数值越大表示表现力越强，但需要更多的内存和时间来训练。而且不要盲目增加此数值。 +* `--network_alpha` + * 指定用于防止下溢并稳定训练的alpha值。默认值为1。如果与`network_dim`指定相同的值，则将获得与以前版本相同的行为。 +* `--persistent_data_loader_workers` + * 在Windows环境中指定可大幅缩短epoch之间的等待时间。 +* `--max_data_loader_n_workers` + * 指定数据读取进程的数量。进程数越多，数据读取速度越快，可以更有效地利用GPU，但会占用主存。默认值为“`8`或`CPU同步执行线程数-1`的最小值”，因此如果主存不足或GPU使用率超过90％，则应将这些数字降低到约`2`或`1`。 +* `--network_weights` + * 在训练之前读取预训练的LoRA权重，并在此基础上进行进一步的训练。 +* `--network_train_unet_only` + * 仅启用与U-Net相关的LoRA模块。在类似fine tuning的学习中指定此选项可能会很有用。 +* `--network_train_text_encoder_only` + * 仅启用与Text Encoder相关的LoRA模块。可能会期望Textual Inversion效果。 +* `--unet_lr` + * 当在U-Net相关的LoRA模块中使用与常规学习率（由`--learning_rate`选项指定）不同的学习率时，应指定此选项。 +* `--text_encoder_lr` + * 当在Text Encoder相关的LoRA模块中使用与常规学习率（由`--learning_rate`选项指定）不同的学习率时，应指定此选项。可能最好将Text Encoder的学习率稍微降低（例如5e-5）。 +* `--network_args` + * 可以指定多个参数。将在下面详细说明。 +* `--alpha_mask` + * 使用图像的 Alpha 值作为遮罩。这在学习透明图像时使用。[PR #1223](https://github.com/kohya-ss/sd-scripts/pull/1223) + +当未指定`--network_train_unet_only`和`--network_train_text_encoder_only`时（默认情况），将启用Text Encoder和U-Net的两个LoRA模块。 + +# 其他的学习方法 + +## 学习 LoRA-C3Lier + +请使用以下方式 + +``` +--network_args "conv_dim=4" +``` + +DyLoRA是在这篇论文中提出的[DyLoRA: Parameter Efficient Tuning of Pre-trained Models using Dynamic Search-Free Low-Rank Adaptation](​https://arxiv.org/abs/2210.07558)， +[其官方实现可在这里找到](​https://github.com/huawei-noah/KD-NLP/tree/main/DyLoRA)。 + +根据论文，LoRA的rank并不是越高越好，而是需要根据模型、数据集、任务等因素来寻找合适的rank。使用DyLoRA，可以同时在指定的维度(rank)下学习多种rank的LoRA，从而省去了寻找最佳rank的麻烦。 + +本存储库的实现基于官方实现进行了自定义扩展（因此可能存在缺陷）。 + +### 本存储库DyLoRA的特点 + +DyLoRA训练后的模型文件与LoRA兼容。此外，可以从模型文件中提取多个低于指定维度(rank)的LoRA。 + +DyLoRA-LierLa和DyLoRA-C3Lier均可训练。 + +### 使用DyLoRA进行训练 + +请指定与DyLoRA相对应的`network.dylora`，例如 `--network_module=networks.dylora`。 + +此外，通过 `--network_args` 指定例如`--network_args "unit=4"`的参数。`unit`是划分rank的单位。例如，可以指定为`--network_dim=16 --network_args "unit=4"`。请将`unit`视为可以被`network_dim`整除的值（`network_dim`是`unit`的倍数）。 + +如果未指定`unit`，则默认为`unit=1`。 + +以下是示例说明。 + +``` +--network_module=networks.dylora --network_dim=16 --network_args "unit=4" + +--network_module=networks.dylora --network_dim=32 --network_alpha=16 --network_args "unit=4" +``` + +对于DyLoRA-C3Lier，需要在 `--network_args` 中指定 `conv_dim`，例如 `conv_dim=4`。与普通的LoRA不同，`conv_dim`必须与`network_dim`具有相同的值。以下是一个示例描述： + +``` +--network_module=networks.dylora --network_dim=16 --network_args "conv_dim=16" "unit=4" + +--network_module=networks.dylora --network_dim=32 --network_alpha=16 --network_args "conv_dim=32" "conv_alpha=16" "unit=8" +``` + +例如，当使用dim=16、unit=4（如下所述）进行学习时，可以学习和提取4个rank的LoRA，即4、8、12和16。通过在每个提取的模型中生成图像并进行比较，可以选择最佳rank的LoRA。 + +其他选项与普通的LoRA相同。 + +*`unit`是本存储库的独有扩展，在DyLoRA中，由于预计相比同维度（rank）的普通LoRA，学习时间更长，因此将分割单位增加。 + +### 从DyLoRA模型中提取LoRA模型 + +请使用`networks`文件夹中的`extract_lora_from_dylora.py`。指定`unit`单位后，从DyLoRA模型中提取LoRA模型。 + +例如，命令行如下： + +```powershell +python networks\extract_lora_from_dylora.py --model "foldername/dylora-model.safetensors" --save_to "foldername/dylora-model-split.safetensors" --unit 4 +``` + +`--model` 参数用于指定DyLoRA模型文件。`--save_to` 参数用于指定要保存提取的模型的文件名（rank值将附加到文件名中）。`--unit` 参数用于指定DyLoRA训练时的`unit`。 + +## 分层学习率 + +请参阅PR＃355了解详细信息。 + +您可以指定完整模型的25个块的权重。虽然第一个块没有对应的LoRA，但为了与分层LoRA应用等的兼容性，将其设为25个。此外，如果不扩展到conv2d3x3，则某些块中可能不存在LoRA，但为了统一描述，请始终指定25个值。 + +请在 `--network_args` 中指定以下参数。 + +- `down_lr_weight`：指定U-Net down blocks的学习率权重。可以指定以下内容： + - 每个块的权重：指定12个数字，例如`"down_lr_weight=0,0,0,0,0,0,1,1,1,1,1,1"` + - 从预设中指定：例如`"down_lr_weight=sine"`（使用正弦曲线指定权重）。可以指定sine、cosine、linear、reverse_linear、zeros。另外，添加 `+数字` 时，可以将指定的数字加上（变为0.25〜1.25）。 +- `mid_lr_weight`：指定U-Net mid block的学习率权重。只需指定一个数字，例如 `"mid_lr_weight=0.5"`。 +- `up_lr_weight`：指定U-Net up blocks的学习率权重。与down_lr_weight相同。 +- 省略指定的部分将被视为1.0。另外，如果将权重设为0，则不会创建该块的LoRA模块。 +- `block_lr_zero_threshold`：如果权重小于此值，则不会创建LoRA模块。默认值为0。 + +### 分层学习率命令行指定示例： + + +```powershell +--network_args "down_lr_weight=0.5,0.5,0.5,0.5,1.0,1.0,1.0,1.0,1.5,1.5,1.5,1.5" "mid_lr_weight=2.0" "up_lr_weight=1.5,1.5,1.5,1.5,1.0,1.0,1.0,1.0,0.5,0.5,0.5,0.5" + +--network_args "block_lr_zero_threshold=0.1" "down_lr_weight=sine+.5" "mid_lr_weight=1.5" "up_lr_weight=cosine+.5" +``` + +### Hierarchical Learning Rate指定的toml文件示例： + +```toml +network_args = [ "down_lr_weight=0.5,0.5,0.5,0.5,1.0,1.0,1.0,1.0,1.5,1.5,1.5,1.5", "mid_lr_weight=2.0", "up_lr_weight=1.5,1.5,1.5,1.5,1.0,1.0,1.0,1.0,0.5,0.5,0.5,0.5",] + +network_args = [ "block_lr_zero_threshold=0.1", "down_lr_weight=sine+.5", "mid_lr_weight=1.5", "up_lr_weight=cosine+.5", ] +``` + +## 层次结构维度（rank） + +您可以指定完整模型的25个块的维度（rank）。与分层学习率一样，某些块可能不存在LoRA，但请始终指定25个值。 + +请在 `--network_args` 中指定以下参数： + +- `block_dims`：指定每个块的维度（rank）。指定25个数字，例如 `"block_dims=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2"`。 +- `block_alphas`：指定每个块的alpha。与block_dims一样，指定25个数字。如果省略，将使用network_alpha的值。 +- `conv_block_dims`：将LoRA扩展到Conv2d 3x3，并指定每个块的维度（rank）。 +- `conv_block_alphas`：在将LoRA扩展到Conv2d 3x3时指定每个块的alpha。如果省略，将使用conv_alpha的值。 + +### 层次结构维度（rank）命令行指定示例： + + +```powershell +--network_args "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2" + +--network_args "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2" "conv_block_dims=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2" + +--network_args "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2" "block_alphas=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2" +``` + +### 层级别dim(rank) toml文件指定示例： + +```toml +network_args = [ "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2",] + +network_args = [ "block_dims=2,4,4,4,8,8,8,8,12,12,12,12,16,12,12,12,12,8,8,8,8,4,4,4,2", "block_alphas=2,2,2,2,4,4,4,4,6,6,6,6,8,6,6,6,6,4,4,4,4,2,2,2,2",] +``` + +# Other scripts +这些是与LoRA相关的脚本，如合并脚本等。 + +关于合并脚本 +您可以使用merge_lora.py脚本将LoRA的训练结果合并到稳定扩散模型中，也可以将多个LoRA模型合并。 + +合并到稳定扩散模型中的LoRA模型 +合并后的模型可以像常规的稳定扩散ckpt一样使用。例如，以下是一个命令行示例： + +``` +python networks\merge_lora.py --sd_model ..\model\model.ckpt + --save_to ..\lora_train1\model-char1-merged.safetensors + --models ..\lora_train1\last.safetensors --ratios 0.8 +``` + +请使用 Stable Diffusion v2.x 模型进行训练并进行合并时，需要指定--v2选项。 + +使用--sd_model选项指定要合并的 Stable Diffusion 模型文件（仅支持 .ckpt 或 .safetensors 格式，目前不支持 Diffusers）。 + +使用--save_to选项指定合并后模型的保存路径（根据扩展名自动判断为 .ckpt 或 .safetensors）。 + +使用--models选项指定已训练的 LoRA 模型文件，也可以指定多个，然后按顺序进行合并。 + +使用--ratios选项以0~1.0的数字指定每个模型的应用率（将多大比例的权重反映到原始模型中）。例如，在接近过度拟合的情况下，降低应用率可能会使结果更好。请指定与模型数量相同的比率。 + +当指定多个模型时，格式如下： + + +``` +python networks\merge_lora.py --sd_model ..\model\model.ckpt + --save_to ..\lora_train1\model-char1-merged.safetensors + --models ..\lora_train1\last.safetensors ..\lora_train2\last.safetensors --ratios 0.8 0.5 +``` + +### 将多个LoRA模型合并 + +将多个LoRA模型逐个应用于SD模型与将多个LoRA模型合并后再应用于SD模型之间，由于计算顺序的不同，会得到微妙不同的结果。 + +例如，下面是一个命令行示例： + +``` +python networks\merge_lora.py + --save_to ..\lora_train1\model-char1-style1-merged.safetensors + --models ..\lora_train1\last.safetensors ..\lora_train2\last.safetensors --ratios 0.6 0.4 +``` + +--sd_model选项不需要指定。 + +通过--save_to选项指定合并后的LoRA模型的保存位置（.ckpt或.safetensors，根据扩展名自动识别）。 + +通过--models选项指定学习的LoRA模型文件。可以指定三个或更多。 + +通过--ratios选项以0~1.0的数字指定每个模型的比率（反映多少权重来自原始模型）。如果将两个模型一对一合并，则比率将是“0.5 0.5”。如果比率为“1.0 1.0”，则总重量将过大，可能会产生不理想的结果。 + +在v1和v2中学习的LoRA，以及rank（维数）或“alpha”不同的LoRA不能合并。仅包含U-Net的LoRA和包含U-Net+文本编码器的LoRA可以合并，但结果未知。 + +### 其他选项 + +* 精度 + * 可以从float、fp16或bf16中选择合并计算时的精度。默认为float以保证精度。如果想减少内存使用量，请指定fp16/bf16。 +* save_precision + * 可以从float、fp16或bf16中选择在保存模型时的精度。默认与精度相同。 + +## 合并多个维度不同的LoRA模型 + +将多个LoRA近似为一个LoRA（无法完全复制）。使用'svd_merge_lora.py'。例如，以下是命令行的示例。 +``` +python networks\svd_merge_lora.py + --save_to ..\lora_train1\model-char1-style1-merged.safetensors + --models ..\lora_train1\last.safetensors ..\lora_train2\last.safetensors + --ratios 0.6 0.4 --new_rank 32 --device cuda +``` +`merge_lora.py`和主要选项相同。以下选项已添加： + +- `--new_rank` + - 指定要创建的LoRA rank。 +- `--new_conv_rank` + - 指定要创建的Conv2d 3x3 LoRA的rank。如果省略，则与`new_rank`相同。 +- `--device` + - 如果指定为`--device cuda`，则在GPU上执行计算。处理速度将更快。 + +## 在此存储库中生成图像的脚本中 + +请在`gen_img_diffusers.py`中添加`--network_module`和`--network_weights`选项。其含义与训练时相同。 + +通过`--network_mul`选项，可以指定0~1.0的数字来改变LoRA的应用率。 + +## 请参考以下示例，在Diffusers的pipeline中生成。 + +所需文件仅为networks/lora.py。请注意，该示例只能在Diffusers版本0.10.2中正常运行。 + +```python +import torch +from diffusers import StableDiffusionPipeline +from networks.lora import LoRAModule, create_network_from_weights +from safetensors.torch import load_file + +# if the ckpt is CompVis based, convert it to Diffusers beforehand with tools/convert_diffusers20_original_sd.py. See --help for more details. + +model_id_or_dir = r"model_id_on_hugging_face_or_dir" +device = "cuda" + +# create pipe +print(f"creating pipe from {model_id_or_dir}...") +pipe = StableDiffusionPipeline.from_pretrained(model_id_or_dir, revision="fp16", torch_dtype=torch.float16) +pipe = pipe.to(device) +vae = pipe.vae +text_encoder = pipe.text_encoder +unet = pipe.unet + +# load lora networks +print(f"loading lora networks...") + +lora_path1 = r"lora1.safetensors" +sd = load_file(lora_path1) # If the file is .ckpt, use torch.load instead. +network1, sd = create_network_from_weights(0.5, None, vae, text_encoder,unet, sd) +network1.apply_to(text_encoder, unet) +network1.load_state_dict(sd) +network1.to(device, dtype=torch.float16) + +# # You can merge weights instead of apply_to+load_state_dict. network.set_multiplier does not work +# network.merge_to(text_encoder, unet, sd) + +lora_path2 = r"lora2.safetensors" +sd = load_file(lora_path2) +network2, sd = create_network_from_weights(0.7, None, vae, text_encoder,unet, sd) +network2.apply_to(text_encoder, unet) +network2.load_state_dict(sd) +network2.to(device, dtype=torch.float16) + +lora_path3 = r"lora3.safetensors" +sd = load_file(lora_path3) +network3, sd = create_network_from_weights(0.5, None, vae, text_encoder,unet, sd) +network3.apply_to(text_encoder, unet) +network3.load_state_dict(sd) +network3.to(device, dtype=torch.float16) + +# prompts +prompt = "masterpiece, best quality, 1girl, in white shirt, looking at viewer" +negative_prompt = "bad quality, worst quality, bad anatomy, bad hands" + +# exec pipe +print("generating image...") +with torch.autocast("cuda"): + image = pipe(prompt, guidance_scale=7.5, negative_prompt=negative_prompt).images[0] + +# if not merged, you can use set_multiplier +# network1.set_multiplier(0.8) +# and generate image again... + +# save image +image.save(r"by_diffusers..png") +``` + +## 从两个模型的差异中创建LoRA模型。 + +[参考讨论链接](https://github.com/cloneofsimo/lora/discussions/56)這是參考實現的結果。數學公式沒有改變（我並不完全理解，但似乎使用奇異值分解進行了近似）。 + +将两个模型（例如微调原始模型和微调后的模型）的差异近似为LoRA。 + +### 脚本执行方法 + +请按以下方式指定。 + +``` +python networks\extract_lora_from_models.py --model_org base-model.ckpt + --model_tuned fine-tuned-model.ckpt + --save_to lora-weights.safetensors --dim 4 +``` + +--model_org 选项指定原始的Stable Diffusion模型。如果要应用创建的LoRA模型，则需要指定该模型并将其应用。可以指定.ckpt或.safetensors文件。 + +--model_tuned 选项指定要提取差分的目标Stable Diffusion模型。例如，可以指定经过Fine Tuning或DreamBooth后的模型。可以指定.ckpt或.safetensors文件。 + +--save_to 指定LoRA模型的保存路径。--dim指定LoRA的维数。 + +生成的LoRA模型可以像已训练的LoRA模型一样使用。 + +当两个模型的文本编码器相同时，LoRA将成为仅包含U-Net的LoRA。 + +### 其他选项 + +- `--v2` + - 如果使用v2.x的稳定扩散模型，请指定此选项。 +- `--device` + - 指定为 ``--device cuda`` 可在GPU上执行计算。这会使处理速度更快（即使在CPU上也不会太慢，大约快几倍）。 +- `--save_precision` + - 指定LoRA的保存格式为“float”、“fp16”、“bf16”。如果省略，将使用float。 +- `--conv_dim` + - 指定后，将扩展LoRA的应用范围到Conv2d 3x3。指定Conv2d 3x3的rank。 + - +## 图像大小调整脚本 + +（稍后将整理文件，但现在先在这里写下说明。） + +在 Aspect Ratio Bucketing 的功能扩展中，现在可以将小图像直接用作教师数据，而无需进行放大。我收到了一个用于前处理的脚本，其中包括将原始教师图像缩小的图像添加到教师数据中可以提高准确性的报告。我整理了这个脚本并加入了感谢 bmaltais 先生。 + +### 执行脚本的方法如下。 +原始图像以及调整大小后的图像将保存到转换目标文件夹中。调整大小后的图像将在文件名中添加“+512x512”之类的调整后的分辨率（与图像大小不同）。小于调整大小后分辨率的图像将不会被放大。 + +``` +python tools\resize_images_to_resolution.py --max_resolution 512x512,384x384,256x256 --save_as_png + --copy_associated_files 源图像文件夹目标文件夹 +``` + +在元画像文件夹中的图像文件将被调整大小以达到指定的分辨率（可以指定多个），并保存到目标文件夹中。除图像外的文件将被保留为原样。 + +请使用“--max_resolution”选项指定调整大小后的大小，使其达到指定的面积大小。如果指定多个，则会在每个分辨率上进行调整大小。例如，“512x512，384x384，256x256”将使目标文件夹中的图像变为原始大小和调整大小后的大小×3共计4张图像。 + +如果使用“--save_as_png”选项，则会以PNG格式保存。如果省略，则默认以JPEG格式（quality=100）保存。 + +如果使用“--copy_associated_files”选项，则会将与图像相同的文件名（例如标题等）的文件复制到调整大小后的图像文件的文件名相同的位置，但不包括扩展名。 + +### 其他选项 + +- divisible_by + - 将图像中心裁剪到能够被该值整除的大小（分别是垂直和水平的大小），以便调整大小后的图像大小可以被该值整除。 +- interpolation + - 指定缩小时的插值方法。可从``area、cubic、lanczos4``中选择，默认为``area``。 + + +# 追加信息 + +## 与cloneofsimo的代码库的区别 + +截至2022年12月25日，本代码库将LoRA应用扩展到了Text Encoder的MLP、U-Net的FFN以及Transformer的输入/输出投影中，从而增强了表现力。但是，内存使用量增加了，接近了8GB的限制。 + +此外，模块交换机制也完全不同。 + +## 关于未来的扩展 + +除了LoRA之外，我们还计划添加其他扩展，以支持更多的功能。 diff --git a/train_ti_README-ja.md b/train_ti_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..86f45a5dcf60331e7551f5da132fa51e2de31ba0 --- /dev/null +++ b/train_ti_README-ja.md @@ -0,0 +1,105 @@ +[Textual Inversion](https://textual-inversion.github.io/) の学習についての説明です。 + +[学習についての共通ドキュメント](./train_README-ja.md) もあわせてご覧ください。 + +実装に当たっては https://github.com/huggingface/diffusers/tree/main/examples/textual_inversion を大いに参考にしました。 + +学習したモデルはWeb UIでもそのまま使えます。 + +# 学習の手順 + +あらかじめこのリポジトリのREADMEを参照し、環境整備を行ってください。 + +## データの準備 + +[学習データの準備について](./train_README-ja.md) を参照してください。 + +## 学習の実行 + +``train_textual_inversion.py`` を用います。以下はコマンドラインの例です（DreamBooth手法）。 + +``` +accelerate launch --num_cpu_threads_per_process 1 train_textual_inversion.py + --dataset_config=<データ準備で作成した.tomlファイル> + --output_dir=<学習したモデルの出力先フォルダ> + --output_name=<学習したモデル出力時のファイル名> + --save_model_as=safetensors + --prior_loss_weight=1.0 + --max_train_steps=1600 + --learning_rate=1e-6 + --optimizer_type="AdamW8bit" + --xformers + --mixed_precision="fp16" + --cache_latents + --gradient_checkpointing + --token_string=mychar4 --init_word=cute --num_vectors_per_token=4 +``` + +``--token_string`` に学習時のトークン文字列を指定します。__学習時のプロンプトは、この文字列を含むようにしてください（token_stringがmychar4なら、``mychar4 1girl`` など）__。プロンプトのこの文字列の部分が、Textual Inversionの新しいtokenに置換されて学習されます。DreamBooth, class+identifier形式のデータセットとして、`token_string` をトークン文字列にするのが最も簡単で確実です。 + +プロンプトにトークン文字列が含まれているかどうかは、``--debug_dataset`` で置換後のtoken idが表示されますので、以下のように ``49408`` 以降のtokenが存在するかどうかで確認できます。 + +``` +input ids: tensor([[49406, 49408, 49409, 49410, 49411, 49412, 49413, 49414, 49415, 49407, + 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, + 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, + 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, + 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, + 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, + 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, + 49407, 49407, 49407, 49407, 49407, 49407, 49407]]) +``` + +tokenizerがすでに持っている単語（一般的な単語）は使用できません。 + +``--init_word`` にembeddingsを初期化するときのコピー元トークンの文字列を指定します。学ばせたい概念が近いものを選ぶとよいようです。二つ以上のトークンになる文字列は指定できません。 + +``--num_vectors_per_token`` にいくつのトークンをこの学習で使うかを指定します。多いほうが表現力が増しますが、その分多くのトークンを消費します。たとえばnum_vectors_per_token=8の場合、指定したトークン文字列は（一般的なプロンプトの77トークン制限のうち）8トークンを消費します。 + +以上がTextual Inversionのための主なオプションです。以降は他の学習スクリプトと同様です。 + +`num_cpu_threads_per_process` には通常は1を指定するとよいようです。 + +`pretrained_model_name_or_path` に追加学習を行う元となるモデルを指定します。Stable Diffusionのcheckpointファイル（.ckptまたは.safetensors）、Diffusersのローカルディスクにあるモデルディレクトリ、DiffusersのモデルID（"stabilityai/stable-diffusion-2"など）が指定できます。 + +`output_dir` に学習後のモデルを保存するフォルダを指定します。`output_name` にモデルのファイル名を拡張子を除いて指定します。`save_model_as` でsafetensors形式での保存を指定しています。 + +`dataset_config` に `.toml` ファイルを指定します。ファイル内でのバッチサイズ指定は、当初はメモリ消費を抑えるために `1` としてください。 + +学習させるステップ数 `max_train_steps` を10000とします。学習率 `learning_rate` はここでは5e-6を指定しています。 + +省メモリ化のため `mixed_precision="fp16"` を指定します（RTX30 シリーズ以降では `bf16` も指定できます。環境整備時にaccelerateに行った設定と合わせてください）。また `gradient_checkpointing` を指定します。 + +オプティマイザ（モデルを学習データにあうように最適化＝学習させるクラス）にメモリ消費の少ない 8bit AdamW を使うため、 `optimizer_type="AdamW8bit"` を指定します。 + +`xformers` オプションを指定し、xformersのCrossAttentionを用います。xformersをインストールしていない場合やエラーとなる場合（環境にもよりますが `mixed_precision="no"` の場合など）、代わりに `mem_eff_attn` オプションを指定すると省メモリ版CrossAttentionを使用します（速度は遅くなります）。 + +ある程度メモリがある場合は、`.toml` ファイルを編集してバッチサイズをたとえば `8` くらいに増やしてください（高速化と精度向上の可能性があります）。 + +### よく使われるオプションについて + +以下の場合にはオプションに関するドキュメントを参照してください。 + +- Stable Diffusion 2.xまたはそこからの派生モデルを学習する +- clip skipを2以上を前提としたモデルを学習する +- 75トークンを超えたキャプションで学習する + +### Textual Inversionでのバッチサイズについて + +モデル全体を学習するDreamBoothやfine tuningに比べてメモリ使用量が少ないため、バッチサイズは大きめにできます。 + +# Textual Inversionのその他の主なオプション + +すべてのオプションについては別文書を参照してください。 + +* `--weights` + * 学習前に学習済みのembeddingsを読み込み、そこから追加で学習します。 +* `--use_object_template` + * キャプションではなく既定の物体用テンプレート文字列（``a photo of a {}``など）で学習します。公式実装と同じになります。キャプションは無視されます。 +* `--use_style_template` + * キャプションではなく既定のスタイル用テンプレート文字列で学習します（``a painting in the style of {}``など）。公式実装と同じになります。キャプションは無視されます。 + +## 当リポジトリ内の画像生成スクリプトで生成する + +gen_img_diffusers.pyに、``--textual_inversion_embeddings`` オプションで学習したembeddingsファイルを指定してください（複数可）。プロンプトでembeddingsファイルのファイル名（拡張子を除く）を使うと、そのembeddingsが適用されます。 + diff --git a/train_util.py b/train_util.py new file mode 100644 index 0000000000000000000000000000000000000000..b9d08f25312fc215d4ffdd594ae29b8946d297f2 --- /dev/null +++ b/train_util.py @@ -0,0 +1,5732 @@ +# common functions for training + +import argparse +import ast +import asyncio +import datetime +import importlib +import json +import logging +import pathlib +import re +import shutil +import time +from typing import ( + Dict, + List, + NamedTuple, + Optional, + Sequence, + Tuple, + Union, +) +from accelerate import Accelerator, InitProcessGroupKwargs, DistributedDataParallelKwargs, PartialState +import glob +import math +import os +import random +import hashlib +import subprocess +from io import BytesIO +import toml + +from tqdm import tqdm + +import torch +from library.device_utils import init_ipex, clean_memory_on_device + +init_ipex() + +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.optim import Optimizer +from torchvision import transforms +from transformers import CLIPTokenizer, CLIPTextModel, CLIPTextModelWithProjection +import transformers +from diffusers.optimization import ( + SchedulerType as DiffusersSchedulerType, + TYPE_TO_SCHEDULER_FUNCTION as DIFFUSERS_TYPE_TO_SCHEDULER_FUNCTION, +) +from transformers.optimization import SchedulerType, TYPE_TO_SCHEDULER_FUNCTION +from diffusers import ( + StableDiffusionPipeline, + DDPMScheduler, + EulerAncestralDiscreteScheduler, + DPMSolverMultistepScheduler, + DPMSolverSinglestepScheduler, + LMSDiscreteScheduler, + PNDMScheduler, + DDIMScheduler, + EulerDiscreteScheduler, + HeunDiscreteScheduler, + KDPM2DiscreteScheduler, + KDPM2AncestralDiscreteScheduler, + AutoencoderKL, +) +from library import custom_train_functions +from library.original_unet import UNet2DConditionModel +from huggingface_hub import hf_hub_download +import numpy as np +from PIL import Image +import imagesize +import cv2 +import safetensors.torch +from library.lpw_stable_diffusion import StableDiffusionLongPromptWeightingPipeline +import library.model_util as model_util +import library.huggingface_util as huggingface_util +import library.sai_model_spec as sai_model_spec +import library.deepspeed_utils as deepspeed_utils +from library.utils import setup_logging, pil_resize + +setup_logging() +import logging + +logger = logging.getLogger(__name__) +# from library.attention_processors import FlashAttnProcessor +# from library.hypernetwork import replace_attentions_for_hypernetwork +from library.original_unet import UNet2DConditionModel + +# Tokenizer: checkpointから読み込むのではなくあらかじめ提供されているものを使う +TOKENIZER_PATH = "openai/clip-vit-large-patch14" +V2_STABLE_DIFFUSION_PATH = "stabilityai/stable-diffusion-2" # ここからtokenizerだけ使う v2とv2.1はtokenizer仕様は同じ + +HIGH_VRAM = False + +# checkpointファイル名 +EPOCH_STATE_NAME = "{}-{:06d}-state" +EPOCH_FILE_NAME = "{}-{:06d}" +EPOCH_DIFFUSERS_DIR_NAME = "{}-{:06d}" +LAST_STATE_NAME = "{}-state" +DEFAULT_EPOCH_NAME = "epoch" +DEFAULT_LAST_OUTPUT_NAME = "last" + +DEFAULT_STEP_NAME = "at" +STEP_STATE_NAME = "{}-step{:08d}-state" +STEP_FILE_NAME = "{}-step{:08d}" +STEP_DIFFUSERS_DIR_NAME = "{}-step{:08d}" + +# region dataset + +IMAGE_EXTENSIONS = [".png", ".jpg", ".jpeg", ".webp", ".bmp", ".PNG", ".JPG", ".JPEG", ".WEBP", ".BMP"] + +try: + import pillow_avif + + IMAGE_EXTENSIONS.extend([".avif", ".AVIF"]) +except: + pass + +# JPEG-XL on Linux +try: + from jxlpy import JXLImagePlugin + + IMAGE_EXTENSIONS.extend([".jxl", ".JXL"]) +except: + pass + +# JPEG-XL on Windows +try: + import pillow_jxl + + IMAGE_EXTENSIONS.extend([".jxl", ".JXL"]) +except: + pass + +IMAGE_TRANSFORMS = transforms.Compose( + [ + transforms.ToTensor(), + transforms.Normalize([0.5], [0.5]), + ] +) + +TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX = "_te_outputs.npz" + + +class ImageInfo: + def __init__(self, image_key: str, num_repeats: int, caption: str, is_reg: bool, absolute_path: str) -> None: + self.image_key: str = image_key + self.num_repeats: int = num_repeats + self.caption: str = caption + self.is_reg: bool = is_reg + self.absolute_path: str = absolute_path + self.image_size: Tuple[int, int] = None + self.resized_size: Tuple[int, int] = None + self.bucket_reso: Tuple[int, int] = None + self.latents: torch.Tensor = None + self.latents_flipped: torch.Tensor = None + self.latents_npz: str = None + self.latents_original_size: Tuple[int, int] = None # original image size, not latents size + self.latents_crop_ltrb: Tuple[int, int] = None # crop left top right bottom in original pixel size, not latents size + self.cond_img_path: str = None + self.image: Optional[Image.Image] = None # optional, original PIL Image + # SDXL, optional + self.text_encoder_outputs_npz: Optional[str] = None + self.text_encoder_outputs1: Optional[torch.Tensor] = None + self.text_encoder_outputs2: Optional[torch.Tensor] = None + self.text_encoder_pool2: Optional[torch.Tensor] = None + self.alpha_mask: Optional[torch.Tensor] = None # alpha mask can be flipped in runtime + + +class BucketManager: + def __init__(self, no_upscale, max_reso, min_size, max_size, reso_steps) -> None: + if max_size is not None: + if max_reso is not None: + assert max_size >= max_reso[0], "the max_size should be larger than the width of max_reso" + assert max_size >= max_reso[1], "the max_size should be larger than the height of max_reso" + if min_size is not None: + assert max_size >= min_size, "the max_size should be larger than the min_size" + + self.no_upscale = no_upscale + if max_reso is None: + self.max_reso = None + self.max_area = None + else: + self.max_reso = max_reso + self.max_area = max_reso[0] * max_reso[1] + self.min_size = min_size + self.max_size = max_size + self.reso_steps = reso_steps + + self.resos = [] + self.reso_to_id = {} + self.buckets = [] # 前処理時は (image_key, image, original size, crop left/top)、学習時は image_key + + def add_image(self, reso, image_or_info): + bucket_id = self.reso_to_id[reso] + self.buckets[bucket_id].append(image_or_info) + + def shuffle(self): + for bucket in self.buckets: + random.shuffle(bucket) + + def sort(self): + # 解像度順にソートする（表示時、メタデータ格納時の見栄えをよくするためだけ）。bucketsも入れ替えてreso_to_idも振り直す + sorted_resos = self.resos.copy() + sorted_resos.sort() + + sorted_buckets = [] + sorted_reso_to_id = {} + for i, reso in enumerate(sorted_resos): + bucket_id = self.reso_to_id[reso] + sorted_buckets.append(self.buckets[bucket_id]) + sorted_reso_to_id[reso] = i + + self.resos = sorted_resos + self.buckets = sorted_buckets + self.reso_to_id = sorted_reso_to_id + + def make_buckets(self): + resos = model_util.make_bucket_resolutions(self.max_reso, self.min_size, self.max_size, self.reso_steps) + self.set_predefined_resos(resos) + + def set_predefined_resos(self, resos): + # 規定サイズから選ぶ場合の解像度、aspect ratioの情報を格納しておく + self.predefined_resos = resos.copy() + self.predefined_resos_set = set(resos) + self.predefined_aspect_ratios = np.array([w / h for w, h in resos]) + + def add_if_new_reso(self, reso): + if reso not in self.reso_to_id: + bucket_id = len(self.resos) + self.reso_to_id[reso] = bucket_id + self.resos.append(reso) + self.buckets.append([]) + # logger.info(reso, bucket_id, len(self.buckets)) + + def round_to_steps(self, x): + x = int(x + 0.5) + return x - x % self.reso_steps + + def select_bucket(self, image_width, image_height): + aspect_ratio = image_width / image_height + if not self.no_upscale: + # 拡大および縮小を行う + # 同じaspect ratioがあるかもしれないので（fine tuningで、no_upscale=Trueで前処理した場合）、解像度が同じものを優先する + reso = (image_width, image_height) + if reso in self.predefined_resos_set: + pass + else: + ar_errors = self.predefined_aspect_ratios - aspect_ratio + predefined_bucket_id = np.abs(ar_errors).argmin() # 当該解像度以外でaspect ratio errorが最も少ないもの + reso = self.predefined_resos[predefined_bucket_id] + + ar_reso = reso[0] / reso[1] + if aspect_ratio > ar_reso: # 横が長い→縦を合わせる + scale = reso[1] / image_height + else: + scale = reso[0] / image_width + + resized_size = (int(image_width * scale + 0.5), int(image_height * scale + 0.5)) + # logger.info(f"use predef, {image_width}, {image_height}, {reso}, {resized_size}") + else: + # 縮小のみを行う + if image_width * image_height > self.max_area: + # 画像が大きすぎるのでアスペクト比を保ったまま縮小することを前提にbucketを決める + resized_width = math.sqrt(self.max_area * aspect_ratio) + resized_height = self.max_area / resized_width + assert abs(resized_width / resized_height - aspect_ratio) < 1e-2, "aspect is illegal" + + # リサイズ後の短辺または長辺をreso_steps単位にする：aspect ratioの差が少ないほうを選ぶ + # 元のbucketingと同じロジック + b_width_rounded = self.round_to_steps(resized_width) + b_height_in_wr = self.round_to_steps(b_width_rounded / aspect_ratio) + ar_width_rounded = b_width_rounded / b_height_in_wr + + b_height_rounded = self.round_to_steps(resized_height) + b_width_in_hr = self.round_to_steps(b_height_rounded * aspect_ratio) + ar_height_rounded = b_width_in_hr / b_height_rounded + + # logger.info(b_width_rounded, b_height_in_wr, ar_width_rounded) + # logger.info(b_width_in_hr, b_height_rounded, ar_height_rounded) + + if abs(ar_width_rounded - aspect_ratio) < abs(ar_height_rounded - aspect_ratio): + resized_size = (b_width_rounded, int(b_width_rounded / aspect_ratio + 0.5)) + else: + resized_size = (int(b_height_rounded * aspect_ratio + 0.5), b_height_rounded) + # logger.info(resized_size) + else: + resized_size = (image_width, image_height) # リサイズは不要 + + # 画像のサイズ未満をbucketのサイズとする（paddingせずにcroppingする） + bucket_width = resized_size[0] - resized_size[0] % self.reso_steps + bucket_height = resized_size[1] - resized_size[1] % self.reso_steps + # logger.info(f"use arbitrary {image_width}, {image_height}, {resized_size}, {bucket_width}, {bucket_height}") + + reso = (bucket_width, bucket_height) + + self.add_if_new_reso(reso) + + ar_error = (reso[0] / reso[1]) - aspect_ratio + return reso, resized_size, ar_error + + @staticmethod + def get_crop_ltrb(bucket_reso: Tuple[int, int], image_size: Tuple[int, int]): + # Stability AIの前処理に合わせてcrop left/topを計算する。crop rightはflipのaugmentationのために求める + # Calculate crop left/top according to the preprocessing of Stability AI. Crop right is calculated for flip augmentation. + + bucket_ar = bucket_reso[0] / bucket_reso[1] + image_ar = image_size[0] / image_size[1] + if bucket_ar > image_ar: + # bucketのほうが横長→縦を合わせる + resized_width = bucket_reso[1] * image_ar + resized_height = bucket_reso[1] + else: + resized_width = bucket_reso[0] + resized_height = bucket_reso[0] / image_ar + crop_left = (bucket_reso[0] - resized_width) // 2 + crop_top = (bucket_reso[1] - resized_height) // 2 + crop_right = crop_left + resized_width + crop_bottom = crop_top + resized_height + return crop_left, crop_top, crop_right, crop_bottom + + +class BucketBatchIndex(NamedTuple): + bucket_index: int + bucket_batch_size: int + batch_index: int + + +class AugHelper: + # albumentationsへの依存をなくしたがとりあえず同じinterfaceを持たせる + + def __init__(self): + pass + + def color_aug(self, image: np.ndarray): + # self.color_aug_method = albu.OneOf( + # [ + # albu.HueSaturationValue(8, 0, 0, p=0.5), + # albu.RandomGamma((95, 105), p=0.5), + # ], + # p=0.33, + # ) + hue_shift_limit = 8 + + # remove dependency to albumentations + if random.random() <= 0.33: + if random.random() > 0.5: + # hue shift + hsv_img = cv2.cvtColor(image, cv2.COLOR_BGR2HSV) + hue_shift = random.uniform(-hue_shift_limit, hue_shift_limit) + if hue_shift < 0: + hue_shift = 180 + hue_shift + hsv_img[:, :, 0] = (hsv_img[:, :, 0] + hue_shift) % 180 + image = cv2.cvtColor(hsv_img, cv2.COLOR_HSV2BGR) + else: + # random gamma + gamma = random.uniform(0.95, 1.05) + image = np.clip(image**gamma, 0, 255).astype(np.uint8) + + return {"image": image} + + def get_augmentor(self, use_color_aug: bool): # -> Optional[Callable[[np.ndarray], Dict[str, np.ndarray]]]: + return self.color_aug if use_color_aug else None + + +class BaseSubset: + def __init__( + self, + image_dir: Optional[str], + alpha_mask: Optional[bool], + num_repeats: int, + shuffle_caption: bool, + caption_separator: str, + keep_tokens: int, + keep_tokens_separator: str, + secondary_separator: Optional[str], + enable_wildcard: bool, + color_aug: bool, + flip_aug: bool, + face_crop_aug_range: Optional[Tuple[float, float]], + random_crop: bool, + caption_dropout_rate: float, + caption_dropout_every_n_epochs: int, + caption_tag_dropout_rate: float, + caption_prefix: Optional[str], + caption_suffix: Optional[str], + token_warmup_min: int, + token_warmup_step: Union[float, int], + ) -> None: + self.image_dir = image_dir + self.alpha_mask = alpha_mask if alpha_mask is not None else False + self.num_repeats = num_repeats + self.shuffle_caption = shuffle_caption + self.caption_separator = caption_separator + self.keep_tokens = keep_tokens + self.keep_tokens_separator = keep_tokens_separator + self.secondary_separator = secondary_separator + self.enable_wildcard = enable_wildcard + self.color_aug = color_aug + self.flip_aug = flip_aug + self.face_crop_aug_range = face_crop_aug_range + self.random_crop = random_crop + self.caption_dropout_rate = caption_dropout_rate + self.caption_dropout_every_n_epochs = caption_dropout_every_n_epochs + self.caption_tag_dropout_rate = caption_tag_dropout_rate + self.caption_prefix = caption_prefix + self.caption_suffix = caption_suffix + + self.token_warmup_min = token_warmup_min # step=0におけるタグの数 + self.token_warmup_step = token_warmup_step # N（N<1ならN*max_train_steps）ステップ目でタグの数が最大になる + + self.img_count = 0 + + +class DreamBoothSubset(BaseSubset): + def __init__( + self, + image_dir: str, + is_reg: bool, + class_tokens: Optional[str], + caption_extension: str, + cache_info: bool, + alpha_mask: bool, + num_repeats, + shuffle_caption, + caption_separator: str, + keep_tokens, + keep_tokens_separator, + secondary_separator, + enable_wildcard, + color_aug, + flip_aug, + face_crop_aug_range, + random_crop, + caption_dropout_rate, + caption_dropout_every_n_epochs, + caption_tag_dropout_rate, + caption_prefix, + caption_suffix, + token_warmup_min, + token_warmup_step, + ) -> None: + assert image_dir is not None, "image_dir must be specified / image_dirは指定が必須です" + + super().__init__( + image_dir, + alpha_mask, + num_repeats, + shuffle_caption, + caption_separator, + keep_tokens, + keep_tokens_separator, + secondary_separator, + enable_wildcard, + color_aug, + flip_aug, + face_crop_aug_range, + random_crop, + caption_dropout_rate, + caption_dropout_every_n_epochs, + caption_tag_dropout_rate, + caption_prefix, + caption_suffix, + token_warmup_min, + token_warmup_step, + ) + + self.is_reg = is_reg + self.class_tokens = class_tokens + self.caption_extension = caption_extension + if self.caption_extension and not self.caption_extension.startswith("."): + self.caption_extension = "." + self.caption_extension + self.cache_info = cache_info + + def __eq__(self, other) -> bool: + if not isinstance(other, DreamBoothSubset): + return NotImplemented + return self.image_dir == other.image_dir + + +class FineTuningSubset(BaseSubset): + def __init__( + self, + image_dir, + metadata_file: str, + alpha_mask: bool, + num_repeats, + shuffle_caption, + caption_separator, + keep_tokens, + keep_tokens_separator, + secondary_separator, + enable_wildcard, + color_aug, + flip_aug, + face_crop_aug_range, + random_crop, + caption_dropout_rate, + caption_dropout_every_n_epochs, + caption_tag_dropout_rate, + caption_prefix, + caption_suffix, + token_warmup_min, + token_warmup_step, + ) -> None: + assert metadata_file is not None, "metadata_file must be specified / metadata_fileは指定が必須です" + + super().__init__( + image_dir, + alpha_mask, + num_repeats, + shuffle_caption, + caption_separator, + keep_tokens, + keep_tokens_separator, + secondary_separator, + enable_wildcard, + color_aug, + flip_aug, + face_crop_aug_range, + random_crop, + caption_dropout_rate, + caption_dropout_every_n_epochs, + caption_tag_dropout_rate, + caption_prefix, + caption_suffix, + token_warmup_min, + token_warmup_step, + ) + + self.metadata_file = metadata_file + + def __eq__(self, other) -> bool: + if not isinstance(other, FineTuningSubset): + return NotImplemented + return self.metadata_file == other.metadata_file + + +class ControlNetSubset(BaseSubset): + def __init__( + self, + image_dir: str, + conditioning_data_dir: str, + caption_extension: str, + cache_info: bool, + num_repeats, + shuffle_caption, + caption_separator, + keep_tokens, + keep_tokens_separator, + secondary_separator, + enable_wildcard, + color_aug, + flip_aug, + face_crop_aug_range, + random_crop, + caption_dropout_rate, + caption_dropout_every_n_epochs, + caption_tag_dropout_rate, + caption_prefix, + caption_suffix, + token_warmup_min, + token_warmup_step, + ) -> None: + assert image_dir is not None, "image_dir must be specified / image_dirは指定が必須です" + + super().__init__( + image_dir, + False, # alpha_mask + num_repeats, + shuffle_caption, + caption_separator, + keep_tokens, + keep_tokens_separator, + secondary_separator, + enable_wildcard, + color_aug, + flip_aug, + face_crop_aug_range, + random_crop, + caption_dropout_rate, + caption_dropout_every_n_epochs, + caption_tag_dropout_rate, + caption_prefix, + caption_suffix, + token_warmup_min, + token_warmup_step, + ) + + self.conditioning_data_dir = conditioning_data_dir + self.caption_extension = caption_extension + if self.caption_extension and not self.caption_extension.startswith("."): + self.caption_extension = "." + self.caption_extension + self.cache_info = cache_info + + def __eq__(self, other) -> bool: + if not isinstance(other, ControlNetSubset): + return NotImplemented + return self.image_dir == other.image_dir and self.conditioning_data_dir == other.conditioning_data_dir + + +class BaseDataset(torch.utils.data.Dataset): + def __init__( + self, + tokenizer: Union[CLIPTokenizer, List[CLIPTokenizer]], + max_token_length: int, + resolution: Optional[Tuple[int, int]], + network_multiplier: float, + debug_dataset: bool, + ) -> None: + super().__init__() + + self.tokenizers = tokenizer if isinstance(tokenizer, list) else [tokenizer] + + self.max_token_length = max_token_length + # width/height is used when enable_bucket==False + self.width, self.height = (None, None) if resolution is None else resolution + self.network_multiplier = network_multiplier + self.debug_dataset = debug_dataset + + self.subsets: List[Union[DreamBoothSubset, FineTuningSubset]] = [] + + self.token_padding_disabled = False + self.tag_frequency = {} + self.XTI_layers = None + self.token_strings = None + + self.enable_bucket = False + self.bucket_manager: BucketManager = None # not initialized + self.min_bucket_reso = None + self.max_bucket_reso = None + self.bucket_reso_steps = None + self.bucket_no_upscale = None + self.bucket_info = None # for metadata + + self.tokenizer_max_length = self.tokenizers[0].model_max_length if max_token_length is None else max_token_length + 2 + + self.current_epoch: int = 0 # インスタンスがepochごとに新しく作られるようなので外側から渡さないとダメ + + self.current_step: int = 0 + self.max_train_steps: int = 0 + self.seed: int = 0 + + # augmentation + self.aug_helper = AugHelper() + + self.image_transforms = IMAGE_TRANSFORMS + + self.image_data: Dict[str, ImageInfo] = {} + self.image_to_subset: Dict[str, Union[DreamBoothSubset, FineTuningSubset]] = {} + + self.replacements = {} + + # caching + self.caching_mode = None # None, 'latents', 'text' + + def adjust_min_max_bucket_reso_by_steps( + self, resolution: Tuple[int, int], min_bucket_reso: int, max_bucket_reso: int, bucket_reso_steps: int + ) -> Tuple[int, int]: + # make min/max bucket reso to be multiple of bucket_reso_steps + if min_bucket_reso % bucket_reso_steps != 0: + adjusted_min_bucket_reso = min_bucket_reso - min_bucket_reso % bucket_reso_steps + logger.warning( + f"min_bucket_reso is adjusted to be multiple of bucket_reso_steps" + f" / min_bucket_resoがbucket_reso_stepsの倍数になるように調整されました: {min_bucket_reso} -> {adjusted_min_bucket_reso}" + ) + min_bucket_reso = adjusted_min_bucket_reso + if max_bucket_reso % bucket_reso_steps != 0: + adjusted_max_bucket_reso = max_bucket_reso + bucket_reso_steps - max_bucket_reso % bucket_reso_steps + logger.warning( + f"max_bucket_reso is adjusted to be multiple of bucket_reso_steps" + f" / max_bucket_resoがbucket_reso_stepsの倍数になるように調整されました: {max_bucket_reso} -> {adjusted_max_bucket_reso}" + ) + max_bucket_reso = adjusted_max_bucket_reso + + assert ( + min(resolution) >= min_bucket_reso + ), f"min_bucket_reso must be equal or less than resolution / min_bucket_resoは最小解像度より大きくできません。解像度を大きくするかmin_bucket_resoを小さくしてください" + assert ( + max(resolution) <= max_bucket_reso + ), f"max_bucket_reso must be equal or greater than resolution / max_bucket_resoは最大解像度より小さくできません。解像度を小さくするかmin_bucket_resoを大きくしてください" + + return min_bucket_reso, max_bucket_reso + + def set_seed(self, seed): + self.seed = seed + + def set_caching_mode(self, mode): + self.caching_mode = mode + + def set_current_epoch(self, epoch): + if not self.current_epoch == epoch: # epochが切り替わったらバケツをシャッフルする + if epoch > self.current_epoch: + logger.info("epoch is incremented. current_epoch: {}, epoch: {}".format(self.current_epoch, epoch)) + num_epochs = epoch - self.current_epoch + for _ in range(num_epochs): + self.current_epoch += 1 + self.shuffle_buckets() + # self.current_epoch seem to be set to 0 again in the next epoch. it may be caused by skipped_dataloader? + else: + logger.warning("epoch is not incremented. current_epoch: {}, epoch: {}".format(self.current_epoch, epoch)) + self.current_epoch = epoch + + def set_current_step(self, step): + self.current_step = step + + def set_max_train_steps(self, max_train_steps): + self.max_train_steps = max_train_steps + + def set_tag_frequency(self, dir_name, captions): + frequency_for_dir = self.tag_frequency.get(dir_name, {}) + self.tag_frequency[dir_name] = frequency_for_dir + for caption in captions: + for tag in caption.split(","): + tag = tag.strip() + if tag: + tag = tag.lower() + frequency = frequency_for_dir.get(tag, 0) + frequency_for_dir[tag] = frequency + 1 + + def disable_token_padding(self): + self.token_padding_disabled = True + + def enable_XTI(self, layers=None, token_strings=None): + self.XTI_layers = layers + self.token_strings = token_strings + + def add_replacement(self, str_from, str_to): + self.replacements[str_from] = str_to + + def process_caption(self, subset: BaseSubset, caption): + # caption に prefix/suffix を付ける + if subset.caption_prefix: + caption = subset.caption_prefix + " " + caption + if subset.caption_suffix: + caption = caption + " " + subset.caption_suffix + + # dropoutの決定：tag dropがこのメソッド内にあるのでここで行うのが良い + is_drop_out = subset.caption_dropout_rate > 0 and random.random() < subset.caption_dropout_rate + is_drop_out = ( + is_drop_out + or subset.caption_dropout_every_n_epochs > 0 + and self.current_epoch % subset.caption_dropout_every_n_epochs == 0 + ) + + if is_drop_out: + caption = "" + else: + # process wildcards + if subset.enable_wildcard: + # if caption is multiline, random choice one line + if "\n" in caption: + caption = random.choice(caption.split("\n")) + + # wildcard is like '{aaa|bbb|ccc...}' + # escape the curly braces like {{ or }} + replacer1 = "⦅" + replacer2 = "⦆" + while replacer1 in caption or replacer2 in caption: + replacer1 += "⦅" + replacer2 += "⦆" + + caption = caption.replace("{{", replacer1).replace("}}", replacer2) + + # replace the wildcard + def replace_wildcard(match): + return random.choice(match.group(1).split("|")) + + caption = re.sub(r"\{([^}]+)\}", replace_wildcard, caption) + + # unescape the curly braces + caption = caption.replace(replacer1, "{").replace(replacer2, "}") + else: + # if caption is multiline, use the first line + caption = caption.split("\n")[0] + + if subset.shuffle_caption or subset.token_warmup_step > 0 or subset.caption_tag_dropout_rate > 0: + fixed_tokens = [] + flex_tokens = [] + fixed_suffix_tokens = [] + if ( + hasattr(subset, "keep_tokens_separator") + and subset.keep_tokens_separator + and subset.keep_tokens_separator in caption + ): + fixed_part, flex_part = caption.split(subset.keep_tokens_separator, 1) + if subset.keep_tokens_separator in flex_part: + flex_part, fixed_suffix_part = flex_part.split(subset.keep_tokens_separator, 1) + fixed_suffix_tokens = [t.strip() for t in fixed_suffix_part.split(subset.caption_separator) if t.strip()] + + fixed_tokens = [t.strip() for t in fixed_part.split(subset.caption_separator) if t.strip()] + flex_tokens = [t.strip() for t in flex_part.split(subset.caption_separator) if t.strip()] + else: + tokens = [t.strip() for t in caption.strip().split(subset.caption_separator)] + flex_tokens = tokens[:] + if subset.keep_tokens > 0: + fixed_tokens = flex_tokens[: subset.keep_tokens] + flex_tokens = tokens[subset.keep_tokens :] + + if subset.token_warmup_step < 1: # 初回に上書きする + subset.token_warmup_step = math.floor(subset.token_warmup_step * self.max_train_steps) + if subset.token_warmup_step and self.current_step < subset.token_warmup_step: + tokens_len = ( + math.floor( + (self.current_step) * ((len(flex_tokens) - subset.token_warmup_min) / (subset.token_warmup_step)) + ) + + subset.token_warmup_min + ) + flex_tokens = flex_tokens[:tokens_len] + + def dropout_tags(tokens): + if subset.caption_tag_dropout_rate <= 0: + return tokens + l = [] + for token in tokens: + if random.random() >= subset.caption_tag_dropout_rate: + l.append(token) + return l + + if subset.shuffle_caption: + random.shuffle(flex_tokens) + + flex_tokens = dropout_tags(flex_tokens) + + caption = ", ".join(fixed_tokens + flex_tokens + fixed_suffix_tokens) + + # process secondary separator + if subset.secondary_separator: + caption = caption.replace(subset.secondary_separator, subset.caption_separator) + + # textual inversion対応 + for str_from, str_to in self.replacements.items(): + if str_from == "": + # replace all + if type(str_to) == list: + caption = random.choice(str_to) + else: + caption = str_to + else: + caption = caption.replace(str_from, str_to) + + return caption + + def get_input_ids(self, caption, tokenizer=None): + if tokenizer is None: + tokenizer = self.tokenizers[0] + + input_ids = tokenizer( + caption, padding="max_length", truncation=True, max_length=self.tokenizer_max_length, return_tensors="pt" + ).input_ids + + if self.tokenizer_max_length > tokenizer.model_max_length: + input_ids = input_ids.squeeze(0) + iids_list = [] + if tokenizer.pad_token_id == tokenizer.eos_token_id: + # v1 + # 77以上の時は " .... " でトータル227とかになっているので、"..."の三連に変換する + # 1111氏のやつは , で区切る、とかしているようだが　とりあえず単純に + for i in range( + 1, self.tokenizer_max_length - tokenizer.model_max_length + 2, tokenizer.model_max_length - 2 + ): # (1, 152, 75) + ids_chunk = ( + input_ids[0].unsqueeze(0), + input_ids[i : i + tokenizer.model_max_length - 2], + input_ids[-1].unsqueeze(0), + ) + ids_chunk = torch.cat(ids_chunk) + iids_list.append(ids_chunk) + else: + # v2 or SDXL + # 77以上の時は " .... ..." でトータル227とかになっているので、"... ..."の三連に変換する + for i in range(1, self.tokenizer_max_length - tokenizer.model_max_length + 2, tokenizer.model_max_length - 2): + ids_chunk = ( + input_ids[0].unsqueeze(0), # BOS + input_ids[i : i + tokenizer.model_max_length - 2], + input_ids[-1].unsqueeze(0), + ) # PAD or EOS + ids_chunk = torch.cat(ids_chunk) + + # 末尾が または の場合は、何もしなくてよい + # 末尾が x の場合は末尾を に変える（x なら結果的に変化なし） + if ids_chunk[-2] != tokenizer.eos_token_id and ids_chunk[-2] != tokenizer.pad_token_id: + ids_chunk[-1] = tokenizer.eos_token_id + # 先頭が ... の場合は ... に変える + if ids_chunk[1] == tokenizer.pad_token_id: + ids_chunk[1] = tokenizer.eos_token_id + + iids_list.append(ids_chunk) + + input_ids = torch.stack(iids_list) # 3,77 + return input_ids + + def register_image(self, info: ImageInfo, subset: BaseSubset): + self.image_data[info.image_key] = info + self.image_to_subset[info.image_key] = subset + + def make_buckets(self): + """ + bucketingを行わない場合も呼び出し必須（ひとつだけbucketを作る） + min_size and max_size are ignored when enable_bucket is False + """ + logger.info("loading image sizes.") + for info in tqdm(self.image_data.values()): + if info.image_size is None: + info.image_size = self.get_image_size(info.absolute_path) + + if self.enable_bucket: + logger.info("make buckets") + else: + logger.info("prepare dataset") + + # bucketを作成し、画像をbucketに振り分ける + if self.enable_bucket: + if self.bucket_manager is None: # fine tuningの場合でmetadataに定義がある場合は、すでに初期化済み + self.bucket_manager = BucketManager( + self.bucket_no_upscale, + (self.width, self.height), + self.min_bucket_reso, + self.max_bucket_reso, + self.bucket_reso_steps, + ) + if not self.bucket_no_upscale: + self.bucket_manager.make_buckets() + else: + logger.warning( + "min_bucket_reso and max_bucket_reso are ignored if bucket_no_upscale is set, because bucket reso is defined by image size automatically / bucket_no_upscaleが指定された場合は、bucketの解像度は画像サイズから自動計算されるため、min_bucket_resoとmax_bucket_resoは無視されます" + ) + + img_ar_errors = [] + for image_info in self.image_data.values(): + image_width, image_height = image_info.image_size + image_info.bucket_reso, image_info.resized_size, ar_error = self.bucket_manager.select_bucket( + image_width, image_height + ) + + # logger.info(image_info.image_key, image_info.bucket_reso) + img_ar_errors.append(abs(ar_error)) + + self.bucket_manager.sort() + else: + self.bucket_manager = BucketManager(False, (self.width, self.height), None, None, None) + self.bucket_manager.set_predefined_resos([(self.width, self.height)]) # ひとつの固定サイズbucketのみ + for image_info in self.image_data.values(): + image_width, image_height = image_info.image_size + image_info.bucket_reso, image_info.resized_size, _ = self.bucket_manager.select_bucket(image_width, image_height) + + for image_info in self.image_data.values(): + for _ in range(image_info.num_repeats): + self.bucket_manager.add_image(image_info.bucket_reso, image_info.image_key) + + # bucket情報を表示、格納する + if self.enable_bucket: + self.bucket_info = {"buckets": {}} + logger.info("number of images (including repeats) / 各bucketの画像枚数（繰り返し回数を含む）") + for i, (reso, bucket) in enumerate(zip(self.bucket_manager.resos, self.bucket_manager.buckets)): + count = len(bucket) + if count > 0: + self.bucket_info["buckets"][i] = {"resolution": reso, "count": len(bucket)} + logger.info(f"bucket {i}: resolution {reso}, count: {len(bucket)}") + + if len(img_ar_errors) == 0: + mean_img_ar_error = 0 # avoid NaN + else: + img_ar_errors = np.array(img_ar_errors) + mean_img_ar_error = np.mean(np.abs(img_ar_errors)) + self.bucket_info["mean_img_ar_error"] = mean_img_ar_error + logger.info(f"mean ar error (without repeats): {mean_img_ar_error}") + + # データ参照用indexを作る。このindexはdatasetのshuffleに用いられる + self.buckets_indices: List[BucketBatchIndex] = [] + for bucket_index, bucket in enumerate(self.bucket_manager.buckets): + batch_count = int(math.ceil(len(bucket) / self.batch_size)) + for batch_index in range(batch_count): + self.buckets_indices.append(BucketBatchIndex(bucket_index, self.batch_size, batch_index)) + + # ↓以下はbucketごとのbatch件数があまりにも増えて混乱を招くので元に戻す + # 　学習時はステップ数がランダムなので、同一画像が同一batch内にあってもそれほど悪影響はないであろう、と考えられる + # + # # bucketが細分化されることにより、ひとつのbucketに一種類の画像のみというケースが増え、つまりそれは + # # ひとつのbatchが同じ画像で占められることになるので、さすがに良くないであろう + # # そのためバッチサイズを画像種類までに制限する + # # ただそれでも同一画像が同一バッチに含まれる可能性はあるので、繰り返し回数が少ないほうがshuffleの品質は良くなることは間違いない？ + # # TO DO 正則化画像をepochまたがりで利用する仕組み + # num_of_image_types = len(set(bucket)) + # bucket_batch_size = min(self.batch_size, num_of_image_types) + # batch_count = int(math.ceil(len(bucket) / bucket_batch_size)) + # # logger.info(bucket_index, num_of_image_types, bucket_batch_size, batch_count) + # for batch_index in range(batch_count): + # self.buckets_indices.append(BucketBatchIndex(bucket_index, bucket_batch_size, batch_index)) + # ↑ここまで + + self.shuffle_buckets() + self._length = len(self.buckets_indices) + + def shuffle_buckets(self): + # set random seed for this epoch + random.seed(self.seed + self.current_epoch) + + random.shuffle(self.buckets_indices) + self.bucket_manager.shuffle() + + def verify_bucket_reso_steps(self, min_steps: int): + assert self.bucket_reso_steps is None or self.bucket_reso_steps % min_steps == 0, ( + f"bucket_reso_steps is {self.bucket_reso_steps}. it must be divisible by {min_steps}.\n" + + f"bucket_reso_stepsが{self.bucket_reso_steps}です。{min_steps}で割り切れる必要があります" + ) + + def is_latent_cacheable(self): + return all([not subset.color_aug and not subset.random_crop for subset in self.subsets]) + + def is_text_encoder_output_cacheable(self): + return all( + [ + not ( + subset.caption_dropout_rate > 0 + or subset.shuffle_caption + or subset.token_warmup_step > 0 + or subset.caption_tag_dropout_rate > 0 + ) + for subset in self.subsets + ] + ) + + def cache_latents(self, vae, vae_batch_size=1, cache_to_disk=False, is_main_process=True): + # マルチGPUには対応していないので、そちらはtools/cache_latents.pyを使うこと + logger.info("caching latents.") + + image_infos = list(self.image_data.values()) + + # sort by resolution + image_infos.sort(key=lambda info: info.bucket_reso[0] * info.bucket_reso[1]) + + # split by resolution and some conditions + class Condition: + def __init__(self, reso, flip_aug, alpha_mask, random_crop): + self.reso = reso + self.flip_aug = flip_aug + self.alpha_mask = alpha_mask + self.random_crop = random_crop + + def __eq__(self, other): + return ( + self.reso == other.reso + and self.flip_aug == other.flip_aug + and self.alpha_mask == other.alpha_mask + and self.random_crop == other.random_crop + ) + + batches: List[Tuple[Condition, List[ImageInfo]]] = [] + batch: List[ImageInfo] = [] + current_condition = None + + logger.info("checking cache validity...") + for info in tqdm(image_infos): + subset = self.image_to_subset[info.image_key] + + if info.latents_npz is not None: # fine tuning dataset + continue + + # check disk cache exists and size of latents + if cache_to_disk: + info.latents_npz = os.path.splitext(info.absolute_path)[0] + ".npz" + if not is_main_process: # store to info only + continue + + cache_available = is_disk_cached_latents_is_expected( + info.bucket_reso, info.latents_npz, subset.flip_aug, subset.alpha_mask + ) + + if cache_available: # do not add to batch + continue + + # if batch is not empty and condition is changed, flush the batch. Note that current_condition is not None if batch is not empty + condition = Condition(info.bucket_reso, subset.flip_aug, subset.alpha_mask, subset.random_crop) + if len(batch) > 0 and current_condition != condition: + batches.append((current_condition, batch)) + batch = [] + + batch.append(info) + current_condition = condition + + # if number of data in batch is enough, flush the batch + if len(batch) >= vae_batch_size: + batches.append((current_condition, batch)) + batch = [] + current_condition = None + + if len(batch) > 0: + batches.append((current_condition, batch)) + + if cache_to_disk and not is_main_process: # if cache to disk, don't cache latents in non-main process, set to info only + return + + # iterate batches: batch doesn't have image, image will be loaded in cache_batch_latents and discarded + logger.info("caching latents...") + for condition, batch in tqdm(batches, smoothing=1, total=len(batches)): + cache_batch_latents(vae, cache_to_disk, batch, condition.flip_aug, condition.alpha_mask, condition.random_crop) + + # weight_dtypeを指定するとText Encoderそのもの、およひ出力がweight_dtypeになる + # SDXLでのみ有効だが、datasetのメソッドとする必要があるので、sdxl_train_util.pyではなくこちらに実装する + # SD1/2に対応するにはv2のフラグを持つ必要があるので後回し + def cache_text_encoder_outputs( + self, tokenizers, text_encoders, device, weight_dtype, cache_to_disk=False, is_main_process=True + ): + assert len(tokenizers) == 2, "only support SDXL" + + # latentsのキャッシュと同様に、ディスクへのキャッシュに対応する + # またマルチGPUには対応していないので、そちらはtools/cache_latents.pyを使うこと + logger.info("caching text encoder outputs.") + image_infos = list(self.image_data.values()) + + logger.info("checking cache existence...") + image_infos_to_cache = [] + for info in tqdm(image_infos): + # subset = self.image_to_subset[info.image_key] + if cache_to_disk: + te_out_npz = os.path.splitext(info.absolute_path)[0] + TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX + info.text_encoder_outputs_npz = te_out_npz + + if not is_main_process: # store to info only + continue + + if os.path.exists(te_out_npz): + continue + + image_infos_to_cache.append(info) + + if cache_to_disk and not is_main_process: # if cache to disk, don't cache latents in non-main process, set to info only + return + + # prepare tokenizers and text encoders + for text_encoder in text_encoders: + text_encoder.to(device) + if weight_dtype is not None: + text_encoder.to(dtype=weight_dtype) + + # create batch + batch = [] + batches = [] + for info in image_infos_to_cache: + input_ids1 = self.get_input_ids(info.caption, tokenizers[0]) + input_ids2 = self.get_input_ids(info.caption, tokenizers[1]) + batch.append((info, input_ids1, input_ids2)) + + if len(batch) >= self.batch_size: + batches.append(batch) + batch = [] + + if len(batch) > 0: + batches.append(batch) + + # iterate batches: call text encoder and cache outputs for memory or disk + logger.info("caching text encoder outputs...") + for batch in tqdm(batches): + infos, input_ids1, input_ids2 = zip(*batch) + input_ids1 = torch.stack(input_ids1, dim=0) + input_ids2 = torch.stack(input_ids2, dim=0) + cache_batch_text_encoder_outputs( + infos, tokenizers, text_encoders, self.max_token_length, cache_to_disk, input_ids1, input_ids2, weight_dtype + ) + + def get_image_size(self, image_path): + return imagesize.get(image_path) + + def load_image_with_face_info(self, subset: BaseSubset, image_path: str, alpha_mask=False): + img = load_image(image_path, alpha_mask) + + face_cx = face_cy = face_w = face_h = 0 + if subset.face_crop_aug_range is not None: + tokens = os.path.splitext(os.path.basename(image_path))[0].split("_") + if len(tokens) >= 5: + face_cx = int(tokens[-4]) + face_cy = int(tokens[-3]) + face_w = int(tokens[-2]) + face_h = int(tokens[-1]) + + return img, face_cx, face_cy, face_w, face_h + + # いい感じに切り出す + def crop_target(self, subset: BaseSubset, image, face_cx, face_cy, face_w, face_h): + height, width = image.shape[0:2] + if height == self.height and width == self.width: + return image + + # 画像サイズはsizeより大きいのでリサイズする + face_size = max(face_w, face_h) + size = min(self.height, self.width) # 短いほう + min_scale = max(self.height / height, self.width / width) # 画像がモデル入力サイズぴったりになる倍率（最小の倍率） + min_scale = min(1.0, max(min_scale, size / (face_size * subset.face_crop_aug_range[1]))) # 指定した顔最小サイズ + max_scale = min(1.0, max(min_scale, size / (face_size * subset.face_crop_aug_range[0]))) # 指定した顔最大サイズ + if min_scale >= max_scale: # range指定がmin==max + scale = min_scale + else: + scale = random.uniform(min_scale, max_scale) + + nh = int(height * scale + 0.5) + nw = int(width * scale + 0.5) + assert nh >= self.height and nw >= self.width, f"internal error. small scale {scale}, {width}*{height}" + image = cv2.resize(image, (nw, nh), interpolation=cv2.INTER_AREA) + face_cx = int(face_cx * scale + 0.5) + face_cy = int(face_cy * scale + 0.5) + height, width = nh, nw + + # 顔を中心として448*640とかへ切り出す + for axis, (target_size, length, face_p) in enumerate(zip((self.height, self.width), (height, width), (face_cy, face_cx))): + p1 = face_p - target_size // 2 # 顔を中心に持ってくるための切り出し位置 + + if subset.random_crop: + # 背景も含めるために顔を中心に置く確率を高めつつずらす + range = max(length - face_p, face_p) # 画像の端から顔中心までの距離の長いほう + p1 = p1 + (random.randint(0, range) + random.randint(0, range)) - range # -range ~ +range までのいい感じの乱数 + else: + # range指定があるときのみ、すこしだけランダムに（わりと適当） + if subset.face_crop_aug_range[0] != subset.face_crop_aug_range[1]: + if face_size > size // 10 and face_size >= 40: + p1 = p1 + random.randint(-face_size // 20, +face_size // 20) + + p1 = max(0, min(p1, length - target_size)) + + if axis == 0: + image = image[p1 : p1 + target_size, :] + else: + image = image[:, p1 : p1 + target_size] + + return image + + def __len__(self): + return self._length + + def __getitem__(self, index): + bucket = self.bucket_manager.buckets[self.buckets_indices[index].bucket_index] + bucket_batch_size = self.buckets_indices[index].bucket_batch_size + image_index = self.buckets_indices[index].batch_index * bucket_batch_size + + if self.caching_mode is not None: # return batch for latents/text encoder outputs caching + return self.get_item_for_caching(bucket, bucket_batch_size, image_index) + + loss_weights = [] + captions = [] + input_ids_list = [] + input_ids2_list = [] + latents_list = [] + alpha_mask_list = [] + images = [] + original_sizes_hw = [] + crop_top_lefts = [] + target_sizes_hw = [] + flippeds = [] # 変数名が微妙 + text_encoder_outputs1_list = [] + text_encoder_outputs2_list = [] + text_encoder_pool2_list = [] + + for image_key in bucket[image_index : image_index + bucket_batch_size]: + image_info = self.image_data[image_key] + subset = self.image_to_subset[image_key] + loss_weights.append( + self.prior_loss_weight if image_info.is_reg else 1.0 + ) # in case of fine tuning, is_reg is always False + + flipped = subset.flip_aug and random.random() < 0.5 # not flipped or flipped with 50% chance + + # image/latentsを処理する + if image_info.latents is not None: # cache_latents=Trueの場合 + original_size = image_info.latents_original_size + crop_ltrb = image_info.latents_crop_ltrb # calc values later if flipped + if not flipped: + latents = image_info.latents + alpha_mask = image_info.alpha_mask + else: + latents = image_info.latents_flipped + alpha_mask = None if image_info.alpha_mask is None else torch.flip(image_info.alpha_mask, [1]) + + image = None + elif image_info.latents_npz is not None: # FineTuningDatasetまたはcache_latents_to_disk=Trueの場合 + latents, original_size, crop_ltrb, flipped_latents, alpha_mask = load_latents_from_disk(image_info.latents_npz) + if flipped: + latents = flipped_latents + alpha_mask = None if alpha_mask is None else alpha_mask[:, ::-1].copy() # copy to avoid negative stride problem + del flipped_latents + latents = torch.FloatTensor(latents) + if alpha_mask is not None: + alpha_mask = torch.FloatTensor(alpha_mask) + + image = None + else: + # 画像を読み込み、必要ならcropする + img, face_cx, face_cy, face_w, face_h = self.load_image_with_face_info( + subset, image_info.absolute_path, subset.alpha_mask + ) + im_h, im_w = img.shape[0:2] + + if self.enable_bucket: + img, original_size, crop_ltrb = trim_and_resize_if_required( + subset.random_crop, img, image_info.bucket_reso, image_info.resized_size + ) + else: + if face_cx > 0: # 顔位置情報あり + img = self.crop_target(subset, img, face_cx, face_cy, face_w, face_h) + elif im_h > self.height or im_w > self.width: + assert ( + subset.random_crop + ), f"image too large, but cropping and bucketing are disabled / 画像サイズが大きいのでface_crop_aug_rangeかrandom_crop、またはbucketを有効にしてください: {image_info.absolute_path}" + if im_h > self.height: + p = random.randint(0, im_h - self.height) + img = img[p : p + self.height] + if im_w > self.width: + p = random.randint(0, im_w - self.width) + img = img[:, p : p + self.width] + + im_h, im_w = img.shape[0:2] + assert ( + im_h == self.height and im_w == self.width + ), f"image size is small / 画像サイズが小さいようです: {image_info.absolute_path}" + + original_size = [im_w, im_h] + crop_ltrb = (0, 0, 0, 0) + + # augmentation + aug = self.aug_helper.get_augmentor(subset.color_aug) + if aug is not None: + # augment RGB channels only + img_rgb = img[:, :, :3] + img_rgb = aug(image=img_rgb)["image"] + img[:, :, :3] = img_rgb + + if flipped: + img = img[:, ::-1, :].copy() # copy to avoid negative stride problem + + if subset.alpha_mask: + if img.shape[2] == 4: + alpha_mask = img[:, :, 3] # [H,W] + alpha_mask = alpha_mask.astype(np.float32) / 255.0 # 0.0~1.0 + alpha_mask = torch.FloatTensor(alpha_mask) + else: + alpha_mask = torch.ones((img.shape[0], img.shape[1]), dtype=torch.float32) + else: + alpha_mask = None + + img = img[:, :, :3] # remove alpha channel + + latents = None + image = self.image_transforms(img) # -1.0~1.0のtorch.Tensorになる + del img + + images.append(image) + latents_list.append(latents) + alpha_mask_list.append(alpha_mask) + + target_size = (image.shape[2], image.shape[1]) if image is not None else (latents.shape[2] * 8, latents.shape[1] * 8) + + if not flipped: + crop_left_top = (crop_ltrb[0], crop_ltrb[1]) + else: + # crop_ltrb[2] is right, so target_size[0] - crop_ltrb[2] is left in flipped image + crop_left_top = (target_size[0] - crop_ltrb[2], crop_ltrb[1]) + + original_sizes_hw.append((int(original_size[1]), int(original_size[0]))) + crop_top_lefts.append((int(crop_left_top[1]), int(crop_left_top[0]))) + target_sizes_hw.append((int(target_size[1]), int(target_size[0]))) + flippeds.append(flipped) + + # captionとtext encoder outputを処理する + caption = image_info.caption # default + if image_info.text_encoder_outputs1 is not None: + text_encoder_outputs1_list.append(image_info.text_encoder_outputs1) + text_encoder_outputs2_list.append(image_info.text_encoder_outputs2) + text_encoder_pool2_list.append(image_info.text_encoder_pool2) + captions.append(caption) + elif image_info.text_encoder_outputs_npz is not None: + text_encoder_outputs1, text_encoder_outputs2, text_encoder_pool2 = load_text_encoder_outputs_from_disk( + image_info.text_encoder_outputs_npz + ) + text_encoder_outputs1_list.append(text_encoder_outputs1) + text_encoder_outputs2_list.append(text_encoder_outputs2) + text_encoder_pool2_list.append(text_encoder_pool2) + captions.append(caption) + else: + caption = self.process_caption(subset, image_info.caption) + if self.XTI_layers: + caption_layer = [] + for layer in self.XTI_layers: + token_strings_from = " ".join(self.token_strings) + token_strings_to = " ".join([f"{x}_{layer}" for x in self.token_strings]) + caption_ = caption.replace(token_strings_from, token_strings_to) + caption_layer.append(caption_) + captions.append(caption_layer) + else: + captions.append(caption) + + if not self.token_padding_disabled: # this option might be omitted in future + if self.XTI_layers: + token_caption = self.get_input_ids(caption_layer, self.tokenizers[0]) + else: + token_caption = self.get_input_ids(caption, self.tokenizers[0]) + input_ids_list.append(token_caption) + + if len(self.tokenizers) > 1: + if self.XTI_layers: + token_caption2 = self.get_input_ids(caption_layer, self.tokenizers[1]) + else: + token_caption2 = self.get_input_ids(caption, self.tokenizers[1]) + input_ids2_list.append(token_caption2) + + example = {} + example["loss_weights"] = torch.FloatTensor(loss_weights) + + if len(text_encoder_outputs1_list) == 0: + if self.token_padding_disabled: + # padding=True means pad in the batch + example["input_ids"] = self.tokenizer[0](captions, padding=True, truncation=True, return_tensors="pt").input_ids + if len(self.tokenizers) > 1: + example["input_ids2"] = self.tokenizer[1]( + captions, padding=True, truncation=True, return_tensors="pt" + ).input_ids + else: + example["input_ids2"] = None + else: + example["input_ids"] = torch.stack(input_ids_list) + example["input_ids2"] = torch.stack(input_ids2_list) if len(self.tokenizers) > 1 else None + example["text_encoder_outputs1_list"] = None + example["text_encoder_outputs2_list"] = None + example["text_encoder_pool2_list"] = None + else: + example["input_ids"] = None + example["input_ids2"] = None + # # for assertion + # example["input_ids"] = torch.stack([self.get_input_ids(cap, self.tokenizers[0]) for cap in captions]) + # example["input_ids2"] = torch.stack([self.get_input_ids(cap, self.tokenizers[1]) for cap in captions]) + example["text_encoder_outputs1_list"] = torch.stack(text_encoder_outputs1_list) + example["text_encoder_outputs2_list"] = torch.stack(text_encoder_outputs2_list) + example["text_encoder_pool2_list"] = torch.stack(text_encoder_pool2_list) + + # if one of alpha_masks is not None, we need to replace None with ones + none_or_not = [x is None for x in alpha_mask_list] + if all(none_or_not): + example["alpha_masks"] = None + elif any(none_or_not): + for i in range(len(alpha_mask_list)): + if alpha_mask_list[i] is None: + if images[i] is not None: + alpha_mask_list[i] = torch.ones((images[i].shape[1], images[i].shape[2]), dtype=torch.float32) + else: + alpha_mask_list[i] = torch.ones( + (latents_list[i].shape[1] * 8, latents_list[i].shape[2] * 8), dtype=torch.float32 + ) + example["alpha_masks"] = torch.stack(alpha_mask_list) + else: + example["alpha_masks"] = torch.stack(alpha_mask_list) + + if images[0] is not None: + images = torch.stack(images) + images = images.to(memory_format=torch.contiguous_format).float() + else: + images = None + example["images"] = images + + example["latents"] = torch.stack(latents_list) if latents_list[0] is not None else None + example["captions"] = captions + + example["original_sizes_hw"] = torch.stack([torch.LongTensor(x) for x in original_sizes_hw]) + example["crop_top_lefts"] = torch.stack([torch.LongTensor(x) for x in crop_top_lefts]) + example["target_sizes_hw"] = torch.stack([torch.LongTensor(x) for x in target_sizes_hw]) + example["flippeds"] = flippeds + + example["network_multipliers"] = torch.FloatTensor([self.network_multiplier] * len(captions)) + + if self.debug_dataset: + example["image_keys"] = bucket[image_index : image_index + self.batch_size] + return example + + def get_item_for_caching(self, bucket, bucket_batch_size, image_index): + captions = [] + images = [] + input_ids1_list = [] + input_ids2_list = [] + absolute_paths = [] + resized_sizes = [] + bucket_reso = None + flip_aug = None + alpha_mask = None + random_crop = None + + for image_key in bucket[image_index : image_index + bucket_batch_size]: + image_info = self.image_data[image_key] + subset = self.image_to_subset[image_key] + + if flip_aug is None: + flip_aug = subset.flip_aug + alpha_mask = subset.alpha_mask + random_crop = subset.random_crop + bucket_reso = image_info.bucket_reso + else: + # TODO そもそも混在してても動くようにしたほうがいい + assert flip_aug == subset.flip_aug, "flip_aug must be same in a batch" + assert alpha_mask == subset.alpha_mask, "alpha_mask must be same in a batch" + assert random_crop == subset.random_crop, "random_crop must be same in a batch" + assert bucket_reso == image_info.bucket_reso, "bucket_reso must be same in a batch" + + caption = image_info.caption # TODO cache some patterns of dropping, shuffling, etc. + + if self.caching_mode == "latents": + image = load_image(image_info.absolute_path) + else: + image = None + + if self.caching_mode == "text": + input_ids1 = self.get_input_ids(caption, self.tokenizers[0]) + input_ids2 = self.get_input_ids(caption, self.tokenizers[1]) + else: + input_ids1 = None + input_ids2 = None + + captions.append(caption) + images.append(image) + input_ids1_list.append(input_ids1) + input_ids2_list.append(input_ids2) + absolute_paths.append(image_info.absolute_path) + resized_sizes.append(image_info.resized_size) + + example = {} + + if images[0] is None: + images = None + example["images"] = images + + example["captions"] = captions + example["input_ids1_list"] = input_ids1_list + example["input_ids2_list"] = input_ids2_list + example["absolute_paths"] = absolute_paths + example["resized_sizes"] = resized_sizes + example["flip_aug"] = flip_aug + example["alpha_mask"] = alpha_mask + example["random_crop"] = random_crop + example["bucket_reso"] = bucket_reso + return example + + +class DreamBoothDataset(BaseDataset): + IMAGE_INFO_CACHE_FILE = "metadata_cache.json" + + def __init__( + self, + subsets: Sequence[DreamBoothSubset], + batch_size: int, + tokenizer, + max_token_length, + resolution, + network_multiplier: float, + enable_bucket: bool, + min_bucket_reso: int, + max_bucket_reso: int, + bucket_reso_steps: int, + bucket_no_upscale: bool, + prior_loss_weight: float, + debug_dataset: bool, + ) -> None: + super().__init__(tokenizer, max_token_length, resolution, network_multiplier, debug_dataset) + + assert resolution is not None, f"resolution is required / resolution（解像度）指定は必須です" + + self.batch_size = batch_size + self.size = min(self.width, self.height) # 短いほう + self.prior_loss_weight = prior_loss_weight + self.latents_cache = None + + self.enable_bucket = enable_bucket + if self.enable_bucket: + min_bucket_reso, max_bucket_reso = self.adjust_min_max_bucket_reso_by_steps( + resolution, min_bucket_reso, max_bucket_reso, bucket_reso_steps + ) + self.min_bucket_reso = min_bucket_reso + self.max_bucket_reso = max_bucket_reso + self.bucket_reso_steps = bucket_reso_steps + self.bucket_no_upscale = bucket_no_upscale + else: + self.min_bucket_reso = None + self.max_bucket_reso = None + self.bucket_reso_steps = None # この情報は使われない + self.bucket_no_upscale = False + + def read_caption(img_path, caption_extension, enable_wildcard): + # captionの候補ファイル名を作る + base_name = os.path.splitext(img_path)[0] + base_name_face_det = base_name + tokens = base_name.split("_") + if len(tokens) >= 5: + base_name_face_det = "_".join(tokens[:-4]) + cap_paths = [base_name + caption_extension, base_name_face_det + caption_extension] + + caption = None + for cap_path in cap_paths: + if os.path.isfile(cap_path): + with open(cap_path, "rt", encoding="utf-8") as f: + try: + lines = f.readlines() + except UnicodeDecodeError as e: + logger.error(f"illegal char in file (not UTF-8) / ファイルにUTF-8以外の文字があります: {cap_path}") + raise e + assert len(lines) > 0, f"caption file is empty / キャプションファイルが空です: {cap_path}" + if enable_wildcard: + caption = "\n".join([line.strip() for line in lines if line.strip() != ""]) # 空行を除く、改行で連結 + else: + caption = lines[0].strip() + break + return caption + + def load_dreambooth_dir(subset: DreamBoothSubset): + if not os.path.isdir(subset.image_dir): + logger.warning(f"not directory: {subset.image_dir}") + return [], [], [] + + info_cache_file = os.path.join(subset.image_dir, self.IMAGE_INFO_CACHE_FILE) + use_cached_info_for_subset = subset.cache_info + if use_cached_info_for_subset: + logger.info( + f"using cached image info for this subset / このサブセットで、キャッシュされた画像情報を使います: {info_cache_file}" + ) + if not os.path.isfile(info_cache_file): + logger.warning( + f"image info file not found. You can ignore this warning if this is the first time to use this subset" + + " / キャッシュファイルが見つかりませんでした。初回実行時はこの警告を無視してください: {metadata_file}" + ) + use_cached_info_for_subset = False + + if use_cached_info_for_subset: + # json: {`img_path`:{"caption": "caption...", "resolution": [width, height]}, ...} + with open(info_cache_file, "r", encoding="utf-8") as f: + metas = json.load(f) + img_paths = list(metas.keys()) + sizes = [meta["resolution"] for meta in metas.values()] + + # we may need to check image size and existence of image files, but it takes time, so user should check it before training + else: + img_paths = glob_images(subset.image_dir, "*") + sizes = [None] * len(img_paths) + + logger.info(f"found directory {subset.image_dir} contains {len(img_paths)} image files") + + if use_cached_info_for_subset: + captions = [meta["caption"] for meta in metas.values()] + missing_captions = [img_path for img_path, caption in zip(img_paths, captions) if caption is None or caption == ""] + else: + # 画像ファイルごとにプロンプトを読み込み、もしあればそちらを使う + captions = [] + missing_captions = [] + for img_path in img_paths: + cap_for_img = read_caption(img_path, subset.caption_extension, subset.enable_wildcard) + if cap_for_img is None and subset.class_tokens is None: + logger.warning( + f"neither caption file nor class tokens are found. use empty caption for {img_path} / キャプションファイルもclass tokenも見つかりませんでした。空のキャプションを使用します: {img_path}" + ) + captions.append("") + missing_captions.append(img_path) + else: + if cap_for_img is None: + captions.append(subset.class_tokens) + missing_captions.append(img_path) + else: + captions.append(cap_for_img) + + self.set_tag_frequency(os.path.basename(subset.image_dir), captions) # タグ頻度を記録 + + if missing_captions: + number_of_missing_captions = len(missing_captions) + number_of_missing_captions_to_show = 5 + remaining_missing_captions = number_of_missing_captions - number_of_missing_captions_to_show + + logger.warning( + f"No caption file found for {number_of_missing_captions} images. Training will continue without captions for these images. If class token exists, it will be used. / {number_of_missing_captions}枚の画像にキャプションファイルが見つかりませんでした。これらの画像についてはキャプションなしで学習を続行します。class tokenが存在する場合はそれを使います。" + ) + for i, missing_caption in enumerate(missing_captions): + if i >= number_of_missing_captions_to_show: + logger.warning(missing_caption + f"... and {remaining_missing_captions} more") + break + logger.warning(missing_caption) + + if not use_cached_info_for_subset and subset.cache_info: + logger.info(f"cache image info for / 画像情報をキャッシュします : {info_cache_file}") + sizes = [self.get_image_size(img_path) for img_path in tqdm(img_paths, desc="get image size")] + matas = {} + for img_path, caption, size in zip(img_paths, captions, sizes): + matas[img_path] = {"caption": caption, "resolution": list(size)} + with open(info_cache_file, "w", encoding="utf-8") as f: + json.dump(matas, f, ensure_ascii=False, indent=2) + logger.info(f"cache image info done for / 画像情報を出力しました : {info_cache_file}") + + # if sizes are not set, image size will be read in make_buckets + return img_paths, captions, sizes + + logger.info("prepare images.") + num_train_images = 0 + num_reg_images = 0 + reg_infos: List[Tuple[ImageInfo, DreamBoothSubset]] = [] + for subset in subsets: + if subset.num_repeats < 1: + logger.warning( + f"ignore subset with image_dir='{subset.image_dir}': num_repeats is less than 1 / num_repeatsが1を下回っているためサブセットを無視します: {subset.num_repeats}" + ) + continue + + if subset in self.subsets: + logger.warning( + f"ignore duplicated subset with image_dir='{subset.image_dir}': use the first one / 既にサブセットが登録されているため、重複した後発のサブセットを無視します" + ) + continue + + img_paths, captions, sizes = load_dreambooth_dir(subset) + if len(img_paths) < 1: + logger.warning( + f"ignore subset with image_dir='{subset.image_dir}': no images found / 画像が見つからないためサブセットを無視します" + ) + continue + + if subset.is_reg: + num_reg_images += subset.num_repeats * len(img_paths) + else: + num_train_images += subset.num_repeats * len(img_paths) + + for img_path, caption, size in zip(img_paths, captions, sizes): + info = ImageInfo(img_path, subset.num_repeats, caption, subset.is_reg, img_path) + if size is not None: + info.image_size = size + if subset.is_reg: + reg_infos.append((info, subset)) + else: + self.register_image(info, subset) + + subset.img_count = len(img_paths) + self.subsets.append(subset) + + logger.info(f"{num_train_images} train images with repeating.") + self.num_train_images = num_train_images + + logger.info(f"{num_reg_images} reg images.") + if num_train_images < num_reg_images: + logger.warning("some of reg images are not used / 正則化画像の数が多いので、一部使用されない正則化画像があります") + + if num_reg_images == 0: + logger.warning("no regularization images / 正則化画像が見つかりませんでした") + else: + # num_repeatsを計算する：どうせ大した数ではないのでループで処理する + n = 0 + first_loop = True + while n < num_train_images: + for info, subset in reg_infos: + if first_loop: + self.register_image(info, subset) + n += info.num_repeats + else: + info.num_repeats += 1 # rewrite registered info + n += 1 + if n >= num_train_images: + break + first_loop = False + + self.num_reg_images = num_reg_images + + +class FineTuningDataset(BaseDataset): + def __init__( + self, + subsets: Sequence[FineTuningSubset], + batch_size: int, + tokenizer, + max_token_length, + resolution, + network_multiplier: float, + enable_bucket: bool, + min_bucket_reso: int, + max_bucket_reso: int, + bucket_reso_steps: int, + bucket_no_upscale: bool, + debug_dataset: bool, + ) -> None: + super().__init__(tokenizer, max_token_length, resolution, network_multiplier, debug_dataset) + + self.batch_size = batch_size + + self.num_train_images = 0 + self.num_reg_images = 0 + + for subset in subsets: + if subset.num_repeats < 1: + logger.warning( + f"ignore subset with metadata_file='{subset.metadata_file}': num_repeats is less than 1 / num_repeatsが1を下回っているためサブセットを無視します: {subset.num_repeats}" + ) + continue + + if subset in self.subsets: + logger.warning( + f"ignore duplicated subset with metadata_file='{subset.metadata_file}': use the first one / 既にサブセットが登録されているため、重複した後発のサブセットを無視します" + ) + continue + + # メタデータを読み込む + if os.path.exists(subset.metadata_file): + logger.info(f"loading existing metadata: {subset.metadata_file}") + with open(subset.metadata_file, "rt", encoding="utf-8") as f: + metadata = json.load(f) + else: + raise ValueError(f"no metadata / メタデータファイルがありません: {subset.metadata_file}") + + if len(metadata) < 1: + logger.warning( + f"ignore subset with '{subset.metadata_file}': no image entries found / 画像に関するデータが見つからないためサブセットを無視します" + ) + continue + + tags_list = [] + for image_key, img_md in metadata.items(): + # path情報を作る + abs_path = None + + # まず画像を優先して探す + if os.path.exists(image_key): + abs_path = image_key + else: + # わりといい加減だがいい方法が思いつかん + paths = glob_images(subset.image_dir, image_key) + if len(paths) > 0: + abs_path = paths[0] + + # なければnpzを探す + if abs_path is None: + if os.path.exists(os.path.splitext(image_key)[0] + ".npz"): + abs_path = os.path.splitext(image_key)[0] + ".npz" + else: + npz_path = os.path.join(subset.image_dir, image_key + ".npz") + if os.path.exists(npz_path): + abs_path = npz_path + + assert abs_path is not None, f"no image / 画像がありません: {image_key}" + + caption = img_md.get("caption") + tags = img_md.get("tags") + if caption is None: + caption = tags # could be multiline + tags = None + + if subset.enable_wildcard: + # tags must be single line + if tags is not None: + tags = tags.replace("\n", subset.caption_separator) + + # add tags to each line of caption + if caption is not None and tags is not None: + caption = "\n".join( + [f"{line}{subset.caption_separator}{tags}" for line in caption.split("\n") if line.strip() != ""] + ) + else: + # use as is + if tags is not None and len(tags) > 0: + caption = caption + subset.caption_separator + tags + tags_list.append(tags) + + if caption is None: + caption = "" + + image_info = ImageInfo(image_key, subset.num_repeats, caption, False, abs_path) + image_info.image_size = img_md.get("train_resolution") + + if not subset.color_aug and not subset.random_crop: + # if npz exists, use them + image_info.latents_npz, image_info.latents_npz_flipped = self.image_key_to_npz_file(subset, image_key) + + self.register_image(image_info, subset) + + self.num_train_images += len(metadata) * subset.num_repeats + + # TODO do not record tag freq when no tag + self.set_tag_frequency(os.path.basename(subset.metadata_file), tags_list) + subset.img_count = len(metadata) + self.subsets.append(subset) + + # check existence of all npz files + use_npz_latents = all([not (subset.color_aug or subset.random_crop) for subset in self.subsets]) + if use_npz_latents: + flip_aug_in_subset = False + npz_any = False + npz_all = True + + for image_info in self.image_data.values(): + subset = self.image_to_subset[image_info.image_key] + + has_npz = image_info.latents_npz is not None + npz_any = npz_any or has_npz + + if subset.flip_aug: + has_npz = has_npz and image_info.latents_npz_flipped is not None + flip_aug_in_subset = True + npz_all = npz_all and has_npz + + if npz_any and not npz_all: + break + + if not npz_any: + use_npz_latents = False + logger.warning(f"npz file does not exist. ignore npz files / npzファイルが見つからないためnpzファイルを無視します") + elif not npz_all: + use_npz_latents = False + logger.warning( + f"some of npz file does not exist. ignore npz files / いくつかのnpzファイルが見つからないためnpzファイルを無視します" + ) + if flip_aug_in_subset: + logger.warning("maybe no flipped files / 反転されたnpzファイルがないのかもしれません") + # else: + # logger.info("npz files are not used with color_aug and/or random_crop / color_augまたはrandom_cropが指定されているためnpzファイルは使用されません") + + # check min/max bucket size + sizes = set() + resos = set() + for image_info in self.image_data.values(): + if image_info.image_size is None: + sizes = None # not calculated + break + sizes.add(image_info.image_size[0]) + sizes.add(image_info.image_size[1]) + resos.add(tuple(image_info.image_size)) + + if sizes is None: + if use_npz_latents: + use_npz_latents = False + logger.warning( + f"npz files exist, but no bucket info in metadata. ignore npz files / メタデータにbucket情報がないためnpzファイルを無視します" + ) + + assert ( + resolution is not None + ), "if metadata doesn't have bucket info, resolution is required / メタデータにbucket情報がない場合はresolutionを指定してください" + + self.enable_bucket = enable_bucket + if self.enable_bucket: + min_bucket_reso, max_bucket_reso = self.adjust_min_max_bucket_reso_by_steps( + resolution, min_bucket_reso, max_bucket_reso, bucket_reso_steps + ) + self.min_bucket_reso = min_bucket_reso + self.max_bucket_reso = max_bucket_reso + self.bucket_reso_steps = bucket_reso_steps + self.bucket_no_upscale = bucket_no_upscale + else: + if not enable_bucket: + logger.info("metadata has bucket info, enable bucketing / メタデータにbucket情報があるためbucketを有効にします") + logger.info("using bucket info in metadata / メタデータ内のbucket情報を使います") + self.enable_bucket = True + + assert ( + not bucket_no_upscale + ), "if metadata has bucket info, bucket reso is precalculated, so bucket_no_upscale cannot be used / メタデータ内にbucket情報がある場合はbucketの解像度は計算済みのため、bucket_no_upscaleは使えません" + + # bucket情報を初期化しておく、make_bucketsで再作成しない + self.bucket_manager = BucketManager(False, None, None, None, None) + self.bucket_manager.set_predefined_resos(resos) + + # npz情報をきれいにしておく + if not use_npz_latents: + for image_info in self.image_data.values(): + image_info.latents_npz = image_info.latents_npz_flipped = None + + def image_key_to_npz_file(self, subset: FineTuningSubset, image_key): + base_name = os.path.splitext(image_key)[0] + npz_file_norm = base_name + ".npz" + + if os.path.exists(npz_file_norm): + # image_key is full path + npz_file_flip = base_name + "_flip.npz" + if not os.path.exists(npz_file_flip): + npz_file_flip = None + return npz_file_norm, npz_file_flip + + # if not full path, check image_dir. if image_dir is None, return None + if subset.image_dir is None: + return None, None + + # image_key is relative path + npz_file_norm = os.path.join(subset.image_dir, image_key + ".npz") + npz_file_flip = os.path.join(subset.image_dir, image_key + "_flip.npz") + + if not os.path.exists(npz_file_norm): + npz_file_norm = None + npz_file_flip = None + elif not os.path.exists(npz_file_flip): + npz_file_flip = None + + return npz_file_norm, npz_file_flip + + +class ControlNetDataset(BaseDataset): + def __init__( + self, + subsets: Sequence[ControlNetSubset], + batch_size: int, + tokenizer, + max_token_length, + resolution, + network_multiplier: float, + enable_bucket: bool, + min_bucket_reso: int, + max_bucket_reso: int, + bucket_reso_steps: int, + bucket_no_upscale: bool, + debug_dataset: float, + ) -> None: + super().__init__(tokenizer, max_token_length, resolution, network_multiplier, debug_dataset) + + db_subsets = [] + for subset in subsets: + assert ( + not subset.random_crop + ), "random_crop is not supported in ControlNetDataset / random_cropはControlNetDatasetではサポートされていません" + db_subset = DreamBoothSubset( + subset.image_dir, + False, + None, + subset.caption_extension, + subset.cache_info, + False, + subset.num_repeats, + subset.shuffle_caption, + subset.caption_separator, + subset.keep_tokens, + subset.keep_tokens_separator, + subset.secondary_separator, + subset.enable_wildcard, + subset.color_aug, + subset.flip_aug, + subset.face_crop_aug_range, + subset.random_crop, + subset.caption_dropout_rate, + subset.caption_dropout_every_n_epochs, + subset.caption_tag_dropout_rate, + subset.caption_prefix, + subset.caption_suffix, + subset.token_warmup_min, + subset.token_warmup_step, + ) + db_subsets.append(db_subset) + + self.dreambooth_dataset_delegate = DreamBoothDataset( + db_subsets, + batch_size, + tokenizer, + max_token_length, + resolution, + network_multiplier, + enable_bucket, + min_bucket_reso, + max_bucket_reso, + bucket_reso_steps, + bucket_no_upscale, + 1.0, + debug_dataset, + ) + + # config_util等から参照される値をいれておく（若干微妙なのでなんとかしたい） + self.image_data = self.dreambooth_dataset_delegate.image_data + self.batch_size = batch_size + self.num_train_images = self.dreambooth_dataset_delegate.num_train_images + self.num_reg_images = self.dreambooth_dataset_delegate.num_reg_images + + # assert all conditioning data exists + missing_imgs = [] + cond_imgs_with_pair = set() + for image_key, info in self.dreambooth_dataset_delegate.image_data.items(): + db_subset = self.dreambooth_dataset_delegate.image_to_subset[image_key] + subset = None + for s in subsets: + if s.image_dir == db_subset.image_dir: + subset = s + break + assert subset is not None, "internal error: subset not found" + + if not os.path.isdir(subset.conditioning_data_dir): + logger.warning(f"not directory: {subset.conditioning_data_dir}") + continue + + img_basename = os.path.splitext(os.path.basename(info.absolute_path))[0] + ctrl_img_path = glob_images(subset.conditioning_data_dir, img_basename) + if len(ctrl_img_path) < 1: + missing_imgs.append(img_basename) + continue + ctrl_img_path = ctrl_img_path[0] + ctrl_img_path = os.path.abspath(ctrl_img_path) # normalize path + + info.cond_img_path = ctrl_img_path + cond_imgs_with_pair.add(os.path.splitext(ctrl_img_path)[0]) # remove extension because Windows is case insensitive + + extra_imgs = [] + for subset in subsets: + conditioning_img_paths = glob_images(subset.conditioning_data_dir, "*") + conditioning_img_paths = [os.path.abspath(p) for p in conditioning_img_paths] # normalize path + extra_imgs.extend([p for p in conditioning_img_paths if os.path.splitext(p)[0] not in cond_imgs_with_pair]) + + assert ( + len(missing_imgs) == 0 + ), f"missing conditioning data for {len(missing_imgs)} images / 制御用画像が見つかりませんでした: {missing_imgs}" + assert ( + len(extra_imgs) == 0 + ), f"extra conditioning data for {len(extra_imgs)} images / 余分な制御用画像があります: {extra_imgs}" + + self.conditioning_image_transforms = IMAGE_TRANSFORMS + + def make_buckets(self): + self.dreambooth_dataset_delegate.make_buckets() + self.bucket_manager = self.dreambooth_dataset_delegate.bucket_manager + self.buckets_indices = self.dreambooth_dataset_delegate.buckets_indices + + def cache_latents(self, vae, vae_batch_size=1, cache_to_disk=False, is_main_process=True): + return self.dreambooth_dataset_delegate.cache_latents(vae, vae_batch_size, cache_to_disk, is_main_process) + + def __len__(self): + return self.dreambooth_dataset_delegate.__len__() + + def __getitem__(self, index): + example = self.dreambooth_dataset_delegate[index] + + bucket = self.dreambooth_dataset_delegate.bucket_manager.buckets[ + self.dreambooth_dataset_delegate.buckets_indices[index].bucket_index + ] + bucket_batch_size = self.dreambooth_dataset_delegate.buckets_indices[index].bucket_batch_size + image_index = self.dreambooth_dataset_delegate.buckets_indices[index].batch_index * bucket_batch_size + + conditioning_images = [] + + for i, image_key in enumerate(bucket[image_index : image_index + bucket_batch_size]): + image_info = self.dreambooth_dataset_delegate.image_data[image_key] + + target_size_hw = example["target_sizes_hw"][i] + original_size_hw = example["original_sizes_hw"][i] + crop_top_left = example["crop_top_lefts"][i] + flipped = example["flippeds"][i] + cond_img = load_image(image_info.cond_img_path) + + if self.dreambooth_dataset_delegate.enable_bucket: + assert ( + cond_img.shape[0] == original_size_hw[0] and cond_img.shape[1] == original_size_hw[1] + ), f"size of conditioning image is not match / 画像サイズが合いません: {image_info.absolute_path}" + cond_img = cv2.resize( + cond_img, image_info.resized_size, interpolation=cv2.INTER_AREA + ) # INTER_AREAでやりたいのでcv2でリサイズ + + # TODO support random crop + # 現在サポートしているcropはrandomではなく中央のみ + h, w = target_size_hw + ct = (cond_img.shape[0] - h) // 2 + cl = (cond_img.shape[1] - w) // 2 + cond_img = cond_img[ct : ct + h, cl : cl + w] + else: + # assert ( + # cond_img.shape[0] == self.height and cond_img.shape[1] == self.width + # ), f"image size is small / 画像サイズが小さいようです: {image_info.absolute_path}" + # resize to target + if cond_img.shape[0] != target_size_hw[0] or cond_img.shape[1] != target_size_hw[1]: + cond_img = pil_resize(cond_img, (int(target_size_hw[1]), int(target_size_hw[0]))) + + if flipped: + cond_img = cond_img[:, ::-1, :].copy() # copy to avoid negative stride + + cond_img = self.conditioning_image_transforms(cond_img) + conditioning_images.append(cond_img) + + example["conditioning_images"] = torch.stack(conditioning_images).to(memory_format=torch.contiguous_format).float() + + return example + + +# behave as Dataset mock +class DatasetGroup(torch.utils.data.ConcatDataset): + def __init__(self, datasets: Sequence[Union[DreamBoothDataset, FineTuningDataset]]): + self.datasets: List[Union[DreamBoothDataset, FineTuningDataset]] + + super().__init__(datasets) + + self.image_data = {} + self.num_train_images = 0 + self.num_reg_images = 0 + + # simply concat together + # TODO: handling image_data key duplication among dataset + # In practical, this is not the big issue because image_data is accessed from outside of dataset only for debug_dataset. + for dataset in datasets: + self.image_data.update(dataset.image_data) + self.num_train_images += dataset.num_train_images + self.num_reg_images += dataset.num_reg_images + + def add_replacement(self, str_from, str_to): + for dataset in self.datasets: + dataset.add_replacement(str_from, str_to) + + # def make_buckets(self): + # for dataset in self.datasets: + # dataset.make_buckets() + + def enable_XTI(self, *args, **kwargs): + for dataset in self.datasets: + dataset.enable_XTI(*args, **kwargs) + + def cache_latents(self, vae, vae_batch_size=1, cache_to_disk=False, is_main_process=True): + for i, dataset in enumerate(self.datasets): + logger.info(f"[Dataset {i}]") + dataset.cache_latents(vae, vae_batch_size, cache_to_disk, is_main_process) + + def cache_text_encoder_outputs( + self, tokenizers, text_encoders, device, weight_dtype, cache_to_disk=False, is_main_process=True + ): + for i, dataset in enumerate(self.datasets): + logger.info(f"[Dataset {i}]") + dataset.cache_text_encoder_outputs(tokenizers, text_encoders, device, weight_dtype, cache_to_disk, is_main_process) + + def set_caching_mode(self, caching_mode): + for dataset in self.datasets: + dataset.set_caching_mode(caching_mode) + + def verify_bucket_reso_steps(self, min_steps: int): + for dataset in self.datasets: + dataset.verify_bucket_reso_steps(min_steps) + + def is_latent_cacheable(self) -> bool: + return all([dataset.is_latent_cacheable() for dataset in self.datasets]) + + def is_text_encoder_output_cacheable(self) -> bool: + return all([dataset.is_text_encoder_output_cacheable() for dataset in self.datasets]) + + def set_current_epoch(self, epoch): + for dataset in self.datasets: + dataset.set_current_epoch(epoch) + + def set_current_step(self, step): + for dataset in self.datasets: + dataset.set_current_step(step) + + def set_max_train_steps(self, max_train_steps): + for dataset in self.datasets: + dataset.set_max_train_steps(max_train_steps) + + def disable_token_padding(self): + for dataset in self.datasets: + dataset.disable_token_padding() + + +def is_disk_cached_latents_is_expected(reso, npz_path: str, flip_aug: bool, alpha_mask: bool): + expected_latents_size = (reso[1] // 8, reso[0] // 8) # bucket_resoはWxHなので注意 + + if not os.path.exists(npz_path): + return False + + try: + npz = np.load(npz_path) + if "latents" not in npz or "original_size" not in npz or "crop_ltrb" not in npz: # old ver? + return False + if npz["latents"].shape[1:3] != expected_latents_size: + return False + + if flip_aug: + if "latents_flipped" not in npz: + return False + if npz["latents_flipped"].shape[1:3] != expected_latents_size: + return False + + if alpha_mask: + if "alpha_mask" not in npz: + return False + if (npz["alpha_mask"].shape[1], npz["alpha_mask"].shape[0]) != reso: # HxW => WxH != reso + return False + else: + if "alpha_mask" in npz: + return False + except Exception as e: + logger.error(f"Error loading file: {npz_path}") + raise e + + return True + + +# 戻り値は、latents_tensor, (original_size width, original_size height), (crop left, crop top) +def load_latents_from_disk( + npz_path, +) -> Tuple[Optional[np.ndarray], Optional[List[int]], Optional[List[int]], Optional[np.ndarray], Optional[np.ndarray]]: + npz = np.load(npz_path) + if "latents" not in npz: + raise ValueError(f"error: npz is old format. please re-generate {npz_path}") + + latents = npz["latents"] + original_size = npz["original_size"].tolist() + crop_ltrb = npz["crop_ltrb"].tolist() + flipped_latents = npz["latents_flipped"] if "latents_flipped" in npz else None + alpha_mask = npz["alpha_mask"] if "alpha_mask" in npz else None + return latents, original_size, crop_ltrb, flipped_latents, alpha_mask + + +def save_latents_to_disk(npz_path, latents_tensor, original_size, crop_ltrb, flipped_latents_tensor=None, alpha_mask=None): + kwargs = {} + if flipped_latents_tensor is not None: + kwargs["latents_flipped"] = flipped_latents_tensor.float().cpu().numpy() + if alpha_mask is not None: + kwargs["alpha_mask"] = alpha_mask.float().cpu().numpy() + np.savez( + npz_path, + latents=latents_tensor.float().cpu().numpy(), + original_size=np.array(original_size), + crop_ltrb=np.array(crop_ltrb), + **kwargs, + ) + + +def debug_dataset(train_dataset, show_input_ids=False): + logger.info(f"Total dataset length (steps) / データセットの長さ（ステップ数）: {len(train_dataset)}") + logger.info( + "`S` for next step, `E` for next epoch no. , Escape for exit. / Sキーで次のステップ、Eキーで次のエポック、Escキーで中断、終了します" + ) + + epoch = 1 + while True: + logger.info(f"") + logger.info(f"epoch: {epoch}") + + steps = (epoch - 1) * len(train_dataset) + 1 + indices = list(range(len(train_dataset))) + random.shuffle(indices) + + k = 0 + for i, idx in enumerate(indices): + train_dataset.set_current_epoch(epoch) + train_dataset.set_current_step(steps) + logger.info(f"steps: {steps} ({i + 1}/{len(train_dataset)})") + + example = train_dataset[idx] + if example["latents"] is not None: + logger.info(f"sample has latents from npz file: {example['latents'].size()}") + for j, (ik, cap, lw, iid, orgsz, crptl, trgsz, flpdz) in enumerate( + zip( + example["image_keys"], + example["captions"], + example["loss_weights"], + example["input_ids"], + example["original_sizes_hw"], + example["crop_top_lefts"], + example["target_sizes_hw"], + example["flippeds"], + ) + ): + logger.info( + f'{ik}, size: {train_dataset.image_data[ik].image_size}, loss weight: {lw}, caption: "{cap}", original size: {orgsz}, crop top left: {crptl}, target size: {trgsz}, flipped: {flpdz}' + ) + if "network_multipliers" in example: + print(f"network multiplier: {example['network_multipliers'][j]}") + + if show_input_ids: + logger.info(f"input ids: {iid}") + if "input_ids2" in example: + logger.info(f"input ids2: {example['input_ids2'][j]}") + if example["images"] is not None: + im = example["images"][j] + logger.info(f"image size: {im.size()}") + im = ((im.numpy() + 1.0) * 127.5).astype(np.uint8) + im = np.transpose(im, (1, 2, 0)) # c,H,W -> H,W,c + im = im[:, :, ::-1] # RGB -> BGR (OpenCV) + + if "conditioning_images" in example: + cond_img = example["conditioning_images"][j] + logger.info(f"conditioning image size: {cond_img.size()}") + cond_img = ((cond_img.numpy() + 1.0) * 127.5).astype(np.uint8) + cond_img = np.transpose(cond_img, (1, 2, 0)) + cond_img = cond_img[:, :, ::-1] + if os.name == "nt": + cv2.imshow("cond_img", cond_img) + + if "alpha_masks" in example and example["alpha_masks"] is not None: + alpha_mask = example["alpha_masks"][j] + logger.info(f"alpha mask size: {alpha_mask.size()}") + alpha_mask = (alpha_mask.numpy() * 255.0).astype(np.uint8) + if os.name == "nt": + cv2.imshow("alpha_mask", alpha_mask) + + if os.name == "nt": # only windows + cv2.imshow("img", im) + k = cv2.waitKey() + cv2.destroyAllWindows() + if k == 27 or k == ord("s") or k == ord("e"): + break + steps += 1 + + if k == ord("e"): + break + if k == 27 or (example["images"] is None and i >= 8): + k = 27 + break + if k == 27: + break + + epoch += 1 + + +def glob_images(directory, base="*"): + img_paths = [] + for ext in IMAGE_EXTENSIONS: + if base == "*": + img_paths.extend(glob.glob(os.path.join(glob.escape(directory), base + ext))) + else: + img_paths.extend(glob.glob(glob.escape(os.path.join(directory, base + ext)))) + img_paths = list(set(img_paths)) # 重複を排除 + img_paths.sort() + return img_paths + + +def glob_images_pathlib(dir_path, recursive): + image_paths = [] + if recursive: + for ext in IMAGE_EXTENSIONS: + image_paths += list(dir_path.rglob("*" + ext)) + else: + for ext in IMAGE_EXTENSIONS: + image_paths += list(dir_path.glob("*" + ext)) + image_paths = list(set(image_paths)) # 重複を排除 + image_paths.sort() + return image_paths + + +class MinimalDataset(BaseDataset): + def __init__(self, tokenizer, max_token_length, resolution, network_multiplier, debug_dataset=False): + super().__init__(tokenizer, max_token_length, resolution, network_multiplier, debug_dataset) + + self.num_train_images = 0 # update in subclass + self.num_reg_images = 0 # update in subclass + self.datasets = [self] + self.batch_size = 1 # update in subclass + + self.subsets = [self] + self.num_repeats = 1 # update in subclass if needed + self.img_count = 1 # update in subclass if needed + self.bucket_info = {} + self.is_reg = False + self.image_dir = "dummy" # for metadata + + def verify_bucket_reso_steps(self, min_steps: int): + pass + + def is_latent_cacheable(self) -> bool: + return False + + def __len__(self): + raise NotImplementedError + + # override to avoid shuffling buckets + def set_current_epoch(self, epoch): + self.current_epoch = epoch + + def __getitem__(self, idx): + r""" + The subclass may have image_data for debug_dataset, which is a dict of ImageInfo objects. + + Returns: example like this: + + for i in range(batch_size): + image_key = ... # whatever hashable + image_keys.append(image_key) + + image = ... # PIL Image + img_tensor = self.image_transforms(img) + images.append(img_tensor) + + caption = ... # str + input_ids = self.get_input_ids(caption) + input_ids_list.append(input_ids) + + captions.append(caption) + + images = torch.stack(images, dim=0) + input_ids_list = torch.stack(input_ids_list, dim=0) + example = { + "images": images, + "input_ids": input_ids_list, + "captions": captions, # for debug_dataset + "latents": None, + "image_keys": image_keys, # for debug_dataset + "loss_weights": torch.ones(batch_size, dtype=torch.float32), + } + return example + """ + raise NotImplementedError + + +def load_arbitrary_dataset(args, tokenizer) -> MinimalDataset: + module = ".".join(args.dataset_class.split(".")[:-1]) + dataset_class = args.dataset_class.split(".")[-1] + module = importlib.import_module(module) + dataset_class = getattr(module, dataset_class) + train_dataset_group: MinimalDataset = dataset_class(tokenizer, args.max_token_length, args.resolution, args.debug_dataset) + return train_dataset_group + + +def load_image(image_path, alpha=False): + try: + with Image.open(image_path) as image: + if alpha: + if not image.mode == "RGBA": + image = image.convert("RGBA") + else: + if not image.mode == "RGB": + image = image.convert("RGB") + img = np.array(image, np.uint8) + return img + except (IOError, OSError) as e: + logger.error(f"Error loading file: {image_path}") + raise e + + +# 画像を読み込む。戻り値はnumpy.ndarray,(original width, original height),(crop left, crop top, crop right, crop bottom) +def trim_and_resize_if_required( + random_crop: bool, image: np.ndarray, reso, resized_size: Tuple[int, int] +) -> Tuple[np.ndarray, Tuple[int, int], Tuple[int, int, int, int]]: + image_height, image_width = image.shape[0:2] + original_size = (image_width, image_height) # size before resize + + if image_width != resized_size[0] or image_height != resized_size[1]: + # リサイズする + if image_width > resized_size[0] and image_height > resized_size[1]: + image = cv2.resize(image, resized_size, interpolation=cv2.INTER_AREA) # INTER_AREAでやりたいのでcv2でリサイズ + else: + image = pil_resize(image, resized_size) + + image_height, image_width = image.shape[0:2] + + if image_width > reso[0]: + trim_size = image_width - reso[0] + p = trim_size // 2 if not random_crop else random.randint(0, trim_size) + # logger.info(f"w {trim_size} {p}") + image = image[:, p : p + reso[0]] + if image_height > reso[1]: + trim_size = image_height - reso[1] + p = trim_size // 2 if not random_crop else random.randint(0, trim_size) + # logger.info(f"h {trim_size} {p}) + image = image[p : p + reso[1]] + + # random cropの場合のcropされた値をどうcrop left/topに反映するべきか全くアイデアがない + # I have no idea how to reflect the cropped value in crop left/top in the case of random crop + + crop_ltrb = BucketManager.get_crop_ltrb(reso, original_size) + + assert image.shape[0] == reso[1] and image.shape[1] == reso[0], f"internal error, illegal trimmed size: {image.shape}, {reso}" + return image, original_size, crop_ltrb + + +def cache_batch_latents( + vae: AutoencoderKL, cache_to_disk: bool, image_infos: List[ImageInfo], flip_aug: bool, use_alpha_mask: bool, random_crop: bool +) -> None: + r""" + requires image_infos to have: absolute_path, bucket_reso, resized_size, latents_npz + optionally requires image_infos to have: image + if cache_to_disk is True, set info.latents_npz + flipped latents is also saved if flip_aug is True + if cache_to_disk is False, set info.latents + latents_flipped is also set if flip_aug is True + latents_original_size and latents_crop_ltrb are also set + """ + images = [] + alpha_masks: List[np.ndarray] = [] + for info in image_infos: + image = load_image(info.absolute_path, use_alpha_mask) if info.image is None else np.array(info.image, np.uint8) + # TODO 画像のメタデータが壊れていて、メタデータから割り当てたbucketと実際の画像サイズが一致しない場合があるのでチェック追加要 + image, original_size, crop_ltrb = trim_and_resize_if_required(random_crop, image, info.bucket_reso, info.resized_size) + + info.latents_original_size = original_size + info.latents_crop_ltrb = crop_ltrb + + if use_alpha_mask: + if image.shape[2] == 4: + alpha_mask = image[:, :, 3] # [H,W] + alpha_mask = alpha_mask.astype(np.float32) / 255.0 + alpha_mask = torch.FloatTensor(alpha_mask) # [H,W] + else: + alpha_mask = torch.ones_like(image[:, :, 0], dtype=torch.float32) # [H,W] + else: + alpha_mask = None + alpha_masks.append(alpha_mask) + + image = image[:, :, :3] # remove alpha channel if exists + image = IMAGE_TRANSFORMS(image) + images.append(image) + + img_tensors = torch.stack(images, dim=0) + img_tensors = img_tensors.to(device=vae.device, dtype=vae.dtype) + + with torch.no_grad(): + latents = vae.encode(img_tensors).latent_dist.sample().to("cpu") + + if flip_aug: + img_tensors = torch.flip(img_tensors, dims=[3]) + with torch.no_grad(): + flipped_latents = vae.encode(img_tensors).latent_dist.sample().to("cpu") + else: + flipped_latents = [None] * len(latents) + + for info, latent, flipped_latent, alpha_mask in zip(image_infos, latents, flipped_latents, alpha_masks): + # check NaN + if torch.isnan(latents).any() or (flipped_latent is not None and torch.isnan(flipped_latent).any()): + raise RuntimeError(f"NaN detected in latents: {info.absolute_path}") + + if cache_to_disk: + save_latents_to_disk( + info.latents_npz, + latent, + info.latents_original_size, + info.latents_crop_ltrb, + flipped_latent, + alpha_mask, + ) + else: + info.latents = latent + if flip_aug: + info.latents_flipped = flipped_latent + info.alpha_mask = alpha_mask + + if not HIGH_VRAM: + clean_memory_on_device(vae.device) + + +def cache_batch_text_encoder_outputs( + image_infos, tokenizers, text_encoders, max_token_length, cache_to_disk, input_ids1, input_ids2, dtype +): + input_ids1 = input_ids1.to(text_encoders[0].device) + input_ids2 = input_ids2.to(text_encoders[1].device) + + with torch.no_grad(): + b_hidden_state1, b_hidden_state2, b_pool2 = get_hidden_states_sdxl( + max_token_length, + input_ids1, + input_ids2, + tokenizers[0], + tokenizers[1], + text_encoders[0], + text_encoders[1], + dtype, + ) + + # ここでcpuに移動しておかないと、上書きされてしまう + b_hidden_state1 = b_hidden_state1.detach().to("cpu") # b,n*75+2,768 + b_hidden_state2 = b_hidden_state2.detach().to("cpu") # b,n*75+2,1280 + b_pool2 = b_pool2.detach().to("cpu") # b,1280 + + for info, hidden_state1, hidden_state2, pool2 in zip(image_infos, b_hidden_state1, b_hidden_state2, b_pool2): + if cache_to_disk: + save_text_encoder_outputs_to_disk(info.text_encoder_outputs_npz, hidden_state1, hidden_state2, pool2) + else: + info.text_encoder_outputs1 = hidden_state1 + info.text_encoder_outputs2 = hidden_state2 + info.text_encoder_pool2 = pool2 + + +def save_text_encoder_outputs_to_disk(npz_path, hidden_state1, hidden_state2, pool2): + np.savez( + npz_path, + hidden_state1=hidden_state1.cpu().float().numpy(), + hidden_state2=hidden_state2.cpu().float().numpy(), + pool2=pool2.cpu().float().numpy(), + ) + + +def load_text_encoder_outputs_from_disk(npz_path): + with np.load(npz_path) as f: + hidden_state1 = torch.from_numpy(f["hidden_state1"]) + hidden_state2 = torch.from_numpy(f["hidden_state2"]) if "hidden_state2" in f else None + pool2 = torch.from_numpy(f["pool2"]) if "pool2" in f else None + return hidden_state1, hidden_state2, pool2 + + +# endregion + +# region モジュール入れ替え部 +""" +高速化のためのモジュール入れ替え +""" + +# FlashAttentionを使うCrossAttention +# based on https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/memory_efficient_attention_pytorch/flash_attention.py +# LICENSE MIT https://github.com/lucidrains/memory-efficient-attention-pytorch/blob/main/LICENSE + +# constants + +EPSILON = 1e-6 + +# helper functions + + +def exists(val): + return val is not None + + +def default(val, d): + return val if exists(val) else d + + +def model_hash(filename): + """Old model hash used by stable-diffusion-webui""" + try: + with open(filename, "rb") as file: + m = hashlib.sha256() + + file.seek(0x100000) + m.update(file.read(0x10000)) + return m.hexdigest()[0:8] + except FileNotFoundError: + return "NOFILE" + except IsADirectoryError: # Linux? + return "IsADirectory" + except PermissionError: # Windows + return "IsADirectory" + + +def calculate_sha256(filename): + """New model hash used by stable-diffusion-webui""" + try: + hash_sha256 = hashlib.sha256() + blksize = 1024 * 1024 + + with open(filename, "rb") as f: + for chunk in iter(lambda: f.read(blksize), b""): + hash_sha256.update(chunk) + + return hash_sha256.hexdigest() + except FileNotFoundError: + return "NOFILE" + except IsADirectoryError: # Linux? + return "IsADirectory" + except PermissionError: # Windows + return "IsADirectory" + + +def precalculate_safetensors_hashes(tensors, metadata): + """Precalculate the model hashes needed by sd-webui-additional-networks to + save time on indexing the model later.""" + + # Because writing user metadata to the file can change the result of + # sd_models.model_hash(), only retain the training metadata for purposes of + # calculating the hash, as they are meant to be immutable + metadata = {k: v for k, v in metadata.items() if k.startswith("ss_")} + + bytes = safetensors.torch.save(tensors, metadata) + b = BytesIO(bytes) + + model_hash = addnet_hash_safetensors(b) + legacy_hash = addnet_hash_legacy(b) + return model_hash, legacy_hash + + +def addnet_hash_legacy(b): + """Old model hash used by sd-webui-additional-networks for .safetensors format files""" + m = hashlib.sha256() + + b.seek(0x100000) + m.update(b.read(0x10000)) + return m.hexdigest()[0:8] + + +def addnet_hash_safetensors(b): + """New model hash used by sd-webui-additional-networks for .safetensors format files""" + hash_sha256 = hashlib.sha256() + blksize = 1024 * 1024 + + b.seek(0) + header = b.read(8) + n = int.from_bytes(header, "little") + + offset = n + 8 + b.seek(offset) + for chunk in iter(lambda: b.read(blksize), b""): + hash_sha256.update(chunk) + + return hash_sha256.hexdigest() + + +def get_git_revision_hash() -> str: + try: + return subprocess.check_output(["git", "rev-parse", "HEAD"], cwd=os.path.dirname(__file__)).decode("ascii").strip() + except: + return "(unknown)" + + +# def replace_unet_modules(unet: diffusers.models.unet_2d_condition.UNet2DConditionModel, mem_eff_attn, xformers): +# replace_attentions_for_hypernetwork() +# # unet is not used currently, but it is here for future use +# unet.enable_xformers_memory_efficient_attention() +# return +# if mem_eff_attn: +# unet.set_attn_processor(FlashAttnProcessor()) +# elif xformers: +# unet.enable_xformers_memory_efficient_attention() + + +# def replace_unet_cross_attn_to_xformers(): +# logger.info("CrossAttention.forward has been replaced to enable xformers.") +# try: +# import xformers.ops +# except ImportError: +# raise ImportError("No xformers / xformersがインストールされていないようです") + +# def forward_xformers(self, x, context=None, mask=None): +# h = self.heads +# q_in = self.to_q(x) + +# context = default(context, x) +# context = context.to(x.dtype) + +# if hasattr(self, "hypernetwork") and self.hypernetwork is not None: +# context_k, context_v = self.hypernetwork.forward(x, context) +# context_k = context_k.to(x.dtype) +# context_v = context_v.to(x.dtype) +# else: +# context_k = context +# context_v = context + +# k_in = self.to_k(context_k) +# v_in = self.to_v(context_v) + +# q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b n h d", h=h), (q_in, k_in, v_in)) +# del q_in, k_in, v_in + +# q = q.contiguous() +# k = k.contiguous() +# v = v.contiguous() +# out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None) # 最適なのを選んでくれる + +# out = rearrange(out, "b n h d -> b n (h d)", h=h) + +# # diffusers 0.7.0~ +# out = self.to_out[0](out) +# out = self.to_out[1](out) +# return out + + +# diffusers.models.attention.CrossAttention.forward = forward_xformers +def replace_unet_modules(unet: UNet2DConditionModel, mem_eff_attn, xformers, sdpa): + if mem_eff_attn: + logger.info("Enable memory efficient attention for U-Net") + unet.set_use_memory_efficient_attention(False, True) + elif xformers: + logger.info("Enable xformers for U-Net") + try: + import xformers.ops + except ImportError: + raise ImportError("No xformers / xformersがインストールされていないようです") + + unet.set_use_memory_efficient_attention(True, False) + elif sdpa: + logger.info("Enable SDPA for U-Net") + unet.set_use_sdpa(True) + + +""" +def replace_vae_modules(vae: diffusers.models.AutoencoderKL, mem_eff_attn, xformers): + # vae is not used currently, but it is here for future use + if mem_eff_attn: + replace_vae_attn_to_memory_efficient() + elif xformers: + # とりあえずDiffusersのxformersを使う。AttentionがあるのはMidBlockのみ + logger.info("Use Diffusers xformers for VAE") + vae.encoder.mid_block.attentions[0].set_use_memory_efficient_attention_xformers(True) + vae.decoder.mid_block.attentions[0].set_use_memory_efficient_attention_xformers(True) + + +def replace_vae_attn_to_memory_efficient(): + logger.info("AttentionBlock.forward has been replaced to FlashAttention (not xformers)") + flash_func = FlashAttentionFunction + + def forward_flash_attn(self, hidden_states): + logger.info("forward_flash_attn") + q_bucket_size = 512 + k_bucket_size = 1024 + + residual = hidden_states + batch, channel, height, width = hidden_states.shape + + # norm + hidden_states = self.group_norm(hidden_states) + + hidden_states = hidden_states.view(batch, channel, height * width).transpose(1, 2) + + # proj to q, k, v + query_proj = self.query(hidden_states) + key_proj = self.key(hidden_states) + value_proj = self.value(hidden_states) + + query_proj, key_proj, value_proj = map( + lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.num_heads), (query_proj, key_proj, value_proj) + ) + + out = flash_func.apply(query_proj, key_proj, value_proj, None, False, q_bucket_size, k_bucket_size) + + out = rearrange(out, "b h n d -> b n (h d)") + + # compute next hidden_states + hidden_states = self.proj_attn(hidden_states) + hidden_states = hidden_states.transpose(-1, -2).reshape(batch, channel, height, width) + + # res connect and rescale + hidden_states = (hidden_states + residual) / self.rescale_output_factor + return hidden_states + + diffusers.models.attention.AttentionBlock.forward = forward_flash_attn +""" + + +# endregion + + +# region arguments + + +def load_metadata_from_safetensors(safetensors_file: str) -> dict: + """r + This method locks the file. see https://github.com/huggingface/safetensors/issues/164 + If the file isn't .safetensors or doesn't have metadata, return empty dict. + """ + if os.path.splitext(safetensors_file)[1] != ".safetensors": + return {} + + with safetensors.safe_open(safetensors_file, framework="pt", device="cpu") as f: + metadata = f.metadata() + if metadata is None: + metadata = {} + return metadata + + +# this metadata is referred from train_network and various scripts, so we wrote here +SS_METADATA_KEY_V2 = "ss_v2" +SS_METADATA_KEY_BASE_MODEL_VERSION = "ss_base_model_version" +SS_METADATA_KEY_NETWORK_MODULE = "ss_network_module" +SS_METADATA_KEY_NETWORK_DIM = "ss_network_dim" +SS_METADATA_KEY_NETWORK_ALPHA = "ss_network_alpha" +SS_METADATA_KEY_NETWORK_ARGS = "ss_network_args" + +SS_METADATA_MINIMUM_KEYS = [ + SS_METADATA_KEY_V2, + SS_METADATA_KEY_BASE_MODEL_VERSION, + SS_METADATA_KEY_NETWORK_MODULE, + SS_METADATA_KEY_NETWORK_DIM, + SS_METADATA_KEY_NETWORK_ALPHA, + SS_METADATA_KEY_NETWORK_ARGS, +] + + +def build_minimum_network_metadata( + v2: Optional[bool], + base_model: Optional[str], + network_module: str, + network_dim: str, + network_alpha: str, + network_args: Optional[dict], +): + # old LoRA doesn't have base_model + metadata = { + SS_METADATA_KEY_NETWORK_MODULE: network_module, + SS_METADATA_KEY_NETWORK_DIM: network_dim, + SS_METADATA_KEY_NETWORK_ALPHA: network_alpha, + } + if v2 is not None: + metadata[SS_METADATA_KEY_V2] = v2 + if base_model is not None: + metadata[SS_METADATA_KEY_BASE_MODEL_VERSION] = base_model + if network_args is not None: + metadata[SS_METADATA_KEY_NETWORK_ARGS] = json.dumps(network_args) + return metadata + + +def get_sai_model_spec( + state_dict: dict, + args: argparse.Namespace, + sdxl: bool, + lora: bool, + textual_inversion: bool, + is_stable_diffusion_ckpt: Optional[bool] = None, # None for TI and LoRA +): + timestamp = time.time() + + v2 = args.v2 + v_parameterization = args.v_parameterization + reso = args.resolution + + title = args.metadata_title if args.metadata_title is not None else args.output_name + + if args.min_timestep is not None or args.max_timestep is not None: + min_time_step = args.min_timestep if args.min_timestep is not None else 0 + max_time_step = args.max_timestep if args.max_timestep is not None else 1000 + timesteps = (min_time_step, max_time_step) + else: + timesteps = None + + metadata = sai_model_spec.build_metadata( + state_dict, + v2, + v_parameterization, + sdxl, + lora, + textual_inversion, + timestamp, + title=title, + reso=reso, + is_stable_diffusion_ckpt=is_stable_diffusion_ckpt, + author=args.metadata_author, + description=args.metadata_description, + license=args.metadata_license, + tags=args.metadata_tags, + timesteps=timesteps, + clip_skip=args.clip_skip, # None or int + ) + return metadata + + +def add_sd_models_arguments(parser: argparse.ArgumentParser): + # for pretrained models + parser.add_argument( + "--v2", action="store_true", help="load Stable Diffusion v2.0 model / Stable Diffusion 2.0のモデルを読み込む" + ) + parser.add_argument( + "--v_parameterization", action="store_true", help="enable v-parameterization training / v-parameterization学習を有効にする" + ) + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + help="pretrained model to train, directory to Diffusers model or StableDiffusion checkpoint / 学習元モデル、Diffusers形式モデルのディレクトリまたはStableDiffusionのckptファイル", + ) + parser.add_argument( + "--tokenizer_cache_dir", + type=str, + default=None, + help="directory for caching Tokenizer (for offline training) / Tokenizerをキャッシュするディレクトリ（ネット接続なしでの学習のため）", + ) + + +def add_optimizer_arguments(parser: argparse.ArgumentParser): + def int_or_float(value): + if value.endswith("%"): + try: + return float(value[:-1]) / 100.0 + except ValueError: + raise argparse.ArgumentTypeError(f"Value '{value}' is not a valid percentage") + try: + float_value = float(value) + if float_value >= 1: + return int(value) + return float(value) + except ValueError: + raise argparse.ArgumentTypeError(f"'{value}' is not an int or float") + + parser.add_argument( + "--optimizer_type", + type=str, + default="", + help="Optimizer to use / オプティマイザの種類: AdamW (default), AdamW8bit, PagedAdamW, PagedAdamW8bit, PagedAdamW32bit, " + "Lion8bit, PagedLion8bit, Lion, SGDNesterov, SGDNesterov8bit, " + "DAdaptation(DAdaptAdamPreprint), DAdaptAdaGrad, DAdaptAdam, DAdaptAdan, DAdaptAdanIP, DAdaptLion, DAdaptSGD, " + "AdaFactor. " + "Also, you can use any optimizer by specifying the full path to the class, like 'bitsandbytes.optim.AdEMAMix8bit' or 'bitsandbytes.optim.PagedAdEMAMix8bit'.", + ) + + # backward compatibility + parser.add_argument( + "--use_8bit_adam", + action="store_true", + help="use 8bit AdamW optimizer (requires bitsandbytes) / 8bit Adamオプティマイザを使う（bitsandbytesのインストールが必要）", + ) + parser.add_argument( + "--use_lion_optimizer", + action="store_true", + help="use Lion optimizer (requires lion-pytorch) / Lionオプティマイザを使う（ lion-pytorch のインストールが必要）", + ) + + parser.add_argument("--learning_rate", type=float, default=2.0e-6, help="learning rate / 学習率") + parser.add_argument( + "--max_grad_norm", + default=1.0, + type=float, + help="Max gradient norm, 0 for no clipping / 勾配正規化の最大norm、0でclippingを行わない", + ) + + parser.add_argument( + "--optimizer_args", + type=str, + default=None, + nargs="*", + help='additional arguments for optimizer (like "weight_decay=0.01 betas=0.9,0.999 ...") / オプティマイザの追加引数（例： "weight_decay=0.01 betas=0.9,0.999 ..."）', + ) + + parser.add_argument("--lr_scheduler_type", type=str, default="", help="custom scheduler module / 使用するスケジューラ") + parser.add_argument( + "--lr_scheduler_args", + type=str, + default=None, + nargs="*", + help='additional arguments for scheduler (like "T_max=100") / スケジューラの追加引数（例： "T_max100"）', + ) + + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help="scheduler to use for learning rate / 学習率のスケジューラ: linear, cosine, cosine_with_restarts, polynomial, constant (default), constant_with_warmup, adafactor", + ) + parser.add_argument( + "--lr_warmup_steps", + type=int_or_float, + default=0, + help="Int number of steps for the warmup in the lr scheduler (default is 0) or float with ratio of train steps" + " / 学習率のスケジューラをウォームアップするステップ数（デフォルト0）、または学習ステップの比率（1未満のfloat値の場合）", + ) + parser.add_argument( + "--lr_decay_steps", + type=int_or_float, + default=0, + help="Int number of steps for the decay in the lr scheduler (default is 0) or float (<1) with ratio of train steps" + " / 学習率のスケジューラを減衰させるステップ数（デフォルト0）、または学習ステップの比率（1未満のfloat値の場合）", + ) + parser.add_argument( + "--lr_scheduler_num_cycles", + type=int, + default=1, + help="Number of restarts for cosine scheduler with restarts / cosine with restartsスケジューラでのリスタート回数", + ) + parser.add_argument( + "--lr_scheduler_power", + type=float, + default=1, + help="Polynomial power for polynomial scheduler / polynomialスケジューラでのpolynomial power", + ) + parser.add_argument( + "--fused_backward_pass", + action="store_true", + help="Combines backward pass and optimizer step to reduce VRAM usage. Only available in SDXL" + + " / バックワードパスとオプティマイザステップを組み合わせてVRAMの使用量を削減します。SDXLでのみ有効", + ) + parser.add_argument( + "--lr_scheduler_timescale", + type=int, + default=None, + help="Inverse sqrt timescale for inverse sqrt scheduler,defaults to `num_warmup_steps`" + + " / 逆平方根スケジューラのタイムスケール、デフォルトは`num_warmup_steps`", + ) + parser.add_argument( + "--lr_scheduler_min_lr_ratio", + type=float, + default=None, + help="The minimum learning rate as a ratio of the initial learning rate for cosine with min lr scheduler and warmup decay scheduler" + + " / 初期学習率の比率としての最小学習率を指定する、cosine with min lr と warmup decay スケジューラ で有効", + ) + + +def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth: bool): + parser.add_argument( + "--output_dir", type=str, default=None, help="directory to output trained model / 学習後のモデル出力先ディレクトリ" + ) + parser.add_argument( + "--output_name", type=str, default=None, help="base name of trained model file / 学習後のモデルの拡張子を除くファイル名" + ) + parser.add_argument( + "--huggingface_repo_id", + type=str, + default=None, + help="huggingface repo name to upload / huggingfaceにアップロードするリポジトリ名", + ) + parser.add_argument( + "--huggingface_repo_type", + type=str, + default=None, + help="huggingface repo type to upload / huggingfaceにアップロードするリポジトリの種類", + ) + parser.add_argument( + "--huggingface_path_in_repo", + type=str, + default=None, + help="huggingface model path to upload files / huggingfaceにアップロードするファイルのパス", + ) + parser.add_argument("--huggingface_token", type=str, default=None, help="huggingface token / huggingfaceのトークン") + parser.add_argument( + "--huggingface_repo_visibility", + type=str, + default=None, + help="huggingface repository visibility ('public' for public, 'private' or None for private) / huggingfaceにアップロードするリポジトリの公開設定（'public'で公開、'private'またはNoneで非公開）", + ) + parser.add_argument( + "--save_state_to_huggingface", action="store_true", help="save state to huggingface / huggingfaceにstateを保存する" + ) + parser.add_argument( + "--resume_from_huggingface", + action="store_true", + help="resume from huggingface (ex: --resume {repo_id}/{path_in_repo}:{revision}:{repo_type}) / huggingfaceから学習を再開する(例: --resume {repo_id}/{path_in_repo}:{revision}:{repo_type})", + ) + parser.add_argument( + "--async_upload", + action="store_true", + help="upload to huggingface asynchronously / huggingfaceに非同期でアップロードする", + ) + parser.add_argument( + "--save_precision", + type=str, + default=None, + choices=[None, "float", "fp16", "bf16"], + help="precision in saving / 保存時に精度を変更して保存する", + ) + parser.add_argument( + "--save_every_n_epochs", + type=int, + default=None, + help="save checkpoint every N epochs / 学習中のモデルを指定エポックごとに保存する", + ) + parser.add_argument( + "--save_every_n_steps", + type=int, + default=None, + help="save checkpoint every N steps / 学習中のモデルを指定ステップごとに保存する", + ) + parser.add_argument( + "--save_n_epoch_ratio", + type=int, + default=None, + help="save checkpoint N epoch ratio (for example 5 means save at least 5 files total) / 学習中のモデルを指定のエポック割合で保存する（たとえば5を指定すると最低5個のファイルが保存される）", + ) + parser.add_argument( + "--save_last_n_epochs", + type=int, + default=None, + help="save last N checkpoints when saving every N epochs (remove older checkpoints) / 指定エポックごとにモデルを保存するとき最大Nエポック保存する（古いチェックポイントは削除する）", + ) + parser.add_argument( + "--save_last_n_epochs_state", + type=int, + default=None, + help="save last N checkpoints of state (overrides the value of --save_last_n_epochs)/ 最大Nエポックstateを保存する（--save_last_n_epochsの指定を上書きする）", + ) + parser.add_argument( + "--save_last_n_steps", + type=int, + default=None, + help="save checkpoints until N steps elapsed (remove older checkpoints if N steps elapsed) / 指定ステップごとにモデルを保存するとき、このステップ数経過するまで保存する（このステップ数経過したら削除する）", + ) + parser.add_argument( + "--save_last_n_steps_state", + type=int, + default=None, + help="save states until N steps elapsed (remove older states if N steps elapsed, overrides --save_last_n_steps) / 指定ステップごとにstateを保存するとき、このステップ数経過するまで保存する（このステップ数経過したら削除する。--save_last_n_stepsを上書きする）", + ) + parser.add_argument( + "--save_state", + action="store_true", + help="save training state additionally (including optimizer states etc.) when saving model / optimizerなど学習状態も含めたstateをモデル保存時に追加で保存する", + ) + parser.add_argument( + "--save_state_on_train_end", + action="store_true", + help="save training state (including optimizer states etc.) on train end / optimizerなど学習状態も含めたstateを学習完了時に保存する", + ) + parser.add_argument("--resume", type=str, default=None, help="saved state to resume training / 学習再開するモデルのstate") + + parser.add_argument("--train_batch_size", type=int, default=1, help="batch size for training / 学習時のバッチサイズ") + parser.add_argument( + "--max_token_length", + type=int, + default=None, + choices=[None, 150, 225], + help="max token length of text encoder (default for 75, 150 or 225) / text encoderのトークンの最大長（未指定で75、150または225が指定可）", + ) + parser.add_argument( + "--mem_eff_attn", + action="store_true", + help="use memory efficient attention for CrossAttention / CrossAttentionに省メモリ版attentionを使う", + ) + parser.add_argument( + "--torch_compile", action="store_true", help="use torch.compile (requires PyTorch 2.0) / torch.compile を使う" + ) + parser.add_argument( + "--dynamo_backend", + type=str, + default="inductor", + # available backends: + # https://github.com/huggingface/accelerate/blob/d1abd59114ada8ba673e1214218cb2878c13b82d/src/accelerate/utils/dataclasses.py#L376-L388C5 + # https://pytorch.org/docs/stable/torch.compiler.html + choices=["eager", "aot_eager", "inductor", "aot_ts_nvfuser", "nvprims_nvfuser", "cudagraphs", "ofi", "fx2trt", "onnxrt"], + help="dynamo backend type (default is inductor) / dynamoのbackendの種類（デフォルトは inductor）", + ) + parser.add_argument("--xformers", action="store_true", help="use xformers for CrossAttention / CrossAttentionにxformersを使う") + parser.add_argument( + "--sdpa", + action="store_true", + help="use sdpa for CrossAttention (requires PyTorch 2.0) / CrossAttentionにsdpaを使う（PyTorch 2.0が必要）", + ) + parser.add_argument( + "--vae", + type=str, + default=None, + help="path to checkpoint of vae to replace / VAEを入れ替える場合、VAEのcheckpointファイルまたはディレクトリ", + ) + + parser.add_argument("--max_train_steps", type=int, default=1600, help="training steps / 学習ステップ数") + parser.add_argument( + "--max_train_epochs", + type=int, + default=None, + help="training epochs (overrides max_train_steps) / 学習エポック数（max_train_stepsを上書きします）", + ) + parser.add_argument( + "--max_data_loader_n_workers", + type=int, + default=8, + help="max num workers for DataLoader (lower is less main RAM usage, faster epoch start and slower data loading) / DataLoaderの最大プロセス数（小さい値ではメインメモリの使用量が減りエポック間の待ち時間が減りますが、データ読み込みは遅くなります）", + ) + parser.add_argument( + "--persistent_data_loader_workers", + action="store_true", + help="persistent DataLoader workers (useful for reduce time gap between epoch, but may use more memory) / DataLoader のワーカーを持続させる (エポック間の時間差を少なくするのに有効だが、より多くのメモリを消費する可能性がある)", + ) + parser.add_argument("--seed", type=int, default=None, help="random seed for training / 学習時の乱数のseed") + parser.add_argument( + "--gradient_checkpointing", action="store_true", help="enable gradient checkpointing / gradient checkpointingを有効にする" + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass / 学習時に逆伝播をする前に勾配を合計するステップ数", + ) + parser.add_argument( + "--mixed_precision", + type=str, + default="no", + choices=["no", "fp16", "bf16"], + help="use mixed precision / 混合精度を使う場合、その精度", + ) + parser.add_argument("--full_fp16", action="store_true", help="fp16 training including gradients / 勾配も含めてfp16で学習する") + parser.add_argument( + "--full_bf16", action="store_true", help="bf16 training including gradients / 勾配も含めてbf16で学習する" + ) # TODO move to SDXL training, because it is not supported by SD1/2 + parser.add_argument("--fp8_base", action="store_true", help="use fp8 for base model / base modelにfp8を使う") + + parser.add_argument( + "--ddp_timeout", + type=int, + default=None, + help="DDP timeout (min, None for default of accelerate) / DDPのタイムアウト（分、Noneでaccelerateのデフォルト）", + ) + parser.add_argument( + "--ddp_gradient_as_bucket_view", + action="store_true", + help="enable gradient_as_bucket_view for DDP / DDPでgradient_as_bucket_viewを有効にする", + ) + parser.add_argument( + "--ddp_static_graph", + action="store_true", + help="enable static_graph for DDP / DDPでstatic_graphを有効にする", + ) + parser.add_argument( + "--clip_skip", + type=int, + default=None, + help="use output of nth layer from back of text encoder (n>=1) / text encoderの後ろからn番目の層の出力を用いる（nは1以上）", + ) + parser.add_argument( + "--logging_dir", + type=str, + default=None, + help="enable logging and output TensorBoard log to this directory / ログ出力を有効にしてこのディレクトリにTensorBoard用のログを出力する", + ) + parser.add_argument( + "--log_with", + type=str, + default=None, + choices=["tensorboard", "wandb", "all"], + help="what logging tool(s) to use (if 'all', TensorBoard and WandB are both used) / ログ出力に使用するツール (allを指定するとTensorBoardとWandBの両方が使用される)", + ) + parser.add_argument( + "--log_prefix", type=str, default=None, help="add prefix for each log directory / ログディレクトリ名の先頭に追加する文字列" + ) + parser.add_argument( + "--log_tracker_name", + type=str, + default=None, + help="name of tracker to use for logging, default is script-specific default name / ログ出力に使用するtrackerの名前、省略時はスクリプトごとのデフォルト名", + ) + parser.add_argument( + "--wandb_run_name", + type=str, + default=None, + help="The name of the specific wandb session / wandb ログに表示される特定の実行の名前", + ) + parser.add_argument( + "--log_tracker_config", + type=str, + default=None, + help="path to tracker config file to use for logging / ログ出力に使用するtrackerの設定ファイルのパス", + ) + parser.add_argument( + "--wandb_api_key", + type=str, + default=None, + help="specify WandB API key to log in before starting training (optional). / WandB APIキーを指定して学習開始前にログインする（オプション）", + ) + parser.add_argument("--log_config", action="store_true", help="log training configuration / 学習設定をログに出力する") + + parser.add_argument( + "--noise_offset", + type=float, + default=None, + help="enable noise offset with this value (if enabled, around 0.1 is recommended) / Noise offsetを有効にしてこの値を設定する（有効にする場合は0.1程度を推奨）", + ) + parser.add_argument( + "--noise_offset_random_strength", + action="store_true", + help="use random strength between 0~noise_offset for noise offset. / noise offsetにおいて、0からnoise_offsetの間でランダムな強度を使用します。", + ) + parser.add_argument( + "--multires_noise_iterations", + type=int, + default=None, + help="enable multires noise with this number of iterations (if enabled, around 6-10 is recommended) / Multires noiseを有効にしてこのイテレーション数を設定する（有効にする場合は6-10程度を推奨）", + ) + parser.add_argument( + "--ip_noise_gamma", + type=float, + default=None, + help="enable input perturbation noise. used for regularization. recommended value: around 0.1 (from arxiv.org/abs/2301.11706) " + + "/ input perturbation noiseを有効にする。正則化に使用される。推奨値: 0.1程度 (arxiv.org/abs/2301.11706 より)", + ) + parser.add_argument( + "--ip_noise_gamma_random_strength", + action="store_true", + help="Use random strength between 0~ip_noise_gamma for input perturbation noise." + + "/ input perturbation noiseにおいて、0からip_noise_gammaの間でランダムな強度を使用します。", + ) + # parser.add_argument( + # "--perlin_noise", + # type=int, + # default=None, + # help="enable perlin noise and set the octaves / perlin noiseを有効にしてoctavesをこの値に設定する", + # ) + parser.add_argument( + "--multires_noise_discount", + type=float, + default=0.3, + help="set discount value for multires noise (has no effect without --multires_noise_iterations) / Multires noiseのdiscount値を設定する（--multires_noise_iterations指定時のみ有効）", + ) + parser.add_argument( + "--adaptive_noise_scale", + type=float, + default=None, + help="add `latent mean absolute value * this value` to noise_offset (disabled if None, default) / latentの平均値の絶対値 * この値をnoise_offsetに加算する（Noneの場合は無効、デフォルト）", + ) + parser.add_argument( + "--zero_terminal_snr", + action="store_true", + help="fix noise scheduler betas to enforce zero terminal SNR / noise schedulerのbetasを修正して、zero terminal SNRを強制する", + ) + parser.add_argument( + "--min_timestep", + type=int, + default=None, + help="set minimum time step for U-Net training (0~999, default is 0) / U-Net学習時のtime stepの最小値を設定する（0~999で指定、省略時はデフォルト値(0)） ", + ) + parser.add_argument( + "--max_timestep", + type=int, + default=None, + help="set maximum time step for U-Net training (1~1000, default is 1000) / U-Net学習時のtime stepの最大値を設定する（1~1000で指定、省略時はデフォルト値(1000)）", + ) + parser.add_argument( + "--loss_type", + type=str, + default="l2", + choices=["l2", "huber", "smooth_l1"], + help="The type of loss function to use (L2, Huber, or smooth L1), default is L2 / 使用する損失関数の種類（L2、Huber、またはsmooth L1）、デフォルトはL2", + ) + parser.add_argument( + "--huber_schedule", + type=str, + default="snr", + choices=["constant", "exponential", "snr"], + help="The scheduling method for Huber loss (constant, exponential, or SNR-based). Only used when loss_type is 'huber' or 'smooth_l1'. default is snr" + + " / Huber損失のスケジューリング方法（constant、exponential、またはSNRベース）。loss_typeが'huber'または'smooth_l1'の場合に有効、デフォルトは snr", + ) + parser.add_argument( + "--huber_c", + type=float, + default=0.1, + help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type. default is 0.1 / Huber損失のパラメータ。loss_typeがhuberまたはsmooth l1の場合に有効。デフォルトは0.1", + ) + + parser.add_argument( + "--lowram", + action="store_true", + help="enable low RAM optimization. e.g. load models to VRAM instead of RAM (for machines which have bigger VRAM than RAM such as Colab and Kaggle) / メインメモリが少ない環境向け最適化を有効にする。たとえばVRAMにモデルを読み込む等（ColabやKaggleなどRAMに比べてVRAMが多い環境向け）", + ) + parser.add_argument( + "--highvram", + action="store_true", + help="disable low VRAM optimization. e.g. do not clear CUDA cache after each latent caching (for machines which have bigger VRAM) " + + "/ VRAMが少ない環境向け最適化を無効にする。たとえば各latentのキャッシュ後のCUDAキャッシュクリアを行わない等（VRAMが多い環境向け）", + ) + + parser.add_argument( + "--sample_every_n_steps", + type=int, + default=None, + help="generate sample images every N steps / 学習中のモデルで指定ステップごとにサンプル出力する", + ) + parser.add_argument( + "--sample_at_first", action="store_true", help="generate sample images before training / 学習前にサンプル出力する" + ) + parser.add_argument( + "--sample_every_n_epochs", + type=int, + default=None, + help="generate sample images every N epochs (overwrites n_steps) / 学習中のモデルで指定エポックごとにサンプル出力する（ステップ数指定を上書きします）", + ) + parser.add_argument( + "--sample_prompts", + type=str, + default=None, + help="file for prompts to generate sample images / 学習中モデルのサンプル出力用プロンプトのファイル", + ) + parser.add_argument( + "--sample_sampler", + type=str, + default="ddim", + choices=[ + "ddim", + "pndm", + "lms", + "euler", + "euler_a", + "heun", + "dpm_2", + "dpm_2_a", + "dpmsolver", + "dpmsolver++", + "dpmsingle", + "k_lms", + "k_euler", + "k_euler_a", + "k_dpm_2", + "k_dpm_2_a", + ], + help=f"sampler (scheduler) type for sample images / サンプル出力時のサンプラー（スケジューラ）の種類", + ) + + parser.add_argument( + "--config_file", + type=str, + default=None, + help="using .toml instead of args to pass hyperparameter / ハイパーパラメータを引数ではなく.tomlファイルで渡す", + ) + parser.add_argument( + "--output_config", action="store_true", help="output command line args to given .toml file / 引数を.tomlファイルに出力する" + ) + + # SAI Model spec + parser.add_argument( + "--metadata_title", + type=str, + default=None, + help="title for model metadata (default is output_name) / メタデータに書き込まれるモデルタイトル、省略時はoutput_name", + ) + parser.add_argument( + "--metadata_author", + type=str, + default=None, + help="author name for model metadata / メタデータに書き込まれるモデル作者名", + ) + parser.add_argument( + "--metadata_description", + type=str, + default=None, + help="description for model metadata / メタデータに書き込まれるモデル説明", + ) + parser.add_argument( + "--metadata_license", + type=str, + default=None, + help="license for model metadata / メタデータに書き込まれるモデルライセンス", + ) + parser.add_argument( + "--metadata_tags", + type=str, + default=None, + help="tags for model metadata, separated by comma / メタデータに書き込まれるモデルタグ、カンマ区切り", + ) + + if support_dreambooth: + # DreamBooth training + parser.add_argument( + "--prior_loss_weight", type=float, default=1.0, help="loss weight for regularization images / 正則化画像のlossの重み" + ) + + +def add_masked_loss_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--conditioning_data_dir", + type=str, + default=None, + help="conditioning data directory / 条件付けデータのディレクトリ", + ) + parser.add_argument( + "--masked_loss", + action="store_true", + help="apply mask for calculating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要", + ) + + +def get_sanitized_config_or_none(args: argparse.Namespace): + # if `--log_config` is enabled, return args for logging. if not, return None. + # when `--log_config is enabled, filter out sensitive values from args + # if wandb is not enabled, the log is not exposed to the public, but it is fine to filter out sensitive values to be safe + + if not args.log_config: + return None + + sensitive_args = ["wandb_api_key", "huggingface_token"] + sensitive_path_args = [ + "pretrained_model_name_or_path", + "vae", + "tokenizer_cache_dir", + "train_data_dir", + "conditioning_data_dir", + "reg_data_dir", + "output_dir", + "logging_dir", + ] + filtered_args = {} + for k, v in vars(args).items(): + # filter out sensitive values and convert to string if necessary + if k not in sensitive_args + sensitive_path_args: + # Accelerate values need to have type `bool`,`str`, `float`, `int`, or `None`. + if v is None or isinstance(v, bool) or isinstance(v, str) or isinstance(v, float) or isinstance(v, int): + filtered_args[k] = v + # accelerate does not support lists + elif isinstance(v, list): + filtered_args[k] = f"{v}" + # accelerate does not support objects + elif isinstance(v, object): + filtered_args[k] = f"{v}" + + return filtered_args + + +# verify command line args for training +def verify_command_line_training_args(args: argparse.Namespace): + # if wandb is enabled, the command line is exposed to the public + # check whether sensitive options are included in the command line arguments + # if so, warn or inform the user to move them to the configuration file + # wandbが有効な場合、コマンドラインが公開される + # 学習用のコマンドライン引数に敏感なオプションが含まれているかどうかを確認し、 + # 含まれている場合は設定ファイルに移動するようにユーザーに警告または通知する + + wandb_enabled = args.log_with is not None and args.log_with != "tensorboard" # "all" or "wandb" + if not wandb_enabled: + return + + sensitive_args = ["wandb_api_key", "huggingface_token"] + sensitive_path_args = [ + "pretrained_model_name_or_path", + "vae", + "tokenizer_cache_dir", + "train_data_dir", + "conditioning_data_dir", + "reg_data_dir", + "output_dir", + "logging_dir", + ] + + for arg in sensitive_args: + if getattr(args, arg, None) is not None: + logger.warning( + f"wandb is enabled, but option `{arg}` is included in the command line. Because the command line is exposed to the public, it is recommended to move it to the `.toml` file." + + f" / wandbが有効で、かつオプション `{arg}` がコマンドラインに含まれています。コマンドラインは公開されるため、`.toml`ファイルに移動することをお勧めします。" + ) + + # if path is absolute, it may include sensitive information + for arg in sensitive_path_args: + if getattr(args, arg, None) is not None and os.path.isabs(getattr(args, arg)): + logger.info( + f"wandb is enabled, but option `{arg}` is included in the command line and it is an absolute path. Because the command line is exposed to the public, it is recommended to move it to the `.toml` file or use relative path." + + f" / wandbが有効で、かつオプション `{arg}` がコマンドラインに含まれており、絶対パスです。コマンドラインは公開されるため、`.toml`ファイルに移動するか、相対パスを使用することをお勧めします。" + ) + + if getattr(args, "config_file", None) is not None: + logger.info( + f"wandb is enabled, but option `config_file` is included in the command line. Because the command line is exposed to the public, please be careful about the information included in the path." + + f" / wandbが有効で、かつオプション `config_file` がコマンドラインに含まれています。コマンドラインは公開されるため、パスに含まれる情報にご注意ください。" + ) + + # other sensitive options + if args.huggingface_repo_id is not None and args.huggingface_repo_visibility != "public": + logger.info( + f"wandb is enabled, but option huggingface_repo_id is included in the command line and huggingface_repo_visibility is not 'public'. Because the command line is exposed to the public, it is recommended to move it to the `.toml` file." + + f" / wandbが有効で、かつオプション huggingface_repo_id がコマンドラインに含まれており、huggingface_repo_visibility が 'public' ではありません。コマンドラインは公開されるため、`.toml`ファイルに移動することをお勧めします。" + ) + + +def verify_training_args(args: argparse.Namespace): + r""" + Verify training arguments. Also reflect highvram option to global variable + 学習用引数を検証する。あわせて highvram オプションの指定をグローバル変数に反映する + """ + if args.highvram: + print("highvram is enabled / highvramが有効です") + global HIGH_VRAM + HIGH_VRAM = True + + if args.v2 and args.clip_skip is not None: + logger.warning("v2 with clip_skip will be unexpected / v2でclip_skipを使用することは想定されていません") + + if args.cache_latents_to_disk and not args.cache_latents: + args.cache_latents = True + logger.warning( + "cache_latents_to_disk is enabled, so cache_latents is also enabled / cache_latents_to_diskが有効なため、cache_latentsを有効にします" + ) + + # noise_offset, perlin_noise, multires_noise_iterations cannot be enabled at the same time + # # Listを使って数えてもいいけど並べてしまえ + # if args.noise_offset is not None and args.multires_noise_iterations is not None: + # raise ValueError( + # "noise_offset and multires_noise_iterations cannot be enabled at the same time / noise_offsetとmultires_noise_iterationsを同時に有効にできません" + # ) + # if args.noise_offset is not None and args.perlin_noise is not None: + # raise ValueError("noise_offset and perlin_noise cannot be enabled at the same time / noise_offsetとperlin_noiseは同時に有効にできません") + # if args.perlin_noise is not None and args.multires_noise_iterations is not None: + # raise ValueError( + # "perlin_noise and multires_noise_iterations cannot be enabled at the same time / perlin_noiseとmultires_noise_iterationsを同時に有効にできません" + # ) + + if args.adaptive_noise_scale is not None and args.noise_offset is None: + raise ValueError("adaptive_noise_scale requires noise_offset / adaptive_noise_scaleを使用するにはnoise_offsetが必要です") + + if args.scale_v_pred_loss_like_noise_pred and not args.v_parameterization: + raise ValueError( + "scale_v_pred_loss_like_noise_pred can be enabled only with v_parameterization / scale_v_pred_loss_like_noise_predはv_parameterizationが有効なときのみ有効にできます" + ) + + if args.v_pred_like_loss and args.v_parameterization: + raise ValueError( + "v_pred_like_loss cannot be enabled with v_parameterization / v_pred_like_lossはv_parameterizationが有効なときには有効にできません" + ) + + if args.zero_terminal_snr and not args.v_parameterization: + logger.warning( + f"zero_terminal_snr is enabled, but v_parameterization is not enabled. training will be unexpected" + + " / zero_terminal_snrが有効ですが、v_parameterizationが有効ではありません。学習結果は想定外になる可能性があります" + ) + + if args.sample_every_n_epochs is not None and args.sample_every_n_epochs <= 0: + logger.warning( + "sample_every_n_epochs is less than or equal to 0, so it will be disabled / sample_every_n_epochsに0以下の値が指定されたため無効になります" + ) + args.sample_every_n_epochs = None + + if args.sample_every_n_steps is not None and args.sample_every_n_steps <= 0: + logger.warning( + "sample_every_n_steps is less than or equal to 0, so it will be disabled / sample_every_n_stepsに0以下の値が指定されたため無効になります" + ) + args.sample_every_n_steps = None + + +def add_dataset_arguments( + parser: argparse.ArgumentParser, support_dreambooth: bool, support_caption: bool, support_caption_dropout: bool +): + # dataset common + parser.add_argument( + "--train_data_dir", type=str, default=None, help="directory for train images / 学習画像データのディレクトリ" + ) + parser.add_argument( + "--cache_info", + action="store_true", + help="cache meta information (caption and image size) for faster dataset loading. only available for DreamBooth" + + " / メタ情報（キャプションとサイズ）をキャッシュしてデータセット読み込みを高速化する。DreamBooth方式のみ有効", + ) + parser.add_argument( + "--shuffle_caption", action="store_true", help="shuffle separated caption / 区切られたcaptionの各要素をshuffleする" + ) + parser.add_argument("--caption_separator", type=str, default=",", help="separator for caption / captionの区切り文字") + parser.add_argument( + "--caption_extension", type=str, default=".caption", help="extension of caption files / 読み込むcaptionファイルの拡張子" + ) + parser.add_argument( + "--caption_extention", + type=str, + default=None, + help="extension of caption files (backward compatibility) / 読み込むcaptionファイルの拡張子（スペルミスを残してあります）", + ) + parser.add_argument( + "--keep_tokens", + type=int, + default=0, + help="keep heading N tokens when shuffling caption tokens (token means comma separated strings) / captionのシャッフル時に、先頭からこの個数のトークンをシャッフルしないで残す（トークンはカンマ区切りの各部分を意味する）", + ) + parser.add_argument( + "--keep_tokens_separator", + type=str, + default="", + help="A custom separator to divide the caption into fixed and flexible parts. Tokens before this separator will not be shuffled. If not specified, '--keep_tokens' will be used to determine the fixed number of tokens." + + " / captionを固定部分と可変部分に分けるためのカスタム区切り文字。この区切り文字より前のトークンはシャッフルされない。指定しない場合、'--keep_tokens'が固定部分のトークン数として使用される。", + ) + parser.add_argument( + "--secondary_separator", + type=str, + default=None, + help="a secondary separator for caption. This separator is replaced to caption_separator after dropping/shuffling caption" + + " / captionのセカンダリ区切り文字。この区切り文字はcaptionのドロップやシャッフル後にcaption_separatorに置き換えられる", + ) + parser.add_argument( + "--enable_wildcard", + action="store_true", + help="enable wildcard for caption (e.g. '{image|picture|rendition}') / captionのワイルドカードを有効にする（例：'{image|picture|rendition}'）", + ) + parser.add_argument( + "--caption_prefix", + type=str, + default=None, + help="prefix for caption text / captionのテキストの先頭に付ける文字列", + ) + parser.add_argument( + "--caption_suffix", + type=str, + default=None, + help="suffix for caption text / captionのテキストの末尾に付ける文字列", + ) + parser.add_argument( + "--color_aug", action="store_true", help="enable weak color augmentation / 学習時に色合いのaugmentationを有効にする" + ) + parser.add_argument( + "--flip_aug", action="store_true", help="enable horizontal flip augmentation / 学習時に左右反転のaugmentationを有効にする" + ) + parser.add_argument( + "--face_crop_aug_range", + type=str, + default=None, + help="enable face-centered crop augmentation and its range (e.g. 2.0,4.0) / 学習時に顔を中心とした切り出しaugmentationを有効にするときは倍率を指定する（例：2.0,4.0）", + ) + parser.add_argument( + "--random_crop", + action="store_true", + help="enable random crop (for style training in face-centered crop augmentation) / ランダムな切り出しを有効にする（顔を中心としたaugmentationを行うときに画風の学習用に指定する）", + ) + parser.add_argument( + "--debug_dataset", + action="store_true", + help="show images for debugging (do not train) / デバッグ用に学習データを画面表示する（学習は行わない）", + ) + parser.add_argument( + "--resolution", + type=str, + default=None, + help="resolution in training ('size' or 'width,height') / 学習時の画像解像度（'サイズ'指定、または'幅,高さ'指定）", + ) + parser.add_argument( + "--cache_latents", + action="store_true", + help="cache latents to main memory to reduce VRAM usage (augmentations must be disabled) / VRAM削減のためにlatentをメインメモリにcacheする（augmentationは使用不可） ", + ) + parser.add_argument( + "--vae_batch_size", type=int, default=1, help="batch size for caching latents / latentのcache時のバッチサイズ" + ) + parser.add_argument( + "--cache_latents_to_disk", + action="store_true", + help="cache latents to disk to reduce VRAM usage (augmentations must be disabled) / VRAM削減のためにlatentをディスクにcacheする（augmentationは使用不可）", + ) + parser.add_argument( + "--enable_bucket", + action="store_true", + help="enable buckets for multi aspect ratio training / 複数解像度学習のためのbucketを有効にする", + ) + parser.add_argument( + "--min_bucket_reso", + type=int, + default=256, + help="minimum resolution for buckets, must be divisible by bucket_reso_steps " + " / bucketの最小解像度、bucket_reso_stepsで割り切れる必要があります", + ) + parser.add_argument( + "--max_bucket_reso", + type=int, + default=1024, + help="maximum resolution for buckets, must be divisible by bucket_reso_steps " + " / bucketの最大解像度、bucket_reso_stepsで割り切れる必要があります", + ) + parser.add_argument( + "--bucket_reso_steps", + type=int, + default=64, + help="steps of resolution for buckets, divisible by 8 is recommended / bucketの解像度の単位、8で割り切れる値を推奨します", + ) + parser.add_argument( + "--bucket_no_upscale", + action="store_true", + help="make bucket for each image without upscaling / 画像を拡大せずbucketを作成します", + ) + + parser.add_argument( + "--token_warmup_min", + type=int, + default=1, + help="start learning at N tags (token means comma separated strinfloatgs) / タグ数をN個から増やしながら学習する", + ) + parser.add_argument( + "--token_warmup_step", + type=float, + default=0, + help="tag length reaches maximum on N steps (or N*max_train_steps if N<1) / N（N<1ならN*max_train_steps）ステップでタグ長が最大になる。デフォルトは0（最初から最大）", + ) + parser.add_argument( + "--alpha_mask", + action="store_true", + help="use alpha channel as mask for training / 画像のアルファチャンネルをlossのマスクに使用する", + ) + + parser.add_argument( + "--dataset_class", + type=str, + default=None, + help="dataset class for arbitrary dataset (package.module.Class) / 任意のデータセットを用いるときのクラス名 (package.module.Class)", + ) + + if support_caption_dropout: + # Textual Inversion はcaptionのdropoutをsupportしない + # いわゆるtensorのDropoutと紛らわしいのでprefixにcaptionを付けておく　every_n_epochsは他と平仄を合わせてdefault Noneに + parser.add_argument( + "--caption_dropout_rate", type=float, default=0.0, help="Rate out dropout caption(0.0~1.0) / captionをdropoutする割合" + ) + parser.add_argument( + "--caption_dropout_every_n_epochs", + type=int, + default=0, + help="Dropout all captions every N epochs / captionを指定エポックごとにdropoutする", + ) + parser.add_argument( + "--caption_tag_dropout_rate", + type=float, + default=0.0, + help="Rate out dropout comma separated tokens(0.0~1.0) / カンマ区切りのタグをdropoutする割合", + ) + + if support_dreambooth: + # DreamBooth dataset + parser.add_argument( + "--reg_data_dir", type=str, default=None, help="directory for regularization images / 正則化画像データのディレクトリ" + ) + + if support_caption: + # caption dataset + parser.add_argument( + "--in_json", type=str, default=None, help="json metadata for dataset / データセットのmetadataのjsonファイル" + ) + parser.add_argument( + "--dataset_repeats", + type=int, + default=1, + help="repeat dataset when training with captions / キャプションでの学習時にデータセットを繰り返す回数", + ) + + +def add_sd_saving_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--save_model_as", + type=str, + default=None, + choices=[None, "ckpt", "safetensors", "diffusers", "diffusers_safetensors"], + help="format to save the model (default is same to original) / モデル保存時の形式（未指定時は元モデルと同じ）", + ) + parser.add_argument( + "--use_safetensors", + action="store_true", + help="use safetensors format to save (if save_model_as is not specified) / checkpoint、モデルをsafetensors形式で保存する（save_model_as未指定時）", + ) + + +def read_config_from_file(args: argparse.Namespace, parser: argparse.ArgumentParser): + if not args.config_file: + return args + + config_path = args.config_file + ".toml" if not args.config_file.endswith(".toml") else args.config_file + + if args.output_config: + # check if config file exists + if os.path.exists(config_path): + logger.error(f"Config file already exists. Aborting... / 出力先の設定ファイルが既に存在します: {config_path}") + exit(1) + + # convert args to dictionary + args_dict = vars(args) + + # remove unnecessary keys + for key in ["config_file", "output_config", "wandb_api_key"]: + if key in args_dict: + del args_dict[key] + + # get default args from parser + default_args = vars(parser.parse_args([])) + + # remove default values: cannot use args_dict.items directly because it will be changed during iteration + for key, value in list(args_dict.items()): + if key in default_args and value == default_args[key]: + del args_dict[key] + + # convert Path to str in dictionary + for key, value in args_dict.items(): + if isinstance(value, pathlib.Path): + args_dict[key] = str(value) + + # convert to toml and output to file + with open(config_path, "w") as f: + toml.dump(args_dict, f) + + logger.info(f"Saved config file / 設定ファイルを保存しました: {config_path}") + exit(0) + + if not os.path.exists(config_path): + logger.info(f"{config_path} not found.") + exit(1) + + logger.info(f"Loading settings from {config_path}...") + with open(config_path, "r", encoding="utf-8") as f: + config_dict = toml.load(f) + + # combine all sections into one + ignore_nesting_dict = {} + for section_name, section_dict in config_dict.items(): + # if value is not dict, save key and value as is + if not isinstance(section_dict, dict): + ignore_nesting_dict[section_name] = section_dict + continue + + # if value is dict, save all key and value into one dict + for key, value in section_dict.items(): + ignore_nesting_dict[key] = value + + config_args = argparse.Namespace(**ignore_nesting_dict) + args = parser.parse_args(namespace=config_args) + args.config_file = os.path.splitext(args.config_file)[0] + logger.info(args.config_file) + + return args + + +# endregion + +# region utils + + +def resume_from_local_or_hf_if_specified(accelerator, args): + if not args.resume: + return + + if not args.resume_from_huggingface: + logger.info(f"resume training from local state: {args.resume}") + accelerator.load_state(args.resume) + return + + logger.info(f"resume training from huggingface state: {args.resume}") + repo_id = args.resume.split("/")[0] + "/" + args.resume.split("/")[1] + path_in_repo = "/".join(args.resume.split("/")[2:]) + revision = None + repo_type = None + if ":" in path_in_repo: + divided = path_in_repo.split(":") + if len(divided) == 2: + path_in_repo, revision = divided + repo_type = "model" + else: + path_in_repo, revision, repo_type = divided + logger.info(f"Downloading state from huggingface: {repo_id}/{path_in_repo}@{revision}") + + list_files = huggingface_util.list_dir( + repo_id=repo_id, + subfolder=path_in_repo, + revision=revision, + token=args.huggingface_token, + repo_type=repo_type, + ) + + async def download(filename) -> str: + def task(): + return hf_hub_download( + repo_id=repo_id, + filename=filename, + revision=revision, + repo_type=repo_type, + token=args.huggingface_token, + ) + + return await asyncio.get_event_loop().run_in_executor(None, task) + + loop = asyncio.get_event_loop() + results = loop.run_until_complete(asyncio.gather(*[download(filename=filename.rfilename) for filename in list_files])) + if len(results) == 0: + raise ValueError( + "No files found in the specified repo id/path/revision / 指定されたリポジトリID/パス/リビジョンにファイルが見つかりませんでした" + ) + dirname = os.path.dirname(results[0]) + accelerator.load_state(dirname) + + +def get_optimizer(args, trainable_params): + # "Optimizer to use: AdamW, AdamW8bit, Lion, SGDNesterov, SGDNesterov8bit, PagedAdamW, PagedAdamW8bit, PagedAdamW32bit, Lion8bit, PagedLion8bit, AdEMAMix8bit, PagedAdEMAMix8bit, DAdaptation(DAdaptAdamPreprint), DAdaptAdaGrad, DAdaptAdam, DAdaptAdan, DAdaptAdanIP, DAdaptLion, DAdaptSGD, Adafactor" + + optimizer_type = args.optimizer_type + if args.use_8bit_adam: + assert ( + not args.use_lion_optimizer + ), "both option use_8bit_adam and use_lion_optimizer are specified / use_8bit_adamとuse_lion_optimizerの両方のオプションが指定されています" + assert ( + optimizer_type is None or optimizer_type == "" + ), "both option use_8bit_adam and optimizer_type are specified / use_8bit_adamとoptimizer_typeの両方のオプションが指定されています" + optimizer_type = "AdamW8bit" + + elif args.use_lion_optimizer: + assert ( + optimizer_type is None or optimizer_type == "" + ), "both option use_lion_optimizer and optimizer_type are specified / use_lion_optimizerとoptimizer_typeの両方のオプションが指定されています" + optimizer_type = "Lion" + + if optimizer_type is None or optimizer_type == "": + optimizer_type = "AdamW" + optimizer_type = optimizer_type.lower() + + if args.fused_backward_pass: + assert ( + optimizer_type == "Adafactor".lower() + ), "fused_backward_pass currently only works with optimizer_type Adafactor / fused_backward_passは現在optimizer_type Adafactorでのみ機能します" + assert ( + args.gradient_accumulation_steps == 1 + ), "fused_backward_pass does not work with gradient_accumulation_steps > 1 / fused_backward_passはgradient_accumulation_steps>1では機能しません" + + # 引数を分解する + optimizer_kwargs = {} + if args.optimizer_args is not None and len(args.optimizer_args) > 0: + for arg in args.optimizer_args: + key, value = arg.split("=") + value = ast.literal_eval(value) + + # value = value.split(",") + # for i in range(len(value)): + # if value[i].lower() == "true" or value[i].lower() == "false": + # value[i] = value[i].lower() == "true" + # else: + # value[i] = ast.float(value[i]) + # if len(value) == 1: + # value = value[0] + # else: + # value = tuple(value) + + optimizer_kwargs[key] = value + # logger.info(f"optkwargs {optimizer}_{kwargs}") + + lr = args.learning_rate + optimizer = None + optimizer_class = None + + if optimizer_type == "Lion".lower(): + try: + import lion_pytorch + except ImportError: + raise ImportError("No lion_pytorch / lion_pytorch がインストールされていないようです") + logger.info(f"use Lion optimizer | {optimizer_kwargs}") + optimizer_class = lion_pytorch.Lion + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + elif optimizer_type.endswith("8bit".lower()): + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError("No bitsandbytes / bitsandbytesがインストールされていないようです") + + if optimizer_type == "AdamW8bit".lower(): + logger.info(f"use 8-bit AdamW optimizer | {optimizer_kwargs}") + optimizer_class = bnb.optim.AdamW8bit + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + elif optimizer_type == "SGDNesterov8bit".lower(): + logger.info(f"use 8-bit SGD with Nesterov optimizer | {optimizer_kwargs}") + if "momentum" not in optimizer_kwargs: + logger.warning( + f"8-bit SGD with Nesterov must be with momentum, set momentum to 0.9 / 8-bit SGD with Nesterovはmomentum指定が必須のため0.9に設定します" + ) + optimizer_kwargs["momentum"] = 0.9 + + optimizer_class = bnb.optim.SGD8bit + optimizer = optimizer_class(trainable_params, lr=lr, nesterov=True, **optimizer_kwargs) + + elif optimizer_type == "Lion8bit".lower(): + logger.info(f"use 8-bit Lion optimizer | {optimizer_kwargs}") + try: + optimizer_class = bnb.optim.Lion8bit + except AttributeError: + raise AttributeError( + "No Lion8bit. The version of bitsandbytes installed seems to be old. Please install 0.38.0 or later. / Lion8bitが定義されていません。インストールされているbitsandbytesのバージョンが古いようです。0.38.0以上をインストールしてください" + ) + elif optimizer_type == "PagedAdamW8bit".lower(): + logger.info(f"use 8-bit PagedAdamW optimizer | {optimizer_kwargs}") + try: + optimizer_class = bnb.optim.PagedAdamW8bit + except AttributeError: + raise AttributeError( + "No PagedAdamW8bit. The version of bitsandbytes installed seems to be old. Please install 0.39.0 or later. / PagedAdamW8bitが定義されていません。インストールされているbitsandbytesのバージョンが古いようです。0.39.0以上をインストールしてください" + ) + elif optimizer_type == "PagedLion8bit".lower(): + logger.info(f"use 8-bit Paged Lion optimizer | {optimizer_kwargs}") + try: + optimizer_class = bnb.optim.PagedLion8bit + except AttributeError: + raise AttributeError( + "No PagedLion8bit. The version of bitsandbytes installed seems to be old. Please install 0.39.0 or later. / PagedLion8bitが定義されていません。インストールされているbitsandbytesのバージョンが古いようです。0.39.0以上をインストールしてください" + ) + + if optimizer_class is not None: + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + elif optimizer_type == "PagedAdamW".lower(): + logger.info(f"use PagedAdamW optimizer | {optimizer_kwargs}") + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError("No bitsandbytes / bitsandbytesがインストールされていないようです") + try: + optimizer_class = bnb.optim.PagedAdamW + except AttributeError: + raise AttributeError( + "No PagedAdamW. The version of bitsandbytes installed seems to be old. Please install 0.39.0 or later. / PagedAdamWが定義されていません。インストールされているbitsandbytesのバージョンが古いようです。0.39.0以上をインストールしてください" + ) + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + elif optimizer_type == "PagedAdamW32bit".lower(): + logger.info(f"use 32-bit PagedAdamW optimizer | {optimizer_kwargs}") + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError("No bitsandbytes / bitsandbytesがインストールされていないようです") + try: + optimizer_class = bnb.optim.PagedAdamW32bit + except AttributeError: + raise AttributeError( + "No PagedAdamW32bit. The version of bitsandbytes installed seems to be old. Please install 0.39.0 or later. / PagedAdamW32bitが定義されていません。インストールされているbitsandbytesのバージョンが古いようです。0.39.0以上をインストールしてください" + ) + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + elif optimizer_type == "SGDNesterov".lower(): + logger.info(f"use SGD with Nesterov optimizer | {optimizer_kwargs}") + if "momentum" not in optimizer_kwargs: + logger.info( + f"SGD with Nesterov must be with momentum, set momentum to 0.9 / SGD with Nesterovはmomentum指定が必須のため0.9に設定します" + ) + optimizer_kwargs["momentum"] = 0.9 + + optimizer_class = torch.optim.SGD + optimizer = optimizer_class(trainable_params, lr=lr, nesterov=True, **optimizer_kwargs) + + elif optimizer_type.startswith("DAdapt".lower()) or optimizer_type == "Prodigy".lower(): + # check lr and lr_count, and logger.info warning + actual_lr = lr + lr_count = 1 + if type(trainable_params) == list and type(trainable_params[0]) == dict: + lrs = set() + actual_lr = trainable_params[0].get("lr", actual_lr) + for group in trainable_params: + lrs.add(group.get("lr", actual_lr)) + lr_count = len(lrs) + + if actual_lr <= 0.1: + logger.warning( + f"learning rate is too low. If using D-Adaptation or Prodigy, set learning rate around 1.0 / 学習率が低すぎるようです。D-AdaptationまたはProdigyの使用時は1.0前後の値を指定してください: lr={actual_lr}" + ) + logger.warning("recommend option: lr=1.0 / 推奨は1.0です") + if lr_count > 1: + logger.warning( + f"when multiple learning rates are specified with dadaptation (e.g. for Text Encoder and U-Net), only the first one will take effect / D-AdaptationまたはProdigyで複数の学習率を指定した場合（Text EncoderとU-Netなど）、最初の学習率のみが有効になります: lr={actual_lr}" + ) + + if optimizer_type.startswith("DAdapt".lower()): + # DAdaptation family + # check dadaptation is installed + try: + import dadaptation + import dadaptation.experimental as experimental + except ImportError: + raise ImportError("No dadaptation / dadaptation がインストールされていないようです") + + # set optimizer + if optimizer_type == "DAdaptation".lower() or optimizer_type == "DAdaptAdamPreprint".lower(): + optimizer_class = experimental.DAdaptAdamPreprint + logger.info(f"use D-Adaptation AdamPreprint optimizer | {optimizer_kwargs}") + elif optimizer_type == "DAdaptAdaGrad".lower(): + optimizer_class = dadaptation.DAdaptAdaGrad + logger.info(f"use D-Adaptation AdaGrad optimizer | {optimizer_kwargs}") + elif optimizer_type == "DAdaptAdam".lower(): + optimizer_class = dadaptation.DAdaptAdam + logger.info(f"use D-Adaptation Adam optimizer | {optimizer_kwargs}") + elif optimizer_type == "DAdaptAdan".lower(): + optimizer_class = dadaptation.DAdaptAdan + logger.info(f"use D-Adaptation Adan optimizer | {optimizer_kwargs}") + elif optimizer_type == "DAdaptAdanIP".lower(): + optimizer_class = experimental.DAdaptAdanIP + logger.info(f"use D-Adaptation AdanIP optimizer | {optimizer_kwargs}") + elif optimizer_type == "DAdaptLion".lower(): + optimizer_class = dadaptation.DAdaptLion + logger.info(f"use D-Adaptation Lion optimizer | {optimizer_kwargs}") + elif optimizer_type == "DAdaptSGD".lower(): + optimizer_class = dadaptation.DAdaptSGD + logger.info(f"use D-Adaptation SGD optimizer | {optimizer_kwargs}") + else: + raise ValueError(f"Unknown optimizer type: {optimizer_type}") + + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + else: + # Prodigy + # check Prodigy is installed + try: + import prodigyopt + except ImportError: + raise ImportError("No Prodigy / Prodigy がインストールされていないようです") + + logger.info(f"use Prodigy optimizer | {optimizer_kwargs}") + optimizer_class = prodigyopt.Prodigy + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + elif optimizer_type == "Adafactor".lower(): + # 引数を確認して適宜補正する + if "relative_step" not in optimizer_kwargs: + optimizer_kwargs["relative_step"] = True # default + if not optimizer_kwargs["relative_step"] and optimizer_kwargs.get("warmup_init", False): + logger.info( + f"set relative_step to True because warmup_init is True / warmup_initがTrueのためrelative_stepをTrueにします" + ) + optimizer_kwargs["relative_step"] = True + logger.info(f"use Adafactor optimizer | {optimizer_kwargs}") + + if optimizer_kwargs["relative_step"]: + logger.info(f"relative_step is true / relative_stepがtrueです") + if lr != 0.0: + logger.warning(f"learning rate is used as initial_lr / 指定したlearning rateはinitial_lrとして使用されます") + args.learning_rate = None + + # trainable_paramsがgroupだった時の処理：lrを削除する + if type(trainable_params) == list and type(trainable_params[0]) == dict: + has_group_lr = False + for group in trainable_params: + p = group.pop("lr", None) + has_group_lr = has_group_lr or (p is not None) + + if has_group_lr: + # 一応argsを無効にしておく TODO 依存関係が逆転してるのであまり望ましくない + logger.warning(f"unet_lr and text_encoder_lr are ignored / unet_lrとtext_encoder_lrは無視されます") + args.unet_lr = None + args.text_encoder_lr = None + + if args.lr_scheduler != "adafactor": + logger.info(f"use adafactor_scheduler / スケジューラにadafactor_schedulerを使用します") + args.lr_scheduler = f"adafactor:{lr}" # ちょっと微妙だけど + + lr = None + else: + if args.max_grad_norm != 0.0: + logger.warning( + f"because max_grad_norm is set, clip_grad_norm is enabled. consider set to 0 / max_grad_normが設定されているためclip_grad_normが有効になります。0に設定して無効にしたほうがいいかもしれません" + ) + if args.lr_scheduler != "constant_with_warmup": + logger.warning(f"constant_with_warmup will be good / スケジューラはconstant_with_warmupが良いかもしれません") + if optimizer_kwargs.get("clip_threshold", 1.0) != 1.0: + logger.warning(f"clip_threshold=1.0 will be good / clip_thresholdは1.0が良いかもしれません") + + optimizer_class = transformers.optimization.Adafactor + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + elif optimizer_type == "AdamW".lower(): + logger.info(f"use AdamW optimizer | {optimizer_kwargs}") + optimizer_class = torch.optim.AdamW + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + if optimizer is None: + # 任意のoptimizerを使う + optimizer_type = args.optimizer_type # lowerでないやつ（微妙） + logger.info(f"use {optimizer_type} | {optimizer_kwargs}") + if "." not in optimizer_type: + optimizer_module = torch.optim + else: + values = optimizer_type.split(".") + optimizer_module = importlib.import_module(".".join(values[:-1])) + optimizer_type = values[-1] + + optimizer_class = getattr(optimizer_module, optimizer_type) + optimizer = optimizer_class(trainable_params, lr=lr, **optimizer_kwargs) + + # for logging + optimizer_name = optimizer_class.__module__ + "." + optimizer_class.__name__ + optimizer_args = ",".join([f"{k}={v}" for k, v in optimizer_kwargs.items()]) + + return optimizer_name, optimizer_args, optimizer + + +# Modified version of get_scheduler() function from diffusers.optimizer.get_scheduler +# Add some checking and features to the original function. + + +def get_scheduler_fix(args, optimizer: Optimizer, num_processes: int): + """ + Unified API to get any scheduler from its name. + """ + name = args.lr_scheduler + num_training_steps = args.max_train_steps * num_processes # * args.gradient_accumulation_steps + num_warmup_steps: Optional[int] = ( + int(args.lr_warmup_steps * num_training_steps) if isinstance(args.lr_warmup_steps, float) else args.lr_warmup_steps + ) + num_decay_steps: Optional[int] = ( + int(args.lr_decay_steps * num_training_steps) if isinstance(args.lr_decay_steps, float) else args.lr_decay_steps + ) + num_stable_steps = num_training_steps - num_warmup_steps - num_decay_steps + num_cycles = args.lr_scheduler_num_cycles + power = args.lr_scheduler_power + timescale = args.lr_scheduler_timescale + min_lr_ratio = args.lr_scheduler_min_lr_ratio + + lr_scheduler_kwargs = {} # get custom lr_scheduler kwargs + if args.lr_scheduler_args is not None and len(args.lr_scheduler_args) > 0: + for arg in args.lr_scheduler_args: + key, value = arg.split("=") + value = ast.literal_eval(value) + lr_scheduler_kwargs[key] = value + + def wrap_check_needless_num_warmup_steps(return_vals): + if num_warmup_steps is not None and num_warmup_steps != 0: + raise ValueError(f"{name} does not require `num_warmup_steps`. Set None or 0.") + return return_vals + + # using any lr_scheduler from other library + if args.lr_scheduler_type: + lr_scheduler_type = args.lr_scheduler_type + logger.info(f"use {lr_scheduler_type} | {lr_scheduler_kwargs} as lr_scheduler") + if "." not in lr_scheduler_type: # default to use torch.optim + lr_scheduler_module = torch.optim.lr_scheduler + else: + values = lr_scheduler_type.split(".") + lr_scheduler_module = importlib.import_module(".".join(values[:-1])) + lr_scheduler_type = values[-1] + lr_scheduler_class = getattr(lr_scheduler_module, lr_scheduler_type) + lr_scheduler = lr_scheduler_class(optimizer, **lr_scheduler_kwargs) + return wrap_check_needless_num_warmup_steps(lr_scheduler) + + if name.startswith("adafactor"): + assert ( + type(optimizer) == transformers.optimization.Adafactor + ), f"adafactor scheduler must be used with Adafactor optimizer / adafactor schedulerはAdafactorオプティマイザと同時に使ってください" + initial_lr = float(name.split(":")[1]) + # logger.info(f"adafactor scheduler init lr {initial_lr}") + return wrap_check_needless_num_warmup_steps(transformers.optimization.AdafactorSchedule(optimizer, initial_lr)) + + if name == DiffusersSchedulerType.PIECEWISE_CONSTANT.value: + name = DiffusersSchedulerType(name) + schedule_func = DIFFUSERS_TYPE_TO_SCHEDULER_FUNCTION[name] + return schedule_func(optimizer, **lr_scheduler_kwargs) # step_rules and last_epoch are given as kwargs + + name = SchedulerType(name) + schedule_func = TYPE_TO_SCHEDULER_FUNCTION[name] + + if name == SchedulerType.CONSTANT: + return wrap_check_needless_num_warmup_steps(schedule_func(optimizer, **lr_scheduler_kwargs)) + + # All other schedulers require `num_warmup_steps` + if num_warmup_steps is None: + raise ValueError(f"{name} requires `num_warmup_steps`, please provide that argument.") + + if name == SchedulerType.CONSTANT_WITH_WARMUP: + return schedule_func(optimizer, num_warmup_steps=num_warmup_steps, **lr_scheduler_kwargs) + + if name == SchedulerType.INVERSE_SQRT: + return schedule_func(optimizer, num_warmup_steps=num_warmup_steps, timescale=timescale, **lr_scheduler_kwargs) + + # All other schedulers require `num_training_steps` + if num_training_steps is None: + raise ValueError(f"{name} requires `num_training_steps`, please provide that argument.") + + if name == SchedulerType.COSINE_WITH_RESTARTS: + return schedule_func( + optimizer, + num_warmup_steps=num_warmup_steps, + num_training_steps=num_training_steps, + num_cycles=num_cycles, + **lr_scheduler_kwargs, + ) + + if name == SchedulerType.POLYNOMIAL: + return schedule_func( + optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps, power=power, **lr_scheduler_kwargs + ) + + if name == SchedulerType.COSINE_WITH_MIN_LR: + return schedule_func( + optimizer, + num_warmup_steps=num_warmup_steps, + num_training_steps=num_training_steps, + num_cycles=num_cycles / 2, + min_lr_rate=min_lr_ratio, + **lr_scheduler_kwargs, + ) + + # these schedulers do not require `num_decay_steps` + if name == SchedulerType.LINEAR or name == SchedulerType.COSINE: + return schedule_func( + optimizer, + num_warmup_steps=num_warmup_steps, + num_training_steps=num_training_steps, + **lr_scheduler_kwargs, + ) + + # All other schedulers require `num_decay_steps` + if num_decay_steps is None: + raise ValueError(f"{name} requires `num_decay_steps`, please provide that argument.") + if name == SchedulerType.WARMUP_STABLE_DECAY: + return schedule_func( + optimizer, + num_warmup_steps=num_warmup_steps, + num_stable_steps=num_stable_steps, + num_decay_steps=num_decay_steps, + num_cycles=num_cycles / 2, + min_lr_ratio=min_lr_ratio if min_lr_ratio is not None else 0.0, + **lr_scheduler_kwargs, + ) + + return schedule_func( + optimizer, + num_warmup_steps=num_warmup_steps, + num_training_steps=num_training_steps, + num_decay_steps=num_decay_steps, + **lr_scheduler_kwargs, + ) + + +def prepare_dataset_args(args: argparse.Namespace, support_metadata: bool): + # backward compatibility + if args.caption_extention is not None: + args.caption_extension = args.caption_extention + args.caption_extention = None + + # assert args.resolution is not None, f"resolution is required / resolution（解像度）を指定してください" + if args.resolution is not None: + args.resolution = tuple([int(r) for r in args.resolution.split(",")]) + if len(args.resolution) == 1: + args.resolution = (args.resolution[0], args.resolution[0]) + assert ( + len(args.resolution) == 2 + ), f"resolution must be 'size' or 'width,height' / resolution（解像度）は'サイズ'または'幅','高さ'で指定してください: {args.resolution}" + + if args.face_crop_aug_range is not None: + args.face_crop_aug_range = tuple([float(r) for r in args.face_crop_aug_range.split(",")]) + assert ( + len(args.face_crop_aug_range) == 2 and args.face_crop_aug_range[0] <= args.face_crop_aug_range[1] + ), f"face_crop_aug_range must be two floats / face_crop_aug_rangeは'下限,上限'で指定してください: {args.face_crop_aug_range}" + else: + args.face_crop_aug_range = None + + if support_metadata: + if args.in_json is not None and (args.color_aug or args.random_crop): + logger.warning( + f"latents in npz is ignored when color_aug or random_crop is True / color_augまたはrandom_cropを有効にした場合、npzファイルのlatentsは無視されます" + ) + + +def load_tokenizer(args: argparse.Namespace): + logger.info("prepare tokenizer") + original_path = V2_STABLE_DIFFUSION_PATH if args.v2 else TOKENIZER_PATH + + tokenizer: CLIPTokenizer = None + if args.tokenizer_cache_dir: + local_tokenizer_path = os.path.join(args.tokenizer_cache_dir, original_path.replace("/", "_")) + if os.path.exists(local_tokenizer_path): + logger.info(f"load tokenizer from cache: {local_tokenizer_path}") + tokenizer = CLIPTokenizer.from_pretrained(local_tokenizer_path) # same for v1 and v2 + + if tokenizer is None: + if args.v2: + tokenizer = CLIPTokenizer.from_pretrained(original_path, subfolder="tokenizer") + else: + tokenizer = CLIPTokenizer.from_pretrained(original_path) + + if hasattr(args, "max_token_length") and args.max_token_length is not None: + logger.info(f"update token length: {args.max_token_length}") + + if args.tokenizer_cache_dir and not os.path.exists(local_tokenizer_path): + logger.info(f"save Tokenizer to cache: {local_tokenizer_path}") + tokenizer.save_pretrained(local_tokenizer_path) + + return tokenizer + + +def prepare_accelerator(args: argparse.Namespace): + """ + this function also prepares deepspeed plugin + """ + + if args.logging_dir is None: + logging_dir = None + else: + log_prefix = "" if args.log_prefix is None else args.log_prefix + logging_dir = args.logging_dir + "/" + log_prefix + time.strftime("%Y%m%d%H%M%S", time.localtime()) + + if args.log_with is None: + if logging_dir is not None: + log_with = "tensorboard" + else: + log_with = None + else: + log_with = args.log_with + if log_with in ["tensorboard", "all"]: + if logging_dir is None: + raise ValueError( + "logging_dir is required when log_with is tensorboard / Tensorboardを使う場合、logging_dirを指定してください" + ) + if log_with in ["wandb", "all"]: + try: + import wandb + except ImportError: + raise ImportError("No wandb / wandb がインストールされていないようです") + if logging_dir is not None: + os.makedirs(logging_dir, exist_ok=True) + os.environ["WANDB_DIR"] = logging_dir + if args.wandb_api_key is not None: + wandb.login(key=args.wandb_api_key) + + # torch.compile のオプション。 NO の場合は torch.compile は使わない + dynamo_backend = "NO" + if args.torch_compile: + dynamo_backend = args.dynamo_backend + + kwargs_handlers = ( + InitProcessGroupKwargs(timeout=datetime.timedelta(minutes=args.ddp_timeout)) if args.ddp_timeout else None, + ( + DistributedDataParallelKwargs( + gradient_as_bucket_view=args.ddp_gradient_as_bucket_view, static_graph=args.ddp_static_graph + ) + if args.ddp_gradient_as_bucket_view or args.ddp_static_graph + else None + ), + ) + kwargs_handlers = list(filter(lambda x: x is not None, kwargs_handlers)) + deepspeed_plugin = deepspeed_utils.prepare_deepspeed_plugin(args) + + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with=log_with, + project_dir=logging_dir, + kwargs_handlers=kwargs_handlers, + dynamo_backend=dynamo_backend, + deepspeed_plugin=deepspeed_plugin, + ) + print("accelerator device:", accelerator.device) + return accelerator + + +def prepare_dtype(args: argparse.Namespace): + weight_dtype = torch.float32 + if args.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif args.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + save_dtype = None + if args.save_precision == "fp16": + save_dtype = torch.float16 + elif args.save_precision == "bf16": + save_dtype = torch.bfloat16 + elif args.save_precision == "float": + save_dtype = torch.float32 + + return weight_dtype, save_dtype + + +def _load_target_model(args: argparse.Namespace, weight_dtype, device="cpu", unet_use_linear_projection_in_v2=False): + name_or_path = args.pretrained_model_name_or_path + name_or_path = os.path.realpath(name_or_path) if os.path.islink(name_or_path) else name_or_path + load_stable_diffusion_format = os.path.isfile(name_or_path) # determine SD or Diffusers + if load_stable_diffusion_format: + logger.info(f"load StableDiffusion checkpoint: {name_or_path}") + text_encoder, vae, unet = model_util.load_models_from_stable_diffusion_checkpoint( + args.v2, name_or_path, device, unet_use_linear_projection_in_v2=unet_use_linear_projection_in_v2 + ) + else: + # Diffusers model is loaded to CPU + logger.info(f"load Diffusers pretrained models: {name_or_path}") + try: + pipe = StableDiffusionPipeline.from_pretrained(name_or_path, tokenizer=None, safety_checker=None) + except EnvironmentError as ex: + logger.error( + f"model is not found as a file or in Hugging Face, perhaps file name is wrong? / 指定したモデル名のファイル、またはHugging Faceのモデルが見つかりません。ファイル名が誤っているかもしれません: {name_or_path}" + ) + raise ex + text_encoder = pipe.text_encoder + vae = pipe.vae + unet = pipe.unet + del pipe + + # Diffusers U-Net to original U-Net + # TODO *.ckpt/*.safetensorsのv2と同じ形式にここで変換すると良さそう + # logger.info(f"unet config: {unet.config}") + original_unet = UNet2DConditionModel( + unet.config.sample_size, + unet.config.attention_head_dim, + unet.config.cross_attention_dim, + unet.config.use_linear_projection, + unet.config.upcast_attention, + ) + original_unet.load_state_dict(unet.state_dict()) + unet = original_unet + logger.info("U-Net converted to original U-Net") + + # VAEを読み込む + if args.vae is not None: + vae = model_util.load_vae(args.vae, weight_dtype) + logger.info("additional VAE loaded") + + return text_encoder, vae, unet, load_stable_diffusion_format + + +def load_target_model(args, weight_dtype, accelerator, unet_use_linear_projection_in_v2=False): + for pi in range(accelerator.state.num_processes): + if pi == accelerator.state.local_process_index: + logger.info(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}") + + text_encoder, vae, unet, load_stable_diffusion_format = _load_target_model( + args, + weight_dtype, + accelerator.device if args.lowram else "cpu", + unet_use_linear_projection_in_v2=unet_use_linear_projection_in_v2, + ) + # work on low-ram device + if args.lowram: + text_encoder.to(accelerator.device) + unet.to(accelerator.device) + vae.to(accelerator.device) + + clean_memory_on_device(accelerator.device) + accelerator.wait_for_everyone() + return text_encoder, vae, unet, load_stable_diffusion_format + + +def patch_accelerator_for_fp16_training(accelerator): + org_unscale_grads = accelerator.scaler._unscale_grads_ + + def _unscale_grads_replacer(optimizer, inv_scale, found_inf, allow_fp16): + return org_unscale_grads(optimizer, inv_scale, found_inf, True) + + accelerator.scaler._unscale_grads_ = _unscale_grads_replacer + + +def get_hidden_states(args: argparse.Namespace, input_ids, tokenizer, text_encoder, weight_dtype=None): + # with no_token_padding, the length is not max length, return result immediately + if input_ids.size()[-1] != tokenizer.model_max_length: + return text_encoder(input_ids)[0] + + # input_ids: b,n,77 + b_size = input_ids.size()[0] + input_ids = input_ids.reshape((-1, tokenizer.model_max_length)) # batch_size*3, 77 + + if args.clip_skip is None: + encoder_hidden_states = text_encoder(input_ids)[0] + else: + enc_out = text_encoder(input_ids, output_hidden_states=True, return_dict=True) + encoder_hidden_states = enc_out["hidden_states"][-args.clip_skip] + encoder_hidden_states = text_encoder.text_model.final_layer_norm(encoder_hidden_states) + + # bs*3, 77, 768 or 1024 + encoder_hidden_states = encoder_hidden_states.reshape((b_size, -1, encoder_hidden_states.shape[-1])) + + if args.max_token_length is not None: + if args.v2: + # v2: ... ... の三連を ... ... へ戻す　正直この実装でいいのかわからん + states_list = [encoder_hidden_states[:, 0].unsqueeze(1)] # + for i in range(1, args.max_token_length, tokenizer.model_max_length): + chunk = encoder_hidden_states[:, i : i + tokenizer.model_max_length - 2] # の後から 最後の前まで + if i > 0: + for j in range(len(chunk)): + if input_ids[j, 1] == tokenizer.eos_token: # 空、つまり ...のパターン + chunk[j, 0] = chunk[j, 1] # 次の の値をコピーする + states_list.append(chunk) # の後から の前まで + states_list.append(encoder_hidden_states[:, -1].unsqueeze(1)) # か のどちらか + encoder_hidden_states = torch.cat(states_list, dim=1) + else: + # v1: ... の三連を ... へ戻す + states_list = [encoder_hidden_states[:, 0].unsqueeze(1)] # + for i in range(1, args.max_token_length, tokenizer.model_max_length): + states_list.append( + encoder_hidden_states[:, i : i + tokenizer.model_max_length - 2] + ) # の後から の前まで + states_list.append(encoder_hidden_states[:, -1].unsqueeze(1)) # + encoder_hidden_states = torch.cat(states_list, dim=1) + + if weight_dtype is not None: + # this is required for additional network training + encoder_hidden_states = encoder_hidden_states.to(weight_dtype) + + return encoder_hidden_states + + +def pool_workaround( + text_encoder: CLIPTextModelWithProjection, last_hidden_state: torch.Tensor, input_ids: torch.Tensor, eos_token_id: int +): + r""" + workaround for CLIP's pooling bug: it returns the hidden states for the max token id as the pooled output + instead of the hidden states for the EOS token + If we use Textual Inversion, we need to use the hidden states for the EOS token as the pooled output + + Original code from CLIP's pooling function: + + \# text_embeds.shape = [batch_size, sequence_length, transformer.width] + \# take features from the eot embedding (eot_token is the highest number in each sequence) + \# casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14 + pooled_output = last_hidden_state[ + torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device), + input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1), + ] + """ + + # input_ids: b*n,77 + # find index for EOS token + + # Following code is not working if one of the input_ids has multiple EOS tokens (very odd case) + # eos_token_index = torch.where(input_ids == eos_token_id)[1] + # eos_token_index = eos_token_index.to(device=last_hidden_state.device) + + # Create a mask where the EOS tokens are + eos_token_mask = (input_ids == eos_token_id).int() + + # Use argmax to find the last index of the EOS token for each element in the batch + eos_token_index = torch.argmax(eos_token_mask, dim=1) # this will be 0 if there is no EOS token, it's fine + eos_token_index = eos_token_index.to(device=last_hidden_state.device) + + # get hidden states for EOS token + pooled_output = last_hidden_state[torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device), eos_token_index] + + # apply projection: projection may be of different dtype than last_hidden_state + pooled_output = text_encoder.text_projection(pooled_output.to(text_encoder.text_projection.weight.dtype)) + pooled_output = pooled_output.to(last_hidden_state.dtype) + + return pooled_output + + +def get_hidden_states_sdxl( + max_token_length: int, + input_ids1: torch.Tensor, + input_ids2: torch.Tensor, + tokenizer1: CLIPTokenizer, + tokenizer2: CLIPTokenizer, + text_encoder1: CLIPTextModel, + text_encoder2: CLIPTextModelWithProjection, + weight_dtype: Optional[str] = None, + accelerator: Optional[Accelerator] = None, +): + # input_ids: b,n,77 -> b*n, 77 + b_size = input_ids1.size()[0] + input_ids1 = input_ids1.reshape((-1, tokenizer1.model_max_length)) # batch_size*n, 77 + input_ids2 = input_ids2.reshape((-1, tokenizer2.model_max_length)) # batch_size*n, 77 + + # text_encoder1 + enc_out = text_encoder1(input_ids1, output_hidden_states=True, return_dict=True) + hidden_states1 = enc_out["hidden_states"][11] + + # text_encoder2 + enc_out = text_encoder2(input_ids2, output_hidden_states=True, return_dict=True) + hidden_states2 = enc_out["hidden_states"][-2] # penuultimate layer + + # pool2 = enc_out["text_embeds"] + unwrapped_text_encoder2 = text_encoder2 if accelerator is None else accelerator.unwrap_model(text_encoder2) + pool2 = pool_workaround(unwrapped_text_encoder2, enc_out["last_hidden_state"], input_ids2, tokenizer2.eos_token_id) + + # b*n, 77, 768 or 1280 -> b, n*77, 768 or 1280 + n_size = 1 if max_token_length is None else max_token_length // 75 + hidden_states1 = hidden_states1.reshape((b_size, -1, hidden_states1.shape[-1])) + hidden_states2 = hidden_states2.reshape((b_size, -1, hidden_states2.shape[-1])) + + if max_token_length is not None: + # bs*3, 77, 768 or 1024 + # encoder1: ... の三連を ... へ戻す + states_list = [hidden_states1[:, 0].unsqueeze(1)] # + for i in range(1, max_token_length, tokenizer1.model_max_length): + states_list.append(hidden_states1[:, i : i + tokenizer1.model_max_length - 2]) # の後から の前まで + states_list.append(hidden_states1[:, -1].unsqueeze(1)) # + hidden_states1 = torch.cat(states_list, dim=1) + + # v2: ... ... の三連を ... ... へ戻す　正直この実装でいいのかわからん + states_list = [hidden_states2[:, 0].unsqueeze(1)] # + for i in range(1, max_token_length, tokenizer2.model_max_length): + chunk = hidden_states2[:, i : i + tokenizer2.model_max_length - 2] # の後から 最後の前まで + # this causes an error: + # RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation + # if i > 1: + # for j in range(len(chunk)): # batch_size + # if input_ids2[n_index + j * n_size, 1] == tokenizer2.eos_token_id: # 空、つまり ...のパターン + # chunk[j, 0] = chunk[j, 1] # 次の の値をコピーする + states_list.append(chunk) # の後から の前まで + states_list.append(hidden_states2[:, -1].unsqueeze(1)) # か のどちらか + hidden_states2 = torch.cat(states_list, dim=1) + + # pool はnの最初のものを使う + pool2 = pool2[::n_size] + + if weight_dtype is not None: + # this is required for additional network training + hidden_states1 = hidden_states1.to(weight_dtype) + hidden_states2 = hidden_states2.to(weight_dtype) + + return hidden_states1, hidden_states2, pool2 + + +def default_if_none(value, default): + return default if value is None else value + + +def get_epoch_ckpt_name(args: argparse.Namespace, ext: str, epoch_no: int): + model_name = default_if_none(args.output_name, DEFAULT_EPOCH_NAME) + return EPOCH_FILE_NAME.format(model_name, epoch_no) + ext + + +def get_step_ckpt_name(args: argparse.Namespace, ext: str, step_no: int): + model_name = default_if_none(args.output_name, DEFAULT_STEP_NAME) + return STEP_FILE_NAME.format(model_name, step_no) + ext + + +def get_last_ckpt_name(args: argparse.Namespace, ext: str): + model_name = default_if_none(args.output_name, DEFAULT_LAST_OUTPUT_NAME) + return model_name + ext + + +def get_remove_epoch_no(args: argparse.Namespace, epoch_no: int): + if args.save_last_n_epochs is None: + return None + + remove_epoch_no = epoch_no - args.save_every_n_epochs * args.save_last_n_epochs + if remove_epoch_no < 0: + return None + return remove_epoch_no + + +def get_remove_step_no(args: argparse.Namespace, step_no: int): + if args.save_last_n_steps is None: + return None + + # last_n_steps前のstep_noから、save_every_n_stepsの倍数のstep_noを計算して削除する + # save_every_n_steps=10, save_last_n_steps=30の場合、50step目には30step分残し、10step目を削除する + remove_step_no = step_no - args.save_last_n_steps - 1 + remove_step_no = remove_step_no - (remove_step_no % args.save_every_n_steps) + if remove_step_no < 0: + return None + return remove_step_no + + +# epochとstepの保存、メタデータにepoch/stepが含まれ引数が同じになるため、統合している +# on_epoch_end: Trueならepoch終了時、Falseならstep経過時 +def save_sd_model_on_epoch_end_or_stepwise( + args: argparse.Namespace, + on_epoch_end: bool, + accelerator, + src_path: str, + save_stable_diffusion_format: bool, + use_safetensors: bool, + save_dtype: torch.dtype, + epoch: int, + num_train_epochs: int, + global_step: int, + text_encoder, + unet, + vae, +): + def sd_saver(ckpt_file, epoch_no, global_step): + sai_metadata = get_sai_model_spec(None, args, False, False, False, is_stable_diffusion_ckpt=True) + model_util.save_stable_diffusion_checkpoint( + args.v2, ckpt_file, text_encoder, unet, src_path, epoch_no, global_step, sai_metadata, save_dtype, vae + ) + + def diffusers_saver(out_dir): + model_util.save_diffusers_checkpoint( + args.v2, out_dir, text_encoder, unet, src_path, vae=vae, use_safetensors=use_safetensors + ) + + save_sd_model_on_epoch_end_or_stepwise_common( + args, + on_epoch_end, + accelerator, + save_stable_diffusion_format, + use_safetensors, + epoch, + num_train_epochs, + global_step, + sd_saver, + diffusers_saver, + ) + + +def save_sd_model_on_epoch_end_or_stepwise_common( + args: argparse.Namespace, + on_epoch_end: bool, + accelerator, + save_stable_diffusion_format: bool, + use_safetensors: bool, + epoch: int, + num_train_epochs: int, + global_step: int, + sd_saver, + diffusers_saver, +): + if on_epoch_end: + epoch_no = epoch + 1 + saving = epoch_no % args.save_every_n_epochs == 0 and epoch_no < num_train_epochs + if not saving: + return + + model_name = default_if_none(args.output_name, DEFAULT_EPOCH_NAME) + remove_no = get_remove_epoch_no(args, epoch_no) + else: + # 保存するか否かは呼び出し側で判断済み + + model_name = default_if_none(args.output_name, DEFAULT_STEP_NAME) + epoch_no = epoch # 例: 最初のepochの途中で保存したら0になる、SDモデルに保存される + remove_no = get_remove_step_no(args, global_step) + + os.makedirs(args.output_dir, exist_ok=True) + if save_stable_diffusion_format: + ext = ".safetensors" if use_safetensors else ".ckpt" + + if on_epoch_end: + ckpt_name = get_epoch_ckpt_name(args, ext, epoch_no) + else: + ckpt_name = get_step_ckpt_name(args, ext, global_step) + + ckpt_file = os.path.join(args.output_dir, ckpt_name) + logger.info("") + logger.info(f"saving checkpoint: {ckpt_file}") + sd_saver(ckpt_file, epoch_no, global_step) + + if args.huggingface_repo_id is not None: + huggingface_util.upload(args, ckpt_file, "/" + ckpt_name) + + # remove older checkpoints + if remove_no is not None: + if on_epoch_end: + remove_ckpt_name = get_epoch_ckpt_name(args, ext, remove_no) + else: + remove_ckpt_name = get_step_ckpt_name(args, ext, remove_no) + + remove_ckpt_file = os.path.join(args.output_dir, remove_ckpt_name) + if os.path.exists(remove_ckpt_file): + logger.info(f"removing old checkpoint: {remove_ckpt_file}") + os.remove(remove_ckpt_file) + + else: + if on_epoch_end: + out_dir = os.path.join(args.output_dir, EPOCH_DIFFUSERS_DIR_NAME.format(model_name, epoch_no)) + else: + out_dir = os.path.join(args.output_dir, STEP_DIFFUSERS_DIR_NAME.format(model_name, global_step)) + + logger.info("") + logger.info(f"saving model: {out_dir}") + diffusers_saver(out_dir) + + if args.huggingface_repo_id is not None: + huggingface_util.upload(args, out_dir, "/" + model_name) + + # remove older checkpoints + if remove_no is not None: + if on_epoch_end: + remove_out_dir = os.path.join(args.output_dir, EPOCH_DIFFUSERS_DIR_NAME.format(model_name, remove_no)) + else: + remove_out_dir = os.path.join(args.output_dir, STEP_DIFFUSERS_DIR_NAME.format(model_name, remove_no)) + + if os.path.exists(remove_out_dir): + logger.info(f"removing old model: {remove_out_dir}") + shutil.rmtree(remove_out_dir) + + if args.save_state: + if on_epoch_end: + save_and_remove_state_on_epoch_end(args, accelerator, epoch_no) + else: + save_and_remove_state_stepwise(args, accelerator, global_step) + + +def save_and_remove_state_on_epoch_end(args: argparse.Namespace, accelerator, epoch_no): + model_name = default_if_none(args.output_name, DEFAULT_EPOCH_NAME) + + logger.info("") + logger.info(f"saving state at epoch {epoch_no}") + os.makedirs(args.output_dir, exist_ok=True) + + state_dir = os.path.join(args.output_dir, EPOCH_STATE_NAME.format(model_name, epoch_no)) + accelerator.save_state(state_dir) + if args.save_state_to_huggingface: + logger.info("uploading state to huggingface.") + huggingface_util.upload(args, state_dir, "/" + EPOCH_STATE_NAME.format(model_name, epoch_no)) + + last_n_epochs = args.save_last_n_epochs_state if args.save_last_n_epochs_state else args.save_last_n_epochs + if last_n_epochs is not None: + remove_epoch_no = epoch_no - args.save_every_n_epochs * last_n_epochs + state_dir_old = os.path.join(args.output_dir, EPOCH_STATE_NAME.format(model_name, remove_epoch_no)) + if os.path.exists(state_dir_old): + logger.info(f"removing old state: {state_dir_old}") + shutil.rmtree(state_dir_old) + + +def save_and_remove_state_stepwise(args: argparse.Namespace, accelerator, step_no): + model_name = default_if_none(args.output_name, DEFAULT_STEP_NAME) + + logger.info("") + logger.info(f"saving state at step {step_no}") + os.makedirs(args.output_dir, exist_ok=True) + + state_dir = os.path.join(args.output_dir, STEP_STATE_NAME.format(model_name, step_no)) + accelerator.save_state(state_dir) + if args.save_state_to_huggingface: + logger.info("uploading state to huggingface.") + huggingface_util.upload(args, state_dir, "/" + STEP_STATE_NAME.format(model_name, step_no)) + + last_n_steps = args.save_last_n_steps_state if args.save_last_n_steps_state else args.save_last_n_steps + if last_n_steps is not None: + # last_n_steps前のstep_noから、save_every_n_stepsの倍数のstep_noを計算して削除する + remove_step_no = step_no - last_n_steps - 1 + remove_step_no = remove_step_no - (remove_step_no % args.save_every_n_steps) + + if remove_step_no > 0: + state_dir_old = os.path.join(args.output_dir, STEP_STATE_NAME.format(model_name, remove_step_no)) + if os.path.exists(state_dir_old): + logger.info(f"removing old state: {state_dir_old}") + shutil.rmtree(state_dir_old) + + +def save_state_on_train_end(args: argparse.Namespace, accelerator): + model_name = default_if_none(args.output_name, DEFAULT_LAST_OUTPUT_NAME) + + logger.info("") + logger.info("saving last state.") + os.makedirs(args.output_dir, exist_ok=True) + + state_dir = os.path.join(args.output_dir, LAST_STATE_NAME.format(model_name)) + accelerator.save_state(state_dir) + + if args.save_state_to_huggingface: + logger.info("uploading last state to huggingface.") + huggingface_util.upload(args, state_dir, "/" + LAST_STATE_NAME.format(model_name)) + + +def save_sd_model_on_train_end( + args: argparse.Namespace, + src_path: str, + save_stable_diffusion_format: bool, + use_safetensors: bool, + save_dtype: torch.dtype, + epoch: int, + global_step: int, + text_encoder, + unet, + vae, +): + def sd_saver(ckpt_file, epoch_no, global_step): + sai_metadata = get_sai_model_spec(None, args, False, False, False, is_stable_diffusion_ckpt=True) + model_util.save_stable_diffusion_checkpoint( + args.v2, ckpt_file, text_encoder, unet, src_path, epoch_no, global_step, sai_metadata, save_dtype, vae + ) + + def diffusers_saver(out_dir): + model_util.save_diffusers_checkpoint( + args.v2, out_dir, text_encoder, unet, src_path, vae=vae, use_safetensors=use_safetensors + ) + + save_sd_model_on_train_end_common( + args, save_stable_diffusion_format, use_safetensors, epoch, global_step, sd_saver, diffusers_saver + ) + + +def save_sd_model_on_train_end_common( + args: argparse.Namespace, + save_stable_diffusion_format: bool, + use_safetensors: bool, + epoch: int, + global_step: int, + sd_saver, + diffusers_saver, +): + model_name = default_if_none(args.output_name, DEFAULT_LAST_OUTPUT_NAME) + + if save_stable_diffusion_format: + os.makedirs(args.output_dir, exist_ok=True) + + ckpt_name = model_name + (".safetensors" if use_safetensors else ".ckpt") + ckpt_file = os.path.join(args.output_dir, ckpt_name) + + logger.info(f"save trained model as StableDiffusion checkpoint to {ckpt_file}") + sd_saver(ckpt_file, epoch, global_step) + + if args.huggingface_repo_id is not None: + huggingface_util.upload(args, ckpt_file, "/" + ckpt_name, force_sync_upload=True) + else: + out_dir = os.path.join(args.output_dir, model_name) + os.makedirs(out_dir, exist_ok=True) + + logger.info(f"save trained model as Diffusers to {out_dir}") + diffusers_saver(out_dir) + + if args.huggingface_repo_id is not None: + huggingface_util.upload(args, out_dir, "/" + model_name, force_sync_upload=True) + + +def get_timesteps_and_huber_c(args, min_timestep, max_timestep, noise_scheduler, b_size, device): + timesteps = torch.randint(min_timestep, max_timestep, (b_size,), device="cpu") + + if args.loss_type == "huber" or args.loss_type == "smooth_l1": + if args.huber_schedule == "exponential": + alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps + huber_c = torch.exp(-alpha * timesteps) + elif args.huber_schedule == "snr": + alphas_cumprod = torch.index_select(noise_scheduler.alphas_cumprod, 0, timesteps) + sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5 + huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c + elif args.huber_schedule == "constant": + huber_c = torch.full((b_size,), args.huber_c) + else: + raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!") + huber_c = huber_c.to(device) + elif args.loss_type == "l2": + huber_c = None # may be anything, as it's not used + else: + raise NotImplementedError(f"Unknown loss type {args.loss_type}") + + timesteps = timesteps.long().to(device) + return timesteps, huber_c + + +def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents): + # Sample noise that we'll add to the latents + noise = torch.randn_like(latents, device=latents.device) + if args.noise_offset: + if args.noise_offset_random_strength: + noise_offset = torch.rand(1, device=latents.device) * args.noise_offset + else: + noise_offset = args.noise_offset + noise = custom_train_functions.apply_noise_offset(latents, noise, noise_offset, args.adaptive_noise_scale) + if args.multires_noise_iterations: + noise = custom_train_functions.pyramid_noise_like( + noise, latents.device, args.multires_noise_iterations, args.multires_noise_discount + ) + + # Sample a random timestep for each image + b_size = latents.shape[0] + min_timestep = 0 if args.min_timestep is None else args.min_timestep + max_timestep = noise_scheduler.config.num_train_timesteps if args.max_timestep is None else args.max_timestep + + timesteps, huber_c = get_timesteps_and_huber_c(args, min_timestep, max_timestep, noise_scheduler, b_size, latents.device) + + # Add noise to the latents according to the noise magnitude at each timestep + # (this is the forward diffusion process) + if args.ip_noise_gamma: + if args.ip_noise_gamma_random_strength: + strength = torch.rand(1, device=latents.device) * args.ip_noise_gamma + else: + strength = args.ip_noise_gamma + noisy_latents = noise_scheduler.add_noise(latents, noise + strength * torch.randn_like(latents), timesteps) + else: + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + return noise, noisy_latents, timesteps, huber_c + + +def conditional_loss( + model_pred: torch.Tensor, target: torch.Tensor, reduction: str, loss_type: str, huber_c: Optional[torch.Tensor] +): + + if loss_type == "l2": + loss = torch.nn.functional.mse_loss(model_pred, target, reduction=reduction) + elif loss_type == "huber": + huber_c = huber_c.view(-1, 1, 1, 1) + loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c) + if reduction == "mean": + loss = torch.mean(loss) + elif reduction == "sum": + loss = torch.sum(loss) + elif loss_type == "smooth_l1": + huber_c = huber_c.view(-1, 1, 1, 1) + loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c) + if reduction == "mean": + loss = torch.mean(loss) + elif reduction == "sum": + loss = torch.sum(loss) + else: + raise NotImplementedError(f"Unsupported Loss Type {loss_type}") + return loss + + +def append_lr_to_logs(logs, lr_scheduler, optimizer_type, including_unet=True): + names = [] + if including_unet: + names.append("unet") + names.append("text_encoder1") + names.append("text_encoder2") + + append_lr_to_logs_with_names(logs, lr_scheduler, optimizer_type, names) + + +def append_lr_to_logs_with_names(logs, lr_scheduler, optimizer_type, names): + lrs = lr_scheduler.get_last_lr() + + for lr_index in range(len(lrs)): + name = names[lr_index] + logs["lr/" + name] = float(lrs[lr_index]) + + if optimizer_type.lower().startswith("DAdapt".lower()) or optimizer_type.lower() == "Prodigy".lower(): + logs["lr/d*lr/" + name] = ( + lr_scheduler.optimizers[-1].param_groups[lr_index]["d"] * lr_scheduler.optimizers[-1].param_groups[lr_index]["lr"] + ) + + +# scheduler: +SCHEDULER_LINEAR_START = 0.00085 +SCHEDULER_LINEAR_END = 0.0120 +SCHEDULER_TIMESTEPS = 1000 +SCHEDLER_SCHEDULE = "scaled_linear" + + +def get_my_scheduler( + *, + sample_sampler: str, + v_parameterization: bool, +): + sched_init_args = {} + if sample_sampler == "ddim": + scheduler_cls = DDIMScheduler + elif sample_sampler == "ddpm": # ddpmはおかしくなるのでoptionから外してある + scheduler_cls = DDPMScheduler + elif sample_sampler == "pndm": + scheduler_cls = PNDMScheduler + elif sample_sampler == "lms" or sample_sampler == "k_lms": + scheduler_cls = LMSDiscreteScheduler + elif sample_sampler == "euler" or sample_sampler == "k_euler": + scheduler_cls = EulerDiscreteScheduler + elif sample_sampler == "euler_a" or sample_sampler == "k_euler_a": + scheduler_cls = EulerAncestralDiscreteScheduler + elif sample_sampler == "dpmsolver" or sample_sampler == "dpmsolver++": + scheduler_cls = DPMSolverMultistepScheduler + sched_init_args["algorithm_type"] = sample_sampler + elif sample_sampler == "dpmsingle": + scheduler_cls = DPMSolverSinglestepScheduler + elif sample_sampler == "heun": + scheduler_cls = HeunDiscreteScheduler + elif sample_sampler == "dpm_2" or sample_sampler == "k_dpm_2": + scheduler_cls = KDPM2DiscreteScheduler + elif sample_sampler == "dpm_2_a" or sample_sampler == "k_dpm_2_a": + scheduler_cls = KDPM2AncestralDiscreteScheduler + else: + scheduler_cls = DDIMScheduler + + if v_parameterization: + sched_init_args["prediction_type"] = "v_prediction" + + scheduler = scheduler_cls( + num_train_timesteps=SCHEDULER_TIMESTEPS, + beta_start=SCHEDULER_LINEAR_START, + beta_end=SCHEDULER_LINEAR_END, + beta_schedule=SCHEDLER_SCHEDULE, + **sched_init_args, + ) + + # clip_sample=Trueにする + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is False: + # logger.info("set clip_sample to True") + scheduler.config.clip_sample = True + + return scheduler + + +def sample_images(*args, **kwargs): + return sample_images_common(StableDiffusionLongPromptWeightingPipeline, *args, **kwargs) + + +def line_to_prompt_dict(line: str) -> dict: + # subset of gen_img_diffusers + prompt_args = line.split(" --") + prompt_dict = {} + prompt_dict["prompt"] = prompt_args[0] + + for parg in prompt_args: + try: + m = re.match(r"w (\d+)", parg, re.IGNORECASE) + if m: + prompt_dict["width"] = int(m.group(1)) + continue + + m = re.match(r"h (\d+)", parg, re.IGNORECASE) + if m: + prompt_dict["height"] = int(m.group(1)) + continue + + m = re.match(r"d (\d+)", parg, re.IGNORECASE) + if m: + prompt_dict["seed"] = int(m.group(1)) + continue + + m = re.match(r"s (\d+)", parg, re.IGNORECASE) + if m: # steps + prompt_dict["sample_steps"] = max(1, min(1000, int(m.group(1)))) + continue + + m = re.match(r"l ([\d\.]+)", parg, re.IGNORECASE) + if m: # scale + prompt_dict["scale"] = float(m.group(1)) + continue + + m = re.match(r"n (.+)", parg, re.IGNORECASE) + if m: # negative prompt + prompt_dict["negative_prompt"] = m.group(1) + continue + + m = re.match(r"ss (.+)", parg, re.IGNORECASE) + if m: + prompt_dict["sample_sampler"] = m.group(1) + continue + + m = re.match(r"cn (.+)", parg, re.IGNORECASE) + if m: + prompt_dict["controlnet_image"] = m.group(1) + continue + + except ValueError as ex: + logger.error(f"Exception in parsing / 解析エラー: {parg}") + logger.error(ex) + + return prompt_dict + + +def sample_images_common( + pipe_class, + accelerator: Accelerator, + args: argparse.Namespace, + epoch, + steps, + device, + vae, + tokenizer, + text_encoder, + unet, + prompt_replacement=None, + controlnet=None, +): + """ + StableDiffusionLongPromptWeightingPipelineの改造版を使うようにしたので、clip skipおよびプロンプトの重みづけに対応した + """ + + if steps == 0: + if not args.sample_at_first: + return + else: + if args.sample_every_n_steps is None and args.sample_every_n_epochs is None: + return + if args.sample_every_n_epochs is not None: + # sample_every_n_steps は無視する + if epoch is None or epoch % args.sample_every_n_epochs != 0: + return + else: + if steps % args.sample_every_n_steps != 0 or epoch is not None: # steps is not divisible or end of epoch + return + + logger.info("") + logger.info(f"generating sample images at step / サンプル画像生成 ステップ: {steps}") + if not os.path.isfile(args.sample_prompts): + logger.error(f"No prompt file / プロンプトファイルがありません: {args.sample_prompts}") + return + + distributed_state = PartialState() # for multi gpu distributed inference. this is a singleton, so it's safe to use it here + + org_vae_device = vae.device # CPUにいるはず + vae.to(distributed_state.device) # distributed_state.device is same as accelerator.device + + # unwrap unet and text_encoder(s) + unet = accelerator.unwrap_model(unet) + if isinstance(text_encoder, (list, tuple)): + text_encoder = [accelerator.unwrap_model(te) for te in text_encoder] + else: + text_encoder = accelerator.unwrap_model(text_encoder) + + # read prompts + if args.sample_prompts.endswith(".txt"): + with open(args.sample_prompts, "r", encoding="utf-8") as f: + lines = f.readlines() + prompts = [line.strip() for line in lines if len(line.strip()) > 0 and line[0] != "#"] + elif args.sample_prompts.endswith(".toml"): + with open(args.sample_prompts, "r", encoding="utf-8") as f: + data = toml.load(f) + prompts = [dict(**data["prompt"], **subset) for subset in data["prompt"]["subset"]] + elif args.sample_prompts.endswith(".json"): + with open(args.sample_prompts, "r", encoding="utf-8") as f: + prompts = json.load(f) + + # schedulers: dict = {} cannot find where this is used + default_scheduler = get_my_scheduler( + sample_sampler=args.sample_sampler, + v_parameterization=args.v_parameterization, + ) + + pipeline = pipe_class( + text_encoder=text_encoder, + vae=vae, + unet=unet, + tokenizer=tokenizer, + scheduler=default_scheduler, + safety_checker=None, + feature_extractor=None, + requires_safety_checker=False, + clip_skip=args.clip_skip, + ) + pipeline.to(distributed_state.device) + save_dir = args.output_dir + "/sample" + os.makedirs(save_dir, exist_ok=True) + + # preprocess prompts + for i in range(len(prompts)): + prompt_dict = prompts[i] + if isinstance(prompt_dict, str): + prompt_dict = line_to_prompt_dict(prompt_dict) + prompts[i] = prompt_dict + assert isinstance(prompt_dict, dict) + + # Adds an enumerator to the dict based on prompt position. Used later to name image files. Also cleanup of extra data in original prompt dict. + prompt_dict["enum"] = i + prompt_dict.pop("subset", None) + + # save random state to restore later + rng_state = torch.get_rng_state() + cuda_rng_state = None + try: + cuda_rng_state = torch.cuda.get_rng_state() if torch.cuda.is_available() else None + except Exception: + pass + + if distributed_state.num_processes <= 1: + # If only one device is available, just use the original prompt list. We don't need to care about the distribution of prompts. + with torch.no_grad(): + for prompt_dict in prompts: + sample_image_inference( + accelerator, args, pipeline, save_dir, prompt_dict, epoch, steps, prompt_replacement, controlnet=controlnet + ) + else: + # Creating list with N elements, where each element is a list of prompt_dicts, and N is the number of processes available (number of devices available) + # prompt_dicts are assigned to lists based on order of processes, to attempt to time the image creation time to match enum order. Probably only works when steps and sampler are identical. + per_process_prompts = [] # list of lists + for i in range(distributed_state.num_processes): + per_process_prompts.append(prompts[i :: distributed_state.num_processes]) + + with torch.no_grad(): + with distributed_state.split_between_processes(per_process_prompts) as prompt_dict_lists: + for prompt_dict in prompt_dict_lists[0]: + sample_image_inference( + accelerator, args, pipeline, save_dir, prompt_dict, epoch, steps, prompt_replacement, controlnet=controlnet + ) + + # clear pipeline and cache to reduce vram usage + del pipeline + + # I'm not sure which of these is the correct way to clear the memory, but accelerator's device is used in the pipeline, so I'm using it here. + # with torch.cuda.device(torch.cuda.current_device()): + # torch.cuda.empty_cache() + clean_memory_on_device(accelerator.device) + + torch.set_rng_state(rng_state) + if torch.cuda.is_available() and cuda_rng_state is not None: + torch.cuda.set_rng_state(cuda_rng_state) + vae.to(org_vae_device) + + +def sample_image_inference( + accelerator: Accelerator, + args: argparse.Namespace, + pipeline, + save_dir, + prompt_dict, + epoch, + steps, + prompt_replacement, + controlnet=None, +): + assert isinstance(prompt_dict, dict) + negative_prompt = prompt_dict.get("negative_prompt") + sample_steps = prompt_dict.get("sample_steps", 30) + width = prompt_dict.get("width", 512) + height = prompt_dict.get("height", 512) + scale = prompt_dict.get("scale", 7.5) + seed = prompt_dict.get("seed") + controlnet_image = prompt_dict.get("controlnet_image") + prompt: str = prompt_dict.get("prompt", "") + sampler_name: str = prompt_dict.get("sample_sampler", args.sample_sampler) + + if prompt_replacement is not None: + prompt = prompt.replace(prompt_replacement[0], prompt_replacement[1]) + if negative_prompt is not None: + negative_prompt = negative_prompt.replace(prompt_replacement[0], prompt_replacement[1]) + + if seed is not None: + torch.manual_seed(seed) + if torch.cuda.is_available(): + torch.cuda.manual_seed(seed) + else: + # True random sample image generation + torch.seed() + if torch.cuda.is_available(): + torch.cuda.seed() + + scheduler = get_my_scheduler( + sample_sampler=sampler_name, + v_parameterization=args.v_parameterization, + ) + pipeline.scheduler = scheduler + + if controlnet_image is not None: + controlnet_image = Image.open(controlnet_image).convert("RGB") + controlnet_image = controlnet_image.resize((width, height), Image.LANCZOS) + + height = max(64, height - height % 8) # round to divisible by 8 + width = max(64, width - width % 8) # round to divisible by 8 + logger.info(f"prompt: {prompt}") + logger.info(f"negative_prompt: {negative_prompt}") + logger.info(f"height: {height}") + logger.info(f"width: {width}") + logger.info(f"sample_steps: {sample_steps}") + logger.info(f"scale: {scale}") + logger.info(f"sample_sampler: {sampler_name}") + if seed is not None: + logger.info(f"seed: {seed}") + with accelerator.autocast(): + latents = pipeline( + prompt=prompt, + height=height, + width=width, + num_inference_steps=sample_steps, + guidance_scale=scale, + negative_prompt=negative_prompt, + controlnet=controlnet, + controlnet_image=controlnet_image, + ) + + if torch.cuda.is_available(): + with torch.cuda.device(torch.cuda.current_device()): + torch.cuda.empty_cache() + + image = pipeline.latents_to_image(latents)[0] + + # adding accelerator.wait_for_everyone() here should sync up and ensure that sample images are saved in the same order as the original prompt list + # but adding 'enum' to the filename should be enough + + ts_str = time.strftime("%Y%m%d%H%M%S", time.localtime()) + num_suffix = f"e{epoch:06d}" if epoch is not None else f"{steps:06d}" + seed_suffix = "" if seed is None else f"_{seed}" + i: int = prompt_dict["enum"] + img_filename = f"{'' if args.output_name is None else args.output_name + '_'}{num_suffix}_{i:02d}_{ts_str}{seed_suffix}.png" + image.save(os.path.join(save_dir, img_filename)) + + # wandb有効時のみログを送信 + try: + wandb_tracker = accelerator.get_tracker("wandb") + try: + import wandb + except ImportError: # 事前に一度確認するのでここはエラー出ないはず + raise ImportError("No wandb / wandb がインストールされていないようです") + + wandb_tracker.log({f"sample_{i}": wandb.Image(image)}) + except: # wandb 無効時 + pass + + +# endregion + + +# region 前処理用 + + +class ImageLoadingDataset(torch.utils.data.Dataset): + def __init__(self, image_paths): + self.images = image_paths + + def __len__(self): + return len(self.images) + + def __getitem__(self, idx): + img_path = self.images[idx] + + try: + image = Image.open(img_path).convert("RGB") + # convert to tensor temporarily so dataloader will accept it + tensor_pil = transforms.functional.pil_to_tensor(image) + except Exception as e: + logger.error(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}") + return None + + return (tensor_pil, img_path) + + +# endregion + + +# collate_fn用 epoch,stepはmultiprocessing.Value +class collator_class: + def __init__(self, epoch, step, dataset): + self.current_epoch = epoch + self.current_step = step + self.dataset = dataset # not used if worker_info is not None, in case of multiprocessing + + def __call__(self, examples): + worker_info = torch.utils.data.get_worker_info() + # worker_info is None in the main process + if worker_info is not None: + dataset = worker_info.dataset + else: + dataset = self.dataset + + # set epoch and step + dataset.set_current_epoch(self.current_epoch.value) + dataset.set_current_step(self.current_step.value) + return examples[0] + + +class LossRecorder: + def __init__(self): + self.loss_list: List[float] = [] + self.loss_total: float = 0.0 + + def add(self, *, epoch: int, step: int, loss: float) -> None: + if epoch == 0: + self.loss_list.append(loss) + else: + while len(self.loss_list) <= step: + self.loss_list.append(0.0) + self.loss_total -= self.loss_list[step] + self.loss_list[step] = loss + self.loss_total += loss + + @property + def moving_average(self) -> float: + return self.loss_total / len(self.loss_list) diff --git a/typos.yml b/typos.yml new file mode 100644 index 0000000000000000000000000000000000000000..0149dcdd366340c4c68de6399bf2b101d654140b --- /dev/null +++ b/typos.yml @@ -0,0 +1,21 @@ +--- +# yamllint disable rule:line-length +name: Typos + +on: # yamllint disable-line rule:truthy + push: + pull_request: + types: + - opened + - synchronize + - reopened + +jobs: + build: + runs-on: ubuntu-latest + + steps: + - uses: actions/checkout@v4 + + - name: typos-action + uses: crate-ci/typos@v1.24.3 diff --git a/utils.py b/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..49d46a54604483aa993364d3ff98aec0bd460b45 --- /dev/null +++ b/utils.py @@ -0,0 +1,287 @@ +import logging +import sys +import threading +import torch +from torchvision import transforms +from typing import * +from diffusers import EulerAncestralDiscreteScheduler +import diffusers.schedulers.scheduling_euler_ancestral_discrete +from diffusers.schedulers.scheduling_euler_ancestral_discrete import EulerAncestralDiscreteSchedulerOutput +import cv2 +from PIL import Image +import numpy as np + + +def fire_in_thread(f, *args, **kwargs): + threading.Thread(target=f, args=args, kwargs=kwargs).start() + + +def add_logging_arguments(parser): + parser.add_argument( + "--console_log_level", + type=str, + default=None, + choices=["DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"], + help="Set the logging level, default is INFO / ログレベルを設定する。デフォルトはINFO", + ) + parser.add_argument( + "--console_log_file", + type=str, + default=None, + help="Log to a file instead of stderr / 標準エラー出力ではなくファイルにログを出力する", + ) + parser.add_argument("--console_log_simple", action="store_true", help="Simple log output / シンプルなログ出力") + + +def setup_logging(args=None, log_level=None, reset=False): + if logging.root.handlers: + if reset: + # remove all handlers + for handler in logging.root.handlers[:]: + logging.root.removeHandler(handler) + else: + return + + # log_level can be set by the caller or by the args, the caller has priority. If not set, use INFO + if log_level is None and args is not None: + log_level = args.console_log_level + if log_level is None: + log_level = "INFO" + log_level = getattr(logging, log_level) + + msg_init = None + if args is not None and args.console_log_file: + handler = logging.FileHandler(args.console_log_file, mode="w") + else: + handler = None + if not args or not args.console_log_simple: + try: + from rich.logging import RichHandler + from rich.console import Console + from rich.logging import RichHandler + + handler = RichHandler(console=Console(stderr=True)) + except ImportError: + # print("rich is not installed, using basic logging") + msg_init = "rich is not installed, using basic logging" + + if handler is None: + handler = logging.StreamHandler(sys.stdout) # same as print + handler.propagate = False + + formatter = logging.Formatter( + fmt="%(message)s", + datefmt="%Y-%m-%d %H:%M:%S", + ) + handler.setFormatter(formatter) + logging.root.setLevel(log_level) + logging.root.addHandler(handler) + + if msg_init is not None: + logger = logging.getLogger(__name__) + logger.info(msg_init) + + +def pil_resize(image, size, interpolation=Image.LANCZOS): + has_alpha = image.shape[2] == 4 if len(image.shape) == 3 else False + + if has_alpha: + pil_image = Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGRA2RGBA)) + else: + pil_image = Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGR2RGB)) + + resized_pil = pil_image.resize(size, interpolation) + + # Convert back to cv2 format + if has_alpha: + resized_cv2 = cv2.cvtColor(np.array(resized_pil), cv2.COLOR_RGBA2BGRA) + else: + resized_cv2 = cv2.cvtColor(np.array(resized_pil), cv2.COLOR_RGB2BGR) + + return resized_cv2 + + +# TODO make inf_utils.py + + +# region Gradual Latent hires fix + + +class GradualLatent: + def __init__( + self, + ratio, + start_timesteps, + every_n_steps, + ratio_step, + s_noise=1.0, + gaussian_blur_ksize=None, + gaussian_blur_sigma=0.5, + gaussian_blur_strength=0.5, + unsharp_target_x=True, + ): + self.ratio = ratio + self.start_timesteps = start_timesteps + self.every_n_steps = every_n_steps + self.ratio_step = ratio_step + self.s_noise = s_noise + self.gaussian_blur_ksize = gaussian_blur_ksize + self.gaussian_blur_sigma = gaussian_blur_sigma + self.gaussian_blur_strength = gaussian_blur_strength + self.unsharp_target_x = unsharp_target_x + + def __str__(self) -> str: + return ( + f"GradualLatent(ratio={self.ratio}, start_timesteps={self.start_timesteps}, " + + f"every_n_steps={self.every_n_steps}, ratio_step={self.ratio_step}, s_noise={self.s_noise}, " + + f"gaussian_blur_ksize={self.gaussian_blur_ksize}, gaussian_blur_sigma={self.gaussian_blur_sigma}, gaussian_blur_strength={self.gaussian_blur_strength}, " + + f"unsharp_target_x={self.unsharp_target_x})" + ) + + def apply_unshark_mask(self, x: torch.Tensor): + if self.gaussian_blur_ksize is None: + return x + blurred = transforms.functional.gaussian_blur(x, self.gaussian_blur_ksize, self.gaussian_blur_sigma) + # mask = torch.sigmoid((x - blurred) * self.gaussian_blur_strength) + mask = (x - blurred) * self.gaussian_blur_strength + sharpened = x + mask + return sharpened + + def interpolate(self, x: torch.Tensor, resized_size, unsharp=True): + org_dtype = x.dtype + if org_dtype == torch.bfloat16: + x = x.float() + + x = torch.nn.functional.interpolate(x, size=resized_size, mode="bicubic", align_corners=False).to(dtype=org_dtype) + + # apply unsharp mask / アンシャープマスクを適用する + if unsharp and self.gaussian_blur_ksize: + x = self.apply_unshark_mask(x) + + return x + + +class EulerAncestralDiscreteSchedulerGL(EulerAncestralDiscreteScheduler): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.resized_size = None + self.gradual_latent = None + + def set_gradual_latent_params(self, size, gradual_latent: GradualLatent): + self.resized_size = size + self.gradual_latent = gradual_latent + + def step( + self, + model_output: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + sample: torch.FloatTensor, + generator: Optional[torch.Generator] = None, + return_dict: bool = True, + ) -> Union[EulerAncestralDiscreteSchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + + """ + + if isinstance(timestep, int) or isinstance(timestep, torch.IntTensor) or isinstance(timestep, torch.LongTensor): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + # logger.warning( + print( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma * model_output + elif self.config.prediction_type == "v_prediction": + # * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError(f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`") + + sigma_from = self.sigmas[self.step_index] + sigma_to = self.sigmas[self.step_index + 1] + sigma_up = (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5 + sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5 + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma + + dt = sigma_down - sigma + + device = model_output.device + if self.resized_size is None: + prev_sample = sample + derivative * dt + + noise = diffusers.schedulers.scheduling_euler_ancestral_discrete.randn_tensor( + model_output.shape, dtype=model_output.dtype, device=device, generator=generator + ) + s_noise = 1.0 + else: + print("resized_size", self.resized_size, "model_output.shape", model_output.shape, "sample.shape", sample.shape) + s_noise = self.gradual_latent.s_noise + + if self.gradual_latent.unsharp_target_x: + prev_sample = sample + derivative * dt + prev_sample = self.gradual_latent.interpolate(prev_sample, self.resized_size) + else: + sample = self.gradual_latent.interpolate(sample, self.resized_size) + derivative = self.gradual_latent.interpolate(derivative, self.resized_size, unsharp=False) + prev_sample = sample + derivative * dt + + noise = diffusers.schedulers.scheduling_euler_ancestral_discrete.randn_tensor( + (model_output.shape[0], model_output.shape[1], self.resized_size[0], self.resized_size[1]), + dtype=model_output.dtype, + device=device, + generator=generator, + ) + + prev_sample = prev_sample + noise * sigma_up * s_noise + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return EulerAncestralDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + +# endregion diff --git a/vit.py b/vit.py new file mode 100644 index 0000000000000000000000000000000000000000..cec3d8e08ed4451d65392feb2e9f4848d1ef3899 --- /dev/null +++ b/vit.py @@ -0,0 +1,305 @@ +''' + * Copyright (c) 2022, salesforce.com, inc. + * All rights reserved. + * SPDX-License-Identifier: BSD-3-Clause + * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause + * By Junnan Li + * Based on timm code base + * https://github.com/rwightman/pytorch-image-models/tree/master/timm +''' + +import torch +import torch.nn as nn +import torch.nn.functional as F +from functools import partial + +from timm.models.vision_transformer import _cfg, PatchEmbed +from timm.models.registry import register_model +from timm.models.layers import trunc_normal_, DropPath +from timm.models.helpers import named_apply, adapt_input_conv + +from fairscale.nn.checkpoint.checkpoint_activations import checkpoint_wrapper + +class Mlp(nn.Module): + """ MLP as used in Vision Transformer, MLP-Mixer and related networks + """ + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.fc2(x) + x = self.drop(x) + return x + + +class Attention(nn.Module): + def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.): + super().__init__() + self.num_heads = num_heads + head_dim = dim // num_heads + # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights + self.scale = qk_scale or head_dim ** -0.5 + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + self.attn_gradients = None + self.attention_map = None + + def save_attn_gradients(self, attn_gradients): + self.attn_gradients = attn_gradients + + def get_attn_gradients(self): + return self.attn_gradients + + def save_attention_map(self, attention_map): + self.attention_map = attention_map + + def get_attention_map(self): + return self.attention_map + + def forward(self, x, register_hook=False): + B, N, C = x.shape + qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + attn = (q @ k.transpose(-2, -1)) * self.scale + attn = attn.softmax(dim=-1) + attn = self.attn_drop(attn) + + if register_hook: + self.save_attention_map(attn) + attn.register_hook(self.save_attn_gradients) + + x = (attn @ v).transpose(1, 2).reshape(B, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class Block(nn.Module): + + def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., + drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_grad_checkpointing=False): + super().__init__() + self.norm1 = norm_layer(dim) + self.attn = Attention( + dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) + # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + if use_grad_checkpointing: + self.attn = checkpoint_wrapper(self.attn) + self.mlp = checkpoint_wrapper(self.mlp) + + def forward(self, x, register_hook=False): + x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook)) + x = x + self.drop_path(self.mlp(self.norm2(x))) + return x + + +class VisionTransformer(nn.Module): + """ Vision Transformer + A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` - + https://arxiv.org/abs/2010.11929 + """ + def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12, + num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None, + drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None, + use_grad_checkpointing=False, ckpt_layer=0): + """ + Args: + img_size (int, tuple): input image size + patch_size (int, tuple): patch size + in_chans (int): number of input channels + num_classes (int): number of classes for classification head + embed_dim (int): embedding dimension + depth (int): depth of transformer + num_heads (int): number of attention heads + mlp_ratio (int): ratio of mlp hidden dim to embedding dim + qkv_bias (bool): enable bias for qkv if True + qk_scale (float): override default qk scale of head_dim ** -0.5 if set + representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set + drop_rate (float): dropout rate + attn_drop_rate (float): attention dropout rate + drop_path_rate (float): stochastic depth rate + norm_layer: (nn.Module): normalization layer + """ + super().__init__() + self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models + norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) + + self.patch_embed = PatchEmbed( + img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) + + num_patches = self.patch_embed.num_patches + + self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) + self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim)) + self.pos_drop = nn.Dropout(p=drop_rate) + + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule + self.blocks = nn.ModuleList([ + Block( + dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, + drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, + use_grad_checkpointing=(use_grad_checkpointing and i>=depth-ckpt_layer) + ) + for i in range(depth)]) + self.norm = norm_layer(embed_dim) + + trunc_normal_(self.pos_embed, std=.02) + trunc_normal_(self.cls_token, std=.02) + self.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + @torch.jit.ignore + def no_weight_decay(self): + return {'pos_embed', 'cls_token'} + + def forward(self, x, register_blk=-1): + B = x.shape[0] + x = self.patch_embed(x) + + cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks + x = torch.cat((cls_tokens, x), dim=1) + + x = x + self.pos_embed[:,:x.size(1),:] + x = self.pos_drop(x) + + for i,blk in enumerate(self.blocks): + x = blk(x, register_blk==i) + x = self.norm(x) + + return x + + @torch.jit.ignore() + def load_pretrained(self, checkpoint_path, prefix=''): + _load_weights(self, checkpoint_path, prefix) + + +@torch.no_grad() +def _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = ''): + """ Load weights from .npz checkpoints for official Google Brain Flax implementation + """ + import numpy as np + + def _n2p(w, t=True): + if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1: + w = w.flatten() + if t: + if w.ndim == 4: + w = w.transpose([3, 2, 0, 1]) + elif w.ndim == 3: + w = w.transpose([2, 0, 1]) + elif w.ndim == 2: + w = w.transpose([1, 0]) + return torch.from_numpy(w) + + w = np.load(checkpoint_path) + if not prefix and 'opt/target/embedding/kernel' in w: + prefix = 'opt/target/' + + if hasattr(model.patch_embed, 'backbone'): + # hybrid + backbone = model.patch_embed.backbone + stem_only = not hasattr(backbone, 'stem') + stem = backbone if stem_only else backbone.stem + stem.conv.weight.copy_(adapt_input_conv(stem.conv.weight.shape[1], _n2p(w[f'{prefix}conv_root/kernel']))) + stem.norm.weight.copy_(_n2p(w[f'{prefix}gn_root/scale'])) + stem.norm.bias.copy_(_n2p(w[f'{prefix}gn_root/bias'])) + if not stem_only: + for i, stage in enumerate(backbone.stages): + for j, block in enumerate(stage.blocks): + bp = f'{prefix}block{i + 1}/unit{j + 1}/' + for r in range(3): + getattr(block, f'conv{r + 1}').weight.copy_(_n2p(w[f'{bp}conv{r + 1}/kernel'])) + getattr(block, f'norm{r + 1}').weight.copy_(_n2p(w[f'{bp}gn{r + 1}/scale'])) + getattr(block, f'norm{r + 1}').bias.copy_(_n2p(w[f'{bp}gn{r + 1}/bias'])) + if block.downsample is not None: + block.downsample.conv.weight.copy_(_n2p(w[f'{bp}conv_proj/kernel'])) + block.downsample.norm.weight.copy_(_n2p(w[f'{bp}gn_proj/scale'])) + block.downsample.norm.bias.copy_(_n2p(w[f'{bp}gn_proj/bias'])) + embed_conv_w = _n2p(w[f'{prefix}embedding/kernel']) + else: + embed_conv_w = adapt_input_conv( + model.patch_embed.proj.weight.shape[1], _n2p(w[f'{prefix}embedding/kernel'])) + model.patch_embed.proj.weight.copy_(embed_conv_w) + model.patch_embed.proj.bias.copy_(_n2p(w[f'{prefix}embedding/bias'])) + model.cls_token.copy_(_n2p(w[f'{prefix}cls'], t=False)) + pos_embed_w = _n2p(w[f'{prefix}Transformer/posembed_input/pos_embedding'], t=False) + if pos_embed_w.shape != model.pos_embed.shape: + pos_embed_w = resize_pos_embed( # resize pos embedding when different size from pretrained weights + pos_embed_w, model.pos_embed, getattr(model, 'num_tokens', 1), model.patch_embed.grid_size) + model.pos_embed.copy_(pos_embed_w) + model.norm.weight.copy_(_n2p(w[f'{prefix}Transformer/encoder_norm/scale'])) + model.norm.bias.copy_(_n2p(w[f'{prefix}Transformer/encoder_norm/bias'])) +# if isinstance(model.head, nn.Linear) and model.head.bias.shape[0] == w[f'{prefix}head/bias'].shape[-1]: +# model.head.weight.copy_(_n2p(w[f'{prefix}head/kernel'])) +# model.head.bias.copy_(_n2p(w[f'{prefix}head/bias'])) +# if isinstance(getattr(model.pre_logits, 'fc', None), nn.Linear) and f'{prefix}pre_logits/bias' in w: +# model.pre_logits.fc.weight.copy_(_n2p(w[f'{prefix}pre_logits/kernel'])) +# model.pre_logits.fc.bias.copy_(_n2p(w[f'{prefix}pre_logits/bias'])) + for i, block in enumerate(model.blocks.children()): + block_prefix = f'{prefix}Transformer/encoderblock_{i}/' + mha_prefix = block_prefix + 'MultiHeadDotProductAttention_1/' + block.norm1.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/scale'])) + block.norm1.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_0/bias'])) + block.attn.qkv.weight.copy_(torch.cat([ + _n2p(w[f'{mha_prefix}{n}/kernel'], t=False).flatten(1).T for n in ('query', 'key', 'value')])) + block.attn.qkv.bias.copy_(torch.cat([ + _n2p(w[f'{mha_prefix}{n}/bias'], t=False).reshape(-1) for n in ('query', 'key', 'value')])) + block.attn.proj.weight.copy_(_n2p(w[f'{mha_prefix}out/kernel']).flatten(1)) + block.attn.proj.bias.copy_(_n2p(w[f'{mha_prefix}out/bias'])) + for r in range(2): + getattr(block.mlp, f'fc{r + 1}').weight.copy_(_n2p(w[f'{block_prefix}MlpBlock_3/Dense_{r}/kernel'])) + getattr(block.mlp, f'fc{r + 1}').bias.copy_(_n2p(w[f'{block_prefix}MlpBlock_3/Dense_{r}/bias'])) + block.norm2.weight.copy_(_n2p(w[f'{block_prefix}LayerNorm_2/scale'])) + block.norm2.bias.copy_(_n2p(w[f'{block_prefix}LayerNorm_2/bias'])) + + +def interpolate_pos_embed(pos_embed_checkpoint, visual_encoder): + # interpolate position embedding + embedding_size = pos_embed_checkpoint.shape[-1] + num_patches = visual_encoder.patch_embed.num_patches + num_extra_tokens = visual_encoder.pos_embed.shape[-2] - num_patches + # height (== width) for the checkpoint position embedding + orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5) + # height (== width) for the new position embedding + new_size = int(num_patches ** 0.5) + + if orig_size!=new_size: + # class_token and dist_token are kept unchanged + extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] + # only the position tokens are interpolated + pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] + pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) + pos_tokens = torch.nn.functional.interpolate( + pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) + pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2) + new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) + print('reshape position embedding from %d to %d'%(orig_size ** 2,new_size ** 2)) + + return new_pos_embed + else: + return pos_embed_checkpoint \ No newline at end of file diff --git a/wd14_tagger_README-en.md b/wd14_tagger_README-en.md new file mode 100644 index 0000000000000000000000000000000000000000..34f4488230f02b4148f8593302120583a79e95d1 --- /dev/null +++ b/wd14_tagger_README-en.md @@ -0,0 +1,88 @@ +# Image Tagging using WD14Tagger + +This document is based on the information from this github page (https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger). + +Using onnx for inference is recommended. Please install onnx with the following command: + +```powershell +pip install onnx==1.15.0 onnxruntime-gpu==1.17.1 +``` + +The model weights will be automatically downloaded from Hugging Face. + +# Usage + +Run the script to perform tagging. + +```powershell +python finetune/tag_images_by_wd14_tagger.py --onnx --repo_id --batch_size +``` + +For example, if using the repository `SmilingWolf/wd-swinv2-tagger-v3` with a batch size of 4, and the training data is located in the parent folder `train_data`, it would be: + +```powershell +python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 --batch_size 4 ..\train_data +``` + +On the first run, the model files will be automatically downloaded to the `wd14_tagger_model` folder (the folder can be changed with an option). + +Tag files will be created in the same directory as the training data images, with the same filename and a `.txt` extension. + +![Generated tag files](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png) + +![Tags and image](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png) + +## Example + +To output in the Animagine XL 3.1 format, it would be as follows (enter on a single line in practice): + +``` +python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 + --batch_size 4 --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive + --use_rating_tags_as_last_tag --character_tags_first --character_tag_expand + --always_first_tags "1girl,1boy" ..\train_data +``` + +## Available Repository IDs + +[SmilingWolf's V2 and V3 models](https://huggingface.co/SmilingWolf) are available for use. Specify them in the format like `SmilingWolf/wd-vit-tagger-v3`. The default when omitted is `SmilingWolf/wd-v1-4-convnext-tagger-v2`. + +# Options + +## General Options + +- `--onnx`: Use ONNX for inference. If not specified, TensorFlow will be used. If using TensorFlow, please install TensorFlow separately. +- `--batch_size`: Number of images to process at once. Default is 1. Adjust according to VRAM capacity. +- `--caption_extension`: File extension for caption files. Default is `.txt`. +- `--max_data_loader_n_workers`: Maximum number of workers for DataLoader. Specifying a value of 1 or more will use DataLoader to speed up image loading. If unspecified, DataLoader will not be used. +- `--thresh`: Confidence threshold for outputting tags. Default is 0.35. Lowering the value will assign more tags but accuracy will decrease. +- `--general_threshold`: Confidence threshold for general tags. If omitted, same as `--thresh`. +- `--character_threshold`: Confidence threshold for character tags. If omitted, same as `--thresh`. +- `--recursive`: If specified, subfolders within the specified folder will also be processed recursively. +- `--append_tags`: Append tags to existing tag files. +- `--frequency_tags`: Output tag frequencies. +- `--debug`: Debug mode. Outputs debug information if specified. + +## Model Download + +- `--model_dir`: Folder to save model files. Default is `wd14_tagger_model`. +- `--force_download`: Re-download model files if specified. + +## Tag Editing + +- `--remove_underscore`: Remove underscores from output tags. +- `--undesired_tags`: Specify tags not to output. Multiple tags can be specified, separated by commas. For example, `black eyes,black hair`. +- `--use_rating_tags`: Output rating tags at the beginning of the tags. +- `--use_rating_tags_as_last_tag`: Add rating tags at the end of the tags. +- `--character_tags_first`: Output character tags first. +- `--character_tag_expand`: Expand character tag series names. For example, split the tag `chara_name_(series)` into `chara_name, series`. +- `--always_first_tags`: Specify tags to always output first when a certain tag appears in an image. Multiple tags can be specified, separated by commas. For example, `1girl,1boy`. +- `--caption_separator`: Separate tags with this string in the output file. Default is `, `. +- `--tag_replacement`: Perform tag replacement. Specify in the format `tag1,tag2;tag3,tag4`. If using `,` and `;`, escape them with `\`. \ + For example, specify `aira tsubase,aira tsubase (uniform)` (when you want to train a specific costume), `aira tsubase,aira tsubase\, heir of shadows` (when the series name is not included in the tag). + +When using `tag_replacement`, it is applied after `character_tag_expand`. + +When specifying `remove_underscore`, specify `undesired_tags`, `always_first_tags`, and `tag_replacement` without including underscores. + +When specifying `caption_separator`, separate `undesired_tags` and `always_first_tags` with `caption_separator`. Always separate `tag_replacement` with `,`. diff --git a/wd14_tagger_README-ja.md b/wd14_tagger_README-ja.md new file mode 100644 index 0000000000000000000000000000000000000000..58e9ede95a941b0d23abaef64d400afc3a2343b4 --- /dev/null +++ b/wd14_tagger_README-ja.md @@ -0,0 +1,88 @@ +# WD14Taggerによるタグ付け + +こちらのgithubページ（https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger ）の情報を参考にさせていただきました。 + +onnx を用いた推論を推奨します。以下のコマンドで onnx をインストールしてください。 + +```powershell +pip install onnx==1.15.0 onnxruntime-gpu==1.17.1 +``` + +モデルの重みはHugging Faceから自動的にダウンロードしてきます。 + +# 使い方 + +スクリプトを実行してタグ付けを行います。 +``` +python fintune/tag_images_by_wd14_tagger.py --onnx --repo_id <モデルのrepo id> --batch_size <バッチサイズ> <教師データフォルダ> +``` + +レポジトリに `SmilingWolf/wd-swinv2-tagger-v3` を使用し、バッチサイズを4にして、教師データを親フォルダの `train_data`に置いた場合、以下のようになります。 + +``` +python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 --batch_size 4 ..\train_data +``` + +初回起動時にはモデルファイルが `wd14_tagger_model` フォルダに自動的にダウンロードされます（フォルダはオプションで変えられます）。 + +タグファイルが教師データ画像と同じディレクトリに、同じファイル名、拡張子.txtで作成されます。 + +![生成されたタグファイル](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png) + +![タグと画像](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png) + +## 記述例 + +Animagine XL 3.1 方式で出力する場合、以下のようになります（実際には 1 行で入力してください）。 + +``` +python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 + --batch_size 4 --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive + --use_rating_tags_as_last_tag --character_tags_first --character_tag_expand + --always_first_tags "1girl,1boy" ..\train_data +``` + +## 使用可能なリポジトリID + +[SmilingWolf 氏の V2、V3 のモデル](https://huggingface.co/SmilingWolf)が使用可能です。`SmilingWolf/wd-vit-tagger-v3` のように指定してください。省略時のデフォルトは `SmilingWolf/wd-v1-4-convnext-tagger-v2` です。 + +# オプション + +## 一般オプション + +- `--onnx` : ONNX を使用して推論します。指定しない場合は TensorFlow を使用します。TensorFlow 使用時は別途 TensorFlow をインストールしてください。 +- `--batch_size` : 一度に処理する画像の数。デフォルトは1です。VRAMの容量に応じて増減してください。 +- `--caption_extension` : キャプションファイルの拡張子。デフォルトは `.txt` です。 +- `--max_data_loader_n_workers` : DataLoader の最大ワーカー数です。このオプションに 1 以上の数値を指定すると、DataLoader を用いて画像読み込みを高速化します。未指定時は DataLoader を用いません。 +- `--thresh` : 出力するタグの信頼度の閾値。デフォルトは0.35です。値を下げるとより多くのタグが付与されますが、精度は下がります。 +- `--general_threshold` : 一般タグの信頼度の閾値。省略時は `--thresh` と同じです。 +- `--character_threshold` : キャラクタータグの信頼度の閾値。省略時は `--thresh` と同じです。 +- `--recursive` : 指定すると、指定したフォルダ内のサブフォルダも再帰的に処理します。 +- `--append_tags` : 既存のタグファイルにタグを追加します。 +- `--frequency_tags` : タグの頻度を出力します。 +- `--debug` : デバッグモード。指定するとデバッグ情報を出力します。 + +## モデルのダウンロード + +- `--model_dir` : モデルファイルの保存先フォルダ。デフォルトは `wd14_tagger_model` です。 +- `--force_download` : 指定するとモデルファイルを再ダウンロードします。 + +## タグ編集関連 + +- `--remove_underscore` : 出力するタグからアンダースコアを削除します。 +- `--undesired_tags` : 出力しないタグを指定します。カンマ区切りで複数指定できます。たとえば `black eyes,black hair` のように指定します。 +- `--use_rating_tags` : タグの最初にレーティングタグを出力します。 +- `--use_rating_tags_as_last_tag` : タグの最後にレーティングタグを追加します。 +- `--character_tags_first` : キャラクタータグを最初に出力します。 +- `--character_tag_expand` : キャラクタータグのシリーズ名を展開します。たとえば `chara_name_(series)` のタグを `chara_name, series` に分割します。 +- `--always_first_tags` : あるタグが画像に出力されたとき、そのタグを最初に出力するタグを指定します。カンマ区切りで複数指定できます。たとえば `1girl,1boy` のように指定します。 +- `--caption_separator` : 出力するファイルでタグをこの文字列で区切ります。デフォルトは `, ` です。 +- `--tag_replacement` : タグの置換を行います。`tag1,tag2;tag3,tag4` のように指定します。`,` および `;` を使う場合は `\` でエスケープしてください。\ + たとえば `aira tsubase,aira tsubase (uniform)` （特定の衣装を学習させたいとき）、`aira tsubase,aira tsubase\, heir of shadows` （シリーズ名がタグに含まれないとき）のように指定します。 + +`tag_replacement` は `character_tag_expand` の後に適用されます。 + +`remove_underscore` 指定時は、`undesired_tags`、`always_first_tags`、`tag_replacement` はアンダースコアを含めずに指定してください。 + +`caption_separator` 指定時は、`undesired_tags`、`always_first_tags` は `caption_separator` で区切ってください。`tag_replacement` は必ず `,` で区切ってください。 +