diff --git a/.gitattributes b/.gitattributes deleted file mode 100644 index c7d9f3332a950355d5a77d85000f05e6f45435ea..0000000000000000000000000000000000000000 --- a/.gitattributes +++ /dev/null @@ -1,34 +0,0 @@ -*.7z filter=lfs diff=lfs merge=lfs -text -*.arrow filter=lfs diff=lfs merge=lfs -text -*.bin filter=lfs diff=lfs merge=lfs -text -*.bz2 filter=lfs diff=lfs merge=lfs -text -*.ckpt filter=lfs diff=lfs merge=lfs -text -*.ftz filter=lfs diff=lfs merge=lfs -text -*.gz filter=lfs diff=lfs merge=lfs -text -*.h5 filter=lfs diff=lfs merge=lfs -text -*.joblib filter=lfs diff=lfs merge=lfs -text -*.lfs.* filter=lfs diff=lfs merge=lfs -text -*.mlmodel filter=lfs diff=lfs merge=lfs -text -*.model filter=lfs diff=lfs merge=lfs -text -*.msgpack filter=lfs diff=lfs merge=lfs -text -*.npy filter=lfs diff=lfs merge=lfs -text -*.npz filter=lfs diff=lfs merge=lfs -text -*.onnx filter=lfs diff=lfs merge=lfs -text -*.ot filter=lfs diff=lfs merge=lfs -text -*.parquet filter=lfs diff=lfs merge=lfs -text -*.pb filter=lfs diff=lfs merge=lfs -text -*.pickle filter=lfs diff=lfs merge=lfs -text -*.pkl filter=lfs diff=lfs merge=lfs -text -*.pt filter=lfs diff=lfs merge=lfs -text -*.pth filter=lfs diff=lfs merge=lfs -text -*.rar filter=lfs diff=lfs merge=lfs -text -*.safetensors filter=lfs diff=lfs merge=lfs -text -saved_model/**/* filter=lfs diff=lfs merge=lfs -text -*.tar.* filter=lfs diff=lfs merge=lfs -text -*.tflite filter=lfs diff=lfs merge=lfs -text -*.tgz filter=lfs diff=lfs merge=lfs -text -*.wasm filter=lfs diff=lfs merge=lfs -text -*.xz 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 diff --git a/README.md b/README.md deleted file mode 100644 index 9d7ae08c451f6cf52192a00f2d713ba6d9d656fe..0000000000000000000000000000000000000000 --- a/README.md +++ /dev/null @@ -1,15 +0,0 @@ ---- -title: EditAnything -emoji: 🦀 -colorFrom: green -colorTo: pink -sdk: gradio -sdk_version: 3.24.1 -app_file: app.py -pinned: false ---- - -Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference - -# Edit Anything by Segment-Anything - diff --git a/annotator/util.py b/annotator/util.py index 5d686dd01f3a38b752319f2097469f4bc2a6e879..553608714bda8fb6e76646e452176783179bb6cf 100644 --- a/annotator/util.py +++ b/annotator/util.py @@ -53,3 +53,21 @@ def resize_points(clicked_points, original_shape, resolution): resized_points.append(resized_point) return resized_points + +def get_bounding_box(mask): + # Convert PIL Image to numpy array + mask = np.array(mask).astype(np.uint8) + + # Take the first channel (R) of the mask + mask = mask[:,:,0] + + # Get the indices of elements that are not zero + rows = np.any(mask, axis=0) + cols = np.any(mask, axis=1) + + # Get the minimum and maximum indices where the elements are not zero + rmin, rmax = np.where(rows)[0][[0, -1]] + cmin, cmax = np.where(cols)[0][[0, -1]] + + # Return as [xmin, ymin, xmax, ymax] + return [rmin, cmin, rmax, cmax] diff --git a/app.py b/app.py index 84f94242f59eb7d242d4fccc038d4261f0d8ae70..5e84e2f0211152263829cd9fc808b0d5035d0986 100644 --- a/app.py +++ b/app.py @@ -1,15 +1,11 @@ -import subprocess -result = subprocess.run( - ['pip', 'install', 'gradio==3.28.3'], check=True) import gradio as gr - import os -from sam2edit import create_demo as create_demo_edit_anything +from editany import create_demo as create_demo_edit_anything from sam2image import create_demo as create_demo_generate_anything -from sam2edit_beauty import create_demo as create_demo_beauty -from sam2edit_handsome import create_demo as create_demo_handsome -from sam2edit_lora import EditAnythingLoraModel, init_sam_model, init_blip_processor, init_blip_model +from editany_beauty import create_demo as create_demo_beauty +from editany_handsome import create_demo as create_demo_handsome +from editany_lora import EditAnythingLoraModel, init_sam_model, init_blip_processor, init_blip_model from huggingface_hub import hf_hub_download, snapshot_download DESCRIPTION = f'''# [Edit Anything](https://github.com/sail-sg/EditAnything) @@ -29,9 +25,9 @@ with gr.Blocks() as demo: gr.Markdown(DESCRIPTION) with gr.Tabs(): with gr.TabItem('🖌Edit Anything'): - model = EditAnythingLoraModel(base_model_path="stabilityai/stable-diffusion-2-inpainting", - controlmodel_name='LAION Pretrained(v0-4)-SD21', - lora_model_path=None, use_blip=True, extra_inpaint=False, + model = EditAnythingLoraModel(base_model_path="runwayml/stable-diffusion-v1-5", + controlmodel_name='LAION Pretrained(v0-4)-SD15', + lora_model_path=None, use_blip=True, extra_inpaint=True, sam_generator=sam_generator, mask_predictor=mask_predictor, blip_processor=blip_processor, @@ -57,6 +53,16 @@ with gr.Blocks() as demo: # blip_processor=blip_processor, # blip_model=blip_model) # create_demo_handsome(model.process, model.process_image_click) + # with gr.TabItem('Edit More'): + # model = EditAnythingLoraModel(base_model_path="andite/anything-v4.0", + # lora_model_path=None, use_blip=True, extra_inpaint=True, + # sam_generator=sam_generator, + # mask_predictor=mask_predictor, + # blip_processor=blip_processor, + # blip_model=blip_model, + # lora_weight=0.5, + # ) + create_demo_beauty(model.process, model.process_image_click) # with gr.TabItem('Generate Anything'): # create_demo_generate_anything() with gr.Tabs(): diff --git a/cldm/cldm.py b/cldm/cldm.py new file mode 100644 index 0000000000000000000000000000000000000000..0b3ac7a575cf4933fc14dfc15dd3cca41cb3f3e8 --- /dev/null +++ b/cldm/cldm.py @@ -0,0 +1,435 @@ +import einops +import torch +import torch as th +import torch.nn as nn + +from ldm.modules.diffusionmodules.util import ( + conv_nd, + linear, + zero_module, + timestep_embedding, +) + +from einops import rearrange, repeat +from torchvision.utils import make_grid +from ldm.modules.attention import SpatialTransformer +from ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock +from ldm.models.diffusion.ddpm import LatentDiffusion +from ldm.util import log_txt_as_img, exists, instantiate_from_config +from ldm.models.diffusion.ddim import DDIMSampler + + +class ControlledUnetModel(UNetModel): + def forward(self, x, timesteps=None, context=None, control=None, only_mid_control=False, **kwargs): + hs = [] + with torch.no_grad(): + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + h = x.type(self.dtype) + for module in self.input_blocks: + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + + if control is not None: + h += control.pop() + + for i, module in enumerate(self.output_blocks): + if only_mid_control or control is None: + h = torch.cat([h, hs.pop()], dim=1) + else: + h = torch.cat([h, hs.pop() + control.pop()], dim=1) + h = module(h, emb, context) + + h = h.type(x.dtype) + return self.out(h) + + +class ControlNet(nn.Module): + def __init__( + self, + image_size, + in_channels, + model_channels, + hint_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + use_checkpoint=False, + use_fp16=False, + num_heads=-1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + resblock_updown=False, + use_new_attention_order=False, + use_spatial_transformer=False, # custom transformer support + transformer_depth=1, # custom transformer support + context_dim=None, # custom transformer support + n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model + legacy=True, + disable_self_attentions=None, + num_attention_blocks=None, + disable_middle_self_attn=False, + use_linear_in_transformer=False, + ): + super().__init__() + if use_spatial_transformer: + assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' + + if context_dim is not None: + assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' + from omegaconf.listconfig import ListConfig + if type(context_dim) == ListConfig: + context_dim = list(context_dim) + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.dims = dims + self.image_size = image_size + self.in_channels = in_channels + self.model_channels = model_channels + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError("provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult") + self.num_res_blocks = num_res_blocks + if disable_self_attentions is not None: + # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) + print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set.") + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.use_checkpoint = use_checkpoint + self.dtype = th.float16 if use_fp16 else th.float32 + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.predict_codebook_ids = n_embed is not None + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)]) + + self.input_hint_block = TimestepEmbedSequential( + conv_nd(dims, hint_channels, 16, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 16, 16, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 16, 32, 3, padding=1, stride=2), + nn.SiLU(), + conv_nd(dims, 32, 32, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 32, 96, 3, padding=1, stride=2), + nn.SiLU(), + conv_nd(dims, 96, 96, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 96, 256, 3, padding=1, stride=2), + nn.SiLU(), + zero_module(conv_nd(dims, 256, model_channels, 3, padding=1)) + ) + + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + # num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self.zero_convs.append(self.make_zero_conv(ch)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + self.zero_convs.append(self.make_zero_conv(ch)) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + # num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self.middle_block_out = self.make_zero_conv(ch) + self._feature_size += ch + + def make_zero_conv(self, channels): + return TimestepEmbedSequential(zero_module(conv_nd(self.dims, channels, channels, 1, padding=0))) + + def forward(self, x, hint, timesteps, context, **kwargs): + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + guided_hint = self.input_hint_block(hint, emb, context) + + outs = [] + + h = x.type(self.dtype) + for module, zero_conv in zip(self.input_blocks, self.zero_convs): + if guided_hint is not None: + h = module(h, emb, context) + h += guided_hint + guided_hint = None + else: + h = module(h, emb, context) + outs.append(zero_conv(h, emb, context)) + + h = self.middle_block(h, emb, context) + outs.append(self.middle_block_out(h, emb, context)) + + return outs + + +class ControlLDM(LatentDiffusion): + + def __init__(self, control_stage_config, control_key, only_mid_control, *args, **kwargs): + super().__init__(*args, **kwargs) + self.control_model = instantiate_from_config(control_stage_config) + self.control_key = control_key + self.only_mid_control = only_mid_control + self.control_scales = [1.0] * 13 + + @torch.no_grad() + def get_input(self, batch, k, bs=None, *args, **kwargs): + x, c = super().get_input(batch, self.first_stage_key, *args, **kwargs) + control = batch[self.control_key] + if bs is not None: + control = control[:bs] + control = control.to(self.device) + control = einops.rearrange(control, 'b h w c -> b c h w') + control = control.to(memory_format=torch.contiguous_format).float() + return x, dict(c_crossattn=[c], c_concat=[control]) + + def apply_model(self, x_noisy, t, cond, *args, **kwargs): + assert isinstance(cond, dict) + diffusion_model = self.model.diffusion_model + + cond_txt = torch.cat(cond['c_crossattn'], 1) + + if cond['c_concat'] is None: + eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=None, only_mid_control=self.only_mid_control) + else: + control = self.control_model(x=x_noisy, hint=torch.cat(cond['c_concat'], 1), timesteps=t, context=cond_txt) + control = [c * scale for c, scale in zip(control, self.control_scales)] + eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=control, only_mid_control=self.only_mid_control) + + return eps + + @torch.no_grad() + def get_unconditional_conditioning(self, N): + return self.get_learned_conditioning([""] * N) + + @torch.no_grad() + def log_images(self, batch, N=4, n_row=2, sample=False, ddim_steps=50, ddim_eta=0.0, return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=False, unconditional_guidance_scale=9.0, unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + use_ddim = ddim_steps is not None + + log = dict() + z, c = self.get_input(batch, self.first_stage_key, bs=N) + c_cat, c = c["c_concat"][0][:N], c["c_crossattn"][0][:N] + N = min(z.shape[0], N) + n_row = min(z.shape[0], n_row) + log["reconstruction"] = self.decode_first_stage(z) + log["control"] = c_cat * 2.0 - 1.0 + log["conditioning"] = log_txt_as_img((512, 512), batch[self.cond_stage_key], size=16) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_cross = self.get_unconditional_conditioning(N) + uc_cat = c_cat # torch.zeros_like(c_cat) + uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} + samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc_full, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + return log + + @torch.no_grad() + def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): + ddim_sampler = DDIMSampler(self) + b, c, h, w = cond["c_concat"][0].shape + shape = (self.channels, h // 8, w // 8) + samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, shape, cond, verbose=False, **kwargs) + return samples, intermediates + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.control_model.parameters()) + if not self.sd_locked: + params += list(self.model.diffusion_model.output_blocks.parameters()) + params += list(self.model.diffusion_model.out.parameters()) + opt = torch.optim.AdamW(params, lr=lr) + return opt + + def low_vram_shift(self, is_diffusing): + if is_diffusing: + self.model = self.model.cuda() + self.control_model = self.control_model.cuda() + self.first_stage_model = self.first_stage_model.cpu() + self.cond_stage_model = self.cond_stage_model.cpu() + else: + self.model = self.model.cpu() + self.control_model = self.control_model.cpu() + self.first_stage_model = self.first_stage_model.cuda() + self.cond_stage_model = self.cond_stage_model.cuda() diff --git a/cldm/ddim_hacked.py b/cldm/ddim_hacked.py new file mode 100644 index 0000000000000000000000000000000000000000..6c040b363ba0705f52509b75437b5ea932c80ec1 --- /dev/null +++ b/cldm/ddim_hacked.py @@ -0,0 +1,316 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + ucg_schedule=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): ctmp = ctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + elif isinstance(conditioning, list): + for ctmp in conditioning: + if ctmp.shape[0] != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ucg_schedule=ucg_schedule + ) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None, + ucg_schedule=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + if ucg_schedule is not None: + assert len(ucg_schedule) == len(time_range) + unconditional_guidance_scale = ucg_schedule[i] + + outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold) + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + model_output = self.model.apply_model(x, t, c) + else: + model_t = self.model.apply_model(x, t, c) + model_uncond = self.model.apply_model(x, t, unconditional_conditioning) + model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond) + + if self.model.parameterization == "v": + e_t = self.model.predict_eps_from_z_and_v(x, t, model_output) + else: + e_t = model_output + + if score_corrector is not None: + assert self.model.parameterization == "eps", 'not implemented' + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if self.model.parameterization != "v": + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output) + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + + if dynamic_threshold is not None: + raise NotImplementedError() + + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + @torch.no_grad() + def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None, + unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None): + num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0] + + assert t_enc <= num_reference_steps + num_steps = t_enc + + if use_original_steps: + alphas_next = self.alphas_cumprod[:num_steps] + alphas = self.alphas_cumprod_prev[:num_steps] + else: + alphas_next = self.ddim_alphas[:num_steps] + alphas = torch.tensor(self.ddim_alphas_prev[:num_steps]) + + x_next = x0 + intermediates = [] + inter_steps = [] + for i in tqdm(range(num_steps), desc='Encoding Image'): + t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long) + if unconditional_guidance_scale == 1.: + noise_pred = self.model.apply_model(x_next, t, c) + else: + assert unconditional_conditioning is not None + e_t_uncond, noise_pred = torch.chunk( + self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)), + torch.cat((unconditional_conditioning, c))), 2) + noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond) + + xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next + weighted_noise_pred = alphas_next[i].sqrt() * ( + (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred + x_next = xt_weighted + weighted_noise_pred + if return_intermediates and i % ( + num_steps // return_intermediates) == 0 and i < num_steps - 1: + intermediates.append(x_next) + inter_steps.append(i) + elif return_intermediates and i >= num_steps - 2: + intermediates.append(x_next) + inter_steps.append(i) + if callback: callback(i) + + out = {'x_encoded': x_next, 'intermediate_steps': inter_steps} + if return_intermediates: + out.update({'intermediates': intermediates}) + return x_next, out + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False, callback=None): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + if callback: callback(i) + return x_dec \ No newline at end of file diff --git a/cldm/hack.py b/cldm/hack.py new file mode 100644 index 0000000000000000000000000000000000000000..454361e9d036cd1a6a79122c2fd16b489e4767b1 --- /dev/null +++ b/cldm/hack.py @@ -0,0 +1,111 @@ +import torch +import einops + +import ldm.modules.encoders.modules +import ldm.modules.attention + +from transformers import logging +from ldm.modules.attention import default + + +def disable_verbosity(): + logging.set_verbosity_error() + print('logging improved.') + return + + +def enable_sliced_attention(): + ldm.modules.attention.CrossAttention.forward = _hacked_sliced_attentin_forward + print('Enabled sliced_attention.') + return + + +def hack_everything(clip_skip=0): + disable_verbosity() + ldm.modules.encoders.modules.FrozenCLIPEmbedder.forward = _hacked_clip_forward + ldm.modules.encoders.modules.FrozenCLIPEmbedder.clip_skip = clip_skip + print('Enabled clip hacks.') + return + + +# Written by Lvmin +def _hacked_clip_forward(self, text): + PAD = self.tokenizer.pad_token_id + EOS = self.tokenizer.eos_token_id + BOS = self.tokenizer.bos_token_id + + def tokenize(t): + return self.tokenizer(t, truncation=False, add_special_tokens=False)["input_ids"] + + def transformer_encode(t): + if self.clip_skip > 1: + rt = self.transformer(input_ids=t, output_hidden_states=True) + return self.transformer.text_model.final_layer_norm(rt.hidden_states[-self.clip_skip]) + else: + return self.transformer(input_ids=t, output_hidden_states=False).last_hidden_state + + def split(x): + return x[75 * 0: 75 * 1], x[75 * 1: 75 * 2], x[75 * 2: 75 * 3] + + def pad(x, p, i): + return x[:i] if len(x) >= i else x + [p] * (i - len(x)) + + raw_tokens_list = tokenize(text) + tokens_list = [] + + for raw_tokens in raw_tokens_list: + raw_tokens_123 = split(raw_tokens) + raw_tokens_123 = [[BOS] + raw_tokens_i + [EOS] for raw_tokens_i in raw_tokens_123] + raw_tokens_123 = [pad(raw_tokens_i, PAD, 77) for raw_tokens_i in raw_tokens_123] + tokens_list.append(raw_tokens_123) + + tokens_list = torch.IntTensor(tokens_list).to(self.device) + + feed = einops.rearrange(tokens_list, 'b f i -> (b f) i') + y = transformer_encode(feed) + z = einops.rearrange(y, '(b f) i c -> b (f i) c', f=3) + + return z + + +# Stolen from https://github.com/basujindal/stable-diffusion/blob/main/optimizedSD/splitAttention.py +def _hacked_sliced_attentin_forward(self, x, context=None, mask=None): + h = self.heads + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + del context, x + + q, k, v = map(lambda t: einops.rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + limit = k.shape[0] + att_step = 1 + q_chunks = list(torch.tensor_split(q, limit // att_step, dim=0)) + k_chunks = list(torch.tensor_split(k, limit // att_step, dim=0)) + v_chunks = list(torch.tensor_split(v, limit // att_step, dim=0)) + + q_chunks.reverse() + k_chunks.reverse() + v_chunks.reverse() + sim = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device) + del k, q, v + for i in range(0, limit, att_step): + q_buffer = q_chunks.pop() + k_buffer = k_chunks.pop() + v_buffer = v_chunks.pop() + sim_buffer = torch.einsum('b i d, b j d -> b i j', q_buffer, k_buffer) * self.scale + + del k_buffer, q_buffer + # attention, what we cannot get enough of, by chunks + + sim_buffer = sim_buffer.softmax(dim=-1) + + sim_buffer = torch.einsum('b i j, b j d -> b i d', sim_buffer, v_buffer) + del v_buffer + sim[i:i + att_step, :, :] = sim_buffer + + del sim_buffer + sim = einops.rearrange(sim, '(b h) n d -> b n (h d)', h=h) + return self.to_out(sim) diff --git a/cldm/logger.py b/cldm/logger.py new file mode 100644 index 0000000000000000000000000000000000000000..6a8803846f2a8979f87f3cf9ea5b12869439e62f --- /dev/null +++ b/cldm/logger.py @@ -0,0 +1,76 @@ +import os + +import numpy as np +import torch +import torchvision +from PIL import Image +from pytorch_lightning.callbacks import Callback +from pytorch_lightning.utilities.distributed import rank_zero_only + + +class ImageLogger(Callback): + def __init__(self, batch_frequency=2000, max_images=4, clamp=True, increase_log_steps=True, + rescale=True, disabled=False, log_on_batch_idx=False, log_first_step=False, + log_images_kwargs=None): + super().__init__() + self.rescale = rescale + self.batch_freq = batch_frequency + self.max_images = max_images + if not increase_log_steps: + self.log_steps = [self.batch_freq] + self.clamp = clamp + self.disabled = disabled + self.log_on_batch_idx = log_on_batch_idx + self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {} + self.log_first_step = log_first_step + + @rank_zero_only + def log_local(self, save_dir, split, images, global_step, current_epoch, batch_idx): + root = os.path.join(save_dir, "image_log", split) + for k in images: + grid = torchvision.utils.make_grid(images[k], nrow=4) + if self.rescale: + grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w + grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1) + grid = grid.numpy() + grid = (grid * 255).astype(np.uint8) + filename = "{}_gs-{:06}_e-{:06}_b-{:06}.png".format(k, global_step, current_epoch, batch_idx) + path = os.path.join(root, filename) + os.makedirs(os.path.split(path)[0], exist_ok=True) + Image.fromarray(grid).save(path) + + def log_img(self, pl_module, batch, batch_idx, split="train"): + check_idx = batch_idx # if self.log_on_batch_idx else pl_module.global_step + if (self.check_frequency(check_idx) and # batch_idx % self.batch_freq == 0 + hasattr(pl_module, "log_images") and + callable(pl_module.log_images) and + self.max_images > 0): + logger = type(pl_module.logger) + + is_train = pl_module.training + if is_train: + pl_module.eval() + + with torch.no_grad(): + images = pl_module.log_images(batch, split=split, **self.log_images_kwargs) + + for k in images: + N = min(images[k].shape[0], self.max_images) + images[k] = images[k][:N] + if isinstance(images[k], torch.Tensor): + images[k] = images[k].detach().cpu() + if self.clamp: + images[k] = torch.clamp(images[k], -1., 1.) + + self.log_local(pl_module.logger.save_dir, split, images, + pl_module.global_step, pl_module.current_epoch, batch_idx) + + if is_train: + pl_module.train() + + def check_frequency(self, check_idx): + return check_idx % self.batch_freq == 0 + + def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx): + if not self.disabled: + self.log_img(pl_module, batch, batch_idx, split="train") diff --git a/cldm/model.py b/cldm/model.py new file mode 100644 index 0000000000000000000000000000000000000000..fed3c31ac145b78907c7f771d1d8db6fb32d92ed --- /dev/null +++ b/cldm/model.py @@ -0,0 +1,28 @@ +import os +import torch + +from omegaconf import OmegaConf +from ldm.util import instantiate_from_config + + +def get_state_dict(d): + return d.get('state_dict', d) + + +def load_state_dict(ckpt_path, location='cpu'): + _, extension = os.path.splitext(ckpt_path) + if extension.lower() == ".safetensors": + import safetensors.torch + state_dict = safetensors.torch.load_file(ckpt_path, device=location) + else: + state_dict = get_state_dict(torch.load(ckpt_path, map_location=torch.device(location))) + state_dict = get_state_dict(state_dict) + print(f'Loaded state_dict from [{ckpt_path}]') + return state_dict + + +def create_model(config_path): + config = OmegaConf.load(config_path) + model = instantiate_from_config(config.model).cpu() + print(f'Loaded model config from [{config_path}]') + return model diff --git a/config.py b/config.py new file mode 100644 index 0000000000000000000000000000000000000000..e0c738d8cbad66bbe1666284aef926c326849701 --- /dev/null +++ b/config.py @@ -0,0 +1 @@ +save_memory = False diff --git a/dataset_build.py b/dataset_build.py new file mode 100644 index 0000000000000000000000000000000000000000..ec0a705c1d3e5fb4cd48eb70b43e1163561d747b --- /dev/null +++ b/dataset_build.py @@ -0,0 +1,45 @@ +from PIL import Image +import json + +from transformers import AutoProcessor, Blip2ForConditionalGeneration +import torch +import os + +processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b") +model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16) + +device = "cuda" if torch.cuda.is_available() else "cpu" +model.to(device) + +def get_blip2_text(image): + inputs = processor(image, return_tensors="pt").to(device, torch.float16) + generated_ids = model.generate(**inputs, max_new_tokens=50) + generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip() + return generated_text + + +data_path = "files" +save_path = "" + +image_names = os.listdir(data_path) +image_names = sorted(image_names) + +text_data = {} +f = open("data.txt","w") +for each in image_names: + if '.jpg' in each: + this_data = {} + this_data['target'] = each + this_data['source'] = each[:-4]+'.json' + this_image = Image.open(os.path.join(data_path, each)) + print(each) + generated_text = get_blip2_text(this_image) + this_data['prompt'] = generated_text + print(this_data) + f.write(str(this_data)+"\n") +f.close() + + + + + diff --git a/editany.py b/editany.py new file mode 100644 index 0000000000000000000000000000000000000000..55aa94cb9d0f4d4f96850af0024fcbbf5f18224c --- /dev/null +++ b/editany.py @@ -0,0 +1,36 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +import os +import gradio as gr +from diffusers.utils import load_image +from editany_lora import EditAnythingLoraModel, config_dict +from editany_demo import create_demo_template +from huggingface_hub import hf_hub_download, snapshot_download + + +def create_demo(process, process_image_click=None): + + examples = None + INFO = f""" + ## EditAnything https://github.com/sail-sg/EditAnything + """ + WARNING_INFO = None + + demo = create_demo_template( + process, + process_image_click, + examples=examples, + INFO=INFO, + WARNING_INFO=WARNING_INFO, + enable_auto_prompt_default=True, + ) + return demo + + +if __name__ == "__main__": + model = EditAnythingLoraModel( + base_model_path="runwayml/stable-diffusion-v1-5", + controlmodel_name='LAION Pretrained(v0-4)-SD15', + lora_model_path=None, use_blip=True, extra_inpaint=True, + ) + demo = create_demo(model.process, model.process_image_click) + demo.queue().launch(server_name="0.0.0.0") diff --git a/editany_beauty.py b/editany_beauty.py new file mode 100644 index 0000000000000000000000000000000000000000..d860898ee2366535872c55bb6f86c146d57fa37c --- /dev/null +++ b/editany_beauty.py @@ -0,0 +1,66 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +import os +import gradio as gr +from diffusers.utils import load_image +from editany_lora import EditAnythingLoraModel, config_dict +from editany_demo import create_demo_template +from huggingface_hub import hf_hub_download, snapshot_download + + +def create_demo(process, process_image_click=None): + + examples = [ + [ + "dudou,1girl, beautiful face, solo, candle, brown hair, long hair, ,ulzzang-6500-v1.1,(raw photo:1.2),((photorealistic:1.4))best quality ,masterpiece, illustration, an extremely delicate and beautiful, extremely detailed ,CG ,unity ,8k wallpaper, Amazing, finely detail, masterpiece,best quality,official art,extremely detailed CG unity 8k wallpaper,absurdres, incredibly absurdres, huge filesize, ultra-detailed, highres, extremely detailed,beautiful detailed girl, extremely detailed eyes and face, beautiful detailed eyes,cinematic lighting,1girl,see-through,looking at viewer,full body,full-body shot,outdoors,arms behind back,(chinese clothes) ", + "(((mole))),sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, bad anatomy,(long hair:1.4),DeepNegative,(fat:1.2),facing away, looking away,tilted head, lowres,bad anatomy,bad hands, text, error, missing fingers,extra digit, fewer digits, cropped, worstquality, low quality, normal quality,jpegartifacts,signature, watermark, username,blurry,bad feet,cropped,poorly drawn hands,poorly drawn face,mutation,deformed,worst quality,low quality,normal quality,jpeg artifacts,signature,watermark,extra fingers,fewer digits,extra limbs,extra arms,extra legs,malformed limbs,fused fingers,too many fingers,long neck,cross-eyed,mutated hands,polar lowres,bad body,bad proportions,gross proportions,text,error,missing fingers,missing arms,missing legs,extra digit, extra arms, extra leg, extra foot,(freckles),(mole:2)", + 5, + ], + [ + "best quality, ultra high res, (photorealistic:1.4), (detailed beautiful girl:1.4), (medium breasts:0.8), looking_at_viewer, Detailed facial details, beautiful detailed eyes, (multicolored|blue|pink hair: 1.2), green eyes, slender, haunting smile, (makeup:0.3), red lips, , highly detailed clothes, (ulzzang-6500-v1.1:0.3)", + "EasyNegative, paintings, sketches, ugly, 3d, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, manboobs, backlight,(ugly:1.3), (duplicate:1.3), (morbid:1.2), (mutilated:1.2), (tranny:1.3), mutated hands, (poorly drawn hands:1.3), blurry, (bad anatomy:1.2), (bad proportions:1.3), extra limbs, (disfigured:1.3), (more than 2 nipples:1.3), (more than 1 navel:1.3), (missing arms:1.3), (extra legs:1.3), (fused fingers:1.6), (too many fingers:1.6), (unclear eyes:1.3), bad hands, missing fingers, extra digit, (futa:1.1), bad body, double navel, mutad arms, hused arms, (puffy nipples, dark areolae, dark nipples, rei no himo, inverted nipples, long nipples), NG_DeepNegative_V1_75t, pubic hair, fat rolls, obese, bad-picture-chill-75v", + 8, + ], + [ + "best quality, ultra high res, (photorealistic:1.4), (detailed beautiful girl:1.4), (medium breasts:0.8), looking_at_viewer, Detailed facial details, beautiful detailed eyes, (blue|pink hair), green eyes, slender, smile, (makeup:0.4), red lips, (full body, sitting, beach), , highly detailed clothes, (ulzzang-6500-v1.1:0.3)", + "asyNegative, paintings, sketches, ugly, 3d, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, manboobs, backlight,(ugly:1.3), (duplicate:1.3), (morbid:1.2), (mutilated:1.2), (tranny:1.3), mutated hands, (poorly drawn hands:1.3), blurry, (bad anatomy:1.2), (bad proportions:1.3), extra limbs, (disfigured:1.3), (more than 2 nipples:1.3), (more than 1 navel:1.3), (missing arms:1.3), (extra legs:1.3), (fused fingers:1.6), (too many fingers:1.6), (unclear eyes:1.3), bad hands, missing fingers, extra digit, (futa:1.1), bad body, double navel, mutad arms, hused arms, (puffy nipples, dark areolae, dark nipples, rei no himo, inverted nipples, long nipples), NG_DeepNegative_V1_75t, pubic hair, fat rolls, obese, bad-picture-chill-75v", + 7, + ], + [ + "mix4, whole body shot, ((8k, RAW photo, highest quality, masterpiece), High detail RAW color photo professional close-up photo, shy expression, cute, beautiful detailed girl, detailed fingers, extremely detailed eyes and face, beautiful detailed nose, beautiful detailed eyes, long eyelashes, light on face, looking at viewer, (closed mouth:1.2), 1girl, cute, young, mature face, (full body:1.3), ((small breasts)), realistic face, realistic body, beautiful detailed thigh,s, same eyes color, (realistic, photo realism:1. 37), (highest quality), (best shadow), (best illustration), ultra high resolution, physics-based rendering, cinematic lighting), solo, 1girl, highly detailed, in office, detailed office, open cardigan, ponytail contorted, beautiful eyes ,sitting in office,dating, business suit, cross-laced clothes, collared shirt, beautiful breast, small breast, Chinese dress, white pantyhose, natural breasts, pink and white hair, ", + "paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), cloth, underwear, bra, low-res, normal quality, ((monochrome)), ((grayscale)), skin spots, acne, skin blemishes, age spots, glans, bad nipples, long nipples, bad vagina, extra fingers,fewer fingers,strange fingers,bad hand, ng_deepnegative_v1_75t, bad-picture-chill-75v", + 7, + ], + ] + INFO = f""" + ## Generate Your Beauty powered by EditAnything https://github.com/sail-sg/EditAnything + This model is good at generating beautiful female. + """ + WARNING_INFO = f"""### [NOTE] the model is collected from the Internet for demo only, please do not use it for commercial purposes. + We are not responsible for possible risks using this model. + Lora model from https://civitai.com/models/14171/cutegirlmix4 Thanks! + """ + demo = create_demo_template( + process, + process_image_click, + examples=examples, + INFO=INFO, + WARNING_INFO=WARNING_INFO, + ) + return demo + + +if __name__ == "__main__": + sd_models_path = snapshot_download("shgao/sdmodels") + lora_model_path = hf_hub_download( + "mlida/Cute_girl_mix4", "cuteGirlMix4_v10.safetensors" + ) + model = EditAnythingLoraModel( + base_model_path=os.path.join( + sd_models_path, "chilloutmix_NiPrunedFp32Fix"), + lora_model_path=lora_model_path, + use_blip=True, + extra_inpaint=True, + lora_weight=0.5, + ) + demo = create_demo(model.process, model.process_image_click) + demo.queue().launch(server_name="0.0.0.0") diff --git a/editany_demo.py b/editany_demo.py new file mode 100644 index 0000000000000000000000000000000000000000..94ba7f537bee4b81652e1f03c91da5fc1fe6c5c9 --- /dev/null +++ b/editany_demo.py @@ -0,0 +1,394 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +import gradio as gr + + +def create_demo_template( + process, + process_image_click=None, + examples=None, + INFO="EditAnything https://github.com/sail-sg/EditAnything", + WARNING_INFO=None, + enable_auto_prompt_default=False, +): + + print("The GUI is not fully tested yet. Please open an issue if you find bugs.") + block = gr.Blocks() + with block as demo: + clicked_points = gr.State([]) + origin_image = gr.State(None) + click_mask = gr.State(None) + ref_clicked_points = gr.State([]) + ref_origin_image = gr.State(None) + ref_click_mask = gr.State(None) + with gr.Row(): + gr.Markdown(INFO) + with gr.Row().style(equal_height=False): + with gr.Column(): + with gr.Tab("Click🖱"): + source_image_click = gr.Image( + type="pil", + interactive=True, + label="Image: Upload an image and click the region you want to edit.", + ) + with gr.Column(): + with gr.Row(): + point_prompt = gr.Radio( + choices=["Foreground Point", + "Background Point"], + value="Foreground Point", + label="Point Label", + interactive=True, + show_label=False, + ) + clear_button_click = gr.Button( + value="Clear Click Points", interactive=True + ) + clear_button_image = gr.Button( + value="Clear Image", interactive=True + ) + with gr.Row(): + run_button_click = gr.Button( + label="Run EditAnying", interactive=True + ) + with gr.Tab("Brush🖌️"): + source_image_brush = gr.Image( + source="upload", + label="Image: Upload an image and cover the region you want to edit with sketch", + type="numpy", + tool="sketch", + ) + run_button = gr.Button( + label="Run EditAnying", interactive=True) + with gr.Column(): + enable_all_generate = gr.Checkbox( + label="Auto generation on all region.", value=False + ) + control_scale = gr.Slider( + label="Mask Align strength", + info="Large value -> strict alignment with SAM mask", + minimum=0, + maximum=1, + value=0.5, + step=0.1, + ) + with gr.Column(): + enable_auto_prompt = gr.Checkbox( + label="Auto generate text prompt from input image with BLIP2", + info="Warning: Enable this may makes your prompt not working.", + value=enable_auto_prompt_default, + ) + a_prompt = gr.Textbox( + label="Positive Prompt", + info="Text in the expected things of edited region", + value="best quality, extremely detailed,", + ) + n_prompt = gr.Textbox( + label="Negative Prompt", + value="longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, NSFW", + ) + with gr.Row(): + num_samples = gr.Slider( + label="Images", minimum=1, maximum=12, value=2, step=1 + ) + seed = gr.Slider( + label="Seed", + minimum=-1, + maximum=2147483647, + step=1, + randomize=True, + ) + with gr.Row(): + enable_tile = gr.Checkbox( + label="Tile refinement for high resolution generation", + info="Slow inference", + value=True, + ) + refine_alignment_ratio = gr.Slider( + label="Alignment Strength", + info="Large value -> strict alignment with input image. Small value -> strong global consistency", + minimum=0.0, + maximum=1.0, + value=0.95, + step=0.05, + ) + + with gr.Accordion("Reference options", open=False): + # ref_image = gr.Image( + # source='upload', label="Upload a reference image", type="pil", value=None) + ref_image = gr.Image( + source="upload", + label="Upload a reference image and cover the region you want to use with sketch", + type="pil", + tool="sketch", + ) + with gr.Column(): + ref_auto_prompt = gr.Checkbox( + label="Ref. Auto Prompt", value=True + ) + ref_prompt = gr.Textbox( + label="Prompt", + info="Text in the prompt of edited region", + value="best quality, extremely detailed, ", + ) + # ref_image = gr.Image( + # type="pil", interactive=True, + # label="Image: Upload an image and click the region you want to use as reference.", + # ) + # with gr.Column(): + # with gr.Row(): + # ref_point_prompt = gr.Radio( + # choices=["Foreground Point", "Background Point"], + # value="Foreground Point", + # label="Point Label", + # interactive=True, show_label=False) + # ref_clear_button_click = gr.Button( + # value="Clear Click Points", interactive=True) + # ref_clear_button_image = gr.Button( + # value="Clear Image", interactive=True) + with gr.Row(): + reference_attn = gr.Checkbox( + label="reference_attn", value=True) + attention_auto_machine_weight = gr.Slider( + label="attention_weight", + minimum=0, + maximum=1.0, + value=0.8, + step=0.01, + ) + with gr.Row(): + reference_adain = gr.Checkbox( + label="reference_adain", value=False + ) + gn_auto_machine_weight = gr.Slider( + label="gn_weight", + minimum=0, + maximum=1.0, + value=0.1, + step=0.01, + ) + style_fidelity = gr.Slider( + label="Style fidelity", + minimum=0, + maximum=1.0, + value=0.5, + step=0.01, + ) + ref_sam_scale = gr.Slider( + label="SAM Control Scale", + minimum=0, + maximum=1.0, + value=0.3, + step=0.1, + ) + ref_inpaint_scale = gr.Slider( + label="Inpaint Control Scale", + minimum=0, + maximum=1.0, + value=0.2, + step=0.1, + ) + with gr.Row(): + ref_textinv = gr.Checkbox( + label="Use textual inversion token", value=False + ) + ref_textinv_path = gr.Textbox( + label="textual inversion token path", + info="Text in the inversion token path", + value=None, + ) + + with gr.Accordion("Advanced options", open=False): + mask_image = gr.Image( + source="upload", + label="Upload a predefined mask of edit region: Switch to Brush mode when using this!", + type="numpy", + value=None, + ) + image_resolution = gr.Slider( + label="Image Resolution", + minimum=256, + maximum=768, + value=512, + step=64, + ) + refine_image_resolution = gr.Slider( + label="Image Resolution", + minimum=256, + maximum=8192, + value=1024, + step=64, + ) + guess_mode = gr.Checkbox(label="Guess Mode", value=False) + detect_resolution = gr.Slider( + label="SAM Resolution", + minimum=128, + maximum=2048, + value=1024, + step=1, + ) + ddim_steps = gr.Slider( + label="Steps", minimum=1, maximum=100, value=30, step=1 + ) + scale = gr.Slider( + label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1) + alpha_weight = gr.Slider( + label="Alpha weight", info="Alpha mixing with original image", minimum=0, + maximum=1, value=0.0, step=0.1) + use_scale_map = gr.Checkbox( + label='Use scale map', value=False) + eta = gr.Number(label="eta (DDIM)", value=0.0) + condition_model = gr.Textbox( + label="Condition model path", + info="Text in the Controlnet model path in hugglingface", + value="EditAnything", + ) + with gr.Column(): + result_gallery_refine = gr.Gallery( + label="Output High quality", show_label=True, elem_id="gallery" + ).style(grid=2, preview=False) + result_gallery_init = gr.Gallery( + label="Output Low quality", show_label=True, elem_id="gallery" + ).style(grid=2, height="auto") + result_gallery_ref = gr.Gallery( + label="Output Ref", show_label=False, elem_id="gallery" + ).style(grid=2, height="auto") + result_text = gr.Text(label="BLIP2+Human Prompt Text") + + ips = [ + source_image_brush, + enable_all_generate, + mask_image, + control_scale, + enable_auto_prompt, + a_prompt, + n_prompt, + num_samples, + image_resolution, + detect_resolution, + ddim_steps, + guess_mode, + scale, + seed, + eta, + enable_tile, + refine_alignment_ratio, + refine_image_resolution, + alpha_weight, + use_scale_map, + condition_model, + ref_image, + attention_auto_machine_weight, + gn_auto_machine_weight, + style_fidelity, + reference_attn, + reference_adain, + ref_prompt, + ref_sam_scale, + ref_inpaint_scale, + ref_auto_prompt, + ref_textinv, + ref_textinv_path, + ] + run_button.click( + fn=process, + inputs=ips, + outputs=[ + result_gallery_refine, + result_gallery_init, + result_gallery_ref, + result_text, + ], + ) + + ip_click = [ + origin_image, + enable_all_generate, + click_mask, + control_scale, + enable_auto_prompt, + a_prompt, + n_prompt, + num_samples, + image_resolution, + detect_resolution, + ddim_steps, + guess_mode, + scale, + seed, + eta, + enable_tile, + refine_alignment_ratio, + refine_image_resolution, + alpha_weight, + use_scale_map, + condition_model, + ref_image, + attention_auto_machine_weight, + gn_auto_machine_weight, + style_fidelity, + reference_attn, + reference_adain, + ref_prompt, + ref_sam_scale, + ref_inpaint_scale, + ref_auto_prompt, + ref_textinv, + ref_textinv_path, + ] + + run_button_click.click( + fn=process, + inputs=ip_click, + outputs=[ + result_gallery_refine, + result_gallery_init, + result_gallery_ref, + result_text, + ], + ) + + source_image_click.upload( + lambda image: image.copy() if image is not None else None, + inputs=[source_image_click], + outputs=[origin_image], + ) + source_image_click.select( + process_image_click, + inputs=[origin_image, point_prompt, + clicked_points, image_resolution], + outputs=[source_image_click, clicked_points, click_mask], + show_progress=True, + queue=True, + ) + clear_button_click.click( + fn=lambda original_image: (original_image.copy(), [], None) + if original_image is not None + else (None, [], None), + inputs=[origin_image], + outputs=[source_image_click, clicked_points, click_mask], + ) + clear_button_image.click( + fn=lambda: (None, [], None, None, None), + inputs=[], + outputs=[ + source_image_click, + clicked_points, + click_mask, + result_gallery_init, + result_text, + ], + ) + + if examples is not None: + with gr.Row(): + ex = gr.Examples( + examples=examples, + fn=process, + inputs=[a_prompt, n_prompt, scale], + outputs=[result_gallery_init], + cache_examples=False, + ) + if WARNING_INFO is not None: + with gr.Row(): + gr.Markdown(WARNING_INFO) + return demo diff --git a/sam2edit_handsome.py b/editany_handsome.py similarity index 57% rename from sam2edit_handsome.py rename to editany_handsome.py index 174e8b4e3dbc7058c57558b5d56c707b6a574d78..25d6d31dd815eb372c0858738390f793bbebfcef 100644 --- a/sam2edit_handsome.py +++ b/editany_handsome.py @@ -2,36 +2,49 @@ import os import gradio as gr from diffusers.utils import load_image -from sam2edit_lora import EditAnythingLoraModel, config_dict -from sam2edit_demo import create_demo_template +from editany_lora import EditAnythingLoraModel, config_dict +from editany_demo import create_demo_template from huggingface_hub import hf_hub_download, snapshot_download def create_demo(process, process_image_click=None): examples = [ - ["1man, muscle,full body, vest, short straight hair, glasses, Gym, barbells, dumbbells, treadmills, boxing rings, squat racks, plates, dumbbell racks soft lighting, masterpiece, best quality, 8k uhd, film grain, Fujifilm XT3 photorealistic painting art by midjourney and greg rutkowski ", - "(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck", 6], - ["1man, 25 years- old, full body, wearing long-sleeve white shirt and tie, muscular rand black suit, soft lighting, masterpiece, best quality, 8k uhd, dslr, film grain, Fujifilm XT3 photorealistic painting art by midjourney and greg rutkowski ", - "(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck", 6], + [ + "1man, muscle,full body, vest, short straight hair, glasses, Gym, barbells, dumbbells, treadmills, boxing rings, squat racks, plates, dumbbell racks soft lighting, masterpiece, best quality, 8k uhd, film grain, Fujifilm XT3 photorealistic painting art by midjourney and greg rutkowski ", + "(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck", + 6, + ], + [ + "1man, 25 years- old, full body, wearing long-sleeve white shirt and tie, muscular rand black suit, soft lighting, masterpiece, best quality, 8k uhd, dslr, film grain, Fujifilm XT3 photorealistic painting art by midjourney and greg rutkowski ", + "(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck", + 6, + ], ] print("The GUI is not fully tested yet. Please open an issue if you find bugs.") - INFO = f''' + INFO = f""" ## Generate Your Handsome powered by EditAnything https://github.com/sail-sg/EditAnything This model is good at generating handsome male. - ''' - WARNING_INFO = f'''### [NOTE] the model is collected from the Internet for demo only, please do not use it for commercial purposes. + """ + WARNING_INFO = f"""### [NOTE] the model is collected from the Internet for demo only, please do not use it for commercial purposes. We are not responsible for possible risks using this model. Base model from https://huggingface.co/SG161222/Realistic_Vision_V2.0 Thanks! - ''' - demo = create_demo_template(process, process_image_click, examples=examples, INFO=INFO, WARNING_INFO=WARNING_INFO) + """ + demo = create_demo_template( + process, + process_image_click, + examples=examples, + INFO=INFO, + WARNING_INFO=WARNING_INFO, + ) return demo -if __name__ == '__main__': - model = EditAnythingLoraModel(base_model_path='Realistic_Vision_V2.0', - lora_model_path=None, use_blip=True) +if __name__ == "__main__": + model = EditAnythingLoraModel( + base_model_path="Realistic_Vision_V2.0", lora_model_path=None, use_blip=True + ) demo = create_demo(model.process, model.process_image_click) - demo.queue().launch(server_name='0.0.0.0') + demo.queue().launch(server_name="0.0.0.0") diff --git a/editany_lora.py b/editany_lora.py new file mode 100644 index 0000000000000000000000000000000000000000..aed7bc1b7f4a2f021d7908dc9722ee175150d33e --- /dev/null +++ b/editany_lora.py @@ -0,0 +1,908 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +from torchvision.utils import save_image +from PIL import Image +from pytorch_lightning import seed_everything +import subprocess +from collections import OrderedDict +import re +import cv2 +import einops +import gradio as gr +import numpy as np +import torch +import random +import os +import requests +from io import BytesIO +from annotator.util import resize_image, HWC3, resize_points, get_bounding_box + +import torch +from safetensors.torch import load_file +from collections import defaultdict +from diffusers import StableDiffusionControlNetPipeline +from diffusers import ControlNetModel, UniPCMultistepScheduler + +from utils.stable_diffusion_controlnet import ControlNetModel2 +from utils.stable_diffusion_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline, \ + StableDiffusionControlNetInpaintMixingPipeline, prepare_mask_image +# need the latest transformers +# pip install git+https://github.com/huggingface/transformers.git +from transformers import AutoProcessor, Blip2ForConditionalGeneration +from diffusers import ControlNetModel, DiffusionPipeline +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput +import PIL.Image + +# Segment-Anything init. +# pip install git+https://github.com/facebookresearch/segment-anything.git +try: + from segment_anything import ( + sam_model_registry, + SamAutomaticMaskGenerator, + SamPredictor, + ) +except ImportError: + print("segment_anything not installed") + result = subprocess.run( + [ + "pip", + "install", + "git+https://github.com/facebookresearch/segment-anything.git", + ], + check=True, + ) + print(f"Install segment_anything {result}") + from segment_anything import ( + sam_model_registry, + SamAutomaticMaskGenerator, + SamPredictor, + ) +if not os.path.exists("./models/sam_vit_h_4b8939.pth"): + result = subprocess.run( + [ + "wget", + "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth", + "-P", + "models", + ], + check=True, + ) + print(f"Download sam_vit_h_4b8939.pth {result}") + +device = "cuda" if torch.cuda.is_available() else "cpu" + +config_dict = OrderedDict( + [ + ("LAION Pretrained(v0-4)-SD15", "shgao/edit-anything-v0-4-sd15"), + ("LAION Pretrained(v0-4)-SD21", "shgao/edit-anything-v0-4-sd21"), + ("LAION Pretrained(v0-3)-SD21", "shgao/edit-anything-v0-3"), + ("SAM Pretrained(v0-1)-SD21", "shgao/edit-anything-v0-1-1"), + ] +) + + +def init_sam_model(sam_generator=None, mask_predictor=None): + if sam_generator is not None and mask_predictor is not None: + return sam_generator, mask_predictor + sam_checkpoint = "models/sam_vit_h_4b8939.pth" + model_type = "default" + sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) + sam.to(device=device) + sam_generator = ( + SamAutomaticMaskGenerator( + sam) if sam_generator is None else sam_generator + ) + mask_predictor = SamPredictor( + sam) if mask_predictor is None else mask_predictor + return sam_generator, mask_predictor + + +def init_blip_processor(): + blip_processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b") + return blip_processor + + +def init_blip_model(): + blip_model = Blip2ForConditionalGeneration.from_pretrained( + "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16, device_map="auto" + ) + return blip_model + + +def get_pipeline_embeds(pipeline, prompt, negative_prompt, device): + # https://github.com/huggingface/diffusers/issues/2136 + """Get pipeline embeds for prompts bigger than the maxlength of the pipe + :param pipeline: + :param prompt: + :param negative_prompt: + :param device: + :return: + """ + max_length = pipeline.tokenizer.model_max_length + + # simple way to determine length of tokens + count_prompt = len(re.split(r", ", prompt)) + count_negative_prompt = len(re.split(r", ", negative_prompt)) + + # create the tensor based on which prompt is longer + if count_prompt >= count_negative_prompt: + input_ids = pipeline.tokenizer( + prompt, return_tensors="pt", truncation=False + ).input_ids.to(device) + shape_max_length = input_ids.shape[-1] + negative_ids = pipeline.tokenizer( + negative_prompt, + truncation=False, + padding="max_length", + max_length=shape_max_length, + return_tensors="pt", + ).input_ids.to(device) + else: + negative_ids = pipeline.tokenizer( + negative_prompt, return_tensors="pt", truncation=False + ).input_ids.to(device) + shape_max_length = negative_ids.shape[-1] + input_ids = pipeline.tokenizer( + prompt, + return_tensors="pt", + truncation=False, + padding="max_length", + max_length=shape_max_length, + ).input_ids.to(device) + + concat_embeds = [] + neg_embeds = [] + for i in range(0, shape_max_length, max_length): + concat_embeds.append(pipeline.text_encoder(input_ids[:, i : i + max_length])[0]) + neg_embeds.append(pipeline.text_encoder(negative_ids[:, i : i + max_length])[0]) + + return torch.cat(concat_embeds, dim=1), torch.cat(neg_embeds, dim=1) + + +def load_lora_weights(pipeline, checkpoint_path, multiplier, device, dtype): + LORA_PREFIX_UNET = "lora_unet" + LORA_PREFIX_TEXT_ENCODER = "lora_te" + # load LoRA weight from .safetensors + print('device: {}'.format(device)) + if isinstance(checkpoint_path, str): + state_dict = load_file(checkpoint_path, device=device) + + updates = defaultdict(dict) + for key, value in state_dict.items(): + # it is suggested to print out the key, it usually will be something like below + # "lora_te_text_model_encoder_layers_0_self_attn_k_proj.lora_down.weight" + + layer, elem = key.split(".", 1) + updates[layer][elem] = value + + # directly update weight in diffusers model + for layer, elems in updates.items(): + + if "text" in layer: + layer_infos = layer.split(LORA_PREFIX_TEXT_ENCODER + "_")[-1].split("_") + curr_layer = pipeline.text_encoder + else: + layer_infos = layer.split(LORA_PREFIX_UNET + "_")[-1].split("_") + curr_layer = pipeline.unet + + # find the target layer + temp_name = layer_infos.pop(0) + while len(layer_infos) > -1: + try: + curr_layer = curr_layer.__getattr__(temp_name) + if len(layer_infos) > 0: + temp_name = layer_infos.pop(0) + elif len(layer_infos) == 0: + break + except Exception: + if len(temp_name) > 0: + temp_name += "_" + layer_infos.pop(0) + else: + temp_name = layer_infos.pop(0) + + # get elements for this layer + weight_up = elems["lora_up.weight"].to(dtype) + weight_down = elems["lora_down.weight"].to(dtype) + alpha = elems["alpha"] + if alpha: + alpha = alpha.item() / weight_up.shape[1] + else: + alpha = 1.0 + + # update weight + if len(weight_up.shape) == 4: + curr_layer.weight.data += ( + multiplier + * alpha + * torch.mm( + weight_up.squeeze(3).squeeze(2), + weight_down.squeeze(3).squeeze(2), + ) + .unsqueeze(2) + .unsqueeze(3) + ) + else: + curr_layer.weight.data += ( + multiplier * alpha * torch.mm(weight_up, weight_down) + ) + else: + for ckptpath in checkpoint_path: + state_dict = load_file(ckptpath, device=device) + + updates = defaultdict(dict) + for key, value in state_dict.items(): + # it is suggested to print out the key, it usually will be something like below + # "lora_te_text_model_encoder_layers_0_self_attn_k_proj.lora_down.weight" + + layer, elem = key.split(".", 1) + updates[layer][elem] = value + + # directly update weight in diffusers model + for layer, elems in updates.items(): + if "text" in layer: + layer_infos = layer.split(LORA_PREFIX_TEXT_ENCODER + "_")[-1].split( + "_" + ) + curr_layer = pipeline.text_encoder + else: + layer_infos = layer.split(LORA_PREFIX_UNET + "_")[-1].split("_") + curr_layer = pipeline.unet + + # find the target layer + temp_name = layer_infos.pop(0) + while len(layer_infos) > -1: + try: + curr_layer = curr_layer.__getattr__(temp_name) + if len(layer_infos) > 0: + temp_name = layer_infos.pop(0) + elif len(layer_infos) == 0: + break + except Exception: + if len(temp_name) > 0: + temp_name += "_" + layer_infos.pop(0) + else: + temp_name = layer_infos.pop(0) + + # get elements for this layer + weight_up = elems["lora_up.weight"].to(dtype) + weight_down = elems["lora_down.weight"].to(dtype) + alpha = elems["alpha"] + if alpha: + alpha = alpha.item() / weight_up.shape[1] + else: + alpha = 1.0 + + # update weight + if len(weight_up.shape) == 4: + curr_layer.weight.data += ( + multiplier + * alpha + * torch.mm( + weight_up.squeeze(3).squeeze(2), + weight_down.squeeze(3).squeeze(2), + ) + .unsqueeze(2) + .unsqueeze(3) + ) + else: + curr_layer.weight.data += ( + multiplier * alpha * torch.mm(weight_up, weight_down) + ) + return pipeline + + +def make_inpaint_condition(image, image_mask): + image = image / 255.0 + assert ( + image.shape[0:1] == image_mask.shape[0:1] + ), "image and image_mask must have the same image size" + image[image_mask > 128] = -1.0 # set as masked pixel + image = np.expand_dims(image, 0).transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + return image + + +def obtain_generation_model( + base_model_path, + lora_model_path, + controlnet_path, + generation_only=False, + extra_inpaint=True, + lora_weight=1.0, +): + controlnet = [] + controlnet.append( + ControlNetModel2.from_pretrained( + controlnet_path, torch_dtype=torch.float16) + ) # sam control + if (not generation_only) and extra_inpaint: # inpainting control + print("Warning: ControlNet based inpainting model only support SD1.5 for now.") + controlnet.append( + ControlNetModel.from_pretrained( + "lllyasviel/control_v11p_sd15_inpaint", torch_dtype=torch.float16 + ) # inpainting controlnet + ) + + if generation_only and extra_inpaint: + pipe = StableDiffusionControlNetPipeline.from_pretrained( + base_model_path, + controlnet=controlnet, + torch_dtype=torch.float16, + safety_checker=None, + ) + else: + pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( + base_model_path, + controlnet=controlnet, + torch_dtype=torch.float16, + safety_checker=None, + ) + if lora_model_path is not None: + pipe = load_lora_weights( + pipe, [lora_model_path], lora_weight, "cpu", torch.float32 + ) + # speed up diffusion process with faster scheduler and memory optimization + pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) + # remove following line if xformers is not installed + pipe.enable_xformers_memory_efficient_attention() + + pipe.enable_model_cpu_offload() + return pipe + + +def obtain_tile_model(base_model_path, lora_model_path, lora_weight=1.0): + controlnet = ControlNetModel2.from_pretrained( + "lllyasviel/control_v11f1e_sd15_tile", torch_dtype=torch.float16 + ) # tile controlnet + if ( + base_model_path == "runwayml/stable-diffusion-v1-5" + or base_model_path == "stabilityai/stable-diffusion-2-inpainting" + ): + print("base_model_path", base_model_path) + pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( + "runwayml/stable-diffusion-v1-5", + controlnet=controlnet, + torch_dtype=torch.float16, + safety_checker=None, + ) + else: + pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( + base_model_path, + controlnet=controlnet, + torch_dtype=torch.float16, + safety_checker=None, + ) + if lora_model_path is not None: + pipe = load_lora_weights( + pipe, [lora_model_path], lora_weight, "cpu", torch.float32 + ) + # speed up diffusion process with faster scheduler and memory optimization + pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) + # remove following line if xformers is not installed + pipe.enable_xformers_memory_efficient_attention() + + pipe.enable_model_cpu_offload() + return pipe + + +def show_anns(anns): + if len(anns) == 0: + return + sorted_anns = sorted(anns, key=(lambda x: x["area"]), reverse=True) + full_img = None + + # for ann in sorted_anns: + for i in range(len(sorted_anns)): + ann = anns[i] + m = ann["segmentation"] + if full_img is None: + full_img = np.zeros((m.shape[0], m.shape[1], 3)) + map = np.zeros((m.shape[0], m.shape[1]), dtype=np.uint16) + map[m != 0] = i + 1 + color_mask = np.random.random((1, 3)).tolist()[0] + full_img[m != 0] = color_mask + full_img = full_img * 255 + # anno encoding from https://github.com/LUSSeg/ImageNet-S + res = np.zeros((map.shape[0], map.shape[1], 3)) + res[:, :, 0] = map % 256 + res[:, :, 1] = map // 256 + res.astype(np.float32) + full_img = Image.fromarray(np.uint8(full_img)) + return full_img, res + + +class EditAnythingLoraModel: + def __init__( + self, + base_model_path="../chilloutmix_NiPrunedFp32Fix", + lora_model_path="../40806/mix4", + use_blip=True, + blip_processor=None, + blip_model=None, + sam_generator=None, + controlmodel_name="LAION Pretrained(v0-4)-SD15", + # used when the base model is not an inpainting model. + extra_inpaint=True, + tile_model=None, + lora_weight=1.0, + alpha_mixing=None, + mask_predictor=None, + ): + self.device = device + self.use_blip = use_blip + + # Diffusion init using diffusers. + self.default_controlnet_path = config_dict[controlmodel_name] + self.base_model_path = base_model_path + self.lora_model_path = lora_model_path + self.defalut_enable_all_generate = False + self.extra_inpaint = extra_inpaint + self.last_ref_infer = False + self.pipe = obtain_generation_model( + base_model_path, + lora_model_path, + self.default_controlnet_path, + generation_only=False, + extra_inpaint=extra_inpaint, + lora_weight=lora_weight, + ) + # self.pipe.load_textual_inversion("textual_inversion_cat/learned_embeds.bin") + # Segment-Anything init. + self.sam_generator, self.mask_predictor = init_sam_model( + sam_generator, mask_predictor + ) + # BLIP2 init. + if use_blip: + if blip_processor is not None: + self.blip_processor = blip_processor + else: + self.blip_processor = init_blip_processor() + + if blip_model is not None: + self.blip_model = blip_model + else: + self.blip_model = init_blip_model() + + # tile model init. + if tile_model is not None: + self.tile_pipe = tile_model + else: + self.tile_pipe = obtain_tile_model( + base_model_path, lora_model_path, lora_weight=lora_weight + ) + + def get_blip2_text(self, image): + inputs = self.blip_processor(image, return_tensors="pt").to( + self.device, torch.float16 + ) + generated_ids = self.blip_model.generate(**inputs, max_new_tokens=50) + generated_text = self.blip_processor.batch_decode( + generated_ids, skip_special_tokens=True + )[0].strip() + return generated_text + + def get_sam_control(self, image): + masks = self.sam_generator.generate(image) + full_img, res = show_anns(masks) + return full_img, res + + def get_click_mask(self, image, clicked_points): + self.mask_predictor.set_image(image) + # Separate the points and labels + points, labels = zip(*[(point[:2], point[2]) + for point in clicked_points]) + + # Convert the points and labels to numpy arrays + input_point = np.array(points) + input_label = np.array(labels) + + masks, _, _ = self.mask_predictor.predict( + point_coords=input_point, + point_labels=input_label, + multimask_output=False, + ) + + return masks + + @torch.inference_mode() + def process_image_click( + self, + original_image: gr.Image, + point_prompt: gr.Radio, + clicked_points: gr.State, + image_resolution, + evt: gr.SelectData, + ): + # Get the clicked coordinates + clicked_coords = evt.index + x, y = clicked_coords + label = point_prompt + lab = 1 if label == "Foreground Point" else 0 + clicked_points.append((x, y, lab)) + + input_image = np.array(original_image, dtype=np.uint8) + H, W, C = input_image.shape + input_image = HWC3(input_image) + img = resize_image(input_image, image_resolution) + + # Update the clicked_points + resized_points = resize_points( + clicked_points, input_image.shape, image_resolution + ) + mask_click_np = self.get_click_mask(img, resized_points) + + # Convert mask_click_np to HWC format + mask_click_np = np.transpose(mask_click_np, (1, 2, 0)) * 255.0 + + mask_image = HWC3(mask_click_np.astype(np.uint8)) + mask_image = cv2.resize(mask_image, (W, H), interpolation=cv2.INTER_LINEAR) + # mask_image = Image.fromarray(mask_image_tmp) + + # Draw circles for all clicked points + edited_image = input_image + for x, y, lab in clicked_points: + # Set the circle color based on the label + color = (255, 0, 0) if lab == 1 else (0, 0, 255) + + # Draw the circle + edited_image = cv2.circle(edited_image, (x, y), 20, color, -1) + + # Set the opacity for the mask_image and edited_image + opacity_mask = 0.75 + opacity_edited = 1.0 + + # Combine the edited_image and the mask_image using cv2.addWeighted() + overlay_image = cv2.addWeighted( + edited_image, + opacity_edited, + (mask_image * + np.array([0 / 255, 255 / 255, 0 / 255])).astype(np.uint8), + opacity_mask, + 0, + ) + + return ( + Image.fromarray(overlay_image), + clicked_points, + Image.fromarray(mask_image), + ) + + @torch.inference_mode() + def process( + self, + source_image, + enable_all_generate, + mask_image, + control_scale, + enable_auto_prompt, + a_prompt, + n_prompt, + num_samples, + image_resolution, + detect_resolution, + ddim_steps, + guess_mode, + scale, + seed, + eta, + enable_tile=True, + refine_alignment_ratio=None, + refine_image_resolution=None, + alpha_weight=0.5, + use_scale_map=False, + condition_model=None, + ref_image=None, + attention_auto_machine_weight=1.0, + gn_auto_machine_weight=1.0, + style_fidelity=0.5, + reference_attn=True, + reference_adain=True, + ref_prompt=None, + ref_sam_scale=None, + ref_inpaint_scale=None, + ref_auto_prompt=False, + ref_textinv=True, + ref_textinv_path=None, + ): + + if condition_model is None or condition_model == "EditAnything": + this_controlnet_path = self.default_controlnet_path + else: + this_controlnet_path = condition_model + input_image = ( + source_image["image"] + if isinstance(source_image, dict) + else np.array(source_image, dtype=np.uint8) + ) + if mask_image is None: + if enable_all_generate != self.defalut_enable_all_generate: + self.pipe = obtain_generation_model( + self.base_model_path, + self.lora_model_path, + this_controlnet_path, + enable_all_generate, + self.extra_inpaint, + ) + self.defalut_enable_all_generate = enable_all_generate + if enable_all_generate: + print( + "source_image", + source_image["mask"].shape, + input_image.shape, + ) + mask_image = ( + np.ones((input_image.shape[0], + input_image.shape[1], 3)) * 255 + ) + else: + mask_image = source_image["mask"] + else: + mask_image = np.array(mask_image, dtype=np.uint8) + if self.default_controlnet_path != this_controlnet_path: + print( + "To Use:", + this_controlnet_path, + "Current:", + self.default_controlnet_path, + ) + print("Change condition model to:", this_controlnet_path) + self.pipe = obtain_generation_model( + self.base_model_path, + self.lora_model_path, + this_controlnet_path, + enable_all_generate, + self.extra_inpaint, + ) + self.default_controlnet_path = this_controlnet_path + torch.cuda.empty_cache() + if self.last_ref_infer: + print("Redefine the model to overwrite the ref mode") + self.pipe = obtain_generation_model( + self.base_model_path, + self.lora_model_path, + this_controlnet_path, + enable_all_generate, + self.extra_inpaint, + ) + self.last_ref_infer = False + + if ref_image is not None: + ref_mask = ref_image["mask"] + ref_image = ref_image["image"] + if ref_auto_prompt or ref_textinv: + bbox = get_bounding_box( + np.array(ref_mask) / 255 + ) # reverse the mask to make 1 the choosen region + cropped_ref_mask = ref_mask.crop( + (bbox[0], bbox[1], bbox[2], bbox[3])) + cropped_ref_image = ref_image.crop( + (bbox[0], bbox[1], bbox[2], bbox[3])) + # cropped_ref_image.save("debug.jpg") + cropped_ref_image = np.array(cropped_ref_image) * ( + np.array(cropped_ref_mask)[:, :, :3] / 255.0 + ) + cropped_ref_image = Image.fromarray( + cropped_ref_image.astype("uint8")) + + if ref_auto_prompt: + generated_prompt = self.get_blip2_text(cropped_ref_image) + ref_prompt += generated_prompt + a_prompt += generated_prompt + print("Generated ref text:", ref_prompt) + print("Generated input text:", a_prompt) + self.last_ref_infer = True + # ref_image = cropped_ref_image + # ref_mask = cropped_ref_mask + if ref_textinv: + try: + self.pipe.load_textual_inversion(ref_textinv_path) + print("Load textinv embedding from:", ref_textinv_path) + except: + print("No textinvert embeddings found.") + ref_data_path = "./utils/tmp/textinv/img" + if not os.path.exists(ref_data_path): + os.makedirs(ref_data_path) + cropped_ref_image.save(os.path.join(ref_data_path, 'ref.png')) + print("Ref image region is save to:", ref_data_path) + print("Plese finetune with run_texutal_inversion.sh in utils folder to get the textinvert embeddings.") + + else: + ref_mask = None + + with torch.no_grad(): + if self.use_blip and enable_auto_prompt: + print("Generating text:") + blip2_prompt = self.get_blip2_text(input_image) + print("Generated text:", blip2_prompt) + if len(a_prompt) > 0: + a_prompt = blip2_prompt + "," + a_prompt + else: + a_prompt = blip2_prompt + + input_image = HWC3(input_image) + + img = resize_image(input_image, image_resolution) + H, W, C = img.shape + + print("Generating SAM seg:") + # the default SAM model is trained with 1024 size. + full_segmask, detected_map = self.get_sam_control( + resize_image(input_image, detect_resolution) + ) + + detected_map = HWC3(detected_map.astype(np.uint8)) + detected_map = cv2.resize( + detected_map, (W, H), interpolation=cv2.INTER_LINEAR + ) + + control = torch.from_numpy(detected_map.copy()).float().cuda() + control = torch.stack([control for _ in range(num_samples)], dim=0) + control = einops.rearrange(control, "b h w c -> b c h w").clone() + + mask_imag_ori = HWC3(mask_image.astype(np.uint8)) + mask_image_tmp = cv2.resize( + mask_imag_ori, (W, H), interpolation=cv2.INTER_LINEAR + ) + mask_image = Image.fromarray(mask_image_tmp) + + if seed == -1: + seed = random.randint(0, 65535) + seed_everything(seed) + generator = torch.manual_seed(seed) + postive_prompt = a_prompt + negative_prompt = n_prompt + prompt_embeds, negative_prompt_embeds = get_pipeline_embeds( + self.pipe, postive_prompt, negative_prompt, "cuda" + ) + prompt_embeds = torch.cat([prompt_embeds] * num_samples, dim=0) + negative_prompt_embeds = torch.cat( + [negative_prompt_embeds] * num_samples, dim=0 + ) + + if enable_all_generate and self.extra_inpaint: + self.pipe.safety_checker = lambda images, clip_input: ( + images, False) + if ref_image is not None: + print("Not support yet.") + return + x_samples = self.pipe( + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=num_samples, + num_inference_steps=ddim_steps, + generator=generator, + height=H, + width=W, + image=[control.type(torch.float16)], + controlnet_conditioning_scale=[float(control_scale)], + guidance_scale=scale, + guess_mode=guess_mode, + ).images + else: + multi_condition_image = [] + multi_condition_scale = [] + multi_condition_image.append(control.type(torch.float16)) + multi_condition_scale.append(float(control_scale)) + ref_multi_condition_scale = [] + if ref_image is not None: + ref_multi_condition_scale.append(float(ref_sam_scale)) + if self.extra_inpaint: + inpaint_image = make_inpaint_condition(img, mask_image_tmp) + multi_condition_image.append( + inpaint_image.type(torch.float16)) + multi_condition_scale.append(1.0) + if ref_image is not None: + ref_multi_condition_scale.append( + float(ref_inpaint_scale)) + if use_scale_map: + scale_map_tmp = source_image["mask"] + tmp = HWC3(scale_map_tmp.astype(np.uint8)) + scale_map_tmp = cv2.resize( + tmp, (W, H), interpolation=cv2.INTER_LINEAR) + scale_map_tmp = Image.fromarray(scale_map_tmp) + controlnet_conditioning_scale_map = 1.0 - \ + prepare_mask_image(scale_map_tmp).float() + print('scale map:', controlnet_conditioning_scale_map.size()) + else: + controlnet_conditioning_scale_map = None + + if isinstance(self.pipe, StableDiffusionControlNetInpaintMixingPipeline): + x_samples = self.pipe( + image=img, + mask_image=mask_image, + prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=num_samples, + num_inference_steps=ddim_steps, + generator=generator, + controlnet_conditioning_image=multi_condition_image, + height=H, + width=W, + controlnet_conditioning_scale=multi_condition_scale, + guidance_scale=scale, + alpha_weight=alpha_weight, + controlnet_conditioning_scale_map=controlnet_conditioning_scale_map + ).images + else: + x_samples = self.pipe( + image=img, + mask_image=mask_image, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=num_samples, + num_inference_steps=ddim_steps, + generator=generator, + controlnet_conditioning_image=multi_condition_image, + height=H, + width=W, + controlnet_conditioning_scale=multi_condition_scale, + guidance_scale=scale, + ref_image=ref_image, + ref_mask=ref_mask, + ref_prompt=ref_prompt, + attention_auto_machine_weight=attention_auto_machine_weight, + gn_auto_machine_weight=gn_auto_machine_weight, + style_fidelity=style_fidelity, + reference_attn=reference_attn, + reference_adain=reference_adain, + ref_controlnet_conditioning_scale=ref_multi_condition_scale, + guess_mode=guess_mode, + ).images + results = [x_samples[i] for i in range(num_samples)] + + results_tile = [] + if enable_tile: + prompt_embeds, negative_prompt_embeds = get_pipeline_embeds( + self.tile_pipe, postive_prompt, negative_prompt, "cuda" + ) + for i in range(num_samples): + img_tile = PIL.Image.fromarray( + resize_image( + np.array(x_samples[i]), refine_image_resolution) + ) + if i == 0: + mask_image_tile = cv2.resize( + mask_imag_ori, + (img_tile.size[0], img_tile.size[1]), + interpolation=cv2.INTER_LINEAR, + ) + mask_image_tile = Image.fromarray(mask_image_tile) + if isinstance(self.pipe, StableDiffusionControlNetInpaintMixingPipeline): + x_samples_tile = self.tile_pipe( + image=img_tile, + mask_image=mask_image_tile, + prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=1, + num_inference_steps=ddim_steps, + generator=generator, + controlnet_conditioning_image=img_tile, + height=img_tile.size[1], + width=img_tile.size[0], + controlnet_conditioning_scale=1.0, + alignment_ratio=refine_alignment_ratio, + guidance_scale=scale, + alpha_weight=alpha_weight, + controlnet_conditioning_scale_map=controlnet_conditioning_scale_map + ).images + else: + x_samples_tile = self.tile_pipe( + image=img_tile, + mask_image=mask_image_tile, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + num_images_per_prompt=1, + num_inference_steps=ddim_steps, + generator=generator, + controlnet_conditioning_image=img_tile, + height=img_tile.size[1], + width=img_tile.size[0], + controlnet_conditioning_scale=1.0, + alignment_ratio=refine_alignment_ratio, + guidance_scale=scale, + guess_mode=guess_mode, + ).images + results_tile += x_samples_tile + + return results_tile, results, [full_segmask, mask_image], postive_prompt + + def download_image(url): + response = requests.get(url) + return Image.open(BytesIO(response.content)).convert("RGB") diff --git a/editany_test.py b/editany_test.py new file mode 100644 index 0000000000000000000000000000000000000000..62f1b4f77cab3d163f8986f673cd919d3e99028c --- /dev/null +++ b/editany_test.py @@ -0,0 +1,73 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +import os +import gradio as gr +from diffusers.utils import load_image +from editany_lora import EditAnythingLoraModel, config_dict +from editany_demo import create_demo_template +from huggingface_hub import hf_hub_download, snapshot_download + + +def create_demo(process, process_image_click=None): + + examples = [ + [ + "dudou,1girl, beautiful face, solo, candle, brown hair, long hair, ,ulzzang-6500-v1.1,(raw photo:1.2),((photorealistic:1.4))best quality ,masterpiece, illustration, an extremely delicate and beautiful, extremely detailed ,CG ,unity ,8k wallpaper, Amazing, finely detail, masterpiece,best quality,official art,extremely detailed CG unity 8k wallpaper,absurdres, incredibly absurdres, huge filesize, ultra-detailed, highres, extremely detailed,beautiful detailed girl, extremely detailed eyes and face, beautiful detailed eyes,cinematic lighting,1girl,see-through,looking at viewer,full body,full-body shot,outdoors,arms behind back,(chinese clothes) ", + "(((mole))),sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, bad anatomy,(long hair:1.4),DeepNegative,(fat:1.2),facing away, looking away,tilted head, lowres,bad anatomy,bad hands, text, error, missing fingers,extra digit, fewer digits, cropped, worstquality, low quality, normal quality,jpegartifacts,signature, watermark, username,blurry,bad feet,cropped,poorly drawn hands,poorly drawn face,mutation,deformed,worst quality,low quality,normal quality,jpeg artifacts,signature,watermark,extra fingers,fewer digits,extra limbs,extra arms,extra legs,malformed limbs,fused fingers,too many fingers,long neck,cross-eyed,mutated hands,polar lowres,bad body,bad proportions,gross proportions,text,error,missing fingers,missing arms,missing legs,extra digit, extra arms, extra leg, extra foot,(freckles),(mole:2)", + 5, + ], + [ + "best quality, ultra high res, (photorealistic:1.4), (detailed beautiful girl:1.4), (medium breasts:0.8), looking_at_viewer, Detailed facial details, beautiful detailed eyes, (multicolored|blue|pink hair: 1.2), green eyes, slender, haunting smile, (makeup:0.3), red lips, , highly detailed clothes, (ulzzang-6500-v1.1:0.3)", + "EasyNegative, paintings, sketches, ugly, 3d, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, manboobs, backlight,(ugly:1.3), (duplicate:1.3), (morbid:1.2), (mutilated:1.2), (tranny:1.3), mutated hands, (poorly drawn hands:1.3), blurry, (bad anatomy:1.2), (bad proportions:1.3), extra limbs, (disfigured:1.3), (more than 2 nipples:1.3), (more than 1 navel:1.3), (missing arms:1.3), (extra legs:1.3), (fused fingers:1.6), (too many fingers:1.6), (unclear eyes:1.3), bad hands, missing fingers, extra digit, (futa:1.1), bad body, double navel, mutad arms, hused arms, (puffy nipples, dark areolae, dark nipples, rei no himo, inverted nipples, long nipples), NG_DeepNegative_V1_75t, pubic hair, fat rolls, obese, bad-picture-chill-75v", + 8, + ], + [ + "best quality, ultra high res, (photorealistic:1.4), (detailed beautiful girl:1.4), (medium breasts:0.8), looking_at_viewer, Detailed facial details, beautiful detailed eyes, (blue|pink hair), green eyes, slender, smile, (makeup:0.4), red lips, (full body, sitting, beach), , highly detailed clothes, (ulzzang-6500-v1.1:0.3)", + "asyNegative, paintings, sketches, ugly, 3d, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, manboobs, backlight,(ugly:1.3), (duplicate:1.3), (morbid:1.2), (mutilated:1.2), (tranny:1.3), mutated hands, (poorly drawn hands:1.3), blurry, (bad anatomy:1.2), (bad proportions:1.3), extra limbs, (disfigured:1.3), (more than 2 nipples:1.3), (more than 1 navel:1.3), (missing arms:1.3), (extra legs:1.3), (fused fingers:1.6), (too many fingers:1.6), (unclear eyes:1.3), bad hands, missing fingers, extra digit, (futa:1.1), bad body, double navel, mutad arms, hused arms, (puffy nipples, dark areolae, dark nipples, rei no himo, inverted nipples, long nipples), NG_DeepNegative_V1_75t, pubic hair, fat rolls, obese, bad-picture-chill-75v", + 7, + ], + [ + "mix4, whole body shot, ((8k, RAW photo, highest quality, masterpiece), High detail RAW color photo professional close-up photo, shy expression, cute, beautiful detailed girl, detailed fingers, extremely detailed eyes and face, beautiful detailed nose, beautiful detailed eyes, long eyelashes, light on face, looking at viewer, (closed mouth:1.2), 1girl, cute, young, mature face, (full body:1.3), ((small breasts)), realistic face, realistic body, beautiful detailed thigh,s, same eyes color, (realistic, photo realism:1. 37), (highest quality), (best shadow), (best illustration), ultra high resolution, physics-based rendering, cinematic lighting), solo, 1girl, highly detailed, in office, detailed office, open cardigan, ponytail contorted, beautiful eyes ,sitting in office,dating, business suit, cross-laced clothes, collared shirt, beautiful breast, small breast, Chinese dress, white pantyhose, natural breasts, pink and white hair, ", + "paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), cloth, underwear, bra, low-res, normal quality, ((monochrome)), ((grayscale)), skin spots, acne, skin blemishes, age spots, glans, bad nipples, long nipples, bad vagina, extra fingers,fewer fingers,strange fingers,bad hand, ng_deepnegative_v1_75t, bad-picture-chill-75v", + 7, + ], + ] + INFO = f""" + ## Generate Your Beauty powered by EditAnything https://github.com/sail-sg/EditAnything + This model is good at generating beautiful female. + """ + WARNING_INFO = f"""### [NOTE] the model is collected from the Internet for demo only, please do not use it for commercial purposes. + We are not responsible for possible risks using this model. + Lora model from https://civitai.com/models/14171/cutegirlmix4 Thanks! + """ + demo = create_demo_template( + process, + process_image_click, + examples=examples, + INFO=INFO, + WARNING_INFO=WARNING_INFO, + ) + return demo + + +if __name__ == "__main__": + # sd_models_path = snapshot_download("shgao/sdmodels") + # lora_model_path = hf_hub_download( + # "mlida/Cute_girl_mix4", "cuteGirlMix4_v10.safetensors") + # model = EditAnythingLoraModel(base_model_path="andite/anything-v4.0", + # lora_model_path=None, use_blip=True, extra_inpaint=True, + # lora_weight=0.5, + # ) + sd_models_path = snapshot_download("shgao/sdmodels") + lora_model_path = hf_hub_download( + "mlida/Cute_girl_mix4", "cuteGirlMix4_v10.safetensors" + ) + model = EditAnythingLoraModel( + base_model_path=os.path.join( + sd_models_path, "chilloutmix_NiPrunedFp32Fix"), + lora_model_path=lora_model_path, + use_blip=True, + extra_inpaint=True, + lora_weight=0.5, + ) + demo = create_demo(model.process, model.process_image_click) + demo.queue().launch(server_name="0.0.0.0") diff --git a/environment.yaml b/environment.yaml new file mode 100644 index 0000000000000000000000000000000000000000..5c6f390270c2031dfa4fa564e6adf1f05f2a56ec --- /dev/null +++ b/environment.yaml @@ -0,0 +1,38 @@ +name: control +channels: + - pytorch + - defaults +dependencies: + - python=3.8.5 + - pip=20.3 + - cudatoolkit=11.3 + - pytorch=1.13.1 + - torchvision=0.14.1 + - numpy=1.23.1 + - pip: + - gradio==3.35.2 + - albumentations==1.3.0 + - opencv-contrib-python==4.3.0.36 + - imageio==2.9.0 + - imageio-ffmpeg==0.4.2 + - pytorch-lightning==1.5.0 + - omegaconf==2.1.1 + - test-tube>=0.7.5 + - streamlit==1.12.1 + - einops==0.3.0 + - webdataset==0.2.5 + - kornia==0.6 + - open_clip_torch==2.0.2 + - invisible-watermark>=0.1.5 + - streamlit-drawable-canvas==0.8.0 + - torchmetrics==0.6.0 + - timm==0.6.12 + - addict==2.4.0 + - yapf==0.32.0 + - prettytable==3.6.0 + - safetensors==0.2.7 + - basicsr==1.4.2 + - diffusers==0.17.1 + - accelerate==0.17.0 + - transformers==4.30.2 + - xformers diff --git a/font/DejaVuSans.ttf b/font/DejaVuSans.ttf new file mode 100644 index 0000000000000000000000000000000000000000..e5f7eecce43be41ff0703ed99e1553029b849f14 Binary files /dev/null and b/font/DejaVuSans.ttf differ diff --git a/ldm/data/__init__.py b/ldm/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/data/util.py b/ldm/data/util.py new file mode 100644 index 0000000000000000000000000000000000000000..5b60ceb2349e3bd7900ff325740e2022d2903b1c --- /dev/null +++ b/ldm/data/util.py @@ -0,0 +1,24 @@ +import torch + +from ldm.modules.midas.api import load_midas_transform + + +class AddMiDaS(object): + def __init__(self, model_type): + super().__init__() + self.transform = load_midas_transform(model_type) + + def pt2np(self, x): + x = ((x + 1.0) * .5).detach().cpu().numpy() + return x + + def np2pt(self, x): + x = torch.from_numpy(x) * 2 - 1. + return x + + def __call__(self, sample): + # sample['jpg'] is tensor hwc in [-1, 1] at this point + x = self.pt2np(sample['jpg']) + x = self.transform({"image": x})["image"] + sample['midas_in'] = x + return sample \ No newline at end of file diff --git a/ldm/models/autoencoder.py b/ldm/models/autoencoder.py new file mode 100644 index 0000000000000000000000000000000000000000..d122549995ce2cd64092c81a58419ed4a15a02fd --- /dev/null +++ b/ldm/models/autoencoder.py @@ -0,0 +1,219 @@ +import torch +import pytorch_lightning as pl +import torch.nn.functional as F +from contextlib import contextmanager + +from ldm.modules.diffusionmodules.model import Encoder, Decoder +from ldm.modules.distributions.distributions import DiagonalGaussianDistribution + +from ldm.util import instantiate_from_config +from ldm.modules.ema import LitEma + + +class AutoencoderKL(pl.LightningModule): + def __init__(self, + ddconfig, + lossconfig, + embed_dim, + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None, + monitor=None, + ema_decay=None, + learn_logvar=False + ): + super().__init__() + self.learn_logvar = learn_logvar + self.image_key = image_key + self.encoder = Encoder(**ddconfig) + self.decoder = Decoder(**ddconfig) + self.loss = instantiate_from_config(lossconfig) + assert ddconfig["double_z"] + self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1) + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) + self.embed_dim = embed_dim + if colorize_nlabels is not None: + assert type(colorize_nlabels)==int + self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) + if monitor is not None: + self.monitor = monitor + + self.use_ema = ema_decay is not None + if self.use_ema: + self.ema_decay = ema_decay + assert 0. < ema_decay < 1. + self.model_ema = LitEma(self, decay=ema_decay) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list()): + sd = torch.load(path, map_location="cpu")["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + self.load_state_dict(sd, strict=False) + print(f"Restored from {path}") + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.parameters()) + self.model_ema.copy_to(self) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self) + + def encode(self, x): + h = self.encoder(x) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + return posterior + + def decode(self, z): + z = self.post_quant_conv(z) + dec = self.decoder(z) + return dec + + def forward(self, input, sample_posterior=True): + posterior = self.encode(input) + if sample_posterior: + z = posterior.sample() + else: + z = posterior.mode() + dec = self.decode(z) + return dec, posterior + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float() + return x + + def training_step(self, batch, batch_idx, optimizer_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + + if optimizer_idx == 0: + # train encoder+decoder+logvar + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return aeloss + + if optimizer_idx == 1: + # train the discriminator + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + + self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return discloss + + def validation_step(self, batch, batch_idx): + log_dict = self._validation_step(batch, batch_idx) + with self.ema_scope(): + log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema") + return log_dict + + def _validation_step(self, batch, batch_idx, postfix=""): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"]) + self.log_dict(log_dict_ae) + self.log_dict(log_dict_disc) + return self.log_dict + + def configure_optimizers(self): + lr = self.learning_rate + ae_params_list = list(self.encoder.parameters()) + list(self.decoder.parameters()) + list( + self.quant_conv.parameters()) + list(self.post_quant_conv.parameters()) + if self.learn_logvar: + print(f"{self.__class__.__name__}: Learning logvar") + ae_params_list.append(self.loss.logvar) + opt_ae = torch.optim.Adam(ae_params_list, + lr=lr, betas=(0.5, 0.9)) + opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), + lr=lr, betas=(0.5, 0.9)) + return [opt_ae, opt_disc], [] + + def get_last_layer(self): + return self.decoder.conv_out.weight + + @torch.no_grad() + def log_images(self, batch, only_inputs=False, log_ema=False, **kwargs): + log = dict() + x = self.get_input(batch, self.image_key) + x = x.to(self.device) + if not only_inputs: + xrec, posterior = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec.shape[1] > 3 + x = self.to_rgb(x) + xrec = self.to_rgb(xrec) + log["samples"] = self.decode(torch.randn_like(posterior.sample())) + log["reconstructions"] = xrec + if log_ema or self.use_ema: + with self.ema_scope(): + xrec_ema, posterior_ema = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec_ema.shape[1] > 3 + xrec_ema = self.to_rgb(xrec_ema) + log["samples_ema"] = self.decode(torch.randn_like(posterior_ema.sample())) + log["reconstructions_ema"] = xrec_ema + log["inputs"] = x + return log + + def to_rgb(self, x): + assert self.image_key == "segmentation" + if not hasattr(self, "colorize"): + self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) + x = F.conv2d(x, weight=self.colorize) + x = 2.*(x-x.min())/(x.max()-x.min()) - 1. + return x + + +class IdentityFirstStage(torch.nn.Module): + def __init__(self, *args, vq_interface=False, **kwargs): + self.vq_interface = vq_interface + super().__init__() + + def encode(self, x, *args, **kwargs): + return x + + def decode(self, x, *args, **kwargs): + return x + + def quantize(self, x, *args, **kwargs): + if self.vq_interface: + return x, None, [None, None, None] + return x + + def forward(self, x, *args, **kwargs): + return x + diff --git a/ldm/models/diffusion/__init__.py b/ldm/models/diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/models/diffusion/ddim.py b/ldm/models/diffusion/ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..27ead0ea914c64c747b64e690662899fb3801144 --- /dev/null +++ b/ldm/models/diffusion/ddim.py @@ -0,0 +1,336 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + ucg_schedule=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): ctmp = ctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + elif isinstance(conditioning, list): + for ctmp in conditioning: + if ctmp.shape[0] != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ucg_schedule=ucg_schedule + ) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None, + ucg_schedule=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + if ucg_schedule is not None: + assert len(ucg_schedule) == len(time_range) + unconditional_guidance_scale = ucg_schedule[i] + + outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold) + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + model_output = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + if isinstance(c, dict): + assert isinstance(unconditional_conditioning, dict) + c_in = dict() + for k in c: + if isinstance(c[k], list): + c_in[k] = [torch.cat([ + unconditional_conditioning[k][i], + c[k][i]]) for i in range(len(c[k]))] + else: + c_in[k] = torch.cat([ + unconditional_conditioning[k], + c[k]]) + elif isinstance(c, list): + c_in = list() + assert isinstance(unconditional_conditioning, list) + for i in range(len(c)): + c_in.append(torch.cat([unconditional_conditioning[i], c[i]])) + else: + c_in = torch.cat([unconditional_conditioning, c]) + model_uncond, model_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond) + + if self.model.parameterization == "v": + e_t = self.model.predict_eps_from_z_and_v(x, t, model_output) + else: + e_t = model_output + + if score_corrector is not None: + assert self.model.parameterization == "eps", 'not implemented' + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if self.model.parameterization != "v": + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output) + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + + if dynamic_threshold is not None: + raise NotImplementedError() + + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + @torch.no_grad() + def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None, + unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None): + num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0] + + assert t_enc <= num_reference_steps + num_steps = t_enc + + if use_original_steps: + alphas_next = self.alphas_cumprod[:num_steps] + alphas = self.alphas_cumprod_prev[:num_steps] + else: + alphas_next = self.ddim_alphas[:num_steps] + alphas = torch.tensor(self.ddim_alphas_prev[:num_steps]) + + x_next = x0 + intermediates = [] + inter_steps = [] + for i in tqdm(range(num_steps), desc='Encoding Image'): + t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long) + if unconditional_guidance_scale == 1.: + noise_pred = self.model.apply_model(x_next, t, c) + else: + assert unconditional_conditioning is not None + e_t_uncond, noise_pred = torch.chunk( + self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)), + torch.cat((unconditional_conditioning, c))), 2) + noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond) + + xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next + weighted_noise_pred = alphas_next[i].sqrt() * ( + (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred + x_next = xt_weighted + weighted_noise_pred + if return_intermediates and i % ( + num_steps // return_intermediates) == 0 and i < num_steps - 1: + intermediates.append(x_next) + inter_steps.append(i) + elif return_intermediates and i >= num_steps - 2: + intermediates.append(x_next) + inter_steps.append(i) + if callback: callback(i) + + out = {'x_encoded': x_next, 'intermediate_steps': inter_steps} + if return_intermediates: + out.update({'intermediates': intermediates}) + return x_next, out + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False, callback=None): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + if callback: callback(i) + return x_dec \ No newline at end of file diff --git a/ldm/models/diffusion/ddpm.py b/ldm/models/diffusion/ddpm.py new file mode 100644 index 0000000000000000000000000000000000000000..f71a44af48c8cba8e97849b7e6813b3e6f9fe83c --- /dev/null +++ b/ldm/models/diffusion/ddpm.py @@ -0,0 +1,1797 @@ +""" +wild mixture of +https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py +https://github.com/CompVis/taming-transformers +-- merci +""" + +import torch +import torch.nn as nn +import numpy as np +import pytorch_lightning as pl +from torch.optim.lr_scheduler import LambdaLR +from einops import rearrange, repeat +from contextlib import contextmanager, nullcontext +from functools import partial +import itertools +from tqdm import tqdm +from torchvision.utils import make_grid +from pytorch_lightning.utilities.distributed import rank_zero_only +from omegaconf import ListConfig + +from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config +from ldm.modules.ema import LitEma +from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution +from ldm.models.autoencoder import IdentityFirstStage, AutoencoderKL +from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like +from ldm.models.diffusion.ddim import DDIMSampler + + +__conditioning_keys__ = {'concat': 'c_concat', + 'crossattn': 'c_crossattn', + 'adm': 'y'} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def uniform_on_device(r1, r2, shape, device): + return (r1 - r2) * torch.rand(*shape, device=device) + r2 + + +class DDPM(pl.LightningModule): + # classic DDPM with Gaussian diffusion, in image space + def __init__(self, + unet_config, + timesteps=1000, + beta_schedule="linear", + loss_type="l2", + ckpt_path=None, + ignore_keys=[], + load_only_unet=False, + monitor="val/loss", + use_ema=True, + first_stage_key="image", + image_size=256, + channels=3, + log_every_t=100, + clip_denoised=True, + linear_start=1e-4, + linear_end=2e-2, + cosine_s=8e-3, + given_betas=None, + original_elbo_weight=0., + v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta + l_simple_weight=1., + conditioning_key=None, + parameterization="eps", # all assuming fixed variance schedules + scheduler_config=None, + use_positional_encodings=False, + learn_logvar=False, + logvar_init=0., + make_it_fit=False, + ucg_training=None, + reset_ema=False, + reset_num_ema_updates=False, + ): + super().__init__() + assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"' + self.parameterization = parameterization + print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode") + self.cond_stage_model = None + self.clip_denoised = clip_denoised + self.log_every_t = log_every_t + self.first_stage_key = first_stage_key + self.image_size = image_size # try conv? + self.channels = channels + self.use_positional_encodings = use_positional_encodings + self.model = DiffusionWrapper(unet_config, conditioning_key) + count_params(self.model, verbose=True) + self.use_ema = use_ema + if self.use_ema: + self.model_ema = LitEma(self.model) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + self.use_scheduler = scheduler_config is not None + if self.use_scheduler: + self.scheduler_config = scheduler_config + + self.v_posterior = v_posterior + self.original_elbo_weight = original_elbo_weight + self.l_simple_weight = l_simple_weight + + if monitor is not None: + self.monitor = monitor + self.make_it_fit = make_it_fit + if reset_ema: assert exists(ckpt_path) + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet) + if reset_ema: + assert self.use_ema + print(f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.") + self.model_ema = LitEma(self.model) + if reset_num_ema_updates: + print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ") + assert self.use_ema + self.model_ema.reset_num_updates() + + self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps, + linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) + + self.loss_type = loss_type + + self.learn_logvar = learn_logvar + logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,)) + if self.learn_logvar: + self.logvar = nn.Parameter(self.logvar, requires_grad=True) + else: + self.register_buffer('logvar', logvar) + + self.ucg_training = ucg_training or dict() + if self.ucg_training: + self.ucg_prng = np.random.RandomState() + + def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if exists(given_betas): + betas = given_betas + else: + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + # calculations for posterior q(x_{t-1} | x_t, x_0) + posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / ( + 1. - alphas_cumprod) + self.v_posterior * betas + # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) + self.register_buffer('posterior_variance', to_torch(posterior_variance)) + # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain + self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) + self.register_buffer('posterior_mean_coef1', to_torch( + betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) + self.register_buffer('posterior_mean_coef2', to_torch( + (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) + + if self.parameterization == "eps": + lvlb_weights = self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)) + elif self.parameterization == "x0": + lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod)) + elif self.parameterization == "v": + lvlb_weights = torch.ones_like(self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))) + else: + raise NotImplementedError("mu not supported") + lvlb_weights[0] = lvlb_weights[1] + self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) + assert not torch.isnan(self.lvlb_weights).all() + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.model.parameters()) + self.model_ema.copy_to(self.model) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.model.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + @torch.no_grad() + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + if self.make_it_fit: + n_params = len([name for name, _ in + itertools.chain(self.named_parameters(), + self.named_buffers())]) + for name, param in tqdm( + itertools.chain(self.named_parameters(), + self.named_buffers()), + desc="Fitting old weights to new weights", + total=n_params + ): + if not name in sd: + continue + old_shape = sd[name].shape + new_shape = param.shape + assert len(old_shape) == len(new_shape) + if len(new_shape) > 2: + # we only modify first two axes + assert new_shape[2:] == old_shape[2:] + # assumes first axis corresponds to output dim + if not new_shape == old_shape: + new_param = param.clone() + old_param = sd[name] + if len(new_shape) == 1: + for i in range(new_param.shape[0]): + new_param[i] = old_param[i % old_shape[0]] + elif len(new_shape) >= 2: + for i in range(new_param.shape[0]): + for j in range(new_param.shape[1]): + new_param[i, j] = old_param[i % old_shape[0], j % old_shape[1]] + + n_used_old = torch.ones(old_shape[1]) + for j in range(new_param.shape[1]): + n_used_old[j % old_shape[1]] += 1 + n_used_new = torch.zeros(new_shape[1]) + for j in range(new_param.shape[1]): + n_used_new[j] = n_used_old[j % old_shape[1]] + + n_used_new = n_used_new[None, :] + while len(n_used_new.shape) < len(new_shape): + n_used_new = n_used_new.unsqueeze(-1) + new_param /= n_used_new + + sd[name] = new_param + + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys:\n {missing}") + if len(unexpected) > 0: + print(f"\nUnexpected Keys:\n {unexpected}") + + def q_mean_variance(self, x_start, t): + """ + Get the distribution q(x_t | x_0). + :param x_start: the [N x C x ...] tensor of noiseless inputs. + :param t: the number of diffusion steps (minus 1). Here, 0 means one step. + :return: A tuple (mean, variance, log_variance), all of x_start's shape. + """ + mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start) + variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape) + log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape) + return mean, variance, log_variance + + def predict_start_from_noise(self, x_t, t, noise): + return ( + extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise + ) + + def predict_start_from_z_and_v(self, x_t, t, v): + # self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + # self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v + ) + + def predict_eps_from_z_and_v(self, x_t, t, v): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * x_t + ) + + def q_posterior(self, x_start, x_t, t): + posterior_mean = ( + extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + + extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t + ) + posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape) + posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape) + return posterior_mean, posterior_variance, posterior_log_variance_clipped + + def p_mean_variance(self, x, t, clip_denoised: bool): + model_out = self.model(x, t) + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + if clip_denoised: + x_recon.clamp_(-1., 1.) + + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, t, clip_denoised=True, repeat_noise=False): + b, *_, device = *x.shape, x.device + model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised) + noise = noise_like(x.shape, device, repeat_noise) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def p_sample_loop(self, shape, return_intermediates=False): + device = self.betas.device + b = shape[0] + img = torch.randn(shape, device=device) + intermediates = [img] + for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps): + img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long), + clip_denoised=self.clip_denoised) + if i % self.log_every_t == 0 or i == self.num_timesteps - 1: + intermediates.append(img) + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, batch_size=16, return_intermediates=False): + image_size = self.image_size + channels = self.channels + return self.p_sample_loop((batch_size, channels, image_size, image_size), + return_intermediates=return_intermediates) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def get_v(self, x, noise, t): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x + ) + + def get_loss(self, pred, target, mean=True): + if self.loss_type == 'l1': + loss = (target - pred).abs() + if mean: + loss = loss.mean() + elif self.loss_type == 'l2': + if mean: + loss = torch.nn.functional.mse_loss(target, pred) + else: + loss = torch.nn.functional.mse_loss(target, pred, reduction='none') + else: + raise NotImplementedError("unknown loss type '{loss_type}'") + + return loss + + def p_losses(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_out = self.model(x_noisy, t) + + loss_dict = {} + if self.parameterization == "eps": + target = noise + elif self.parameterization == "x0": + target = x_start + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError(f"Parameterization {self.parameterization} not yet supported") + + loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3]) + + log_prefix = 'train' if self.training else 'val' + + loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()}) + loss_simple = loss.mean() * self.l_simple_weight + + loss_vlb = (self.lvlb_weights[t] * loss).mean() + loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb}) + + loss = loss_simple + self.original_elbo_weight * loss_vlb + + loss_dict.update({f'{log_prefix}/loss': loss}) + + return loss, loss_dict + + def forward(self, x, *args, **kwargs): + # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size + # assert h == img_size and w == img_size, f'height and width of image must be {img_size}' + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + return self.p_losses(x, t, *args, **kwargs) + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = rearrange(x, 'b h w c -> b c h w') + x = x.to(memory_format=torch.contiguous_format).float() + return x + + def shared_step(self, batch): + x = self.get_input(batch, self.first_stage_key) + loss, loss_dict = self(x) + return loss, loss_dict + + def training_step(self, batch, batch_idx): + for k in self.ucg_training: + p = self.ucg_training[k]["p"] + val = self.ucg_training[k]["val"] + if val is None: + val = "" + for i in range(len(batch[k])): + if self.ucg_prng.choice(2, p=[1 - p, p]): + batch[k][i] = val + + loss, loss_dict = self.shared_step(batch) + + self.log_dict(loss_dict, prog_bar=True, + logger=True, on_step=True, on_epoch=True) + + self.log("global_step", self.global_step, + prog_bar=True, logger=True, on_step=True, on_epoch=False) + + if self.use_scheduler: + lr = self.optimizers().param_groups[0]['lr'] + self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False) + + return loss + + @torch.no_grad() + def validation_step(self, batch, batch_idx): + _, loss_dict_no_ema = self.shared_step(batch) + with self.ema_scope(): + _, loss_dict_ema = self.shared_step(batch) + loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema} + self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self.model) + + def _get_rows_from_list(self, samples): + n_imgs_per_row = len(samples) + denoise_grid = rearrange(samples, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs): + log = dict() + x = self.get_input(batch, self.first_stage_key) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + x = x.to(self.device)[:N] + log["inputs"] = x + + # get diffusion row + diffusion_row = list() + x_start = x[:n_row] + + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(x_start) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + diffusion_row.append(x_noisy) + + log["diffusion_row"] = self._get_rows_from_list(diffusion_row) + + if sample: + # get denoise row + with self.ema_scope("Plotting"): + samples, denoise_row = self.sample(batch_size=N, return_intermediates=True) + + log["samples"] = samples + log["denoise_row"] = self._get_rows_from_list(denoise_row) + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + if self.learn_logvar: + params = params + [self.logvar] + opt = torch.optim.AdamW(params, lr=lr) + return opt + + +class LatentDiffusion(DDPM): + """main class""" + + def __init__(self, + first_stage_config, + cond_stage_config, + num_timesteps_cond=None, + cond_stage_key="image", + cond_stage_trainable=False, + concat_mode=True, + cond_stage_forward=None, + conditioning_key=None, + scale_factor=1.0, + scale_by_std=False, + force_null_conditioning=False, + *args, **kwargs): + self.force_null_conditioning = force_null_conditioning + self.num_timesteps_cond = default(num_timesteps_cond, 1) + self.scale_by_std = scale_by_std + assert self.num_timesteps_cond <= kwargs['timesteps'] + # for backwards compatibility after implementation of DiffusionWrapper + if conditioning_key is None: + conditioning_key = 'concat' if concat_mode else 'crossattn' + if cond_stage_config == '__is_unconditional__' and not self.force_null_conditioning: + conditioning_key = None + ckpt_path = kwargs.pop("ckpt_path", None) + reset_ema = kwargs.pop("reset_ema", False) + reset_num_ema_updates = kwargs.pop("reset_num_ema_updates", False) + ignore_keys = kwargs.pop("ignore_keys", []) + super().__init__(conditioning_key=conditioning_key, *args, **kwargs) + self.concat_mode = concat_mode + self.cond_stage_trainable = cond_stage_trainable + self.cond_stage_key = cond_stage_key + try: + self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1 + except: + self.num_downs = 0 + if not scale_by_std: + self.scale_factor = scale_factor + else: + self.register_buffer('scale_factor', torch.tensor(scale_factor)) + self.instantiate_first_stage(first_stage_config) + self.instantiate_cond_stage(cond_stage_config) + self.cond_stage_forward = cond_stage_forward + self.clip_denoised = False + self.bbox_tokenizer = None + + self.restarted_from_ckpt = False + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys) + self.restarted_from_ckpt = True + if reset_ema: + assert self.use_ema + print( + f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.") + self.model_ema = LitEma(self.model) + if reset_num_ema_updates: + print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ") + assert self.use_ema + self.model_ema.reset_num_updates() + + def make_cond_schedule(self, ): + self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long) + ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long() + self.cond_ids[:self.num_timesteps_cond] = ids + + @rank_zero_only + @torch.no_grad() + def on_train_batch_start(self, batch, batch_idx, dataloader_idx): + # only for very first batch + if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt: + assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously' + # set rescale weight to 1./std of encodings + print("### USING STD-RESCALING ###") + x = super().get_input(batch, self.first_stage_key) + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + del self.scale_factor + self.register_buffer('scale_factor', 1. / z.flatten().std()) + print(f"setting self.scale_factor to {self.scale_factor}") + print("### USING STD-RESCALING ###") + + def register_schedule(self, + given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s) + + self.shorten_cond_schedule = self.num_timesteps_cond > 1 + if self.shorten_cond_schedule: + self.make_cond_schedule() + + def instantiate_first_stage(self, config): + model = instantiate_from_config(config) + self.first_stage_model = model.eval() + self.first_stage_model.train = disabled_train + for param in self.first_stage_model.parameters(): + param.requires_grad = False + + def instantiate_cond_stage(self, config): + if not self.cond_stage_trainable: + if config == "__is_first_stage__": + print("Using first stage also as cond stage.") + self.cond_stage_model = self.first_stage_model + elif config == "__is_unconditional__": + print(f"Training {self.__class__.__name__} as an unconditional model.") + self.cond_stage_model = None + # self.be_unconditional = True + else: + model = instantiate_from_config(config) + self.cond_stage_model = model.eval() + self.cond_stage_model.train = disabled_train + for param in self.cond_stage_model.parameters(): + param.requires_grad = False + else: + assert config != '__is_first_stage__' + assert config != '__is_unconditional__' + model = instantiate_from_config(config) + self.cond_stage_model = model + + def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False): + denoise_row = [] + for zd in tqdm(samples, desc=desc): + denoise_row.append(self.decode_first_stage(zd.to(self.device), + force_not_quantize=force_no_decoder_quantization)) + n_imgs_per_row = len(denoise_row) + denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W + denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + def get_first_stage_encoding(self, encoder_posterior): + if isinstance(encoder_posterior, DiagonalGaussianDistribution): + z = encoder_posterior.sample() + elif isinstance(encoder_posterior, torch.Tensor): + z = encoder_posterior + else: + raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") + return self.scale_factor * z + + def get_learned_conditioning(self, c): + if self.cond_stage_forward is None: + if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): + c = self.cond_stage_model.encode(c) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + else: + c = self.cond_stage_model(c) + else: + assert hasattr(self.cond_stage_model, self.cond_stage_forward) + c = getattr(self.cond_stage_model, self.cond_stage_forward)(c) + return c + + def meshgrid(self, h, w): + y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1) + x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1) + + arr = torch.cat([y, x], dim=-1) + return arr + + def delta_border(self, h, w): + """ + :param h: height + :param w: width + :return: normalized distance to image border, + wtith min distance = 0 at border and max dist = 0.5 at image center + """ + lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2) + arr = self.meshgrid(h, w) / lower_right_corner + dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0] + dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0] + edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0] + return edge_dist + + def get_weighting(self, h, w, Ly, Lx, device): + weighting = self.delta_border(h, w) + weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"], + self.split_input_params["clip_max_weight"], ) + weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device) + + if self.split_input_params["tie_braker"]: + L_weighting = self.delta_border(Ly, Lx) + L_weighting = torch.clip(L_weighting, + self.split_input_params["clip_min_tie_weight"], + self.split_input_params["clip_max_tie_weight"]) + + L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device) + weighting = weighting * L_weighting + return weighting + + def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code + """ + :param x: img of size (bs, c, h, w) + :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1]) + """ + bs, nc, h, w = x.shape + + # number of crops in image + Ly = (h - kernel_size[0]) // stride[0] + 1 + Lx = (w - kernel_size[1]) // stride[1] + 1 + + if uf == 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params) + + weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx)) + + elif uf > 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf), + dilation=1, padding=0, + stride=(stride[0] * uf, stride[1] * uf)) + fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx)) + + elif df > 1 and uf == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df), + dilation=1, padding=0, + stride=(stride[0] // df, stride[1] // df)) + fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx)) + + else: + raise NotImplementedError + + return fold, unfold, normalization, weighting + + @torch.no_grad() + def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False, + cond_key=None, return_original_cond=False, bs=None, return_x=False): + x = super().get_input(batch, k) + if bs is not None: + x = x[:bs] + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + + if self.model.conditioning_key is not None and not self.force_null_conditioning: + if cond_key is None: + cond_key = self.cond_stage_key + if cond_key != self.first_stage_key: + if cond_key in ['caption', 'coordinates_bbox', "txt"]: + xc = batch[cond_key] + elif cond_key in ['class_label', 'cls']: + xc = batch + else: + xc = super().get_input(batch, cond_key).to(self.device) + else: + xc = x + if not self.cond_stage_trainable or force_c_encode: + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + c = self.get_learned_conditioning(xc.to(self.device)) + else: + c = xc + if bs is not None: + c = c[:bs] + + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + ckey = __conditioning_keys__[self.model.conditioning_key] + c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y} + + else: + c = None + xc = None + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + c = {'pos_x': pos_x, 'pos_y': pos_y} + out = [z, c] + if return_first_stage_outputs: + xrec = self.decode_first_stage(z) + out.extend([x, xrec]) + if return_x: + out.extend([x]) + if return_original_cond: + out.append(xc) + return out + + @torch.no_grad() + def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False): + if predict_cids: + if z.dim() == 4: + z = torch.argmax(z.exp(), dim=1).long() + z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None) + z = rearrange(z, 'b h w c -> b c h w').contiguous() + + z = 1. / self.scale_factor * z + return self.first_stage_model.decode(z) + + @torch.no_grad() + def encode_first_stage(self, x): + return self.first_stage_model.encode(x) + + def shared_step(self, batch, **kwargs): + x, c = self.get_input(batch, self.first_stage_key) + loss = self(x, c) + return loss + + def forward(self, x, c, *args, **kwargs): + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + if self.model.conditioning_key is not None: + assert c is not None + if self.cond_stage_trainable: + c = self.get_learned_conditioning(c) + if self.shorten_cond_schedule: # TODO: drop this option + tc = self.cond_ids[t].to(self.device) + c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float())) + return self.p_losses(x, c, t, *args, **kwargs) + + def apply_model(self, x_noisy, t, cond, return_ids=False): + if isinstance(cond, dict): + # hybrid case, cond is expected to be a dict + pass + else: + if not isinstance(cond, list): + cond = [cond] + key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn' + cond = {key: cond} + + x_recon = self.model(x_noisy, t, **cond) + + if isinstance(x_recon, tuple) and not return_ids: + return x_recon[0] + else: + return x_recon + + def _predict_eps_from_xstart(self, x_t, t, pred_xstart): + return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \ + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) + + def _prior_bpd(self, x_start): + """ + Get the prior KL term for the variational lower-bound, measured in + bits-per-dim. + This term can't be optimized, as it only depends on the encoder. + :param x_start: the [N x C x ...] tensor of inputs. + :return: a batch of [N] KL values (in bits), one per batch element. + """ + batch_size = x_start.shape[0] + t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device) + qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t) + kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0) + return mean_flat(kl_prior) / np.log(2.0) + + def p_losses(self, x_start, cond, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_output = self.apply_model(x_noisy, t, cond) + + loss_dict = {} + prefix = 'train' if self.training else 'val' + + if self.parameterization == "x0": + target = x_start + elif self.parameterization == "eps": + target = noise + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError() + + loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3]) + loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()}) + + logvar_t = self.logvar[t].to(self.device) + loss = loss_simple / torch.exp(logvar_t) + logvar_t + # loss = loss_simple / torch.exp(self.logvar) + self.logvar + if self.learn_logvar: + loss_dict.update({f'{prefix}/loss_gamma': loss.mean()}) + loss_dict.update({'logvar': self.logvar.data.mean()}) + + loss = self.l_simple_weight * loss.mean() + + loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3)) + loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean() + loss_dict.update({f'{prefix}/loss_vlb': loss_vlb}) + loss += (self.original_elbo_weight * loss_vlb) + loss_dict.update({f'{prefix}/loss': loss}) + + return loss, loss_dict + + def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False, + return_x0=False, score_corrector=None, corrector_kwargs=None): + t_in = t + model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids) + + if score_corrector is not None: + assert self.parameterization == "eps" + model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs) + + if return_codebook_ids: + model_out, logits = model_out + + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + else: + raise NotImplementedError() + + if clip_denoised: + x_recon.clamp_(-1., 1.) + if quantize_denoised: + x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon) + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + if return_codebook_ids: + return model_mean, posterior_variance, posterior_log_variance, logits + elif return_x0: + return model_mean, posterior_variance, posterior_log_variance, x_recon + else: + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, + return_codebook_ids=False, quantize_denoised=False, return_x0=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None): + b, *_, device = *x.shape, x.device + outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, + return_codebook_ids=return_codebook_ids, + quantize_denoised=quantize_denoised, + return_x0=return_x0, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + if return_codebook_ids: + raise DeprecationWarning("Support dropped.") + model_mean, _, model_log_variance, logits = outputs + elif return_x0: + model_mean, _, model_log_variance, x0 = outputs + else: + model_mean, _, model_log_variance = outputs + + noise = noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + + if return_codebook_ids: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1) + if return_x0: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0 + else: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False, + img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0., + score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None, + log_every_t=None): + if not log_every_t: + log_every_t = self.log_every_t + timesteps = self.num_timesteps + if batch_size is not None: + b = batch_size if batch_size is not None else shape[0] + shape = [batch_size] + list(shape) + else: + b = batch_size = shape[0] + if x_T is None: + img = torch.randn(shape, device=self.device) + else: + img = x_T + intermediates = [] + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation', + total=timesteps) if verbose else reversed( + range(0, timesteps)) + if type(temperature) == float: + temperature = [temperature] * timesteps + + for i in iterator: + ts = torch.full((b,), i, device=self.device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img, x0_partial = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised, return_x0=True, + temperature=temperature[i], noise_dropout=noise_dropout, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + if mask is not None: + assert x0 is not None + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(x0_partial) + if callback: callback(i) + if img_callback: img_callback(img, i) + return img, intermediates + + @torch.no_grad() + def p_sample_loop(self, cond, shape, return_intermediates=False, + x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, start_T=None, + log_every_t=None): + + if not log_every_t: + log_every_t = self.log_every_t + device = self.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + intermediates = [img] + if timesteps is None: + timesteps = self.num_timesteps + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed( + range(0, timesteps)) + + if mask is not None: + assert x0 is not None + assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match + + for i in iterator: + ts = torch.full((b,), i, device=device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised) + if mask is not None: + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(img) + if callback: callback(i) + if img_callback: img_callback(img, i) + + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None, + verbose=True, timesteps=None, quantize_denoised=False, + mask=None, x0=None, shape=None, **kwargs): + if shape is None: + shape = (batch_size, self.channels, self.image_size, self.image_size) + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + return self.p_sample_loop(cond, + shape, + return_intermediates=return_intermediates, x_T=x_T, + verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised, + mask=mask, x0=x0) + + @torch.no_grad() + def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): + if ddim: + ddim_sampler = DDIMSampler(self) + shape = (self.channels, self.image_size, self.image_size) + samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, + shape, cond, verbose=False, **kwargs) + + else: + samples, intermediates = self.sample(cond=cond, batch_size=batch_size, + return_intermediates=True, **kwargs) + + return samples, intermediates + + @torch.no_grad() + def get_unconditional_conditioning(self, batch_size, null_label=None): + if null_label is not None: + xc = null_label + if isinstance(xc, ListConfig): + xc = list(xc) + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + if hasattr(xc, "to"): + xc = xc.to(self.device) + c = self.get_learned_conditioning(xc) + else: + if self.cond_stage_key in ["class_label", "cls"]: + xc = self.cond_stage_model.get_unconditional_conditioning(batch_size, device=self.device) + return self.get_learned_conditioning(xc) + else: + raise NotImplementedError("todo") + if isinstance(c, list): # in case the encoder gives us a list + for i in range(len(c)): + c[i] = repeat(c[i], '1 ... -> b ...', b=batch_size).to(self.device) + else: + c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device) + return c + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=50, ddim_eta=0., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, + return_first_stage_outputs=True, + force_c_encode=True, + return_original_cond=True, + bs=N) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', "cls"]: + try: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + except KeyError: + # probably no "human_label" in batch + pass + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance( + self.first_stage_model, IdentityFirstStage): + # also display when quantizing x0 while sampling + with ema_scope("Plotting Quantized Denoised"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + quantize_denoised=True) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True, + # quantize_denoised=True) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_x0_quantized"] = x_samples + + if unconditional_guidance_scale > 1.0: + uc = self.get_unconditional_conditioning(N, unconditional_guidance_label) + if self.model.conditioning_key == "crossattn-adm": + uc = {"c_crossattn": [uc], "c_adm": c["c_adm"]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if inpaint: + # make a simple center square + b, h, w = z.shape[0], z.shape[2], z.shape[3] + mask = torch.ones(N, h, w).to(self.device) + # zeros will be filled in + mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0. + mask = mask[:, None, ...] + with ema_scope("Plotting Inpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_inpainting"] = x_samples + log["mask"] = mask + + # outpaint + mask = 1. - mask + with ema_scope("Plotting Outpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_outpainting"] = x_samples + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + if self.cond_stage_trainable: + print(f"{self.__class__.__name__}: Also optimizing conditioner params!") + params = params + list(self.cond_stage_model.parameters()) + if self.learn_logvar: + print('Diffusion model optimizing logvar') + params.append(self.logvar) + opt = torch.optim.AdamW(params, lr=lr) + if self.use_scheduler: + assert 'target' in self.scheduler_config + scheduler = instantiate_from_config(self.scheduler_config) + + print("Setting up LambdaLR scheduler...") + scheduler = [ + { + 'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule), + 'interval': 'step', + 'frequency': 1 + }] + return [opt], scheduler + return opt + + @torch.no_grad() + def to_rgb(self, x): + x = x.float() + if not hasattr(self, "colorize"): + self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x) + x = nn.functional.conv2d(x, weight=self.colorize) + x = 2. * (x - x.min()) / (x.max() - x.min()) - 1. + return x + + +class DiffusionWrapper(pl.LightningModule): + def __init__(self, diff_model_config, conditioning_key): + super().__init__() + self.sequential_cross_attn = diff_model_config.pop("sequential_crossattn", False) + self.diffusion_model = instantiate_from_config(diff_model_config) + self.conditioning_key = conditioning_key + assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm', 'hybrid-adm', 'crossattn-adm'] + + def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None): + if self.conditioning_key is None: + out = self.diffusion_model(x, t) + elif self.conditioning_key == 'concat': + xc = torch.cat([x] + c_concat, dim=1) + out = self.diffusion_model(xc, t) + elif self.conditioning_key == 'crossattn': + if not self.sequential_cross_attn: + cc = torch.cat(c_crossattn, 1) + else: + cc = c_crossattn + out = self.diffusion_model(x, t, context=cc) + elif self.conditioning_key == 'hybrid': + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc) + elif self.conditioning_key == 'hybrid-adm': + assert c_adm is not None + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, y=c_adm) + elif self.conditioning_key == 'crossattn-adm': + assert c_adm is not None + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(x, t, context=cc, y=c_adm) + elif self.conditioning_key == 'adm': + cc = c_crossattn[0] + out = self.diffusion_model(x, t, y=cc) + else: + raise NotImplementedError() + + return out + + +class LatentUpscaleDiffusion(LatentDiffusion): + def __init__(self, *args, low_scale_config, low_scale_key="LR", noise_level_key=None, **kwargs): + super().__init__(*args, **kwargs) + # assumes that neither the cond_stage nor the low_scale_model contain trainable params + assert not self.cond_stage_trainable + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + self.noise_level_key = noise_level_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False): + if not log_mode: + z, c = super().get_input(batch, k, force_c_encode=True, bs=bs) + else: + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + x_low = batch[self.low_scale_key][:bs] + x_low = rearrange(x_low, 'b h w c -> b c h w') + x_low = x_low.to(memory_format=torch.contiguous_format).float() + zx, noise_level = self.low_scale_model(x_low) + if self.noise_level_key is not None: + # get noise level from batch instead, e.g. when extracting a custom noise level for bsr + raise NotImplementedError('TODO') + + all_conds = {"c_concat": [zx], "c_crossattn": [c], "c_adm": noise_level} + if log_mode: + # TODO: maybe disable if too expensive + x_low_rec = self.low_scale_model.decode(zx) + return z, all_conds, x, xrec, xc, x_low, x_low_rec, noise_level + return z, all_conds + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True, + unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc, x_low, x_low_rec, noise_level = self.get_input(batch, self.first_stage_key, bs=N, + log_mode=True) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + log["x_lr"] = x_low + log[f"x_lr_rec_@noise_levels{'-'.join(map(lambda x: str(x), list(noise_level.cpu().numpy())))}"] = x_low_rec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label) + # TODO explore better "unconditional" choices for the other keys + # maybe guide away from empty text label and highest noise level and maximally degraded zx? + uc = dict() + for k in c: + if k == "c_crossattn": + assert isinstance(c[k], list) and len(c[k]) == 1 + uc[k] = [uc_tmp] + elif k == "c_adm": # todo: only run with text-based guidance? + assert isinstance(c[k], torch.Tensor) + #uc[k] = torch.ones_like(c[k]) * self.low_scale_model.max_noise_level + uc[k] = c[k] + elif isinstance(c[k], list): + uc[k] = [c[k][i] for i in range(len(c[k]))] + else: + uc[k] = c[k] + + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + return log + + +class LatentFinetuneDiffusion(LatentDiffusion): + """ + Basis for different finetunas, such as inpainting or depth2image + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys: tuple, + finetune_keys=("model.diffusion_model.input_blocks.0.0.weight", + "model_ema.diffusion_modelinput_blocks00weight" + ), + keep_finetune_dims=4, + # if model was trained without concat mode before and we would like to keep these channels + c_concat_log_start=None, # to log reconstruction of c_concat codes + c_concat_log_end=None, + *args, **kwargs + ): + ckpt_path = kwargs.pop("ckpt_path", None) + ignore_keys = kwargs.pop("ignore_keys", list()) + super().__init__(*args, **kwargs) + self.finetune_keys = finetune_keys + self.concat_keys = concat_keys + self.keep_dims = keep_finetune_dims + self.c_concat_log_start = c_concat_log_start + self.c_concat_log_end = c_concat_log_end + if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint' + if exists(ckpt_path): + self.init_from_ckpt(ckpt_path, ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + + # make it explicit, finetune by including extra input channels + if exists(self.finetune_keys) and k in self.finetune_keys: + new_entry = None + for name, param in self.named_parameters(): + if name in self.finetune_keys: + print( + f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only") + new_entry = torch.zeros_like(param) # zero init + assert exists(new_entry), 'did not find matching parameter to modify' + new_entry[:, :self.keep_dims, ...] = sd[k] + sd[k] = new_entry + + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys: {missing}") + if len(unexpected) > 0: + print(f"Unexpected Keys: {unexpected}") + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True) + c_cat, c = c["c_concat"][0], c["c_crossattn"][0] + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if not (self.c_concat_log_start is None and self.c_concat_log_end is None): + log["c_concat_decoded"] = self.decode_first_stage(c_cat[:, self.c_concat_log_start:self.c_concat_log_end]) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label) + uc_cat = c_cat + uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc_full, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + return log + + +class LatentInpaintDiffusion(LatentFinetuneDiffusion): + """ + can either run as pure inpainting model (only concat mode) or with mixed conditionings, + e.g. mask as concat and text via cross-attn. + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys=("mask", "masked_image"), + masked_image_key="masked_image", + *args, **kwargs + ): + super().__init__(concat_keys, *args, **kwargs) + self.masked_image_key = masked_image_key + assert self.masked_image_key in concat_keys + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + c_cat = list() + for ck in self.concat_keys: + cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + bchw = z.shape + if ck != self.masked_image_key: + cc = torch.nn.functional.interpolate(cc, size=bchw[-2:]) + else: + cc = self.get_first_stage_encoding(self.encode_first_stage(cc)) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super(LatentInpaintDiffusion, self).log_images(*args, **kwargs) + log["masked_image"] = rearrange(args[0]["masked_image"], + 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + return log + + +class LatentDepth2ImageDiffusion(LatentFinetuneDiffusion): + """ + condition on monocular depth estimation + """ + + def __init__(self, depth_stage_config, concat_keys=("midas_in",), *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.depth_model = instantiate_from_config(depth_stage_config) + self.depth_stage_key = concat_keys[0] + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for depth2img' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + c_cat = list() + for ck in self.concat_keys: + cc = batch[ck] + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + cc = self.depth_model(cc) + cc = torch.nn.functional.interpolate( + cc, + size=z.shape[2:], + mode="bicubic", + align_corners=False, + ) + + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + cc = 2. * (cc - depth_min) / (depth_max - depth_min + 0.001) - 1. + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + depth = self.depth_model(args[0][self.depth_stage_key]) + depth_min, depth_max = torch.amin(depth, dim=[1, 2, 3], keepdim=True), \ + torch.amax(depth, dim=[1, 2, 3], keepdim=True) + log["depth"] = 2. * (depth - depth_min) / (depth_max - depth_min) - 1. + return log + + +class LatentUpscaleFinetuneDiffusion(LatentFinetuneDiffusion): + """ + condition on low-res image (and optionally on some spatial noise augmentation) + """ + def __init__(self, concat_keys=("lr",), reshuffle_patch_size=None, + low_scale_config=None, low_scale_key=None, *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.reshuffle_patch_size = reshuffle_patch_size + self.low_scale_model = None + if low_scale_config is not None: + print("Initializing a low-scale model") + assert exists(low_scale_key) + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for upscaling-ft' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + # optionally make spatial noise_level here + c_cat = list() + noise_level = None + for ck in self.concat_keys: + cc = batch[ck] + cc = rearrange(cc, 'b h w c -> b c h w') + if exists(self.reshuffle_patch_size): + assert isinstance(self.reshuffle_patch_size, int) + cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w', + p1=self.reshuffle_patch_size, p2=self.reshuffle_patch_size) + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + if exists(self.low_scale_model) and ck == self.low_scale_key: + cc, noise_level = self.low_scale_model(cc) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + if exists(noise_level): + all_conds = {"c_concat": [c_cat], "c_crossattn": [c], "c_adm": noise_level} + else: + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + log["lr"] = rearrange(args[0]["lr"], 'b h w c -> b c h w') + return log diff --git a/ldm/models/diffusion/dpm_solver/__init__.py b/ldm/models/diffusion/dpm_solver/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7427f38c07530afbab79154ea8aaf88c4bf70a08 --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/__init__.py @@ -0,0 +1 @@ +from .sampler import DPMSolverSampler \ No newline at end of file diff --git a/ldm/models/diffusion/dpm_solver/dpm_solver.py b/ldm/models/diffusion/dpm_solver/dpm_solver.py new file mode 100644 index 0000000000000000000000000000000000000000..095e5ba3ce0b1aa7f4b3f1e2e5d8fff7cfe6dc8c --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/dpm_solver.py @@ -0,0 +1,1154 @@ +import torch +import torch.nn.functional as F +import math +from tqdm import tqdm + + +class NoiseScheduleVP: + def __init__( + self, + schedule='discrete', + betas=None, + alphas_cumprod=None, + continuous_beta_0=0.1, + continuous_beta_1=20., + ): + """Create a wrapper class for the forward SDE (VP type). + *** + Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t. + We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images. + *** + The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ). + We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper). + Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have: + log_alpha_t = self.marginal_log_mean_coeff(t) + sigma_t = self.marginal_std(t) + lambda_t = self.marginal_lambda(t) + Moreover, as lambda(t) is an invertible function, we also support its inverse function: + t = self.inverse_lambda(lambda_t) + =============================================================== + We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]). + 1. For discrete-time DPMs: + For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by: + t_i = (i + 1) / N + e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1. + We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3. + Args: + betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details) + alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details) + Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`. + **Important**: Please pay special attention for the args for `alphas_cumprod`: + The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that + q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ). + Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have + alpha_{t_n} = \sqrt{\hat{alpha_n}}, + and + log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}). + 2. For continuous-time DPMs: + We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise + schedule are the default settings in DDPM and improved-DDPM: + Args: + beta_min: A `float` number. The smallest beta for the linear schedule. + beta_max: A `float` number. The largest beta for the linear schedule. + cosine_s: A `float` number. The hyperparameter in the cosine schedule. + cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule. + T: A `float` number. The ending time of the forward process. + =============================================================== + Args: + schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs, + 'linear' or 'cosine' for continuous-time DPMs. + Returns: + A wrapper object of the forward SDE (VP type). + + =============================================================== + Example: + # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', betas=betas) + # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod) + # For continuous-time DPMs (VPSDE), linear schedule: + >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.) + """ + + if schedule not in ['discrete', 'linear', 'cosine']: + raise ValueError( + "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format( + schedule)) + + self.schedule = schedule + if schedule == 'discrete': + if betas is not None: + log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0) + else: + assert alphas_cumprod is not None + log_alphas = 0.5 * torch.log(alphas_cumprod) + self.total_N = len(log_alphas) + self.T = 1. + self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1)) + self.log_alpha_array = log_alphas.reshape((1, -1,)) + else: + self.total_N = 1000 + self.beta_0 = continuous_beta_0 + self.beta_1 = continuous_beta_1 + self.cosine_s = 0.008 + self.cosine_beta_max = 999. + self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.)) + self.schedule = schedule + if schedule == 'cosine': + # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T. + # Note that T = 0.9946 may be not the optimal setting. However, we find it works well. + self.T = 0.9946 + else: + self.T = 1. + + def marginal_log_mean_coeff(self, t): + """ + Compute log(alpha_t) of a given continuous-time label t in [0, T]. + """ + if self.schedule == 'discrete': + return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device), + self.log_alpha_array.to(t.device)).reshape((-1)) + elif self.schedule == 'linear': + return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0 + elif self.schedule == 'cosine': + log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.)) + log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0 + return log_alpha_t + + def marginal_alpha(self, t): + """ + Compute alpha_t of a given continuous-time label t in [0, T]. + """ + return torch.exp(self.marginal_log_mean_coeff(t)) + + def marginal_std(self, t): + """ + Compute sigma_t of a given continuous-time label t in [0, T]. + """ + return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t))) + + def marginal_lambda(self, t): + """ + Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T]. + """ + log_mean_coeff = self.marginal_log_mean_coeff(t) + log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff)) + return log_mean_coeff - log_std + + def inverse_lambda(self, lamb): + """ + Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t. + """ + if self.schedule == 'linear': + tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + Delta = self.beta_0 ** 2 + tmp + return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0) + elif self.schedule == 'discrete': + log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb) + t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]), + torch.flip(self.t_array.to(lamb.device), [1])) + return t.reshape((-1,)) + else: + log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + t = t_fn(log_alpha) + return t + + +def model_wrapper( + model, + noise_schedule, + model_type="noise", + model_kwargs={}, + guidance_type="uncond", + condition=None, + unconditional_condition=None, + guidance_scale=1., + classifier_fn=None, + classifier_kwargs={}, +): + """Create a wrapper function for the noise prediction model. + DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to + firstly wrap the model function to a noise prediction model that accepts the continuous time as the input. + We support four types of the diffusion model by setting `model_type`: + 1. "noise": noise prediction model. (Trained by predicting noise). + 2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0). + 3. "v": velocity prediction model. (Trained by predicting the velocity). + The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2]. + [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models." + arXiv preprint arXiv:2202.00512 (2022). + [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models." + arXiv preprint arXiv:2210.02303 (2022). + + 4. "score": marginal score function. (Trained by denoising score matching). + Note that the score function and the noise prediction model follows a simple relationship: + ``` + noise(x_t, t) = -sigma_t * score(x_t, t) + ``` + We support three types of guided sampling by DPMs by setting `guidance_type`: + 1. "uncond": unconditional sampling by DPMs. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + 2. "classifier": classifier guidance sampling [3] by DPMs and another classifier. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + The input `classifier_fn` has the following format: + `` + classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond) + `` + [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis," + in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794. + 3. "classifier-free": classifier-free guidance sampling by conditional DPMs. + The input `model` has the following format: + `` + model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score + `` + And if cond == `unconditional_condition`, the model output is the unconditional DPM output. + [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance." + arXiv preprint arXiv:2207.12598 (2022). + + The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999) + or continuous-time labels (i.e. epsilon to T). + We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise: + `` + def model_fn(x, t_continuous) -> noise: + t_input = get_model_input_time(t_continuous) + return noise_pred(model, x, t_input, **model_kwargs) + `` + where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver. + =============================================================== + Args: + model: A diffusion model with the corresponding format described above. + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + model_type: A `str`. The parameterization type of the diffusion model. + "noise" or "x_start" or "v" or "score". + model_kwargs: A `dict`. A dict for the other inputs of the model function. + guidance_type: A `str`. The type of the guidance for sampling. + "uncond" or "classifier" or "classifier-free". + condition: A pytorch tensor. The condition for the guided sampling. + Only used for "classifier" or "classifier-free" guidance type. + unconditional_condition: A pytorch tensor. The condition for the unconditional sampling. + Only used for "classifier-free" guidance type. + guidance_scale: A `float`. The scale for the guided sampling. + classifier_fn: A classifier function. Only used for the classifier guidance. + classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function. + Returns: + A noise prediction model that accepts the noised data and the continuous time as the inputs. + """ + + def get_model_input_time(t_continuous): + """ + Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time. + For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N]. + For continuous-time DPMs, we just use `t_continuous`. + """ + if noise_schedule.schedule == 'discrete': + return (t_continuous - 1. / noise_schedule.total_N) * 1000. + else: + return t_continuous + + def noise_pred_fn(x, t_continuous, cond=None): + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + t_input = get_model_input_time(t_continuous) + if cond is None: + output = model(x, t_input, **model_kwargs) + else: + output = model(x, t_input, cond, **model_kwargs) + if model_type == "noise": + return output + elif model_type == "x_start": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims) + elif model_type == "v": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x + elif model_type == "score": + sigma_t = noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return -expand_dims(sigma_t, dims) * output + + def cond_grad_fn(x, t_input): + """ + Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t). + """ + with torch.enable_grad(): + x_in = x.detach().requires_grad_(True) + log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs) + return torch.autograd.grad(log_prob.sum(), x_in)[0] + + def model_fn(x, t_continuous): + """ + The noise predicition model function that is used for DPM-Solver. + """ + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + if guidance_type == "uncond": + return noise_pred_fn(x, t_continuous) + elif guidance_type == "classifier": + assert classifier_fn is not None + t_input = get_model_input_time(t_continuous) + cond_grad = cond_grad_fn(x, t_input) + sigma_t = noise_schedule.marginal_std(t_continuous) + noise = noise_pred_fn(x, t_continuous) + return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad + elif guidance_type == "classifier-free": + if guidance_scale == 1. or unconditional_condition is None: + return noise_pred_fn(x, t_continuous, cond=condition) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t_continuous] * 2) + c_in = torch.cat([unconditional_condition, condition]) + noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2) + return noise_uncond + guidance_scale * (noise - noise_uncond) + + assert model_type in ["noise", "x_start", "v"] + assert guidance_type in ["uncond", "classifier", "classifier-free"] + return model_fn + + +class DPM_Solver: + def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.): + """Construct a DPM-Solver. + We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0"). + If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver). + If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++). + In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True. + The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales. + Args: + model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]): + `` + def model_fn(x, t_continuous): + return noise + `` + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model. + thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1]. + max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding. + + [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b. + """ + self.model = model_fn + self.noise_schedule = noise_schedule + self.predict_x0 = predict_x0 + self.thresholding = thresholding + self.max_val = max_val + + def noise_prediction_fn(self, x, t): + """ + Return the noise prediction model. + """ + return self.model(x, t) + + def data_prediction_fn(self, x, t): + """ + Return the data prediction model (with thresholding). + """ + noise = self.noise_prediction_fn(x, t) + dims = x.dim() + alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t) + x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims) + if self.thresholding: + p = 0.995 # A hyperparameter in the paper of "Imagen" [1]. + s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1) + s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims) + x0 = torch.clamp(x0, -s, s) / s + return x0 + + def model_fn(self, x, t): + """ + Convert the model to the noise prediction model or the data prediction model. + """ + if self.predict_x0: + return self.data_prediction_fn(x, t) + else: + return self.noise_prediction_fn(x, t) + + def get_time_steps(self, skip_type, t_T, t_0, N, device): + """Compute the intermediate time steps for sampling. + Args: + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + N: A `int`. The total number of the spacing of the time steps. + device: A torch device. + Returns: + A pytorch tensor of the time steps, with the shape (N + 1,). + """ + if skip_type == 'logSNR': + lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device)) + lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device)) + logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device) + return self.noise_schedule.inverse_lambda(logSNR_steps) + elif skip_type == 'time_uniform': + return torch.linspace(t_T, t_0, N + 1).to(device) + elif skip_type == 'time_quadratic': + t_order = 2 + t = torch.linspace(t_T ** (1. / t_order), t_0 ** (1. / t_order), N + 1).pow(t_order).to(device) + return t + else: + raise ValueError( + "Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type)) + + def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device): + """ + Get the order of each step for sampling by the singlestep DPM-Solver. + We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast". + Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is: + - If order == 1: + We take `steps` of DPM-Solver-1 (i.e. DDIM). + - If order == 2: + - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of DPM-Solver-2. + - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If order == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2. + ============================================ + Args: + order: A `int`. The max order for the solver (2 or 3). + steps: A `int`. The total number of function evaluations (NFE). + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + device: A torch device. + Returns: + orders: A list of the solver order of each step. + """ + if order == 3: + K = steps // 3 + 1 + if steps % 3 == 0: + orders = [3, ] * (K - 2) + [2, 1] + elif steps % 3 == 1: + orders = [3, ] * (K - 1) + [1] + else: + orders = [3, ] * (K - 1) + [2] + elif order == 2: + if steps % 2 == 0: + K = steps // 2 + orders = [2, ] * K + else: + K = steps // 2 + 1 + orders = [2, ] * (K - 1) + [1] + elif order == 1: + K = 1 + orders = [1, ] * steps + else: + raise ValueError("'order' must be '1' or '2' or '3'.") + if skip_type == 'logSNR': + # To reproduce the results in DPM-Solver paper + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device) + else: + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[ + torch.cumsum(torch.tensor([0, ] + orders)).to(device)] + return timesteps_outer, orders + + def denoise_to_zero_fn(self, x, s): + """ + Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization. + """ + return self.data_prediction_fn(x, s) + + def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False): + """ + DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + if self.predict_x0: + phi_1 = torch.expm1(-h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + else: + phi_1 = torch.expm1(h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + + def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, + solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-2 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the second-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 0.5 + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + s1 = ns.inverse_lambda(lambda_s1) + log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff( + s1), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t) + alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_1 = torch.expm1(-h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * ( + model_s1 - model_s) + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_1 = torch.expm1(h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s) + ) + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1} + else: + return x_t + + def singlestep_dpm_solver_third_update(self, x, s, t, r1=1. / 3., r2=2. / 3., model_s=None, model_s1=None, + return_intermediate=False, solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-3 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`). + If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 1. / 3. + if r2 is None: + r2 = 2. / 3. + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + lambda_s2 = lambda_s + r2 * h + s1 = ns.inverse_lambda(lambda_s1) + s2 = ns.inverse_lambda(lambda_s2) + log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff( + s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std( + s2), ns.marginal_std(t) + alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_12 = torch.expm1(-r2 * h) + phi_1 = torch.expm1(-h) + phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1. + phi_2 = phi_1 / h + 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(sigma_s2 / sigma_s, dims) * x + - expand_dims(alpha_s2 * phi_12, dims) * model_s + + r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + expand_dims(alpha_t * phi_2, dims) * D1 + - expand_dims(alpha_t * phi_3, dims) * D2 + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_12 = torch.expm1(r2 * h) + phi_1 = torch.expm1(h) + phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1. + phi_2 = phi_1 / h - 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x + - expand_dims(sigma_s2 * phi_12, dims) * model_s + - r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - expand_dims(sigma_t * phi_2, dims) * D1 + - expand_dims(sigma_t * phi_3, dims) * D2 + ) + + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2} + else: + return x_t + + def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"): + """ + Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + ns = self.noise_schedule + dims = x.dim() + model_prev_1, model_prev_0 = model_prev_list + t_prev_1, t_prev_0 = t_prev_list + lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda( + t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0 = h_0 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + if self.predict_x0: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0 + ) + else: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0 + ) + return x_t + + def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'): + """ + Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + model_prev_2, model_prev_1, model_prev_0 = model_prev_list + t_prev_2, t_prev_1, t_prev_0 = t_prev_list + lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda( + t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_1 = lambda_prev_1 - lambda_prev_2 + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0, r1 = h_0 / h, h_1 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2) + D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1) + D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1) + if self.predict_x0: + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1 + - expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h ** 2 - 0.5), dims) * D2 + ) + else: + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1 + - expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h ** 2 - 0.5), dims) * D2 + ) + return x_t + + def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None, + r2=None): + """ + Singlestep DPM-Solver with the order `order` from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + r1: A `float`. The hyperparameter of the second-order or third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate) + elif order == 2: + return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1) + elif order == 3: + return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1, r2=r2) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'): + """ + Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1]) + elif order == 2: + return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + elif order == 3: + return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, + solver_type='dpm_solver'): + """ + The adaptive step size solver based on singlestep DPM-Solver. + Args: + x: A pytorch tensor. The initial value at time `t_T`. + order: A `int`. The (higher) order of the solver. We only support order == 2 or 3. + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + h_init: A `float`. The initial step size (for logSNR). + atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1]. + rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05. + theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1]. + t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the + current time and `t_0` is less than `t_err`. The default setting is 1e-5. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_0: A pytorch tensor. The approximated solution at time `t_0`. + [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021. + """ + ns = self.noise_schedule + s = t_T * torch.ones((x.shape[0],)).to(x) + lambda_s = ns.marginal_lambda(s) + lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x)) + h = h_init * torch.ones_like(s).to(x) + x_prev = x + nfe = 0 + if order == 2: + r1 = 0.5 + lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + solver_type=solver_type, + **kwargs) + elif order == 3: + r1, r2 = 1. / 3., 2. / 3. + lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + return_intermediate=True, + solver_type=solver_type) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2, + solver_type=solver_type, + **kwargs) + else: + raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order)) + while torch.abs((s - t_0)).mean() > t_err: + t = ns.inverse_lambda(lambda_s + h) + x_lower, lower_noise_kwargs = lower_update(x, s, t) + x_higher = higher_update(x, s, t, **lower_noise_kwargs) + delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev))) + norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)) + E = norm_fn((x_higher - x_lower) / delta).max() + if torch.all(E <= 1.): + x = x_higher + s = t + x_prev = x_lower + lambda_s = ns.marginal_lambda(s) + h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s) + nfe += order + print('adaptive solver nfe', nfe) + return x + + def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform', + method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver', + atol=0.0078, rtol=0.05, + ): + """ + Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`. + ===================================================== + We support the following algorithms for both noise prediction model and data prediction model: + - 'singlestep': + Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver. + We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps). + The total number of function evaluations (NFE) == `steps`. + Given a fixed NFE == `steps`, the sampling procedure is: + - If `order` == 1: + - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2. + - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If `order` == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2. + - 'multistep': + Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`. + We initialize the first `order` values by lower order multistep solvers. + Given a fixed NFE == `steps`, the sampling procedure is: + Denote K = steps. + - If `order` == 1: + - We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2. + - If `order` == 3: + - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3. + - 'singlestep_fixed': + Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3). + We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE. + - 'adaptive': + Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper). + We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`. + You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs + (NFE) and the sample quality. + - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2. + - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3. + ===================================================== + Some advices for choosing the algorithm: + - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs: + Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3, + skip_type='time_uniform', method='singlestep') + - For **guided sampling with large guidance scale** by DPMs: + Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2, + skip_type='time_uniform', method='multistep') + We support three types of `skip_type`: + - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images** + - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**. + - 'time_quadratic': quadratic time for the time steps. + ===================================================== + Args: + x: A pytorch tensor. The initial value at time `t_start` + e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution. + steps: A `int`. The total number of function evaluations (NFE). + t_start: A `float`. The starting time of the sampling. + If `T` is None, we use self.noise_schedule.T (default is 1.0). + t_end: A `float`. The ending time of the sampling. + If `t_end` is None, we use 1. / self.noise_schedule.total_N. + e.g. if total_N == 1000, we have `t_end` == 1e-3. + For discrete-time DPMs: + - We recommend `t_end` == 1. / self.noise_schedule.total_N. + For continuous-time DPMs: + - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15. + order: A `int`. The order of DPM-Solver. + skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'. + method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'. + denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step. + Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1). + This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and + score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID + for diffusion models sampling by diffusion SDEs for low-resolutional images + (such as CIFAR-10). However, we observed that such trick does not matter for + high-resolutional images. As it needs an additional NFE, we do not recommend + it for high-resolutional images. + lower_order_final: A `bool`. Whether to use lower order solvers at the final steps. + Only valid for `method=multistep` and `steps < 15`. We empirically find that + this trick is a key to stabilizing the sampling by DPM-Solver with very few steps + (especially for steps <= 10). So we recommend to set it to be `True`. + solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`. + atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + Returns: + x_end: A pytorch tensor. The approximated solution at time `t_end`. + """ + t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end + t_T = self.noise_schedule.T if t_start is None else t_start + device = x.device + if method == 'adaptive': + with torch.no_grad(): + x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, + solver_type=solver_type) + elif method == 'multistep': + assert steps >= order + timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device) + assert timesteps.shape[0] - 1 == steps + with torch.no_grad(): + vec_t = timesteps[0].expand((x.shape[0])) + model_prev_list = [self.model_fn(x, vec_t)] + t_prev_list = [vec_t] + # Init the first `order` values by lower order multistep DPM-Solver. + for init_order in tqdm(range(1, order), desc="DPM init order"): + vec_t = timesteps[init_order].expand(x.shape[0]) + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order, + solver_type=solver_type) + model_prev_list.append(self.model_fn(x, vec_t)) + t_prev_list.append(vec_t) + # Compute the remaining values by `order`-th order multistep DPM-Solver. + for step in tqdm(range(order, steps + 1), desc="DPM multistep"): + vec_t = timesteps[step].expand(x.shape[0]) + if lower_order_final and steps < 15: + step_order = min(order, steps + 1 - step) + else: + step_order = order + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order, + solver_type=solver_type) + for i in range(order - 1): + t_prev_list[i] = t_prev_list[i + 1] + model_prev_list[i] = model_prev_list[i + 1] + t_prev_list[-1] = vec_t + # We do not need to evaluate the final model value. + if step < steps: + model_prev_list[-1] = self.model_fn(x, vec_t) + elif method in ['singlestep', 'singlestep_fixed']: + if method == 'singlestep': + timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order, + skip_type=skip_type, + t_T=t_T, t_0=t_0, + device=device) + elif method == 'singlestep_fixed': + K = steps // order + orders = [order, ] * K + timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device) + for i, order in enumerate(orders): + t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1] + timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(), + N=order, device=device) + lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner) + vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0]) + h = lambda_inner[-1] - lambda_inner[0] + r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h + r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h + x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2) + if denoise_to_zero: + x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0) + return x + + +############################################################# +# other utility functions +############################################################# + +def interpolate_fn(x, xp, yp): + """ + A piecewise linear function y = f(x), using xp and yp as keypoints. + We implement f(x) in a differentiable way (i.e. applicable for autograd). + The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.) + Args: + x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver). + xp: PyTorch tensor with shape [C, K], where K is the number of keypoints. + yp: PyTorch tensor with shape [C, K]. + Returns: + The function values f(x), with shape [N, C]. + """ + N, K = x.shape[0], xp.shape[1] + all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2) + sorted_all_x, x_indices = torch.sort(all_x, dim=2) + x_idx = torch.argmin(x_indices, dim=2) + cand_start_idx = x_idx - 1 + start_idx = torch.where( + torch.eq(x_idx, 0), + torch.tensor(1, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1) + start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2) + end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2) + start_idx2 = torch.where( + torch.eq(x_idx, 0), + torch.tensor(0, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1) + start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2) + end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2) + cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x) + return cand + + +def expand_dims(v, dims): + """ + Expand the tensor `v` to the dim `dims`. + Args: + `v`: a PyTorch tensor with shape [N]. + `dim`: a `int`. + Returns: + a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`. + """ + return v[(...,) + (None,) * (dims - 1)] \ No newline at end of file diff --git a/ldm/models/diffusion/dpm_solver/sampler.py b/ldm/models/diffusion/dpm_solver/sampler.py new file mode 100644 index 0000000000000000000000000000000000000000..7d137b8cf36718c1c58faa09f9dd919e5fb2977b --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/sampler.py @@ -0,0 +1,87 @@ +"""SAMPLING ONLY.""" +import torch + +from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver + + +MODEL_TYPES = { + "eps": "noise", + "v": "v" +} + + +class DPMSolverSampler(object): + def __init__(self, model, **kwargs): + super().__init__() + self.model = model + to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device) + self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod)) + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + + print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}') + + device = self.model.betas.device + if x_T is None: + img = torch.randn(size, device=device) + else: + img = x_T + + ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod) + + model_fn = model_wrapper( + lambda x, t, c: self.model.apply_model(x, t, c), + ns, + model_type=MODEL_TYPES[self.model.parameterization], + guidance_type="classifier-free", + condition=conditioning, + unconditional_condition=unconditional_conditioning, + guidance_scale=unconditional_guidance_scale, + ) + + dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False) + x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True) + + return x.to(device), None \ No newline at end of file diff --git a/ldm/models/diffusion/plms.py b/ldm/models/diffusion/plms.py new file mode 100644 index 0000000000000000000000000000000000000000..7002a365d27168ced0a04e9a4d83e088f8284eae --- /dev/null +++ b/ldm/models/diffusion/plms.py @@ -0,0 +1,244 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm +from functools import partial + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like +from ldm.models.diffusion.sampling_util import norm_thresholding + + +class PLMSSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + if ddim_eta != 0: + raise ValueError('ddim_eta must be 0 for PLMS') + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for PLMS sampling is {size}') + + samples, intermediates = self.plms_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ) + return samples, intermediates + + @torch.no_grad() + def plms_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running PLMS Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps) + old_eps = [] + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + old_eps=old_eps, t_next=ts_next, + dynamic_threshold=dynamic_threshold) + img, pred_x0, e_t = outs + old_eps.append(e_t) + if len(old_eps) >= 4: + old_eps.pop(0) + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + def get_model_output(x, t): + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + c_in = torch.cat([unconditional_conditioning, c]) + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + return e_t + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + + def get_x_prev_and_pred_x0(e_t, index): + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + if dynamic_threshold is not None: + pred_x0 = norm_thresholding(pred_x0, dynamic_threshold) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + e_t = get_model_output(x, t) + if len(old_eps) == 0: + # Pseudo Improved Euler (2nd order) + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index) + e_t_next = get_model_output(x_prev, t_next) + e_t_prime = (e_t + e_t_next) / 2 + elif len(old_eps) == 1: + # 2nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (3 * e_t - old_eps[-1]) / 2 + elif len(old_eps) == 2: + # 3nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12 + elif len(old_eps) >= 3: + # 4nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24 + + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) + + return x_prev, pred_x0, e_t diff --git a/ldm/models/diffusion/sampling_util.py b/ldm/models/diffusion/sampling_util.py new file mode 100644 index 0000000000000000000000000000000000000000..7eff02be6d7c54d43ee6680636ac0698dd3b3f33 --- /dev/null +++ b/ldm/models/diffusion/sampling_util.py @@ -0,0 +1,22 @@ +import torch +import numpy as np + + +def append_dims(x, target_dims): + """Appends dimensions to the end of a tensor until it has target_dims dimensions. + From https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/utils.py""" + dims_to_append = target_dims - x.ndim + if dims_to_append < 0: + raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less') + return x[(...,) + (None,) * dims_to_append] + + +def norm_thresholding(x0, value): + s = append_dims(x0.pow(2).flatten(1).mean(1).sqrt().clamp(min=value), x0.ndim) + return x0 * (value / s) + + +def spatial_norm_thresholding(x0, value): + # b c h w + s = x0.pow(2).mean(1, keepdim=True).sqrt().clamp(min=value) + return x0 * (value / s) \ No newline at end of file diff --git a/ldm/modules/attention.py b/ldm/modules/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..509cd873768f0dd75a75ab3fcdd652822b12b59f --- /dev/null +++ b/ldm/modules/attention.py @@ -0,0 +1,341 @@ +from inspect import isfunction +import math +import torch +import torch.nn.functional as F +from torch import nn, einsum +from einops import rearrange, repeat +from typing import Optional, Any + +from ldm.modules.diffusionmodules.util import checkpoint + + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + +# CrossAttn precision handling +import os +_ATTN_PRECISION = os.environ.get("ATTN_PRECISION", "fp32") + +def exists(val): + return val is not None + + +def uniq(arr): + return{el: True for el in arr}.keys() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + + +# feedforward +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def Normalize(in_channels): + return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = rearrange(q, 'b c h w -> b (h w) c') + k = rearrange(k, 'b c h w -> b c (h w)') + w_ = torch.einsum('bij,bjk->bik', q, k) + + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, 'b c h w -> b c (h w)') + w_ = rearrange(w_, 'b i j -> b j i') + h_ = torch.einsum('bij,bjk->bik', v, w_) + h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h) + h_ = self.proj_out(h_) + + return x+h_ + + +class CrossAttention(nn.Module): + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + 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(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), + nn.Dropout(dropout) + ) + + def forward(self, x, context=None, mask=None): + h = self.heads + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + # force cast to fp32 to avoid overflowing + if _ATTN_PRECISION =="fp32": + with torch.autocast(enabled=False, device_type = 'cuda'): + q, k = q.float(), k.float() + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + else: + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + + del q, k + + if exists(mask): + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', sim, v) + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + return self.to_out(out) + + +class MemoryEfficientCrossAttention(nn.Module): + # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0): + super().__init__() + print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using " + f"{heads} heads.") + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.heads = heads + self.dim_head = dim_head + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)) + self.attention_op: Optional[Any] = None + + def forward(self, x, context=None, mask=None): + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + return self.to_out(out) + + +class BasicTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention + } + def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, + disable_self_attn=False): + super().__init__() + attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax" + assert attn_mode in self.ATTENTION_MODES + attn_cls = self.ATTENTION_MODES[attn_mode] + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None) # is a self-attention if not self.disable_self_attn + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, + heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + + def forward(self, x, context=None): + return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint) + + def _forward(self, x, context=None): + x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x + x = self.attn2(self.norm2(x), context=context) + x + x = self.ff(self.norm3(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + def __init__(self, in_channels, n_heads, d_head, + depth=1, dropout=0., context_dim=None, + disable_self_attn=False, use_linear=False, + use_checkpoint=True): + super().__init__() + if exists(context_dim) and not isinstance(context_dim, list): + context_dim = [context_dim] + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = Normalize(in_channels) + if not use_linear: + self.proj_in = nn.Conv2d(in_channels, + inner_dim, + kernel_size=1, + stride=1, + padding=0) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + self.transformer_blocks = nn.ModuleList( + [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], + disable_self_attn=disable_self_attn, checkpoint=use_checkpoint) + for d in range(depth)] + ) + if not use_linear: + self.proj_out = zero_module(nn.Conv2d(inner_dim, + in_channels, + kernel_size=1, + stride=1, + padding=0)) + else: + self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) + self.use_linear = use_linear + + def forward(self, x, context=None): + # note: if no context is given, cross-attention defaults to self-attention + if not isinstance(context, list): + context = [context] + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'b c h w -> b (h w) c').contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + x = block(x, context=context[i]) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in + diff --git a/ldm/modules/diffusionmodules/__init__.py b/ldm/modules/diffusionmodules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/diffusionmodules/model.py b/ldm/modules/diffusionmodules/model.py new file mode 100644 index 0000000000000000000000000000000000000000..b089eebbe1676d8249005bb9def002ff5180715b --- /dev/null +++ b/ldm/modules/diffusionmodules/model.py @@ -0,0 +1,852 @@ +# pytorch_diffusion + derived encoder decoder +import math +import torch +import torch.nn as nn +import numpy as np +from einops import rearrange +from typing import Optional, Any + +from ldm.modules.attention import MemoryEfficientCrossAttention + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + print("No module 'xformers'. Proceeding without it.") + + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0,1,0,0)) + return emb + + +def nonlinearity(x): + # swish + return x*torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=2, + padding=0) + + def forward(self, x): + if self.with_conv: + pad = (0,1,0,1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + + +class ResnetBlock(nn.Module): + def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, + dropout, temb_channels=512): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, + out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + else: + self.nin_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x+h + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = q.reshape(b,c,h*w) + q = q.permute(0,2,1) # b,hw,c + k = k.reshape(b,c,h*w) # b,c,hw + w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b,c,h*w) + w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b,c,h,w) + + h_ = self.proj_out(h_) + + return x+h_ + +class MemoryEfficientAttnBlock(nn.Module): + """ + Uses xformers efficient implementation, + see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + Note: this is a single-head self-attention operation + """ + # + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.attention_op: Optional[Any] = None + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + B, C, H, W = q.shape + q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v)) + + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(B, t.shape[1], 1, C) + .permute(0, 2, 1, 3) + .reshape(B * 1, t.shape[1], C) + .contiguous(), + (q, k, v), + ) + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + out = ( + out.unsqueeze(0) + .reshape(B, 1, out.shape[1], C) + .permute(0, 2, 1, 3) + .reshape(B, out.shape[1], C) + ) + out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C) + out = self.proj_out(out) + return x+out + + +class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention): + def forward(self, x, context=None, mask=None): + b, c, h, w = x.shape + x = rearrange(x, 'b c h w -> b (h w) c') + out = super().forward(x, context=context, mask=mask) + out = rearrange(out, 'b (h w) c -> b c h w', h=h, w=w, c=c) + return x + out + + +def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None): + assert attn_type in ["vanilla", "vanilla-xformers", "memory-efficient-cross-attn", "linear", "none"], f'attn_type {attn_type} unknown' + if XFORMERS_IS_AVAILBLE and attn_type == "vanilla": + attn_type = "vanilla-xformers" + print(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + assert attn_kwargs is None + return AttnBlock(in_channels) + elif attn_type == "vanilla-xformers": + print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...") + return MemoryEfficientAttnBlock(in_channels) + elif type == "memory-efficient-cross-attn": + attn_kwargs["query_dim"] = in_channels + return MemoryEfficientCrossAttentionWrapper(**attn_kwargs) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + raise NotImplementedError() + + +class Model(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch*4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList([ + torch.nn.Linear(self.ch, + self.temb_ch), + torch.nn.Linear(self.temb_ch, + self.temb_ch), + ]) + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + skip_in = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + if i_block == self.num_res_blocks: + skip_in = ch*in_ch_mult[i_level] + block.append(ResnetBlock(in_channels=block_in+skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x, t=None, context=None): + #assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla", + **ignore_kwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + 2*z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # timestep embedding + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class Decoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + attn_type="vanilla", **ignorekwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,)+tuple(ch_mult) + block_in = ch*ch_mult[self.num_resolutions-1] + curr_res = resolution // 2**(self.num_resolutions-1) + self.z_shape = (1,z_channels,curr_res,curr_res) + print("Working with z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape))) + + # z to block_in + self.conv_in = torch.nn.Conv2d(z_channels, + block_in, + kernel_size=3, + stride=1, + padding=1) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, z): + #assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block](h, temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + if self.tanh_out: + h = torch.tanh(h) + return h + + +class SimpleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, *args, **kwargs): + super().__init__() + self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), + ResnetBlock(in_channels=in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=2 * in_channels, + out_channels=4 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=4 * in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + nn.Conv2d(2*in_channels, in_channels, 1), + Upsample(in_channels, with_conv=True)]) + # end + self.norm_out = Normalize(in_channels) + self.conv_out = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + for i, layer in enumerate(self.model): + if i in [1,2,3]: + x = layer(x, None) + else: + x = layer(x) + + h = self.norm_out(x) + h = nonlinearity(h) + x = self.conv_out(h) + return x + + +class UpsampleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, + ch_mult=(2,2), dropout=0.0): + super().__init__() + # upsampling + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + block_in = in_channels + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.res_blocks = nn.ModuleList() + self.upsample_blocks = nn.ModuleList() + for i_level in range(self.num_resolutions): + res_block = [] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + res_block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + self.res_blocks.append(nn.ModuleList(res_block)) + if i_level != self.num_resolutions - 1: + self.upsample_blocks.append(Upsample(block_in, True)) + curr_res = curr_res * 2 + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # upsampling + h = x + for k, i_level in enumerate(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.res_blocks[i_level][i_block](h, None) + if i_level != self.num_resolutions - 1: + h = self.upsample_blocks[k](h) + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class LatentRescaler(nn.Module): + def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2): + super().__init__() + # residual block, interpolate, residual block + self.factor = factor + self.conv_in = nn.Conv2d(in_channels, + mid_channels, + kernel_size=3, + stride=1, + padding=1) + self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + self.attn = AttnBlock(mid_channels) + self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + + self.conv_out = nn.Conv2d(mid_channels, + out_channels, + kernel_size=1, + ) + + def forward(self, x): + x = self.conv_in(x) + for block in self.res_block1: + x = block(x, None) + x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor)))) + x = self.attn(x) + for block in self.res_block2: + x = block(x, None) + x = self.conv_out(x) + return x + + +class MergedRescaleEncoder(nn.Module): + def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, + ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + intermediate_chn = ch * ch_mult[-1] + self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult, + z_channels=intermediate_chn, double_z=False, resolution=resolution, + attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv, + out_ch=None) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn, + mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth) + + def forward(self, x): + x = self.encoder(x) + x = self.rescaler(x) + return x + + +class MergedRescaleDecoder(nn.Module): + def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8), + dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + tmp_chn = z_channels*ch_mult[-1] + self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout, + resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks, + ch_mult=ch_mult, resolution=resolution, ch=ch) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn, + out_channels=tmp_chn, depth=rescale_module_depth) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Upsampler(nn.Module): + def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2): + super().__init__() + assert out_size >= in_size + num_blocks = int(np.log2(out_size//in_size))+1 + factor_up = 1.+ (out_size % in_size) + print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}") + self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels, + out_channels=in_channels) + self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2, + attn_resolutions=[], in_channels=None, ch=in_channels, + ch_mult=[ch_mult for _ in range(num_blocks)]) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Resize(nn.Module): + def __init__(self, in_channels=None, learned=False, mode="bilinear"): + super().__init__() + self.with_conv = learned + self.mode = mode + if self.with_conv: + print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode") + raise NotImplementedError() + assert in_channels is not None + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=4, + stride=2, + padding=1) + + def forward(self, x, scale_factor=1.0): + if scale_factor==1.0: + return x + else: + x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor) + return x diff --git a/ldm/modules/diffusionmodules/openaimodel.py b/ldm/modules/diffusionmodules/openaimodel.py new file mode 100644 index 0000000000000000000000000000000000000000..7df6b5abfe8eff07f0c8e8703ba8aee90d45984b --- /dev/null +++ b/ldm/modules/diffusionmodules/openaimodel.py @@ -0,0 +1,786 @@ +from abc import abstractmethod +import math + +import numpy as np +import torch as th +import torch.nn as nn +import torch.nn.functional as F + +from ldm.modules.diffusionmodules.util import ( + checkpoint, + conv_nd, + linear, + avg_pool_nd, + zero_module, + normalization, + timestep_embedding, +) +from ldm.modules.attention import SpatialTransformer +from ldm.util import exists + + +# dummy replace +def convert_module_to_f16(x): + pass + +def convert_module_to_f32(x): + pass + + +## go +class AttentionPool2d(nn.Module): + """ + Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py + """ + + def __init__( + self, + spacial_dim: int, + embed_dim: int, + num_heads_channels: int, + output_dim: int = None, + ): + super().__init__() + self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5) + self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) + self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) + self.num_heads = embed_dim // num_heads_channels + self.attention = QKVAttention(self.num_heads) + + def forward(self, x): + b, c, *_spatial = x.shape + x = x.reshape(b, c, -1) # NC(HW) + x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1) + x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1) + x = self.qkv_proj(x) + x = self.attention(x) + x = self.c_proj(x) + return x[:, :, 0] + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward(self, x, emb, context=None): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate( + x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" + ) + else: + x = F.interpolate(x, scale_factor=2, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + +class TransposedUpsample(nn.Module): + 'Learned 2x upsampling without padding' + def __init__(self, channels, out_channels=None, ks=5): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + + self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) + + def forward(self,x): + return self.up(x) + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=padding + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + return checkpoint( + self._forward, (x, emb), self.parameters(), self.use_checkpoint + ) + + + def _forward(self, x, emb): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels, + num_heads=1, + num_head_channels=-1, + use_checkpoint=False, + use_new_attention_order=False, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels) + self.qkv = conv_nd(1, channels, channels * 3, 1) + if use_new_attention_order: + # split qkv before split heads + self.attention = QKVAttention(self.num_heads) + else: + # split heads before split qkv + self.attention = QKVAttentionLegacy(self.num_heads) + + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x): + return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! + #return pt_checkpoint(self._forward, x) # pytorch + + def _forward(self, x): + b, c, *spatial = x.shape + x = x.reshape(b, c, -1) + qkv = self.qkv(self.norm(x)) + h = self.attention(qkv) + h = self.proj_out(h) + return (x + h).reshape(b, c, *spatial) + + +def count_flops_attn(model, _x, y): + """ + A counter for the `thop` package to count the operations in an + attention operation. + Meant to be used like: + macs, params = thop.profile( + model, + inputs=(inputs, timestamps), + custom_ops={QKVAttention: QKVAttention.count_flops}, + ) + """ + b, c, *spatial = y[0].shape + num_spatial = int(np.prod(spatial)) + # We perform two matmuls with the same number of ops. + # The first computes the weight matrix, the second computes + # the combination of the value vectors. + matmul_ops = 2 * b * (num_spatial ** 2) * c + model.total_ops += th.DoubleTensor([matmul_ops]) + + +class QKVAttentionLegacy(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention and splits in a different order. + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.chunk(3, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", + (q * scale).view(bs * self.n_heads, ch, length), + (k * scale).view(bs * self.n_heads, ch, length), + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param use_new_attention_order: use a different attention pattern for potentially + increased efficiency. + """ + + def __init__( + self, + image_size, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + num_classes=None, + use_checkpoint=False, + use_fp16=False, + num_heads=-1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + resblock_updown=False, + use_new_attention_order=False, + use_spatial_transformer=False, # custom transformer support + transformer_depth=1, # custom transformer support + context_dim=None, # custom transformer support + n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model + legacy=True, + disable_self_attentions=None, + num_attention_blocks=None, + disable_middle_self_attn=False, + use_linear_in_transformer=False, + ): + super().__init__() + if use_spatial_transformer: + assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' + + if context_dim is not None: + assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' + from omegaconf.listconfig import ListConfig + if type(context_dim) == ListConfig: + context_dim = list(context_dim) + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.image_size = image_size + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError("provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult") + self.num_res_blocks = num_res_blocks + if disable_self_attentions is not None: + # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) + print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set.") + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.dtype = th.float16 if use_fp16 else th.float32 + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.predict_codebook_ids = n_embed is not None + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + print("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + else: + raise ValueError() + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(self.num_res_blocks[level] + 1): + ich = input_block_chans.pop() + layers = [ + ResBlock( + ch + ich, + time_embed_dim, + dropout, + out_channels=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or i < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads_upsample, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + if level and i == self.num_res_blocks[level]: + out_ch = ch + layers.append( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + if self.predict_codebook_ids: + self.id_predictor = nn.Sequential( + normalization(ch), + conv_nd(dims, model_channels, n_embed, 1), + #nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits + ) + + def convert_to_fp16(self): + """ + Convert the torso of the model to float16. + """ + self.input_blocks.apply(convert_module_to_f16) + self.middle_block.apply(convert_module_to_f16) + self.output_blocks.apply(convert_module_to_f16) + + def convert_to_fp32(self): + """ + Convert the torso of the model to float32. + """ + self.input_blocks.apply(convert_module_to_f32) + self.middle_block.apply(convert_module_to_f32) + self.output_blocks.apply(convert_module_to_f32) + + def forward(self, x, timesteps=None, context=None, y=None,**kwargs): + """ + Apply the model to an input batch. + :param x: an [N x C x ...] Tensor of inputs. + :param timesteps: a 1-D batch of timesteps. + :param context: conditioning plugged in via crossattn + :param y: an [N] Tensor of labels, if class-conditional. + :return: an [N x C x ...] Tensor of outputs. + """ + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) + + h = x.type(self.dtype) + for module in self.input_blocks: + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, context) + h = h.type(x.dtype) + if self.predict_codebook_ids: + return self.id_predictor(h) + else: + return self.out(h) diff --git a/ldm/modules/diffusionmodules/upscaling.py b/ldm/modules/diffusionmodules/upscaling.py new file mode 100644 index 0000000000000000000000000000000000000000..03816662098ce1ffac79bd939b892e867ab91988 --- /dev/null +++ b/ldm/modules/diffusionmodules/upscaling.py @@ -0,0 +1,81 @@ +import torch +import torch.nn as nn +import numpy as np +from functools import partial + +from ldm.modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule +from ldm.util import default + + +class AbstractLowScaleModel(nn.Module): + # for concatenating a downsampled image to the latent representation + def __init__(self, noise_schedule_config=None): + super(AbstractLowScaleModel, self).__init__() + if noise_schedule_config is not None: + self.register_schedule(**noise_schedule_config) + + def register_schedule(self, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def forward(self, x): + return x, None + + def decode(self, x): + return x + + +class SimpleImageConcat(AbstractLowScaleModel): + # no noise level conditioning + def __init__(self): + super(SimpleImageConcat, self).__init__(noise_schedule_config=None) + self.max_noise_level = 0 + + def forward(self, x): + # fix to constant noise level + return x, torch.zeros(x.shape[0], device=x.device).long() + + +class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel): + def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False): + super().__init__(noise_schedule_config=noise_schedule_config) + self.max_noise_level = max_noise_level + + def forward(self, x, noise_level=None): + if noise_level is None: + noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long() + else: + assert isinstance(noise_level, torch.Tensor) + z = self.q_sample(x, noise_level) + return z, noise_level + + + diff --git a/ldm/modules/diffusionmodules/util.py b/ldm/modules/diffusionmodules/util.py new file mode 100644 index 0000000000000000000000000000000000000000..637363dfe34799e70cfdbcd11445212df9d9ca1f --- /dev/null +++ b/ldm/modules/diffusionmodules/util.py @@ -0,0 +1,270 @@ +# adopted from +# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# and +# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +# and +# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py +# +# thanks! + + +import os +import math +import torch +import torch.nn as nn +import numpy as np +from einops import repeat + +from ldm.util import instantiate_from_config + + +def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if schedule == "linear": + betas = ( + torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2 + ) + + elif schedule == "cosine": + timesteps = ( + torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s + ) + alphas = timesteps / (1 + cosine_s) * np.pi / 2 + alphas = torch.cos(alphas).pow(2) + alphas = alphas / alphas[0] + betas = 1 - alphas[1:] / alphas[:-1] + betas = np.clip(betas, a_min=0, a_max=0.999) + + elif schedule == "sqrt_linear": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) + elif schedule == "sqrt": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5 + else: + raise ValueError(f"schedule '{schedule}' unknown.") + return betas.numpy() + + +def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): + if ddim_discr_method == 'uniform': + c = num_ddpm_timesteps // num_ddim_timesteps + ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c))) + elif ddim_discr_method == 'quad': + ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int) + else: + raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"') + + # assert ddim_timesteps.shape[0] == num_ddim_timesteps + # add one to get the final alpha values right (the ones from first scale to data during sampling) + steps_out = ddim_timesteps + 1 + if verbose: + print(f'Selected timesteps for ddim sampler: {steps_out}') + return steps_out + + +def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): + # select alphas for computing the variance schedule + alphas = alphacums[ddim_timesteps] + alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) + + # according the the formula provided in https://arxiv.org/abs/2010.02502 + sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) + if verbose: + print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') + print(f'For the chosen value of eta, which is {eta}, ' + f'this results in the following sigma_t schedule for ddim sampler {sigmas}') + return sigmas, alphas, alphas_prev + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + + +def extract_into_tensor(a, t, x_shape): + b, *_ = t.shape + out = a.gather(-1, t) + return out.reshape(b, *((1,) * (len(x_shape) - 1))) + + +def checkpoint(func, inputs, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. + :param func: the function to evaluate. + :param inputs: the argument sequence to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + args = tuple(inputs) + tuple(params) + return CheckpointFunction.apply(func, len(inputs), *args) + else: + return func(*inputs) + + +class CheckpointFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, run_function, length, *args): + ctx.run_function = run_function + ctx.input_tensors = list(args[:length]) + ctx.input_params = list(args[length:]) + ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(), + "dtype": torch.get_autocast_gpu_dtype(), + "cache_enabled": torch.is_autocast_cache_enabled()} + with torch.no_grad(): + output_tensors = ctx.run_function(*ctx.input_tensors) + return output_tensors + + @staticmethod + def backward(ctx, *output_grads): + ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors] + with torch.enable_grad(), \ + torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs): + # Fixes a bug where the first op in run_function modifies the + # Tensor storage in place, which is not allowed for detach()'d + # Tensors. + shallow_copies = [x.view_as(x) for x in ctx.input_tensors] + output_tensors = ctx.run_function(*shallow_copies) + input_grads = torch.autograd.grad( + output_tensors, + ctx.input_tensors + ctx.input_params, + output_grads, + allow_unused=True, + ) + del ctx.input_tensors + del ctx.input_params + del output_tensors + return (None, None) + input_grads + + +def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): + """ + 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. + """ + if not repeat_only: + 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) + else: + embedding = repeat(timesteps, 'b -> b d', d=dim) + return embedding + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def normalization(channels): + """ + Make a standard normalization layer. + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNorm32(32, channels) + + +# PyTorch 1.7 has SiLU, but we support PyTorch 1.5. +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + + +class GroupNorm32(nn.GroupNorm): + def forward(self, x): + return super().forward(x.float()).type(x.dtype) + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +class HybridConditioner(nn.Module): + + def __init__(self, c_concat_config, c_crossattn_config): + super().__init__() + self.concat_conditioner = instantiate_from_config(c_concat_config) + self.crossattn_conditioner = instantiate_from_config(c_crossattn_config) + + def forward(self, c_concat, c_crossattn): + c_concat = self.concat_conditioner(c_concat) + c_crossattn = self.crossattn_conditioner(c_crossattn) + return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]} + + +def noise_like(shape, device, repeat=False): + repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1))) + noise = lambda: torch.randn(shape, device=device) + return repeat_noise() if repeat else noise() \ No newline at end of file diff --git a/ldm/modules/distributions/__init__.py b/ldm/modules/distributions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/distributions/distributions.py b/ldm/modules/distributions/distributions.py new file mode 100644 index 0000000000000000000000000000000000000000..f2b8ef901130efc171aa69742ca0244d94d3f2e9 --- /dev/null +++ b/ldm/modules/distributions/distributions.py @@ -0,0 +1,92 @@ +import torch +import numpy as np + + +class AbstractDistribution: + def sample(self): + raise NotImplementedError() + + def mode(self): + raise NotImplementedError() + + +class DiracDistribution(AbstractDistribution): + def __init__(self, value): + self.value = value + + def sample(self): + return self.value + + def mode(self): + return self.value + + +class DiagonalGaussianDistribution(object): + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) + + def sample(self): + x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device) + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.]) + else: + if other is None: + return 0.5 * torch.sum(torch.pow(self.mean, 2) + + self.var - 1.0 - self.logvar, + dim=[1, 2, 3]) + else: + return 0.5 * torch.sum( + torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var - 1.0 - self.logvar + other.logvar, + dim=[1, 2, 3]) + + def nll(self, sample, dims=[1,2,3]): + if self.deterministic: + return torch.Tensor([0.]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum( + logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims) + + def mode(self): + return self.mean + + +def normal_kl(mean1, logvar1, mean2, logvar2): + """ + source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12 + Compute the KL divergence between two gaussians. + Shapes are automatically broadcasted, so batches can be compared to + scalars, among other use cases. + """ + tensor = None + for obj in (mean1, logvar1, mean2, logvar2): + if isinstance(obj, torch.Tensor): + tensor = obj + break + assert tensor is not None, "at least one argument must be a Tensor" + + # Force variances to be Tensors. Broadcasting helps convert scalars to + # Tensors, but it does not work for torch.exp(). + logvar1, logvar2 = [ + x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) + for x in (logvar1, logvar2) + ] + + return 0.5 * ( + -1.0 + + logvar2 + - logvar1 + + torch.exp(logvar1 - logvar2) + + ((mean1 - mean2) ** 2) * torch.exp(-logvar2) + ) diff --git a/ldm/modules/ema.py b/ldm/modules/ema.py new file mode 100644 index 0000000000000000000000000000000000000000..bded25019b9bcbcd0260f0b8185f8c7859ca58c4 --- /dev/null +++ b/ldm/modules/ema.py @@ -0,0 +1,80 @@ +import torch +from torch import nn + + +class LitEma(nn.Module): + def __init__(self, model, decay=0.9999, use_num_upates=True): + super().__init__() + if decay < 0.0 or decay > 1.0: + raise ValueError('Decay must be between 0 and 1') + + self.m_name2s_name = {} + self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates + else torch.tensor(-1, dtype=torch.int)) + + for name, p in model.named_parameters(): + if p.requires_grad: + # remove as '.'-character is not allowed in buffers + s_name = name.replace('.', '') + self.m_name2s_name.update({name: s_name}) + self.register_buffer(s_name, p.clone().detach().data) + + self.collected_params = [] + + def reset_num_updates(self): + del self.num_updates + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int)) + + def forward(self, model): + decay = self.decay + + if self.num_updates >= 0: + self.num_updates += 1 + decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates)) + + one_minus_decay = 1.0 - decay + + with torch.no_grad(): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + + for key in m_param: + if m_param[key].requires_grad: + sname = self.m_name2s_name[key] + shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) + shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) + else: + assert not key in self.m_name2s_name + + def copy_to(self, model): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + for key in m_param: + if m_param[key].requires_grad: + m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) + else: + assert not key in self.m_name2s_name + + def store(self, parameters): + """ + Save the current parameters for restoring later. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + temporarily stored. + """ + self.collected_params = [param.clone() for param in parameters] + + def restore(self, parameters): + """ + Restore the parameters stored with the `store` method. + Useful to validate the model with EMA parameters without affecting the + original optimization process. Store the parameters before the + `copy_to` method. After validation (or model saving), use this to + restore the former parameters. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. + """ + for c_param, param in zip(self.collected_params, parameters): + param.data.copy_(c_param.data) diff --git a/ldm/modules/encoders/__init__.py b/ldm/modules/encoders/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/encoders/modules.py b/ldm/modules/encoders/modules.py new file mode 100644 index 0000000000000000000000000000000000000000..4edd5496b9e668ea72a5be39db9cca94b6a42f9b --- /dev/null +++ b/ldm/modules/encoders/modules.py @@ -0,0 +1,213 @@ +import torch +import torch.nn as nn +from torch.utils.checkpoint import checkpoint + +from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel + +import open_clip +from ldm.util import default, count_params + + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class IdentityEncoder(AbstractEncoder): + + def encode(self, x): + return x + + +class ClassEmbedder(nn.Module): + def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1): + super().__init__() + self.key = key + self.embedding = nn.Embedding(n_classes, embed_dim) + self.n_classes = n_classes + self.ucg_rate = ucg_rate + + def forward(self, batch, key=None, disable_dropout=False): + if key is None: + key = self.key + # this is for use in crossattn + c = batch[key][:, None] + if self.ucg_rate > 0. and not disable_dropout: + mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate) + c = mask * c + (1-mask) * torch.ones_like(c)*(self.n_classes-1) + c = c.long() + c = self.embedding(c) + return c + + def get_unconditional_conditioning(self, bs, device="cuda"): + uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) + uc = torch.ones((bs,), device=device) * uc_class + uc = {self.key: uc} + return uc + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +class FrozenT5Embedder(AbstractEncoder): + """Uses the T5 transformer encoder for text""" + def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = T5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length # TODO: typical value? + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + #self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens) + + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenCLIPEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for text (from huggingface)""" + LAYERS = [ + "last", + "pooled", + "hidden" + ] + def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, + freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32 + super().__init__() + assert layer in self.LAYERS + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.transformer = CLIPTextModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + self.layer_idx = layer_idx + if layer == "hidden": + assert layer_idx is not None + assert 0 <= abs(layer_idx) <= 12 + + def freeze(self): + self.transformer = self.transformer.eval() + #self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer=="hidden") + if self.layer == "last": + z = outputs.last_hidden_state + elif self.layer == "pooled": + z = outputs.pooler_output[:, None, :] + else: + z = outputs.hidden_states[self.layer_idx] + return z + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP transformer encoder for text + """ + LAYERS = [ + #"pooled", + "last", + "penultimate" + ] + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="last"): + super().__init__() + assert layer in self.LAYERS + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + tokens = open_clip.tokenize(text) + z = self.encode_with_transformer(tokens.to(self.device)) + return z + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.model.ln_final(x) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask = None): + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - self.layer_idx: + break + if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting(): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + return x + + def encode(self, text): + return self(text) + + +class FrozenCLIPT5Encoder(AbstractEncoder): + def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda", + clip_max_length=77, t5_max_length=77): + super().__init__() + self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length) + self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) + print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder)*1.e-6:.2f} M parameters, " + f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder)*1.e-6:.2f} M params.") + + def encode(self, text): + return self(text) + + def forward(self, text): + clip_z = self.clip_encoder.encode(text) + t5_z = self.t5_encoder.encode(text) + return [clip_z, t5_z] + + diff --git a/ldm/modules/image_degradation/__init__.py b/ldm/modules/image_degradation/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7836cada81f90ded99c58d5942eea4c3477f58fc --- /dev/null +++ b/ldm/modules/image_degradation/__init__.py @@ -0,0 +1,2 @@ +from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr +from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light diff --git a/ldm/modules/image_degradation/bsrgan.py b/ldm/modules/image_degradation/bsrgan.py new file mode 100644 index 0000000000000000000000000000000000000000..32ef56169978e550090261cddbcf5eb611a6173b --- /dev/null +++ b/ldm/modules/image_degradation/bsrgan.py @@ -0,0 +1,730 @@ +# -*- coding: utf-8 -*- +""" +# -------------------------------------------- +# Super-Resolution +# -------------------------------------------- +# +# Kai Zhang (cskaizhang@gmail.com) +# https://github.com/cszn +# From 2019/03--2021/08 +# -------------------------------------------- +""" + +import numpy as np +import cv2 +import torch + +from functools import partial +import random +from scipy import ndimage +import scipy +import scipy.stats as ss +from scipy.interpolate import interp2d +from scipy.linalg import orth +import albumentations + +import ldm.modules.image_degradation.utils_image as util + + +def modcrop_np(img, sf): + ''' + Args: + img: numpy image, WxH or WxHxC + sf: scale factor + Return: + cropped image + ''' + w, h = img.shape[:2] + im = np.copy(img) + return im[:w - w % sf, :h - h % sf, ...] + + +""" +# -------------------------------------------- +# anisotropic Gaussian kernels +# -------------------------------------------- +""" + + +def analytic_kernel(k): + """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)""" + k_size = k.shape[0] + # Calculate the big kernels size + big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2)) + # Loop over the small kernel to fill the big one + for r in range(k_size): + for c in range(k_size): + big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k + # Crop the edges of the big kernel to ignore very small values and increase run time of SR + crop = k_size // 2 + cropped_big_k = big_k[crop:-crop, crop:-crop] + # Normalize to 1 + return cropped_big_k / cropped_big_k.sum() + + +def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6): + """ generate an anisotropic Gaussian kernel + Args: + ksize : e.g., 15, kernel size + theta : [0, pi], rotation angle range + l1 : [0.1,50], scaling of eigenvalues + l2 : [0.1,l1], scaling of eigenvalues + If l1 = l2, will get an isotropic Gaussian kernel. + Returns: + k : kernel + """ + + v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.])) + V = np.array([[v[0], v[1]], [v[1], -v[0]]]) + D = np.array([[l1, 0], [0, l2]]) + Sigma = np.dot(np.dot(V, D), np.linalg.inv(V)) + k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize) + + return k + + +def gm_blur_kernel(mean, cov, size=15): + center = size / 2.0 + 0.5 + k = np.zeros([size, size]) + for y in range(size): + for x in range(size): + cy = y - center + 1 + cx = x - center + 1 + k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov) + + k = k / np.sum(k) + return k + + +def shift_pixel(x, sf, upper_left=True): + """shift pixel for super-resolution with different scale factors + Args: + x: WxHxC or WxH + sf: scale factor + upper_left: shift direction + """ + h, w = x.shape[:2] + shift = (sf - 1) * 0.5 + xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0) + if upper_left: + x1 = xv + shift + y1 = yv + shift + else: + x1 = xv - shift + y1 = yv - shift + + x1 = np.clip(x1, 0, w - 1) + y1 = np.clip(y1, 0, h - 1) + + if x.ndim == 2: + x = interp2d(xv, yv, x)(x1, y1) + if x.ndim == 3: + for i in range(x.shape[-1]): + x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1) + + return x + + +def blur(x, k): + ''' + x: image, NxcxHxW + k: kernel, Nx1xhxw + ''' + n, c = x.shape[:2] + p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2 + x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate') + k = k.repeat(1, c, 1, 1) + k = k.view(-1, 1, k.shape[2], k.shape[3]) + x = x.view(1, -1, x.shape[2], x.shape[3]) + x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c) + x = x.view(n, c, x.shape[2], x.shape[3]) + + return x + + +def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0): + """" + # modified version of https://github.com/assafshocher/BlindSR_dataset_generator + # Kai Zhang + # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var + # max_var = 2.5 * sf + """ + # Set random eigen-vals (lambdas) and angle (theta) for COV matrix + lambda_1 = min_var + np.random.rand() * (max_var - min_var) + lambda_2 = min_var + np.random.rand() * (max_var - min_var) + theta = np.random.rand() * np.pi # random theta + noise = -noise_level + np.random.rand(*k_size) * noise_level * 2 + + # Set COV matrix using Lambdas and Theta + LAMBDA = np.diag([lambda_1, lambda_2]) + Q = np.array([[np.cos(theta), -np.sin(theta)], + [np.sin(theta), np.cos(theta)]]) + SIGMA = Q @ LAMBDA @ Q.T + INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :] + + # Set expectation position (shifting kernel for aligned image) + MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2) + MU = MU[None, None, :, None] + + # Create meshgrid for Gaussian + [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1])) + Z = np.stack([X, Y], 2)[:, :, :, None] + + # Calcualte Gaussian for every pixel of the kernel + ZZ = Z - MU + ZZ_t = ZZ.transpose(0, 1, 3, 2) + raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise) + + # shift the kernel so it will be centered + # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor) + + # Normalize the kernel and return + # kernel = raw_kernel_centered / np.sum(raw_kernel_centered) + kernel = raw_kernel / np.sum(raw_kernel) + return kernel + + +def fspecial_gaussian(hsize, sigma): + hsize = [hsize, hsize] + siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0] + std = sigma + [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)) + arg = -(x * x + y * y) / (2 * std * std) + h = np.exp(arg) + h[h < scipy.finfo(float).eps * h.max()] = 0 + sumh = h.sum() + if sumh != 0: + h = h / sumh + return h + + +def fspecial_laplacian(alpha): + alpha = max([0, min([alpha, 1])]) + h1 = alpha / (alpha + 1) + h2 = (1 - alpha) / (alpha + 1) + h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]] + h = np.array(h) + return h + + +def fspecial(filter_type, *args, **kwargs): + ''' + python code from: + https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py + ''' + if filter_type == 'gaussian': + return fspecial_gaussian(*args, **kwargs) + if filter_type == 'laplacian': + return fspecial_laplacian(*args, **kwargs) + + +""" +# -------------------------------------------- +# degradation models +# -------------------------------------------- +""" + + +def bicubic_degradation(x, sf=3): + ''' + Args: + x: HxWxC image, [0, 1] + sf: down-scale factor + Return: + bicubicly downsampled LR image + ''' + x = util.imresize_np(x, scale=1 / sf) + return x + + +def srmd_degradation(x, k, sf=3): + ''' blur + bicubic downsampling + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2018learning, + title={Learning a single convolutional super-resolution network for multiple degradations}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={3262--3271}, + year={2018} + } + ''' + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror' + x = bicubic_degradation(x, sf=sf) + return x + + +def dpsr_degradation(x, k, sf=3): + ''' bicubic downsampling + blur + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2019deep, + title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={1671--1681}, + year={2019} + } + ''' + x = bicubic_degradation(x, sf=sf) + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + return x + + +def classical_degradation(x, k, sf=3): + ''' blur + downsampling + Args: + x: HxWxC image, [0, 1]/[0, 255] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + ''' + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2)) + st = 0 + return x[st::sf, st::sf, ...] + + +def add_sharpening(img, weight=0.5, radius=50, threshold=10): + """USM sharpening. borrowed from real-ESRGAN + Input image: I; Blurry image: B. + 1. K = I + weight * (I - B) + 2. Mask = 1 if abs(I - B) > threshold, else: 0 + 3. Blur mask: + 4. Out = Mask * K + (1 - Mask) * I + Args: + img (Numpy array): Input image, HWC, BGR; float32, [0, 1]. + weight (float): Sharp weight. Default: 1. + radius (float): Kernel size of Gaussian blur. Default: 50. + threshold (int): + """ + if radius % 2 == 0: + radius += 1 + blur = cv2.GaussianBlur(img, (radius, radius), 0) + residual = img - blur + mask = np.abs(residual) * 255 > threshold + mask = mask.astype('float32') + soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0) + + K = img + weight * residual + K = np.clip(K, 0, 1) + return soft_mask * K + (1 - soft_mask) * img + + +def add_blur(img, sf=4): + wd2 = 4.0 + sf + wd = 2.0 + 0.2 * sf + if random.random() < 0.5: + l1 = wd2 * random.random() + l2 = wd2 * random.random() + k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2) + else: + k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random()) + img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror') + + return img + + +def add_resize(img, sf=4): + rnum = np.random.rand() + if rnum > 0.8: # up + sf1 = random.uniform(1, 2) + elif rnum < 0.7: # down + sf1 = random.uniform(0.5 / sf, 1) + else: + sf1 = 1.0 + img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + return img + + +# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): +# noise_level = random.randint(noise_level1, noise_level2) +# rnum = np.random.rand() +# if rnum > 0.6: # add color Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) +# elif rnum < 0.4: # add grayscale Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) +# else: # add noise +# L = noise_level2 / 255. +# D = np.diag(np.random.rand(3)) +# U = orth(np.random.rand(3, 3)) +# conv = np.dot(np.dot(np.transpose(U), D), U) +# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) +# img = np.clip(img, 0.0, 1.0) +# return img + +def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + rnum = np.random.rand() + if rnum > 0.6: # add color Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: # add grayscale Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: # add noise + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_speckle_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + img = np.clip(img, 0.0, 1.0) + rnum = random.random() + if rnum > 0.6: + img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: + img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_Poisson_noise(img): + img = np.clip((img * 255.0).round(), 0, 255) / 255. + vals = 10 ** (2 * random.random() + 2.0) # [2, 4] + if random.random() < 0.5: + img = np.random.poisson(img * vals).astype(np.float32) / vals + else: + img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114]) + img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255. + noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray + img += noise_gray[:, :, np.newaxis] + img = np.clip(img, 0.0, 1.0) + return img + + +def add_JPEG_noise(img): + quality_factor = random.randint(30, 95) + img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR) + result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]) + img = cv2.imdecode(encimg, 1) + img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB) + return img + + +def random_crop(lq, hq, sf=4, lq_patchsize=64): + h, w = lq.shape[:2] + rnd_h = random.randint(0, h - lq_patchsize) + rnd_w = random.randint(0, w - lq_patchsize) + lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :] + + rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf) + hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :] + return lq, hq + + +def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + hq = img.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + img = util.imresize_np(img, 1 / 2, True) + img = np.clip(img, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + img = add_blur(img, sf=sf) + + elif i == 1: + img = add_blur(img, sf=sf) + + elif i == 2: + a, b = img.shape[1], img.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror') + img = img[0::sf, 0::sf, ...] # nearest downsampling + img = np.clip(img, 0.0, 1.0) + + elif i == 3: + # downsample3 + img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + img = add_JPEG_noise(img) + + elif i == 6: + # add processed camera sensor noise + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf_ori, lq_patchsize) + + return img, hq + + +# todo no isp_model? +def degradation_bsrgan_variant(image, sf=4, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + image = util.uint2single(image) + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = image.shape[:2] + image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = image.shape[:2] + + hq = image.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + image = util.imresize_np(image, 1 / 2, True) + image = np.clip(image, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + image = add_blur(image, sf=sf) + + elif i == 1: + image = add_blur(image, sf=sf) + + elif i == 2: + a, b = image.shape[1], image.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror') + image = image[0::sf, 0::sf, ...] # nearest downsampling + image = np.clip(image, 0.0, 1.0) + + elif i == 3: + # downsample3 + image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + image = np.clip(image, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + image = add_JPEG_noise(image) + + # elif i == 6: + # # add processed camera sensor noise + # if random.random() < isp_prob and isp_model is not None: + # with torch.no_grad(): + # img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + image = add_JPEG_noise(image) + image = util.single2uint(image) + example = {"image":image} + return example + + +# TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc... +def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None): + """ + This is an extended degradation model by combining + the degradation models of BSRGAN and Real-ESRGAN + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + use_shuffle: the degradation shuffle + use_sharp: sharpening the img + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + if use_sharp: + img = add_sharpening(img) + hq = img.copy() + + if random.random() < shuffle_prob: + shuffle_order = random.sample(range(13), 13) + else: + shuffle_order = list(range(13)) + # local shuffle for noise, JPEG is always the last one + shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6))) + shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13))) + + poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1 + + for i in shuffle_order: + if i == 0: + img = add_blur(img, sf=sf) + elif i == 1: + img = add_resize(img, sf=sf) + elif i == 2: + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + elif i == 3: + if random.random() < poisson_prob: + img = add_Poisson_noise(img) + elif i == 4: + if random.random() < speckle_prob: + img = add_speckle_noise(img) + elif i == 5: + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + elif i == 6: + img = add_JPEG_noise(img) + elif i == 7: + img = add_blur(img, sf=sf) + elif i == 8: + img = add_resize(img, sf=sf) + elif i == 9: + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + elif i == 10: + if random.random() < poisson_prob: + img = add_Poisson_noise(img) + elif i == 11: + if random.random() < speckle_prob: + img = add_speckle_noise(img) + elif i == 12: + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + else: + print('check the shuffle!') + + # resize to desired size + img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])), + interpolation=random.choice([1, 2, 3])) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf, lq_patchsize) + + return img, hq + + +if __name__ == '__main__': + print("hey") + img = util.imread_uint('utils/test.png', 3) + print(img) + img = util.uint2single(img) + print(img) + img = img[:448, :448] + h = img.shape[0] // 4 + print("resizing to", h) + sf = 4 + deg_fn = partial(degradation_bsrgan_variant, sf=sf) + for i in range(20): + print(i) + img_lq = deg_fn(img) + print(img_lq) + img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"] + print(img_lq.shape) + print("bicubic", img_lq_bicubic.shape) + print(img_hq.shape) + lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1) + util.imsave(img_concat, str(i) + '.png') + + diff --git a/ldm/modules/image_degradation/bsrgan_light.py b/ldm/modules/image_degradation/bsrgan_light.py new file mode 100644 index 0000000000000000000000000000000000000000..808c7f882cb75e2ba2340d5b55881d11927351f0 --- /dev/null +++ b/ldm/modules/image_degradation/bsrgan_light.py @@ -0,0 +1,651 @@ +# -*- coding: utf-8 -*- +import numpy as np +import cv2 +import torch + +from functools import partial +import random +from scipy import ndimage +import scipy +import scipy.stats as ss +from scipy.interpolate import interp2d +from scipy.linalg import orth +import albumentations + +import ldm.modules.image_degradation.utils_image as util + +""" +# -------------------------------------------- +# Super-Resolution +# -------------------------------------------- +# +# Kai Zhang (cskaizhang@gmail.com) +# https://github.com/cszn +# From 2019/03--2021/08 +# -------------------------------------------- +""" + +def modcrop_np(img, sf): + ''' + Args: + img: numpy image, WxH or WxHxC + sf: scale factor + Return: + cropped image + ''' + w, h = img.shape[:2] + im = np.copy(img) + return im[:w - w % sf, :h - h % sf, ...] + + +""" +# -------------------------------------------- +# anisotropic Gaussian kernels +# -------------------------------------------- +""" + + +def analytic_kernel(k): + """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)""" + k_size = k.shape[0] + # Calculate the big kernels size + big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2)) + # Loop over the small kernel to fill the big one + for r in range(k_size): + for c in range(k_size): + big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k + # Crop the edges of the big kernel to ignore very small values and increase run time of SR + crop = k_size // 2 + cropped_big_k = big_k[crop:-crop, crop:-crop] + # Normalize to 1 + return cropped_big_k / cropped_big_k.sum() + + +def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6): + """ generate an anisotropic Gaussian kernel + Args: + ksize : e.g., 15, kernel size + theta : [0, pi], rotation angle range + l1 : [0.1,50], scaling of eigenvalues + l2 : [0.1,l1], scaling of eigenvalues + If l1 = l2, will get an isotropic Gaussian kernel. + Returns: + k : kernel + """ + + v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.])) + V = np.array([[v[0], v[1]], [v[1], -v[0]]]) + D = np.array([[l1, 0], [0, l2]]) + Sigma = np.dot(np.dot(V, D), np.linalg.inv(V)) + k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize) + + return k + + +def gm_blur_kernel(mean, cov, size=15): + center = size / 2.0 + 0.5 + k = np.zeros([size, size]) + for y in range(size): + for x in range(size): + cy = y - center + 1 + cx = x - center + 1 + k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov) + + k = k / np.sum(k) + return k + + +def shift_pixel(x, sf, upper_left=True): + """shift pixel for super-resolution with different scale factors + Args: + x: WxHxC or WxH + sf: scale factor + upper_left: shift direction + """ + h, w = x.shape[:2] + shift = (sf - 1) * 0.5 + xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0) + if upper_left: + x1 = xv + shift + y1 = yv + shift + else: + x1 = xv - shift + y1 = yv - shift + + x1 = np.clip(x1, 0, w - 1) + y1 = np.clip(y1, 0, h - 1) + + if x.ndim == 2: + x = interp2d(xv, yv, x)(x1, y1) + if x.ndim == 3: + for i in range(x.shape[-1]): + x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1) + + return x + + +def blur(x, k): + ''' + x: image, NxcxHxW + k: kernel, Nx1xhxw + ''' + n, c = x.shape[:2] + p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2 + x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate') + k = k.repeat(1, c, 1, 1) + k = k.view(-1, 1, k.shape[2], k.shape[3]) + x = x.view(1, -1, x.shape[2], x.shape[3]) + x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c) + x = x.view(n, c, x.shape[2], x.shape[3]) + + return x + + +def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0): + """" + # modified version of https://github.com/assafshocher/BlindSR_dataset_generator + # Kai Zhang + # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var + # max_var = 2.5 * sf + """ + # Set random eigen-vals (lambdas) and angle (theta) for COV matrix + lambda_1 = min_var + np.random.rand() * (max_var - min_var) + lambda_2 = min_var + np.random.rand() * (max_var - min_var) + theta = np.random.rand() * np.pi # random theta + noise = -noise_level + np.random.rand(*k_size) * noise_level * 2 + + # Set COV matrix using Lambdas and Theta + LAMBDA = np.diag([lambda_1, lambda_2]) + Q = np.array([[np.cos(theta), -np.sin(theta)], + [np.sin(theta), np.cos(theta)]]) + SIGMA = Q @ LAMBDA @ Q.T + INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :] + + # Set expectation position (shifting kernel for aligned image) + MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2) + MU = MU[None, None, :, None] + + # Create meshgrid for Gaussian + [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1])) + Z = np.stack([X, Y], 2)[:, :, :, None] + + # Calcualte Gaussian for every pixel of the kernel + ZZ = Z - MU + ZZ_t = ZZ.transpose(0, 1, 3, 2) + raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise) + + # shift the kernel so it will be centered + # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor) + + # Normalize the kernel and return + # kernel = raw_kernel_centered / np.sum(raw_kernel_centered) + kernel = raw_kernel / np.sum(raw_kernel) + return kernel + + +def fspecial_gaussian(hsize, sigma): + hsize = [hsize, hsize] + siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0] + std = sigma + [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)) + arg = -(x * x + y * y) / (2 * std * std) + h = np.exp(arg) + h[h < scipy.finfo(float).eps * h.max()] = 0 + sumh = h.sum() + if sumh != 0: + h = h / sumh + return h + + +def fspecial_laplacian(alpha): + alpha = max([0, min([alpha, 1])]) + h1 = alpha / (alpha + 1) + h2 = (1 - alpha) / (alpha + 1) + h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]] + h = np.array(h) + return h + + +def fspecial(filter_type, *args, **kwargs): + ''' + python code from: + https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py + ''' + if filter_type == 'gaussian': + return fspecial_gaussian(*args, **kwargs) + if filter_type == 'laplacian': + return fspecial_laplacian(*args, **kwargs) + + +""" +# -------------------------------------------- +# degradation models +# -------------------------------------------- +""" + + +def bicubic_degradation(x, sf=3): + ''' + Args: + x: HxWxC image, [0, 1] + sf: down-scale factor + Return: + bicubicly downsampled LR image + ''' + x = util.imresize_np(x, scale=1 / sf) + return x + + +def srmd_degradation(x, k, sf=3): + ''' blur + bicubic downsampling + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2018learning, + title={Learning a single convolutional super-resolution network for multiple degradations}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={3262--3271}, + year={2018} + } + ''' + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror' + x = bicubic_degradation(x, sf=sf) + return x + + +def dpsr_degradation(x, k, sf=3): + ''' bicubic downsampling + blur + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2019deep, + title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={1671--1681}, + year={2019} + } + ''' + x = bicubic_degradation(x, sf=sf) + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + return x + + +def classical_degradation(x, k, sf=3): + ''' blur + downsampling + Args: + x: HxWxC image, [0, 1]/[0, 255] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + ''' + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2)) + st = 0 + return x[st::sf, st::sf, ...] + + +def add_sharpening(img, weight=0.5, radius=50, threshold=10): + """USM sharpening. borrowed from real-ESRGAN + Input image: I; Blurry image: B. + 1. K = I + weight * (I - B) + 2. Mask = 1 if abs(I - B) > threshold, else: 0 + 3. Blur mask: + 4. Out = Mask * K + (1 - Mask) * I + Args: + img (Numpy array): Input image, HWC, BGR; float32, [0, 1]. + weight (float): Sharp weight. Default: 1. + radius (float): Kernel size of Gaussian blur. Default: 50. + threshold (int): + """ + if radius % 2 == 0: + radius += 1 + blur = cv2.GaussianBlur(img, (radius, radius), 0) + residual = img - blur + mask = np.abs(residual) * 255 > threshold + mask = mask.astype('float32') + soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0) + + K = img + weight * residual + K = np.clip(K, 0, 1) + return soft_mask * K + (1 - soft_mask) * img + + +def add_blur(img, sf=4): + wd2 = 4.0 + sf + wd = 2.0 + 0.2 * sf + + wd2 = wd2/4 + wd = wd/4 + + if random.random() < 0.5: + l1 = wd2 * random.random() + l2 = wd2 * random.random() + k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2) + else: + k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random()) + img = ndimage.convolve(img, np.expand_dims(k, axis=2), mode='mirror') + + return img + + +def add_resize(img, sf=4): + rnum = np.random.rand() + if rnum > 0.8: # up + sf1 = random.uniform(1, 2) + elif rnum < 0.7: # down + sf1 = random.uniform(0.5 / sf, 1) + else: + sf1 = 1.0 + img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + return img + + +# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): +# noise_level = random.randint(noise_level1, noise_level2) +# rnum = np.random.rand() +# if rnum > 0.6: # add color Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) +# elif rnum < 0.4: # add grayscale Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) +# else: # add noise +# L = noise_level2 / 255. +# D = np.diag(np.random.rand(3)) +# U = orth(np.random.rand(3, 3)) +# conv = np.dot(np.dot(np.transpose(U), D), U) +# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) +# img = np.clip(img, 0.0, 1.0) +# return img + +def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + rnum = np.random.rand() + if rnum > 0.6: # add color Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: # add grayscale Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: # add noise + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_speckle_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + img = np.clip(img, 0.0, 1.0) + rnum = random.random() + if rnum > 0.6: + img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: + img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_Poisson_noise(img): + img = np.clip((img * 255.0).round(), 0, 255) / 255. + vals = 10 ** (2 * random.random() + 2.0) # [2, 4] + if random.random() < 0.5: + img = np.random.poisson(img * vals).astype(np.float32) / vals + else: + img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114]) + img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255. + noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray + img += noise_gray[:, :, np.newaxis] + img = np.clip(img, 0.0, 1.0) + return img + + +def add_JPEG_noise(img): + quality_factor = random.randint(80, 95) + img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR) + result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]) + img = cv2.imdecode(encimg, 1) + img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB) + return img + + +def random_crop(lq, hq, sf=4, lq_patchsize=64): + h, w = lq.shape[:2] + rnd_h = random.randint(0, h - lq_patchsize) + rnd_w = random.randint(0, w - lq_patchsize) + lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :] + + rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf) + hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :] + return lq, hq + + +def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + hq = img.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + img = util.imresize_np(img, 1 / 2, True) + img = np.clip(img, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + img = add_blur(img, sf=sf) + + elif i == 1: + img = add_blur(img, sf=sf) + + elif i == 2: + a, b = img.shape[1], img.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + img = ndimage.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror') + img = img[0::sf, 0::sf, ...] # nearest downsampling + img = np.clip(img, 0.0, 1.0) + + elif i == 3: + # downsample3 + img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + img = add_JPEG_noise(img) + + elif i == 6: + # add processed camera sensor noise + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf_ori, lq_patchsize) + + return img, hq + + +# todo no isp_model? +def degradation_bsrgan_variant(image, sf=4, isp_model=None, up=False): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + image = util.uint2single(image) + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = image.shape[:2] + image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = image.shape[:2] + + hq = image.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + image = util.imresize_np(image, 1 / 2, True) + image = np.clip(image, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + image = add_blur(image, sf=sf) + + # elif i == 1: + # image = add_blur(image, sf=sf) + + if i == 0: + pass + + elif i == 2: + a, b = image.shape[1], image.shape[0] + # downsample2 + if random.random() < 0.8: + sf1 = random.uniform(1, 2 * sf) + image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + image = ndimage.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror') + image = image[0::sf, 0::sf, ...] # nearest downsampling + + image = np.clip(image, 0.0, 1.0) + + elif i == 3: + # downsample3 + image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + image = np.clip(image, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + image = add_JPEG_noise(image) + # + # elif i == 6: + # # add processed camera sensor noise + # if random.random() < isp_prob and isp_model is not None: + # with torch.no_grad(): + # img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + image = add_JPEG_noise(image) + image = util.single2uint(image) + if up: + image = cv2.resize(image, (w1, h1), interpolation=cv2.INTER_CUBIC) # todo: random, as above? want to condition on it then + example = {"image": image} + return example + + + + +if __name__ == '__main__': + print("hey") + img = util.imread_uint('utils/test.png', 3) + img = img[:448, :448] + h = img.shape[0] // 4 + print("resizing to", h) + sf = 4 + deg_fn = partial(degradation_bsrgan_variant, sf=sf) + for i in range(20): + print(i) + img_hq = img + img_lq = deg_fn(img)["image"] + img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq) + print(img_lq) + img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"] + print(img_lq.shape) + print("bicubic", img_lq_bicubic.shape) + print(img_hq.shape) + lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), + (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1) + util.imsave(img_concat, str(i) + '.png') diff --git a/ldm/modules/image_degradation/utils/test.png b/ldm/modules/image_degradation/utils/test.png new file mode 100644 index 0000000000000000000000000000000000000000..4249b43de0f22707758d13c240268a401642f6e6 Binary files /dev/null and b/ldm/modules/image_degradation/utils/test.png differ diff --git a/ldm/modules/image_degradation/utils_image.py b/ldm/modules/image_degradation/utils_image.py new file mode 100644 index 0000000000000000000000000000000000000000..0175f155ad900ae33c3c46ed87f49b352e3faf98 --- /dev/null +++ b/ldm/modules/image_degradation/utils_image.py @@ -0,0 +1,916 @@ +import os +import math +import random +import numpy as np +import torch +import cv2 +from torchvision.utils import make_grid +from datetime import datetime +#import matplotlib.pyplot as plt # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py + + +os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" + + +''' +# -------------------------------------------- +# Kai Zhang (github: https://github.com/cszn) +# 03/Mar/2019 +# -------------------------------------------- +# https://github.com/twhui/SRGAN-pyTorch +# https://github.com/xinntao/BasicSR +# -------------------------------------------- +''' + + +IMG_EXTENSIONS = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tif'] + + +def is_image_file(filename): + return any(filename.endswith(extension) for extension in IMG_EXTENSIONS) + + +def get_timestamp(): + return datetime.now().strftime('%y%m%d-%H%M%S') + + +def imshow(x, title=None, cbar=False, figsize=None): + plt.figure(figsize=figsize) + plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray') + if title: + plt.title(title) + if cbar: + plt.colorbar() + plt.show() + + +def surf(Z, cmap='rainbow', figsize=None): + plt.figure(figsize=figsize) + ax3 = plt.axes(projection='3d') + + w, h = Z.shape[:2] + xx = np.arange(0,w,1) + yy = np.arange(0,h,1) + X, Y = np.meshgrid(xx, yy) + ax3.plot_surface(X,Y,Z,cmap=cmap) + #ax3.contour(X,Y,Z, zdim='z',offset=-2，cmap=cmap) + plt.show() + + +''' +# -------------------------------------------- +# get image pathes +# -------------------------------------------- +''' + + +def get_image_paths(dataroot): + paths = None # return None if dataroot is None + if dataroot is not None: + paths = sorted(_get_paths_from_images(dataroot)) + return paths + + +def _get_paths_from_images(path): + assert os.path.isdir(path), '{:s} is not a valid directory'.format(path) + images = [] + for dirpath, _, fnames in sorted(os.walk(path)): + for fname in sorted(fnames): + if is_image_file(fname): + img_path = os.path.join(dirpath, fname) + images.append(img_path) + assert images, '{:s} has no valid image file'.format(path) + return images + + +''' +# -------------------------------------------- +# split large images into small images +# -------------------------------------------- +''' + + +def patches_from_image(img, p_size=512, p_overlap=64, p_max=800): + w, h = img.shape[:2] + patches = [] + if w > p_max and h > p_max: + w1 = list(np.arange(0, w-p_size, p_size-p_overlap, dtype=np.int)) + h1 = list(np.arange(0, h-p_size, p_size-p_overlap, dtype=np.int)) + w1.append(w-p_size) + h1.append(h-p_size) +# print(w1) +# print(h1) + for i in w1: + for j in h1: + patches.append(img[i:i+p_size, j:j+p_size,:]) + else: + patches.append(img) + + return patches + + +def imssave(imgs, img_path): + """ + imgs: list, N images of size WxHxC + """ + img_name, ext = os.path.splitext(os.path.basename(img_path)) + + for i, img in enumerate(imgs): + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + new_path = os.path.join(os.path.dirname(img_path), img_name+str('_s{:04d}'.format(i))+'.png') + cv2.imwrite(new_path, img) + + +def split_imageset(original_dataroot, taget_dataroot, n_channels=3, p_size=800, p_overlap=96, p_max=1000): + """ + split the large images from original_dataroot into small overlapped images with size (p_size)x(p_size), + and save them into taget_dataroot; only the images with larger size than (p_max)x(p_max) + will be splitted. + Args: + original_dataroot: + taget_dataroot: + p_size: size of small images + p_overlap: patch size in training is a good choice + p_max: images with smaller size than (p_max)x(p_max) keep unchanged. + """ + paths = get_image_paths(original_dataroot) + for img_path in paths: + # img_name, ext = os.path.splitext(os.path.basename(img_path)) + img = imread_uint(img_path, n_channels=n_channels) + patches = patches_from_image(img, p_size, p_overlap, p_max) + imssave(patches, os.path.join(taget_dataroot,os.path.basename(img_path))) + #if original_dataroot == taget_dataroot: + #del img_path + +''' +# -------------------------------------------- +# makedir +# -------------------------------------------- +''' + + +def mkdir(path): + if not os.path.exists(path): + os.makedirs(path) + + +def mkdirs(paths): + if isinstance(paths, str): + mkdir(paths) + else: + for path in paths: + mkdir(path) + + +def mkdir_and_rename(path): + if os.path.exists(path): + new_name = path + '_archived_' + get_timestamp() + print('Path already exists. Rename it to [{:s}]'.format(new_name)) + os.rename(path, new_name) + os.makedirs(path) + + +''' +# -------------------------------------------- +# read image from path +# opencv is fast, but read BGR numpy image +# -------------------------------------------- +''' + + +# -------------------------------------------- +# get uint8 image of size HxWxn_channles (RGB) +# -------------------------------------------- +def imread_uint(path, n_channels=3): + # input: path + # output: HxWx3(RGB or GGG), or HxWx1 (G) + if n_channels == 1: + img = cv2.imread(path, 0) # cv2.IMREAD_GRAYSCALE + img = np.expand_dims(img, axis=2) # HxWx1 + elif n_channels == 3: + img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # BGR or G + if img.ndim == 2: + img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) # GGG + else: + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # RGB + return img + + +# -------------------------------------------- +# matlab's imwrite +# -------------------------------------------- +def imsave(img, img_path): + img = np.squeeze(img) + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + cv2.imwrite(img_path, img) + +def imwrite(img, img_path): + img = np.squeeze(img) + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + cv2.imwrite(img_path, img) + + + +# -------------------------------------------- +# get single image of size HxWxn_channles (BGR) +# -------------------------------------------- +def read_img(path): + # read image by cv2 + # return: Numpy float32, HWC, BGR, [0,1] + img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # cv2.IMREAD_GRAYSCALE + img = img.astype(np.float32) / 255. + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + # some images have 4 channels + if img.shape[2] > 3: + img = img[:, :, :3] + return img + + +''' +# -------------------------------------------- +# image format conversion +# -------------------------------------------- +# numpy(single) <---> numpy(unit) +# numpy(single) <---> tensor +# numpy(unit) <---> tensor +# -------------------------------------------- +''' + + +# -------------------------------------------- +# numpy(single) [0, 1] <---> numpy(unit) +# -------------------------------------------- + + +def uint2single(img): + + return np.float32(img/255.) + + +def single2uint(img): + + return np.uint8((img.clip(0, 1)*255.).round()) + + +def uint162single(img): + + return np.float32(img/65535.) + + +def single2uint16(img): + + return np.uint16((img.clip(0, 1)*65535.).round()) + + +# -------------------------------------------- +# numpy(unit) (HxWxC or HxW) <---> tensor +# -------------------------------------------- + + +# convert uint to 4-dimensional torch tensor +def uint2tensor4(img): + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.).unsqueeze(0) + + +# convert uint to 3-dimensional torch tensor +def uint2tensor3(img): + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.) + + +# convert 2/3/4-dimensional torch tensor to uint +def tensor2uint(img): + img = img.data.squeeze().float().clamp_(0, 1).cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + return np.uint8((img*255.0).round()) + + +# -------------------------------------------- +# numpy(single) (HxWxC) <---> tensor +# -------------------------------------------- + + +# convert single (HxWxC) to 3-dimensional torch tensor +def single2tensor3(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float() + + +# convert single (HxWxC) to 4-dimensional torch tensor +def single2tensor4(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().unsqueeze(0) + + +# convert torch tensor to single +def tensor2single(img): + img = img.data.squeeze().float().cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + + return img + +# convert torch tensor to single +def tensor2single3(img): + img = img.data.squeeze().float().cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + elif img.ndim == 2: + img = np.expand_dims(img, axis=2) + return img + + +def single2tensor5(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float().unsqueeze(0) + + +def single32tensor5(img): + return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0) + + +def single42tensor4(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float() + + +# from skimage.io import imread, imsave +def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)): + ''' + Converts a torch Tensor into an image Numpy array of BGR channel order + Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order + Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default) + ''' + tensor = tensor.squeeze().float().cpu().clamp_(*min_max) # squeeze first, then clamp + tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0]) # to range [0,1] + n_dim = tensor.dim() + if n_dim == 4: + n_img = len(tensor) + img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy() + img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR + elif n_dim == 3: + img_np = tensor.numpy() + img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR + elif n_dim == 2: + img_np = tensor.numpy() + else: + raise TypeError( + 'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(n_dim)) + if out_type == np.uint8: + img_np = (img_np * 255.0).round() + # Important. Unlike matlab, numpy.unit8() WILL NOT round by default. + return img_np.astype(out_type) + + +''' +# -------------------------------------------- +# Augmentation, flipe and/or rotate +# -------------------------------------------- +# The following two are enough. +# (1) augmet_img: numpy image of WxHxC or WxH +# (2) augment_img_tensor4: tensor image 1xCxWxH +# -------------------------------------------- +''' + + +def augment_img(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + if mode == 0: + return img + elif mode == 1: + return np.flipud(np.rot90(img)) + elif mode == 2: + return np.flipud(img) + elif mode == 3: + return np.rot90(img, k=3) + elif mode == 4: + return np.flipud(np.rot90(img, k=2)) + elif mode == 5: + return np.rot90(img) + elif mode == 6: + return np.rot90(img, k=2) + elif mode == 7: + return np.flipud(np.rot90(img, k=3)) + + +def augment_img_tensor4(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + if mode == 0: + return img + elif mode == 1: + return img.rot90(1, [2, 3]).flip([2]) + elif mode == 2: + return img.flip([2]) + elif mode == 3: + return img.rot90(3, [2, 3]) + elif mode == 4: + return img.rot90(2, [2, 3]).flip([2]) + elif mode == 5: + return img.rot90(1, [2, 3]) + elif mode == 6: + return img.rot90(2, [2, 3]) + elif mode == 7: + return img.rot90(3, [2, 3]).flip([2]) + + +def augment_img_tensor(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + img_size = img.size() + img_np = img.data.cpu().numpy() + if len(img_size) == 3: + img_np = np.transpose(img_np, (1, 2, 0)) + elif len(img_size) == 4: + img_np = np.transpose(img_np, (2, 3, 1, 0)) + img_np = augment_img(img_np, mode=mode) + img_tensor = torch.from_numpy(np.ascontiguousarray(img_np)) + if len(img_size) == 3: + img_tensor = img_tensor.permute(2, 0, 1) + elif len(img_size) == 4: + img_tensor = img_tensor.permute(3, 2, 0, 1) + + return img_tensor.type_as(img) + + +def augment_img_np3(img, mode=0): + if mode == 0: + return img + elif mode == 1: + return img.transpose(1, 0, 2) + elif mode == 2: + return img[::-1, :, :] + elif mode == 3: + img = img[::-1, :, :] + img = img.transpose(1, 0, 2) + return img + elif mode == 4: + return img[:, ::-1, :] + elif mode == 5: + img = img[:, ::-1, :] + img = img.transpose(1, 0, 2) + return img + elif mode == 6: + img = img[:, ::-1, :] + img = img[::-1, :, :] + return img + elif mode == 7: + img = img[:, ::-1, :] + img = img[::-1, :, :] + img = img.transpose(1, 0, 2) + return img + + +def augment_imgs(img_list, hflip=True, rot=True): + # horizontal flip OR rotate + hflip = hflip and random.random() < 0.5 + vflip = rot and random.random() < 0.5 + rot90 = rot and random.random() < 0.5 + + def _augment(img): + if hflip: + img = img[:, ::-1, :] + if vflip: + img = img[::-1, :, :] + if rot90: + img = img.transpose(1, 0, 2) + return img + + return [_augment(img) for img in img_list] + + +''' +# -------------------------------------------- +# modcrop and shave +# -------------------------------------------- +''' + + +def modcrop(img_in, scale): + # img_in: Numpy, HWC or HW + img = np.copy(img_in) + if img.ndim == 2: + H, W = img.shape + H_r, W_r = H % scale, W % scale + img = img[:H - H_r, :W - W_r] + elif img.ndim == 3: + H, W, C = img.shape + H_r, W_r = H % scale, W % scale + img = img[:H - H_r, :W - W_r, :] + else: + raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim)) + return img + + +def shave(img_in, border=0): + # img_in: Numpy, HWC or HW + img = np.copy(img_in) + h, w = img.shape[:2] + img = img[border:h-border, border:w-border] + return img + + +''' +# -------------------------------------------- +# image processing process on numpy image +# channel_convert(in_c, tar_type, img_list): +# rgb2ycbcr(img, only_y=True): +# bgr2ycbcr(img, only_y=True): +# ycbcr2rgb(img): +# -------------------------------------------- +''' + + +def rgb2ycbcr(img, only_y=True): + '''same as matlab rgb2ycbcr + only_y: only return Y channel + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + if only_y: + rlt = np.dot(img, [65.481, 128.553, 24.966]) / 255.0 + 16.0 + else: + rlt = np.matmul(img, [[65.481, -37.797, 112.0], [128.553, -74.203, -93.786], + [24.966, 112.0, -18.214]]) / 255.0 + [16, 128, 128] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def ycbcr2rgb(img): + '''same as matlab ycbcr2rgb + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + rlt = np.matmul(img, [[0.00456621, 0.00456621, 0.00456621], [0, -0.00153632, 0.00791071], + [0.00625893, -0.00318811, 0]]) * 255.0 + [-222.921, 135.576, -276.836] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def bgr2ycbcr(img, only_y=True): + '''bgr version of rgb2ycbcr + only_y: only return Y channel + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + if only_y: + rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0 + else: + rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786], + [65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def channel_convert(in_c, tar_type, img_list): + # conversion among BGR, gray and y + if in_c == 3 and tar_type == 'gray': # BGR to gray + gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list] + return [np.expand_dims(img, axis=2) for img in gray_list] + elif in_c == 3 and tar_type == 'y': # BGR to y + y_list = [bgr2ycbcr(img, only_y=True) for img in img_list] + return [np.expand_dims(img, axis=2) for img in y_list] + elif in_c == 1 and tar_type == 'RGB': # gray/y to BGR + return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list] + else: + return img_list + + +''' +# -------------------------------------------- +# metric, PSNR and SSIM +# -------------------------------------------- +''' + + +# -------------------------------------------- +# PSNR +# -------------------------------------------- +def calculate_psnr(img1, img2, border=0): + # img1 and img2 have range [0, 255] + #img1 = img1.squeeze() + #img2 = img2.squeeze() + if not img1.shape == img2.shape: + raise ValueError('Input images must have the same dimensions.') + h, w = img1.shape[:2] + img1 = img1[border:h-border, border:w-border] + img2 = img2[border:h-border, border:w-border] + + img1 = img1.astype(np.float64) + img2 = img2.astype(np.float64) + mse = np.mean((img1 - img2)**2) + if mse == 0: + return float('inf') + return 20 * math.log10(255.0 / math.sqrt(mse)) + + +# -------------------------------------------- +# SSIM +# -------------------------------------------- +def calculate_ssim(img1, img2, border=0): + '''calculate SSIM + the same outputs as MATLAB's + img1, img2: [0, 255] + ''' + #img1 = img1.squeeze() + #img2 = img2.squeeze() + if not img1.shape == img2.shape: + raise ValueError('Input images must have the same dimensions.') + h, w = img1.shape[:2] + img1 = img1[border:h-border, border:w-border] + img2 = img2[border:h-border, border:w-border] + + if img1.ndim == 2: + return ssim(img1, img2) + elif img1.ndim == 3: + if img1.shape[2] == 3: + ssims = [] + for i in range(3): + ssims.append(ssim(img1[:,:,i], img2[:,:,i])) + return np.array(ssims).mean() + elif img1.shape[2] == 1: + return ssim(np.squeeze(img1), np.squeeze(img2)) + else: + raise ValueError('Wrong input image dimensions.') + + +def ssim(img1, img2): + C1 = (0.01 * 255)**2 + C2 = (0.03 * 255)**2 + + img1 = img1.astype(np.float64) + img2 = img2.astype(np.float64) + kernel = cv2.getGaussianKernel(11, 1.5) + window = np.outer(kernel, kernel.transpose()) + + mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid + mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5] + mu1_sq = mu1**2 + mu2_sq = mu2**2 + mu1_mu2 = mu1 * mu2 + sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq + sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq + sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2 + + ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * + (sigma1_sq + sigma2_sq + C2)) + return ssim_map.mean() + + +''' +# -------------------------------------------- +# matlab's bicubic imresize (numpy and torch) [0, 1] +# -------------------------------------------- +''' + + +# matlab 'imresize' function, now only support 'bicubic' +def cubic(x): + absx = torch.abs(x) + absx2 = absx**2 + absx3 = absx**3 + return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \ + (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx)) + + +def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing): + if (scale < 1) and (antialiasing): + # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width + kernel_width = kernel_width / scale + + # Output-space coordinates + x = torch.linspace(1, out_length, out_length) + + # Input-space coordinates. Calculate the inverse mapping such that 0.5 + # in output space maps to 0.5 in input space, and 0.5+scale in output + # space maps to 1.5 in input space. + u = x / scale + 0.5 * (1 - 1 / scale) + + # What is the left-most pixel that can be involved in the computation? + left = torch.floor(u - kernel_width / 2) + + # What is the maximum number of pixels that can be involved in the + # computation? Note: it's OK to use an extra pixel here; if the + # corresponding weights are all zero, it will be eliminated at the end + # of this function. + P = math.ceil(kernel_width) + 2 + + # The indices of the input pixels involved in computing the k-th output + # pixel are in row k of the indices matrix. + indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view( + 1, P).expand(out_length, P) + + # The weights used to compute the k-th output pixel are in row k of the + # weights matrix. + distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices + # apply cubic kernel + if (scale < 1) and (antialiasing): + weights = scale * cubic(distance_to_center * scale) + else: + weights = cubic(distance_to_center) + # Normalize the weights matrix so that each row sums to 1. + weights_sum = torch.sum(weights, 1).view(out_length, 1) + weights = weights / weights_sum.expand(out_length, P) + + # If a column in weights is all zero, get rid of it. only consider the first and last column. + weights_zero_tmp = torch.sum((weights == 0), 0) + if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6): + indices = indices.narrow(1, 1, P - 2) + weights = weights.narrow(1, 1, P - 2) + if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6): + indices = indices.narrow(1, 0, P - 2) + weights = weights.narrow(1, 0, P - 2) + weights = weights.contiguous() + indices = indices.contiguous() + sym_len_s = -indices.min() + 1 + sym_len_e = indices.max() - in_length + indices = indices + sym_len_s - 1 + return weights, indices, int(sym_len_s), int(sym_len_e) + + +# -------------------------------------------- +# imresize for tensor image [0, 1] +# -------------------------------------------- +def imresize(img, scale, antialiasing=True): + # Now the scale should be the same for H and W + # input: img: pytorch tensor, CHW or HW [0,1] + # output: CHW or HW [0,1] w/o round + need_squeeze = True if img.dim() == 2 else False + if need_squeeze: + img.unsqueeze_(0) + in_C, in_H, in_W = img.size() + out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) + kernel_width = 4 + kernel = 'cubic' + + # Return the desired dimension order for performing the resize. The + # strategy is to perform the resize first along the dimension with the + # smallest scale factor. + # Now we do not support this. + + # get weights and indices + weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( + in_H, out_H, scale, kernel, kernel_width, antialiasing) + weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( + in_W, out_W, scale, kernel, kernel_width, antialiasing) + # process H dimension + # symmetric copying + img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W) + img_aug.narrow(1, sym_len_Hs, in_H).copy_(img) + + sym_patch = img[:, :sym_len_Hs, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv) + + sym_patch = img[:, -sym_len_He:, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) + + out_1 = torch.FloatTensor(in_C, out_H, in_W) + kernel_width = weights_H.size(1) + for i in range(out_H): + idx = int(indices_H[i][0]) + for j in range(out_C): + out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i]) + + # process W dimension + # symmetric copying + out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We) + out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1) + + sym_patch = out_1[:, :, :sym_len_Ws] + inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(2, inv_idx) + out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv) + + sym_patch = out_1[:, :, -sym_len_We:] + inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(2, inv_idx) + out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) + + out_2 = torch.FloatTensor(in_C, out_H, out_W) + kernel_width = weights_W.size(1) + for i in range(out_W): + idx = int(indices_W[i][0]) + for j in range(out_C): + out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i]) + if need_squeeze: + out_2.squeeze_() + return out_2 + + +# -------------------------------------------- +# imresize for numpy image [0, 1] +# -------------------------------------------- +def imresize_np(img, scale, antialiasing=True): + # Now the scale should be the same for H and W + # input: img: Numpy, HWC or HW [0,1] + # output: HWC or HW [0,1] w/o round + img = torch.from_numpy(img) + need_squeeze = True if img.dim() == 2 else False + if need_squeeze: + img.unsqueeze_(2) + + in_H, in_W, in_C = img.size() + out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) + kernel_width = 4 + kernel = 'cubic' + + # Return the desired dimension order for performing the resize. The + # strategy is to perform the resize first along the dimension with the + # smallest scale factor. + # Now we do not support this. + + # get weights and indices + weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( + in_H, out_H, scale, kernel, kernel_width, antialiasing) + weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( + in_W, out_W, scale, kernel, kernel_width, antialiasing) + # process H dimension + # symmetric copying + img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C) + img_aug.narrow(0, sym_len_Hs, in_H).copy_(img) + + sym_patch = img[:sym_len_Hs, :, :] + inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(0, inv_idx) + img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv) + + sym_patch = img[-sym_len_He:, :, :] + inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(0, inv_idx) + img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) + + out_1 = torch.FloatTensor(out_H, in_W, in_C) + kernel_width = weights_H.size(1) + for i in range(out_H): + idx = int(indices_H[i][0]) + for j in range(out_C): + out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i]) + + # process W dimension + # symmetric copying + out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C) + out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1) + + sym_patch = out_1[:, :sym_len_Ws, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv) + + sym_patch = out_1[:, -sym_len_We:, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) + + out_2 = torch.FloatTensor(out_H, out_W, in_C) + kernel_width = weights_W.size(1) + for i in range(out_W): + idx = int(indices_W[i][0]) + for j in range(out_C): + out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i]) + if need_squeeze: + out_2.squeeze_() + + return out_2.numpy() + + +if __name__ == '__main__': + print('---') +# img = imread_uint('test.bmp', 3) +# img = uint2single(img) +# img_bicubic = imresize_np(img, 1/4) \ No newline at end of file diff --git a/ldm/modules/midas/__init__.py b/ldm/modules/midas/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/midas/api.py b/ldm/modules/midas/api.py new file mode 100644 index 0000000000000000000000000000000000000000..b58ebbffd942a2fc22264f0ab47e400c26b9f41c --- /dev/null +++ b/ldm/modules/midas/api.py @@ -0,0 +1,170 @@ +# based on https://github.com/isl-org/MiDaS + +import cv2 +import torch +import torch.nn as nn +from torchvision.transforms import Compose + +from ldm.modules.midas.midas.dpt_depth import DPTDepthModel +from ldm.modules.midas.midas.midas_net import MidasNet +from ldm.modules.midas.midas.midas_net_custom import MidasNet_small +from ldm.modules.midas.midas.transforms import Resize, NormalizeImage, PrepareForNet + + +ISL_PATHS = { + "dpt_large": "midas_models/dpt_large-midas-2f21e586.pt", + "dpt_hybrid": "midas_models/dpt_hybrid-midas-501f0c75.pt", + "midas_v21": "", + "midas_v21_small": "", +} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def load_midas_transform(model_type): + # https://github.com/isl-org/MiDaS/blob/master/run.py + # load transform only + if model_type == "dpt_large": # DPT-Large + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "dpt_hybrid": # DPT-Hybrid + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "midas_v21": + net_w, net_h = 384, 384 + resize_mode = "upper_bound" + normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) + + elif model_type == "midas_v21_small": + net_w, net_h = 256, 256 + resize_mode = "upper_bound" + normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) + + else: + assert False, f"model_type '{model_type}' not implemented, use: --model_type large" + + transform = Compose( + [ + Resize( + net_w, + net_h, + resize_target=None, + keep_aspect_ratio=True, + ensure_multiple_of=32, + resize_method=resize_mode, + image_interpolation_method=cv2.INTER_CUBIC, + ), + normalization, + PrepareForNet(), + ] + ) + + return transform + + +def load_model(model_type): + # https://github.com/isl-org/MiDaS/blob/master/run.py + # load network + model_path = ISL_PATHS[model_type] + if model_type == "dpt_large": # DPT-Large + model = DPTDepthModel( + path=model_path, + backbone="vitl16_384", + non_negative=True, + ) + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "dpt_hybrid": # DPT-Hybrid + model = DPTDepthModel( + path=model_path, + backbone="vitb_rn50_384", + non_negative=True, + ) + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "midas_v21": + model = MidasNet(model_path, non_negative=True) + net_w, net_h = 384, 384 + resize_mode = "upper_bound" + normalization = NormalizeImage( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] + ) + + elif model_type == "midas_v21_small": + model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True, + non_negative=True, blocks={'expand': True}) + net_w, net_h = 256, 256 + resize_mode = "upper_bound" + normalization = NormalizeImage( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] + ) + + else: + print(f"model_type '{model_type}' not implemented, use: --model_type large") + assert False + + transform = Compose( + [ + Resize( + net_w, + net_h, + resize_target=None, + keep_aspect_ratio=True, + ensure_multiple_of=32, + resize_method=resize_mode, + image_interpolation_method=cv2.INTER_CUBIC, + ), + normalization, + PrepareForNet(), + ] + ) + + return model.eval(), transform + + +class MiDaSInference(nn.Module): + MODEL_TYPES_TORCH_HUB = [ + "DPT_Large", + "DPT_Hybrid", + "MiDaS_small" + ] + MODEL_TYPES_ISL = [ + "dpt_large", + "dpt_hybrid", + "midas_v21", + "midas_v21_small", + ] + + def __init__(self, model_type): + super().__init__() + assert (model_type in self.MODEL_TYPES_ISL) + model, _ = load_model(model_type) + self.model = model + self.model.train = disabled_train + + def forward(self, x): + # x in 0..1 as produced by calling self.transform on a 0..1 float64 numpy array + # NOTE: we expect that the correct transform has been called during dataloading. + with torch.no_grad(): + prediction = self.model(x) + prediction = torch.nn.functional.interpolate( + prediction.unsqueeze(1), + size=x.shape[2:], + mode="bicubic", + align_corners=False, + ) + assert prediction.shape == (x.shape[0], 1, x.shape[2], x.shape[3]) + return prediction + diff --git a/ldm/modules/midas/midas/__init__.py b/ldm/modules/midas/midas/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/midas/midas/base_model.py b/ldm/modules/midas/midas/base_model.py new file mode 100644 index 0000000000000000000000000000000000000000..5cf430239b47ec5ec07531263f26f5c24a2311cd --- /dev/null +++ b/ldm/modules/midas/midas/base_model.py @@ -0,0 +1,16 @@ +import torch + + +class BaseModel(torch.nn.Module): + def load(self, path): + """Load model from file. + + Args: + path (str): file path + """ + parameters = torch.load(path, map_location=torch.device('cpu')) + + if "optimizer" in parameters: + parameters = parameters["model"] + + self.load_state_dict(parameters) diff --git a/ldm/modules/midas/midas/blocks.py b/ldm/modules/midas/midas/blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..2145d18fa98060a618536d9a64fe6589e9be4f78 --- /dev/null +++ b/ldm/modules/midas/midas/blocks.py @@ -0,0 +1,342 @@ +import torch +import torch.nn as nn + +from .vit import ( + _make_pretrained_vitb_rn50_384, + _make_pretrained_vitl16_384, + _make_pretrained_vitb16_384, + forward_vit, +) + +def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",): + if backbone == "vitl16_384": + pretrained = _make_pretrained_vitl16_384( + use_pretrained, hooks=hooks, use_readout=use_readout + ) + scratch = _make_scratch( + [256, 512, 1024, 1024], features, groups=groups, expand=expand + ) # ViT-L/16 - 85.0% Top1 (backbone) + elif backbone == "vitb_rn50_384": + pretrained = _make_pretrained_vitb_rn50_384( + use_pretrained, + hooks=hooks, + use_vit_only=use_vit_only, + use_readout=use_readout, + ) + scratch = _make_scratch( + [256, 512, 768, 768], features, groups=groups, expand=expand + ) # ViT-H/16 - 85.0% Top1 (backbone) + elif backbone == "vitb16_384": + pretrained = _make_pretrained_vitb16_384( + use_pretrained, hooks=hooks, use_readout=use_readout + ) + scratch = _make_scratch( + [96, 192, 384, 768], features, groups=groups, expand=expand + ) # ViT-B/16 - 84.6% Top1 (backbone) + elif backbone == "resnext101_wsl": + pretrained = _make_pretrained_resnext101_wsl(use_pretrained) + scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand) # efficientnet_lite3 + elif backbone == "efficientnet_lite3": + pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable) + scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand) # efficientnet_lite3 + else: + print(f"Backbone '{backbone}' not implemented") + assert False + + return pretrained, scratch + + +def _make_scratch(in_shape, out_shape, groups=1, expand=False): + scratch = nn.Module() + + out_shape1 = out_shape + out_shape2 = out_shape + out_shape3 = out_shape + out_shape4 = out_shape + if expand==True: + out_shape1 = out_shape + out_shape2 = out_shape*2 + out_shape3 = out_shape*4 + out_shape4 = out_shape*8 + + scratch.layer1_rn = nn.Conv2d( + in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer2_rn = nn.Conv2d( + in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer3_rn = nn.Conv2d( + in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer4_rn = nn.Conv2d( + in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + + return scratch + + +def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False): + efficientnet = torch.hub.load( + "rwightman/gen-efficientnet-pytorch", + "tf_efficientnet_lite3", + pretrained=use_pretrained, + exportable=exportable + ) + return _make_efficientnet_backbone(efficientnet) + + +def _make_efficientnet_backbone(effnet): + pretrained = nn.Module() + + pretrained.layer1 = nn.Sequential( + effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2] + ) + pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3]) + pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5]) + pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9]) + + return pretrained + + +def _make_resnet_backbone(resnet): + pretrained = nn.Module() + pretrained.layer1 = nn.Sequential( + resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1 + ) + + pretrained.layer2 = resnet.layer2 + pretrained.layer3 = resnet.layer3 + pretrained.layer4 = resnet.layer4 + + return pretrained + + +def _make_pretrained_resnext101_wsl(use_pretrained): + resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl") + return _make_resnet_backbone(resnet) + + + +class Interpolate(nn.Module): + """Interpolation module. + """ + + def __init__(self, scale_factor, mode, align_corners=False): + """Init. + + Args: + scale_factor (float): scaling + mode (str): interpolation mode + """ + super(Interpolate, self).__init__() + + self.interp = nn.functional.interpolate + self.scale_factor = scale_factor + self.mode = mode + self.align_corners = align_corners + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: interpolated data + """ + + x = self.interp( + x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners + ) + + return x + + +class ResidualConvUnit(nn.Module): + """Residual convolution module. + """ + + def __init__(self, features): + """Init. + + Args: + features (int): number of features + """ + super().__init__() + + self.conv1 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True + ) + + self.conv2 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True + ) + + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: output + """ + out = self.relu(x) + out = self.conv1(out) + out = self.relu(out) + out = self.conv2(out) + + return out + x + + +class FeatureFusionBlock(nn.Module): + """Feature fusion block. + """ + + def __init__(self, features): + """Init. + + Args: + features (int): number of features + """ + super(FeatureFusionBlock, self).__init__() + + self.resConfUnit1 = ResidualConvUnit(features) + self.resConfUnit2 = ResidualConvUnit(features) + + def forward(self, *xs): + """Forward pass. + + Returns: + tensor: output + """ + output = xs[0] + + if len(xs) == 2: + output += self.resConfUnit1(xs[1]) + + output = self.resConfUnit2(output) + + output = nn.functional.interpolate( + output, scale_factor=2, mode="bilinear", align_corners=True + ) + + return output + + + + +class ResidualConvUnit_custom(nn.Module): + """Residual convolution module. + """ + + def __init__(self, features, activation, bn): + """Init. + + Args: + features (int): number of features + """ + super().__init__() + + self.bn = bn + + self.groups=1 + + self.conv1 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups + ) + + self.conv2 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups + ) + + if self.bn==True: + self.bn1 = nn.BatchNorm2d(features) + self.bn2 = nn.BatchNorm2d(features) + + self.activation = activation + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: output + """ + + out = self.activation(x) + out = self.conv1(out) + if self.bn==True: + out = self.bn1(out) + + out = self.activation(out) + out = self.conv2(out) + if self.bn==True: + out = self.bn2(out) + + if self.groups > 1: + out = self.conv_merge(out) + + return self.skip_add.add(out, x) + + # return out + x + + +class FeatureFusionBlock_custom(nn.Module): + """Feature fusion block. + """ + + def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True): + """Init. + + Args: + features (int): number of features + """ + super(FeatureFusionBlock_custom, self).__init__() + + self.deconv = deconv + self.align_corners = align_corners + + self.groups=1 + + self.expand = expand + out_features = features + if self.expand==True: + out_features = features//2 + + self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1) + + self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn) + self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn) + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, *xs): + """Forward pass. + + Returns: + tensor: output + """ + output = xs[0] + + if len(xs) == 2: + res = self.resConfUnit1(xs[1]) + output = self.skip_add.add(output, res) + # output += res + + output = self.resConfUnit2(output) + + output = nn.functional.interpolate( + output, scale_factor=2, mode="bilinear", align_corners=self.align_corners + ) + + output = self.out_conv(output) + + return output + diff --git a/ldm/modules/midas/midas/dpt_depth.py b/ldm/modules/midas/midas/dpt_depth.py new file mode 100644 index 0000000000000000000000000000000000000000..4e9aab5d2767dffea39da5b3f30e2798688216f1 --- /dev/null +++ b/ldm/modules/midas/midas/dpt_depth.py @@ -0,0 +1,109 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F + +from .base_model import BaseModel +from .blocks import ( + FeatureFusionBlock, + FeatureFusionBlock_custom, + Interpolate, + _make_encoder, + forward_vit, +) + + +def _make_fusion_block(features, use_bn): + return FeatureFusionBlock_custom( + features, + nn.ReLU(False), + deconv=False, + bn=use_bn, + expand=False, + align_corners=True, + ) + + +class DPT(BaseModel): + def __init__( + self, + head, + features=256, + backbone="vitb_rn50_384", + readout="project", + channels_last=False, + use_bn=False, + ): + + super(DPT, self).__init__() + + self.channels_last = channels_last + + hooks = { + "vitb_rn50_384": [0, 1, 8, 11], + "vitb16_384": [2, 5, 8, 11], + "vitl16_384": [5, 11, 17, 23], + } + + # Instantiate backbone and reassemble blocks + self.pretrained, self.scratch = _make_encoder( + backbone, + features, + False, # Set to true of you want to train from scratch, uses ImageNet weights + groups=1, + expand=False, + exportable=False, + hooks=hooks[backbone], + use_readout=readout, + ) + + self.scratch.refinenet1 = _make_fusion_block(features, use_bn) + self.scratch.refinenet2 = _make_fusion_block(features, use_bn) + self.scratch.refinenet3 = _make_fusion_block(features, use_bn) + self.scratch.refinenet4 = _make_fusion_block(features, use_bn) + + self.scratch.output_conv = head + + + def forward(self, x): + if self.channels_last == True: + x.contiguous(memory_format=torch.channels_last) + + layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return out + + +class DPTDepthModel(DPT): + def __init__(self, path=None, non_negative=True, **kwargs): + features = kwargs["features"] if "features" in kwargs else 256 + + head = nn.Sequential( + nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1), + Interpolate(scale_factor=2, mode="bilinear", align_corners=True), + nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1), + nn.ReLU(True), + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + nn.Identity(), + ) + + super().__init__(head, **kwargs) + + if path is not None: + self.load(path) + + def forward(self, x): + return super().forward(x).squeeze(dim=1) + diff --git a/ldm/modules/midas/midas/midas_net.py b/ldm/modules/midas/midas/midas_net.py new file mode 100644 index 0000000000000000000000000000000000000000..8a954977800b0a0f48807e80fa63041910e33c1f --- /dev/null +++ b/ldm/modules/midas/midas/midas_net.py @@ -0,0 +1,76 @@ +"""MidashNet: Network for monocular depth estimation trained by mixing several datasets. +This file contains code that is adapted from +https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py +""" +import torch +import torch.nn as nn + +from .base_model import BaseModel +from .blocks import FeatureFusionBlock, Interpolate, _make_encoder + + +class MidasNet(BaseModel): + """Network for monocular depth estimation. + """ + + def __init__(self, path=None, features=256, non_negative=True): + """Init. + + Args: + path (str, optional): Path to saved model. Defaults to None. + features (int, optional): Number of features. Defaults to 256. + backbone (str, optional): Backbone network for encoder. Defaults to resnet50 + """ + print("Loading weights: ", path) + + super(MidasNet, self).__init__() + + use_pretrained = False if path is None else True + + self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained) + + self.scratch.refinenet4 = FeatureFusionBlock(features) + self.scratch.refinenet3 = FeatureFusionBlock(features) + self.scratch.refinenet2 = FeatureFusionBlock(features) + self.scratch.refinenet1 = FeatureFusionBlock(features) + + self.scratch.output_conv = nn.Sequential( + nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1), + Interpolate(scale_factor=2, mode="bilinear"), + nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1), + nn.ReLU(True), + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + ) + + if path: + self.load(path) + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input data (image) + + Returns: + tensor: depth + """ + + layer_1 = self.pretrained.layer1(x) + layer_2 = self.pretrained.layer2(layer_1) + layer_3 = self.pretrained.layer3(layer_2) + layer_4 = self.pretrained.layer4(layer_3) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return torch.squeeze(out, dim=1) diff --git a/ldm/modules/midas/midas/midas_net_custom.py b/ldm/modules/midas/midas/midas_net_custom.py new file mode 100644 index 0000000000000000000000000000000000000000..50e4acb5e53d5fabefe3dde16ab49c33c2b7797c --- /dev/null +++ b/ldm/modules/midas/midas/midas_net_custom.py @@ -0,0 +1,128 @@ +"""MidashNet: Network for monocular depth estimation trained by mixing several datasets. +This file contains code that is adapted from +https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py +""" +import torch +import torch.nn as nn + +from .base_model import BaseModel +from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder + + +class MidasNet_small(BaseModel): + """Network for monocular depth estimation. + """ + + def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True, + blocks={'expand': True}): + """Init. + + Args: + path (str, optional): Path to saved model. Defaults to None. + features (int, optional): Number of features. Defaults to 256. + backbone (str, optional): Backbone network for encoder. Defaults to resnet50 + """ + print("Loading weights: ", path) + + super(MidasNet_small, self).__init__() + + use_pretrained = False if path else True + + self.channels_last = channels_last + self.blocks = blocks + self.backbone = backbone + + self.groups = 1 + + features1=features + features2=features + features3=features + features4=features + self.expand = False + if "expand" in self.blocks and self.blocks['expand'] == True: + self.expand = True + features1=features + features2=features*2 + features3=features*4 + features4=features*8 + + self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable) + + self.scratch.activation = nn.ReLU(False) + + self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners) + + + self.scratch.output_conv = nn.Sequential( + nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups), + Interpolate(scale_factor=2, mode="bilinear"), + nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1), + self.scratch.activation, + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + nn.Identity(), + ) + + if path: + self.load(path) + + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input data (image) + + Returns: + tensor: depth + """ + if self.channels_last==True: + print("self.channels_last = ", self.channels_last) + x.contiguous(memory_format=torch.channels_last) + + + layer_1 = self.pretrained.layer1(x) + layer_2 = self.pretrained.layer2(layer_1) + layer_3 = self.pretrained.layer3(layer_2) + layer_4 = self.pretrained.layer4(layer_3) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return torch.squeeze(out, dim=1) + + + +def fuse_model(m): + prev_previous_type = nn.Identity() + prev_previous_name = '' + previous_type = nn.Identity() + previous_name = '' + for name, module in m.named_modules(): + if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU: + # print("FUSED ", prev_previous_name, previous_name, name) + torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True) + elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d: + # print("FUSED ", prev_previous_name, previous_name) + torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True) + # elif previous_type == nn.Conv2d and type(module) == nn.ReLU: + # print("FUSED ", previous_name, name) + # torch.quantization.fuse_modules(m, [previous_name, name], inplace=True) + + prev_previous_type = previous_type + prev_previous_name = previous_name + previous_type = type(module) + previous_name = name \ No newline at end of file diff --git a/ldm/modules/midas/midas/transforms.py b/ldm/modules/midas/midas/transforms.py new file mode 100644 index 0000000000000000000000000000000000000000..350cbc11662633ad7f8968eb10be2e7de6e384e9 --- /dev/null +++ b/ldm/modules/midas/midas/transforms.py @@ -0,0 +1,234 @@ +import numpy as np +import cv2 +import math + + +def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA): + """Rezise the sample to ensure the given size. Keeps aspect ratio. + + Args: + sample (dict): sample + size (tuple): image size + + Returns: + tuple: new size + """ + shape = list(sample["disparity"].shape) + + if shape[0] >= size[0] and shape[1] >= size[1]: + return sample + + scale = [0, 0] + scale[0] = size[0] / shape[0] + scale[1] = size[1] / shape[1] + + scale = max(scale) + + shape[0] = math.ceil(scale * shape[0]) + shape[1] = math.ceil(scale * shape[1]) + + # resize + sample["image"] = cv2.resize( + sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method + ) + + sample["disparity"] = cv2.resize( + sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST + ) + sample["mask"] = cv2.resize( + sample["mask"].astype(np.float32), + tuple(shape[::-1]), + interpolation=cv2.INTER_NEAREST, + ) + sample["mask"] = sample["mask"].astype(bool) + + return tuple(shape) + + +class Resize(object): + """Resize sample to given size (width, height). + """ + + def __init__( + self, + width, + height, + resize_target=True, + keep_aspect_ratio=False, + ensure_multiple_of=1, + resize_method="lower_bound", + image_interpolation_method=cv2.INTER_AREA, + ): + """Init. + + Args: + width (int): desired output width + height (int): desired output height + resize_target (bool, optional): + True: Resize the full sample (image, mask, target). + False: Resize image only. + Defaults to True. + keep_aspect_ratio (bool, optional): + True: Keep the aspect ratio of the input sample. + Output sample might not have the given width and height, and + resize behaviour depends on the parameter 'resize_method'. + Defaults to False. + ensure_multiple_of (int, optional): + Output width and height is constrained to be multiple of this parameter. + Defaults to 1. + resize_method (str, optional): + "lower_bound": Output will be at least as large as the given size. + "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.) + "minimal": Scale as least as possible. (Output size might be smaller than given size.) + Defaults to "lower_bound". + """ + self.__width = width + self.__height = height + + self.__resize_target = resize_target + self.__keep_aspect_ratio = keep_aspect_ratio + self.__multiple_of = ensure_multiple_of + self.__resize_method = resize_method + self.__image_interpolation_method = image_interpolation_method + + def constrain_to_multiple_of(self, x, min_val=0, max_val=None): + y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int) + + if max_val is not None and y > max_val: + y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int) + + if y < min_val: + y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int) + + return y + + def get_size(self, width, height): + # determine new height and width + scale_height = self.__height / height + scale_width = self.__width / width + + if self.__keep_aspect_ratio: + if self.__resize_method == "lower_bound": + # scale such that output size is lower bound + if scale_width > scale_height: + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + elif self.__resize_method == "upper_bound": + # scale such that output size is upper bound + if scale_width < scale_height: + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + elif self.__resize_method == "minimal": + # scale as least as possbile + if abs(1 - scale_width) < abs(1 - scale_height): + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + else: + raise ValueError( + f"resize_method {self.__resize_method} not implemented" + ) + + if self.__resize_method == "lower_bound": + new_height = self.constrain_to_multiple_of( + scale_height * height, min_val=self.__height + ) + new_width = self.constrain_to_multiple_of( + scale_width * width, min_val=self.__width + ) + elif self.__resize_method == "upper_bound": + new_height = self.constrain_to_multiple_of( + scale_height * height, max_val=self.__height + ) + new_width = self.constrain_to_multiple_of( + scale_width * width, max_val=self.__width + ) + elif self.__resize_method == "minimal": + new_height = self.constrain_to_multiple_of(scale_height * height) + new_width = self.constrain_to_multiple_of(scale_width * width) + else: + raise ValueError(f"resize_method {self.__resize_method} not implemented") + + return (new_width, new_height) + + def __call__(self, sample): + width, height = self.get_size( + sample["image"].shape[1], sample["image"].shape[0] + ) + + # resize sample + sample["image"] = cv2.resize( + sample["image"], + (width, height), + interpolation=self.__image_interpolation_method, + ) + + if self.__resize_target: + if "disparity" in sample: + sample["disparity"] = cv2.resize( + sample["disparity"], + (width, height), + interpolation=cv2.INTER_NEAREST, + ) + + if "depth" in sample: + sample["depth"] = cv2.resize( + sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST + ) + + sample["mask"] = cv2.resize( + sample["mask"].astype(np.float32), + (width, height), + interpolation=cv2.INTER_NEAREST, + ) + sample["mask"] = sample["mask"].astype(bool) + + return sample + + +class NormalizeImage(object): + """Normlize image by given mean and std. + """ + + def __init__(self, mean, std): + self.__mean = mean + self.__std = std + + def __call__(self, sample): + sample["image"] = (sample["image"] - self.__mean) / self.__std + + return sample + + +class PrepareForNet(object): + """Prepare sample for usage as network input. + """ + + def __init__(self): + pass + + def __call__(self, sample): + image = np.transpose(sample["image"], (2, 0, 1)) + sample["image"] = np.ascontiguousarray(image).astype(np.float32) + + if "mask" in sample: + sample["mask"] = sample["mask"].astype(np.float32) + sample["mask"] = np.ascontiguousarray(sample["mask"]) + + if "disparity" in sample: + disparity = sample["disparity"].astype(np.float32) + sample["disparity"] = np.ascontiguousarray(disparity) + + if "depth" in sample: + depth = sample["depth"].astype(np.float32) + sample["depth"] = np.ascontiguousarray(depth) + + return sample diff --git a/ldm/modules/midas/midas/vit.py b/ldm/modules/midas/midas/vit.py new file mode 100644 index 0000000000000000000000000000000000000000..ea46b1be88b261b0dec04f3da0256f5f66f88a74 --- /dev/null +++ b/ldm/modules/midas/midas/vit.py @@ -0,0 +1,491 @@ +import torch +import torch.nn as nn +import timm +import types +import math +import torch.nn.functional as F + + +class Slice(nn.Module): + def __init__(self, start_index=1): + super(Slice, self).__init__() + self.start_index = start_index + + def forward(self, x): + return x[:, self.start_index :] + + +class AddReadout(nn.Module): + def __init__(self, start_index=1): + super(AddReadout, self).__init__() + self.start_index = start_index + + def forward(self, x): + if self.start_index == 2: + readout = (x[:, 0] + x[:, 1]) / 2 + else: + readout = x[:, 0] + return x[:, self.start_index :] + readout.unsqueeze(1) + + +class ProjectReadout(nn.Module): + def __init__(self, in_features, start_index=1): + super(ProjectReadout, self).__init__() + self.start_index = start_index + + self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU()) + + def forward(self, x): + readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :]) + features = torch.cat((x[:, self.start_index :], readout), -1) + + return self.project(features) + + +class Transpose(nn.Module): + def __init__(self, dim0, dim1): + super(Transpose, self).__init__() + self.dim0 = dim0 + self.dim1 = dim1 + + def forward(self, x): + x = x.transpose(self.dim0, self.dim1) + return x + + +def forward_vit(pretrained, x): + b, c, h, w = x.shape + + glob = pretrained.model.forward_flex(x) + + layer_1 = pretrained.activations["1"] + layer_2 = pretrained.activations["2"] + layer_3 = pretrained.activations["3"] + layer_4 = pretrained.activations["4"] + + layer_1 = pretrained.act_postprocess1[0:2](layer_1) + layer_2 = pretrained.act_postprocess2[0:2](layer_2) + layer_3 = pretrained.act_postprocess3[0:2](layer_3) + layer_4 = pretrained.act_postprocess4[0:2](layer_4) + + unflatten = nn.Sequential( + nn.Unflatten( + 2, + torch.Size( + [ + h // pretrained.model.patch_size[1], + w // pretrained.model.patch_size[0], + ] + ), + ) + ) + + if layer_1.ndim == 3: + layer_1 = unflatten(layer_1) + if layer_2.ndim == 3: + layer_2 = unflatten(layer_2) + if layer_3.ndim == 3: + layer_3 = unflatten(layer_3) + if layer_4.ndim == 3: + layer_4 = unflatten(layer_4) + + layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1) + layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2) + layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3) + layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4) + + return layer_1, layer_2, layer_3, layer_4 + + +def _resize_pos_embed(self, posemb, gs_h, gs_w): + posemb_tok, posemb_grid = ( + posemb[:, : self.start_index], + posemb[0, self.start_index :], + ) + + gs_old = int(math.sqrt(len(posemb_grid))) + + posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2) + posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear") + posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1) + + posemb = torch.cat([posemb_tok, posemb_grid], dim=1) + + return posemb + + +def forward_flex(self, x): + b, c, h, w = x.shape + + pos_embed = self._resize_pos_embed( + self.pos_embed, h // self.patch_size[1], w // self.patch_size[0] + ) + + B = x.shape[0] + + if hasattr(self.patch_embed, "backbone"): + x = self.patch_embed.backbone(x) + if isinstance(x, (list, tuple)): + x = x[-1] # last feature if backbone outputs list/tuple of features + + x = self.patch_embed.proj(x).flatten(2).transpose(1, 2) + + if getattr(self, "dist_token", None) is not None: + cls_tokens = self.cls_token.expand( + B, -1, -1 + ) # stole cls_tokens impl from Phil Wang, thanks + dist_token = self.dist_token.expand(B, -1, -1) + x = torch.cat((cls_tokens, dist_token, x), dim=1) + else: + 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 + pos_embed + x = self.pos_drop(x) + + for blk in self.blocks: + x = blk(x) + + x = self.norm(x) + + return x + + +activations = {} + + +def get_activation(name): + def hook(model, input, output): + activations[name] = output + + return hook + + +def get_readout_oper(vit_features, features, use_readout, start_index=1): + if use_readout == "ignore": + readout_oper = [Slice(start_index)] * len(features) + elif use_readout == "add": + readout_oper = [AddReadout(start_index)] * len(features) + elif use_readout == "project": + readout_oper = [ + ProjectReadout(vit_features, start_index) for out_feat in features + ] + else: + assert ( + False + ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'" + + return readout_oper + + +def _make_vit_b16_backbone( + model, + features=[96, 192, 384, 768], + size=[384, 384], + hooks=[2, 5, 8, 11], + vit_features=768, + use_readout="ignore", + start_index=1, +): + pretrained = nn.Module() + + pretrained.model = model + pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) + pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) + pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) + pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) + + pretrained.activations = activations + + readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) + + # 32, 48, 136, 384 + pretrained.act_postprocess1 = nn.Sequential( + readout_oper[0], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[0], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[0], + out_channels=features[0], + kernel_size=4, + stride=4, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess2 = nn.Sequential( + readout_oper[1], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[1], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[1], + out_channels=features[1], + kernel_size=2, + stride=2, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess3 = nn.Sequential( + readout_oper[2], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[2], + kernel_size=1, + stride=1, + padding=0, + ), + ) + + pretrained.act_postprocess4 = nn.Sequential( + readout_oper[3], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[3], + kernel_size=1, + stride=1, + padding=0, + ), + nn.Conv2d( + in_channels=features[3], + out_channels=features[3], + kernel_size=3, + stride=2, + padding=1, + ), + ) + + pretrained.model.start_index = start_index + pretrained.model.patch_size = [16, 16] + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) + pretrained.model._resize_pos_embed = types.MethodType( + _resize_pos_embed, pretrained.model + ) + + return pretrained + + +def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_large_patch16_384", pretrained=pretrained) + + hooks = [5, 11, 17, 23] if hooks == None else hooks + return _make_vit_b16_backbone( + model, + features=[256, 512, 1024, 1024], + hooks=hooks, + vit_features=1024, + use_readout=use_readout, + ) + + +def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_base_patch16_384", pretrained=pretrained) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout + ) + + +def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout + ) + + +def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model( + "vit_deit_base_distilled_patch16_384", pretrained=pretrained + ) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, + features=[96, 192, 384, 768], + hooks=hooks, + use_readout=use_readout, + start_index=2, + ) + + +def _make_vit_b_rn50_backbone( + model, + features=[256, 512, 768, 768], + size=[384, 384], + hooks=[0, 1, 8, 11], + vit_features=768, + use_vit_only=False, + use_readout="ignore", + start_index=1, +): + pretrained = nn.Module() + + pretrained.model = model + + if use_vit_only == True: + pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) + pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) + else: + pretrained.model.patch_embed.backbone.stages[0].register_forward_hook( + get_activation("1") + ) + pretrained.model.patch_embed.backbone.stages[1].register_forward_hook( + get_activation("2") + ) + + pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) + pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) + + pretrained.activations = activations + + readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) + + if use_vit_only == True: + pretrained.act_postprocess1 = nn.Sequential( + readout_oper[0], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[0], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[0], + out_channels=features[0], + kernel_size=4, + stride=4, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess2 = nn.Sequential( + readout_oper[1], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[1], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[1], + out_channels=features[1], + kernel_size=2, + stride=2, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + else: + pretrained.act_postprocess1 = nn.Sequential( + nn.Identity(), nn.Identity(), nn.Identity() + ) + pretrained.act_postprocess2 = nn.Sequential( + nn.Identity(), nn.Identity(), nn.Identity() + ) + + pretrained.act_postprocess3 = nn.Sequential( + readout_oper[2], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[2], + kernel_size=1, + stride=1, + padding=0, + ), + ) + + pretrained.act_postprocess4 = nn.Sequential( + readout_oper[3], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[3], + kernel_size=1, + stride=1, + padding=0, + ), + nn.Conv2d( + in_channels=features[3], + out_channels=features[3], + kernel_size=3, + stride=2, + padding=1, + ), + ) + + pretrained.model.start_index = start_index + pretrained.model.patch_size = [16, 16] + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model._resize_pos_embed = types.MethodType( + _resize_pos_embed, pretrained.model + ) + + return pretrained + + +def _make_pretrained_vitb_rn50_384( + pretrained, use_readout="ignore", hooks=None, use_vit_only=False +): + model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained) + + hooks = [0, 1, 8, 11] if hooks == None else hooks + return _make_vit_b_rn50_backbone( + model, + features=[256, 512, 768, 768], + size=[384, 384], + hooks=hooks, + use_vit_only=use_vit_only, + use_readout=use_readout, + ) diff --git a/ldm/modules/midas/utils.py b/ldm/modules/midas/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..9a9d3b5b66370fa98da9e067ba53ead848ea9a59 --- /dev/null +++ b/ldm/modules/midas/utils.py @@ -0,0 +1,189 @@ +"""Utils for monoDepth.""" +import sys +import re +import numpy as np +import cv2 +import torch + + +def read_pfm(path): + """Read pfm file. + + Args: + path (str): path to file + + Returns: + tuple: (data, scale) + """ + with open(path, "rb") as file: + + color = None + width = None + height = None + scale = None + endian = None + + header = file.readline().rstrip() + if header.decode("ascii") == "PF": + color = True + elif header.decode("ascii") == "Pf": + color = False + else: + raise Exception("Not a PFM file: " + path) + + dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii")) + if dim_match: + width, height = list(map(int, dim_match.groups())) + else: + raise Exception("Malformed PFM header.") + + scale = float(file.readline().decode("ascii").rstrip()) + if scale < 0: + # little-endian + endian = "<" + scale = -scale + else: + # big-endian + endian = ">" + + data = np.fromfile(file, endian + "f") + shape = (height, width, 3) if color else (height, width) + + data = np.reshape(data, shape) + data = np.flipud(data) + + return data, scale + + +def write_pfm(path, image, scale=1): + """Write pfm file. + + Args: + path (str): pathto file + image (array): data + scale (int, optional): Scale. Defaults to 1. + """ + + with open(path, "wb") as file: + color = None + + if image.dtype.name != "float32": + raise Exception("Image dtype must be float32.") + + image = np.flipud(image) + + if len(image.shape) == 3 and image.shape[2] == 3: # color image + color = True + elif ( + len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1 + ): # greyscale + color = False + else: + raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.") + + file.write("PF\n" if color else "Pf\n".encode()) + file.write("%d %d\n".encode() % (image.shape[1], image.shape[0])) + + endian = image.dtype.byteorder + + if endian == "<" or endian == "=" and sys.byteorder == "little": + scale = -scale + + file.write("%f\n".encode() % scale) + + image.tofile(file) + + +def read_image(path): + """Read image and output RGB image (0-1). + + Args: + path (str): path to file + + Returns: + array: RGB image (0-1) + """ + img = cv2.imread(path) + + if img.ndim == 2: + img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) + + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0 + + return img + + +def resize_image(img): + """Resize image and make it fit for network. + + Args: + img (array): image + + Returns: + tensor: data ready for network + """ + height_orig = img.shape[0] + width_orig = img.shape[1] + + if width_orig > height_orig: + scale = width_orig / 384 + else: + scale = height_orig / 384 + + height = (np.ceil(height_orig / scale / 32) * 32).astype(int) + width = (np.ceil(width_orig / scale / 32) * 32).astype(int) + + img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA) + + img_resized = ( + torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float() + ) + img_resized = img_resized.unsqueeze(0) + + return img_resized + + +def resize_depth(depth, width, height): + """Resize depth map and bring to CPU (numpy). + + Args: + depth (tensor): depth + width (int): image width + height (int): image height + + Returns: + array: processed depth + """ + depth = torch.squeeze(depth[0, :, :, :]).to("cpu") + + depth_resized = cv2.resize( + depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC + ) + + return depth_resized + +def write_depth(path, depth, bits=1): + """Write depth map to pfm and png file. + + Args: + path (str): filepath without extension + depth (array): depth + """ + write_pfm(path + ".pfm", depth.astype(np.float32)) + + depth_min = depth.min() + depth_max = depth.max() + + max_val = (2**(8*bits))-1 + + if depth_max - depth_min > np.finfo("float").eps: + out = max_val * (depth - depth_min) / (depth_max - depth_min) + else: + out = np.zeros(depth.shape, dtype=depth.type) + + if bits == 1: + cv2.imwrite(path + ".png", out.astype("uint8")) + elif bits == 2: + cv2.imwrite(path + ".png", out.astype("uint16")) + + return diff --git a/ldm/util.py b/ldm/util.py new file mode 100644 index 0000000000000000000000000000000000000000..45cb050ece6f401a22dde098ce3f1ff663c5eb6a --- /dev/null +++ b/ldm/util.py @@ -0,0 +1,197 @@ +import importlib + +import torch +from torch import optim +import numpy as np + +from inspect import isfunction +from PIL import Image, ImageDraw, ImageFont + + +def log_txt_as_img(wh, xc, size=10): + # wh a tuple of (width, height) + # xc a list of captions to plot + b = len(xc) + txts = list() + for bi in range(b): + txt = Image.new("RGB", wh, color="white") + draw = ImageDraw.Draw(txt) + font = ImageFont.truetype('font/DejaVuSans.ttf', size=size) + nc = int(40 * (wh[0] / 256)) + lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc)) + + try: + draw.text((0, 0), lines, fill="black", font=font) + except UnicodeEncodeError: + print("Cant encode string for logging. Skipping.") + + txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0 + txts.append(txt) + txts = np.stack(txts) + txts = torch.tensor(txts) + return txts + + +def ismap(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] > 3) + + +def isimage(x): + if not isinstance(x,torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1) + + +def exists(x): + return x is not None + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def mean_flat(tensor): + """ + https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86 + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def count_params(model, verbose=False): + total_params = sum(p.numel() for p in model.parameters()) + if verbose: + print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.") + return total_params + + +def instantiate_from_config(config): + if not "target" in config: + if config == '__is_first_stage__': + return None + elif config == "__is_unconditional__": + return None + raise KeyError("Expected key `target` to instantiate.") + return get_obj_from_str(config["target"])(**config.get("params", dict())) + + +def get_obj_from_str(string, reload=False): + module, cls = string.rsplit(".", 1) + if reload: + module_imp = importlib.import_module(module) + importlib.reload(module_imp) + return getattr(importlib.import_module(module, package=None), cls) + + +class AdamWwithEMAandWings(optim.Optimizer): + # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298 + def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8, # TODO: check hyperparameters before using + weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999, # ema decay to match previous code + ema_power=1., param_names=()): + """AdamW that saves EMA versions of the parameters.""" + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0.0 <= weight_decay: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + if not 0.0 <= ema_decay <= 1.0: + raise ValueError("Invalid ema_decay value: {}".format(ema_decay)) + defaults = dict(lr=lr, betas=betas, eps=eps, + weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay, + ema_power=ema_power, param_names=param_names) + super().__init__(params, defaults) + + def __setstate__(self, state): + super().__setstate__(state) + for group in self.param_groups: + group.setdefault('amsgrad', False) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + params_with_grad = [] + grads = [] + exp_avgs = [] + exp_avg_sqs = [] + ema_params_with_grad = [] + state_sums = [] + max_exp_avg_sqs = [] + state_steps = [] + amsgrad = group['amsgrad'] + beta1, beta2 = group['betas'] + ema_decay = group['ema_decay'] + ema_power = group['ema_power'] + + for p in group['params']: + if p.grad is None: + continue + params_with_grad.append(p) + if p.grad.is_sparse: + raise RuntimeError('AdamW does not support sparse gradients') + grads.append(p.grad) + + state = self.state[p] + + # State initialization + if len(state) == 0: + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of squared gradient values + state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of parameter values + state['param_exp_avg'] = p.detach().float().clone() + + exp_avgs.append(state['exp_avg']) + exp_avg_sqs.append(state['exp_avg_sq']) + ema_params_with_grad.append(state['param_exp_avg']) + + if amsgrad: + max_exp_avg_sqs.append(state['max_exp_avg_sq']) + + # update the steps for each param group update + state['step'] += 1 + # record the step after step update + state_steps.append(state['step']) + + optim._functional.adamw(params_with_grad, + grads, + exp_avgs, + exp_avg_sqs, + max_exp_avg_sqs, + state_steps, + amsgrad=amsgrad, + beta1=beta1, + beta2=beta2, + lr=group['lr'], + weight_decay=group['weight_decay'], + eps=group['eps'], + maximize=False) + + cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power) + for param, ema_param in zip(params_with_grad, ema_params_with_grad): + ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay) + + return loss \ No newline at end of file diff --git a/models/cldm_v21.yaml b/models/cldm_v21.yaml new file mode 100644 index 0000000000000000000000000000000000000000..fc65193647e476e108fce5977f11250d55919106 --- /dev/null +++ b/models/cldm_v21.yaml @@ -0,0 +1,85 @@ +model: + target: cldm.cldm.ControlLDM + params: + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + control_key: "hint" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn + monitor: val/loss_simple_ema + scale_factor: 0.18215 + use_ema: False + only_mid_control: False + + control_stage_config: + target: cldm.cldm.ControlNet + params: + use_checkpoint: True + image_size: 32 # unused + in_channels: 4 + hint_channels: 3 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + unet_config: + target: cldm.cldm.ControlledUnetModel + params: + use_checkpoint: True + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/sam2edit.py b/sam2edit.py deleted file mode 100644 index 1ac319d8a1b1753256fe69e9d7321d57a387e0c0..0000000000000000000000000000000000000000 --- a/sam2edit.py +++ /dev/null @@ -1,28 +0,0 @@ -# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 -import os -import gradio as gr -from diffusers.utils import load_image -from sam2edit_lora import EditAnythingLoraModel, config_dict -from sam2edit_demo import create_demo_template -from huggingface_hub import hf_hub_download, snapshot_download - - -def create_demo(process, process_image_click=None): - - examples = None - INFO = f''' - ## EditAnything https://github.com/sail-sg/EditAnything - ''' - WARNING_INFO = None - - demo = create_demo_template(process, process_image_click, examples=examples, - INFO=INFO, WARNING_INFO=WARNING_INFO, enable_auto_prompt_default=True) - return demo - - -if __name__ == '__main__': - model = EditAnythingLoraModel(base_model_path="stabilityai/stable-diffusion-2", - controlmodel_name='LAION Pretrained(v0-4)-SD21', extra_inpaint=False, - lora_model_path=None, use_blip=True) - demo = create_demo(model.process, model.process_image_click) - demo.queue().launch(server_name='0.0.0.0') diff --git a/sam2edit_beauty.py b/sam2edit_beauty.py deleted file mode 100644 index 07e2448aa3950b8628e2bea6ad052a9b61101b31..0000000000000000000000000000000000000000 --- a/sam2edit_beauty.py +++ /dev/null @@ -1,44 +0,0 @@ -# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 -import os -import gradio as gr -from diffusers.utils import load_image -from sam2edit_lora import EditAnythingLoraModel, config_dict -from sam2edit_demo import create_demo_template -from huggingface_hub import hf_hub_download, snapshot_download - - -def create_demo(process, process_image_click=None): - - examples = [ - ["dudou,1girl, beautiful face, solo, candle, brown hair, long hair, ,ulzzang-6500-v1.1,(raw photo:1.2),((photorealistic:1.4))best quality ,masterpiece, illustration, an extremely delicate and beautiful, extremely detailed ,CG ,unity ,8k wallpaper, Amazing, finely detail, masterpiece,best quality,official art,extremely detailed CG unity 8k wallpaper,absurdres, incredibly absurdres, huge filesize, ultra-detailed, highres, extremely detailed,beautiful detailed girl, extremely detailed eyes and face, beautiful detailed eyes,cinematic lighting,1girl,see-through,looking at viewer,full body,full-body shot,outdoors,arms behind back,(chinese clothes) ", - "(((mole))),sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, bad anatomy,(long hair:1.4),DeepNegative,(fat:1.2),facing away, looking away,tilted head, lowres,bad anatomy,bad hands, text, error, missing fingers,extra digit, fewer digits, cropped, worstquality, low quality, normal quality,jpegartifacts,signature, watermark, username,blurry,bad feet,cropped,poorly drawn hands,poorly drawn face,mutation,deformed,worst quality,low quality,normal quality,jpeg artifacts,signature,watermark,extra fingers,fewer digits,extra limbs,extra arms,extra legs,malformed limbs,fused fingers,too many fingers,long neck,cross-eyed,mutated hands,polar lowres,bad body,bad proportions,gross proportions,text,error,missing fingers,missing arms,missing legs,extra digit, extra arms, extra leg, extra foot,(freckles),(mole:2)", 5], - ["best quality, ultra high res, (photorealistic:1.4), (detailed beautiful girl:1.4), (medium breasts:0.8), looking_at_viewer, Detailed facial details, beautiful detailed eyes, (multicolored|blue|pink hair: 1.2), green eyes, slender, haunting smile, (makeup:0.3), red lips, , highly detailed clothes, (ulzzang-6500-v1.1:0.3)", - "EasyNegative, paintings, sketches, ugly, 3d, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, manboobs, backlight,(ugly:1.3), (duplicate:1.3), (morbid:1.2), (mutilated:1.2), (tranny:1.3), mutated hands, (poorly drawn hands:1.3), blurry, (bad anatomy:1.2), (bad proportions:1.3), extra limbs, (disfigured:1.3), (more than 2 nipples:1.3), (more than 1 navel:1.3), (missing arms:1.3), (extra legs:1.3), (fused fingers:1.6), (too many fingers:1.6), (unclear eyes:1.3), bad hands, missing fingers, extra digit, (futa:1.1), bad body, double navel, mutad arms, hused arms, (puffy nipples, dark areolae, dark nipples, rei no himo, inverted nipples, long nipples), NG_DeepNegative_V1_75t, pubic hair, fat rolls, obese, bad-picture-chill-75v", 8], - ["best quality, ultra high res, (photorealistic:1.4), (detailed beautiful girl:1.4), (medium breasts:0.8), looking_at_viewer, Detailed facial details, beautiful detailed eyes, (blue|pink hair), green eyes, slender, smile, (makeup:0.4), red lips, (full body, sitting, beach), , highly detailed clothes, (ulzzang-6500-v1.1:0.3)", - "asyNegative, paintings, sketches, ugly, 3d, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, manboobs, backlight,(ugly:1.3), (duplicate:1.3), (morbid:1.2), (mutilated:1.2), (tranny:1.3), mutated hands, (poorly drawn hands:1.3), blurry, (bad anatomy:1.2), (bad proportions:1.3), extra limbs, (disfigured:1.3), (more than 2 nipples:1.3), (more than 1 navel:1.3), (missing arms:1.3), (extra legs:1.3), (fused fingers:1.6), (too many fingers:1.6), (unclear eyes:1.3), bad hands, missing fingers, extra digit, (futa:1.1), bad body, double navel, mutad arms, hused arms, (puffy nipples, dark areolae, dark nipples, rei no himo, inverted nipples, long nipples), NG_DeepNegative_V1_75t, pubic hair, fat rolls, obese, bad-picture-chill-75v", 7], - ["mix4, whole body shot, ((8k, RAW photo, highest quality, masterpiece), High detail RAW color photo professional close-up photo, shy expression, cute, beautiful detailed girl, detailed fingers, extremely detailed eyes and face, beautiful detailed nose, beautiful detailed eyes, long eyelashes, light on face, looking at viewer, (closed mouth:1.2), 1girl, cute, young, mature face, (full body:1.3), ((small breasts)), realistic face, realistic body, beautiful detailed thigh,s, same eyes color, (realistic, photo realism:1. 37), (highest quality), (best shadow), (best illustration), ultra high resolution, physics-based rendering, cinematic lighting), solo, 1girl, highly detailed, in office, detailed office, open cardigan, ponytail contorted, beautiful eyes ,sitting in office,dating, business suit, cross-laced clothes, collared shirt, beautiful breast, small breast, Chinese dress, white pantyhose, natural breasts, pink and white hair, ", - "paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), cloth, underwear, bra, low-res, normal quality, ((monochrome)), ((grayscale)), skin spots, acne, skin blemishes, age spots, glans, bad nipples, long nipples, bad vagina, extra fingers,fewer fingers,strange fingers,bad hand, ng_deepnegative_v1_75t, bad-picture-chill-75v", 7] - ] - INFO = f''' - ## Generate Your Beauty powered by EditAnything https://github.com/sail-sg/EditAnything - This model is good at generating beautiful female. - ''' - WARNING_INFO = f'''### [NOTE] the model is collected from the Internet for demo only, please do not use it for commercial purposes. - We are not responsible for possible risks using this model. - Lora model from https://civitai.com/models/14171/cutegirlmix4 Thanks! - ''' - demo = create_demo_template(process, process_image_click, - examples=examples, INFO=INFO, WARNING_INFO=WARNING_INFO) - return demo - - -if __name__ == '__main__': - sd_models_path = snapshot_download("shgao/sdmodels") - lora_model_path = hf_hub_download( - "mlida/Cute_girl_mix4", "cuteGirlMix4_v10.safetensors") - model = EditAnythingLoraModel(base_model_path=os.path.join(sd_models_path, "chilloutmix_NiPrunedFp32Fix"), - lora_model_path=lora_model_path, use_blip=True, extra_inpaint=True, - lora_weight=0.5, - ) - demo = create_demo(model.process, model.process_image_click) - demo.queue().launch(server_name='0.0.0.0') diff --git a/sam2edit_demo.py b/sam2edit_demo.py deleted file mode 100644 index 2cb12ed3a3887ba96d8f107fac1ea3cd3fe4b0b5..0000000000000000000000000000000000000000 --- a/sam2edit_demo.py +++ /dev/null @@ -1,138 +0,0 @@ -# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 -import gradio as gr - -def create_demo_template(process, process_image_click=None, examples=None, - INFO='EditAnything https://github.com/sail-sg/EditAnything', WARNING_INFO=None, - enable_auto_prompt_default=False, - ): - - print("The GUI is not fully tested yet. Please open an issue if you find bugs.") - block = gr.Blocks() - with block as demo: - clicked_points = gr.State([]) - origin_image = gr.State(None) - click_mask = gr.State(None) - with gr.Row(): - gr.Markdown(INFO) - with gr.Row().style(equal_height=False): - with gr.Column(): - with gr.Tab("Click🖱"): - source_image_click = gr.Image( - type="pil", interactive=True, - label="Image: Upload an image and click the region you want to edit.", - ) - with gr.Column(): - with gr.Row(): - point_prompt = gr.Radio( - choices=["Foreground Point", "Background Point"], - value="Foreground Point", - label="Point Label", - interactive=True, show_label=False) - clear_button_click = gr.Button( - value="Clear Click Points", interactive=True) - clear_button_image = gr.Button( - value="Clear Image", interactive=True) - with gr.Row(): - run_button_click = gr.Button( - label="Run EditAnying", interactive=True) - with gr.Tab("Brush🖌️"): - source_image_brush = gr.Image( - source='upload', - label="Image: Upload an image and cover the region you want to edit with sketch", - type="numpy", tool="sketch" - ) - run_button = gr.Button(label="Run EditAnying", interactive=True) - with gr.Column(): - enable_all_generate = gr.Checkbox( - label='Auto generation on all region.', value=False) - control_scale = gr.Slider( - label="Mask Align strength", info="Large value -> strict alignment with SAM mask", minimum=0, maximum=1, value=1, step=0.1) - with gr.Column(): - enable_auto_prompt = gr.Checkbox( - label='Auto generate text prompt from input image with BLIP2', info='Warning: Enable this may makes your prompt not working.', value=enable_auto_prompt_default) - a_prompt = gr.Textbox( - label="Positive Prompt", info='Text in the expected things of edited region', value='best quality, extremely detailed') - n_prompt = gr.Textbox(label="Negative Prompt", - value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, NSFW') - with gr.Row(): - num_samples = gr.Slider( - label="Images", minimum=1, maximum=12, value=2, step=1) - seed = gr.Slider(label="Seed", minimum=-1, - maximum=2147483647, step=1, randomize=True) - with gr.Row(): - enable_tile = gr.Checkbox( - label='Tile refinement for high resolution generation', info='Slow inference', value=True) - refine_alignment_ratio = gr.Slider( - label="Alignment Strength", info='Large value -> strict alignment with input image. Small value -> strong global consistency', minimum=0.0, maximum=1.0, value=0.95, step=0.05) - - with gr.Accordion("Advanced options", open=False): - mask_image = gr.Image( - source='upload', label="Upload a predefined mask of edit region if you do not want to write your prompt.", info="(Optional:Switch to Brush mode when using this!) ", type="numpy", value=None) - image_resolution = gr.Slider( - label="Image Resolution", minimum=256, maximum=768, value=512, step=64) - refine_image_resolution = gr.Slider( - label="Image Resolution", minimum=256, maximum=8192, value=1024, step=64) - guess_mode = gr.Checkbox( - label='Guess Mode', value=False) - detect_resolution = gr.Slider( - label="SAM Resolution", minimum=128, maximum=2048, value=1024, step=1) - ddim_steps = gr.Slider( - label="Steps", minimum=1, maximum=100, value=30, step=1) - scale = gr.Slider( - label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1) - eta = gr.Number(label="eta (DDIM)", value=0.0) - with gr.Column(): - result_gallery_refine = gr.Gallery( - label='Output High quality', show_label=True, elem_id="gallery").style(grid=2, preview=False) - result_gallery_init = gr.Gallery( - label='Output Low quality', show_label=True, elem_id="gallery").style(grid=2, height='auto') - result_gallery_ref = gr.Gallery( - label='Output Ref', show_label=False, elem_id="gallery").style(grid=2, height='auto') - result_text = gr.Text(label='BLIP2+Human Prompt Text') - - ips = [source_image_brush, enable_all_generate, mask_image, control_scale, enable_auto_prompt, a_prompt, n_prompt, num_samples, image_resolution, - detect_resolution, ddim_steps, guess_mode, scale, seed, eta, enable_tile, refine_alignment_ratio, refine_image_resolution] - run_button.click(fn=process, inputs=ips, outputs=[ - result_gallery_refine, result_gallery_init, result_gallery_ref, result_text]) - - ip_click = [origin_image, enable_all_generate, click_mask, control_scale, enable_auto_prompt, a_prompt, n_prompt, num_samples, image_resolution, - detect_resolution, ddim_steps, guess_mode, scale, seed, eta, enable_tile, refine_alignment_ratio, refine_image_resolution] - - run_button_click.click(fn=process, - inputs=ip_click, - outputs=[result_gallery_refine, result_gallery_init, result_gallery_ref, result_text]) - - source_image_click.upload( - lambda image: image.copy() if image is not None else None, - inputs=[source_image_click], - outputs=[origin_image] - ) - source_image_click.select( - process_image_click, - inputs=[origin_image, point_prompt, - clicked_points, image_resolution], - outputs=[source_image_click, clicked_points, click_mask], - show_progress=True, queue=True - ) - clear_button_click.click( - fn=lambda original_image: (original_image.copy(), [], None) - if original_image is not None else (None, [], None), - inputs=[origin_image], - outputs=[source_image_click, clicked_points, click_mask] - ) - clear_button_image.click( - fn=lambda: (None, [], None, None, None), - inputs=[], - outputs=[source_image_click, clicked_points, - click_mask, result_gallery_init, result_text] - ) - if examples is not None: - with gr.Row(): - ex = gr.Examples(examples=examples, fn=process, - inputs=[a_prompt, n_prompt, scale], - outputs=[result_gallery_init], - cache_examples=False) - if WARNING_INFO is not None: - with gr.Row(): - gr.Markdown(WARNING_INFO) - return demo diff --git a/sam2edit_lora.py b/sam2edit_lora.py deleted file mode 100644 index bface86d2b99a0cb510d39cb9096e3b85ab64c36..0000000000000000000000000000000000000000 --- a/sam2edit_lora.py +++ /dev/null @@ -1,618 +0,0 @@ -# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 -from torchvision.utils import save_image -from PIL import Image -from pytorch_lightning import seed_everything -import subprocess -from collections import OrderedDict -import re -import cv2 -import einops -import gradio as gr -import numpy as np -import torch -import random -import os -import requests -from io import BytesIO -from annotator.util import resize_image, HWC3, resize_points - -import torch -from safetensors.torch import load_file -from collections import defaultdict -from diffusers import StableDiffusionControlNetPipeline -from diffusers import ControlNetModel, UniPCMultistepScheduler -from utils.stable_diffusion_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline -# need the latest transformers -# pip install git+https://github.com/huggingface/transformers.git -from transformers import AutoProcessor, Blip2ForConditionalGeneration -from diffusers import ControlNetModel, DiffusionPipeline -import PIL.Image - -# Segment-Anything init. -# pip install git+https://github.com/facebookresearch/segment-anything.git -try: - from segment_anything import sam_model_registry, SamAutomaticMaskGenerator, SamPredictor -except ImportError: - print('segment_anything not installed') - result = subprocess.run( - ['pip', 'install', 'git+https://github.com/facebookresearch/segment-anything.git'], check=True) - print(f'Install segment_anything {result}') - from segment_anything import sam_model_registry, SamAutomaticMaskGenerator, SamPredictor -if not os.path.exists('./models/sam_vit_h_4b8939.pth'): - result = subprocess.run( - ['wget', 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth', '-P', 'models'], check=True) - print(f'Download sam_vit_h_4b8939.pth {result}') - -device = "cuda" if torch.cuda.is_available() else "cpu" - -config_dict = OrderedDict([ - ('LAION Pretrained(v0-4)-SD15', 'shgao/edit-anything-v0-4-sd15'), - ('LAION Pretrained(v0-4)-SD21', 'shgao/edit-anything-v0-4-sd21'), -]) - - -def init_sam_model(sam_generator=None, mask_predictor=None): - if sam_generator is not None and mask_predictor is not None: - return sam_generator, mask_predictor - sam_checkpoint = "models/sam_vit_h_4b8939.pth" - model_type = "default" - sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) - sam.to(device=device) - sam_generator = SamAutomaticMaskGenerator( - sam) if sam_generator is None else sam_generator - mask_predictor = SamPredictor( - sam) if mask_predictor is None else mask_predictor - return sam_generator, mask_predictor - - -def init_blip_processor(): - blip_processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b") - return blip_processor - - -def init_blip_model(): - blip_model = Blip2ForConditionalGeneration.from_pretrained( - "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16, device_map="auto") - return blip_model - - -def get_pipeline_embeds(pipeline, prompt, negative_prompt, device): - # https://github.com/huggingface/diffusers/issues/2136 - """ Get pipeline embeds for prompts bigger than the maxlength of the pipe - :param pipeline: - :param prompt: - :param negative_prompt: - :param device: - :return: - """ - max_length = pipeline.tokenizer.model_max_length - - # simple way to determine length of tokens - count_prompt = len(re.split(r', ', prompt)) - count_negative_prompt = len(re.split(r', ', negative_prompt)) - - # create the tensor based on which prompt is longer - if count_prompt >= count_negative_prompt: - input_ids = pipeline.tokenizer( - prompt, return_tensors="pt", truncation=False).input_ids.to(device) - shape_max_length = input_ids.shape[-1] - negative_ids = pipeline.tokenizer(negative_prompt, truncation=False, padding="max_length", - max_length=shape_max_length, return_tensors="pt").input_ids.to(device) - else: - negative_ids = pipeline.tokenizer( - negative_prompt, return_tensors="pt", truncation=False).input_ids.to(device) - shape_max_length = negative_ids.shape[-1] - input_ids = pipeline.tokenizer(prompt, return_tensors="pt", truncation=False, padding="max_length", - max_length=shape_max_length).input_ids.to(device) - - concat_embeds = [] - neg_embeds = [] - for i in range(0, shape_max_length, max_length): - concat_embeds.append(pipeline.text_encoder( - input_ids[:, i: i + max_length])[0]) - neg_embeds.append(pipeline.text_encoder( - negative_ids[:, i: i + max_length])[0]) - - return torch.cat(concat_embeds, dim=1), torch.cat(neg_embeds, dim=1) - - -def load_lora_weights(pipeline, checkpoint_path, multiplier, device, dtype): - LORA_PREFIX_UNET = "lora_unet" - LORA_PREFIX_TEXT_ENCODER = "lora_te" - # load LoRA weight from .safetensors - if isinstance(checkpoint_path, str): - - state_dict = load_file(checkpoint_path, device=device) - - updates = defaultdict(dict) - for key, value in state_dict.items(): - # it is suggested to print out the key, it usually will be something like below - # "lora_te_text_model_encoder_layers_0_self_attn_k_proj.lora_down.weight" - - layer, elem = key.split('.', 1) - updates[layer][elem] = value - - # directly update weight in diffusers model - for layer, elems in updates.items(): - - if "text" in layer: - layer_infos = layer.split( - LORA_PREFIX_TEXT_ENCODER + "_")[-1].split("_") - curr_layer = pipeline.text_encoder - else: - layer_infos = layer.split( - LORA_PREFIX_UNET + "_")[-1].split("_") - curr_layer = pipeline.unet - - # find the target layer - temp_name = layer_infos.pop(0) - while len(layer_infos) > -1: - try: - curr_layer = curr_layer.__getattr__(temp_name) - if len(layer_infos) > 0: - temp_name = layer_infos.pop(0) - elif len(layer_infos) == 0: - break - except Exception: - if len(temp_name) > 0: - temp_name += "_" + layer_infos.pop(0) - else: - temp_name = layer_infos.pop(0) - - # get elements for this layer - weight_up = elems['lora_up.weight'].to(dtype) - weight_down = elems['lora_down.weight'].to(dtype) - alpha = elems['alpha'] - if alpha: - alpha = alpha.item() / weight_up.shape[1] - else: - alpha = 1.0 - - # update weight - if len(weight_up.shape) == 4: - curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up.squeeze( - 3).squeeze(2), weight_down.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) - else: - curr_layer.weight.data += multiplier * \ - alpha * torch.mm(weight_up, weight_down) - else: - for ckptpath in checkpoint_path: - state_dict = load_file(ckptpath, device=device) - - updates = defaultdict(dict) - for key, value in state_dict.items(): - # it is suggested to print out the key, it usually will be something like below - # "lora_te_text_model_encoder_layers_0_self_attn_k_proj.lora_down.weight" - - layer, elem = key.split('.', 1) - updates[layer][elem] = value - - # directly update weight in diffusers model - for layer, elems in updates.items(): - if "text" in layer: - layer_infos = layer.split( - LORA_PREFIX_TEXT_ENCODER + "_")[-1].split("_") - curr_layer = pipeline.text_encoder - else: - layer_infos = layer.split( - LORA_PREFIX_UNET + "_")[-1].split("_") - curr_layer = pipeline.unet - - # find the target layer - temp_name = layer_infos.pop(0) - while len(layer_infos) > -1: - try: - curr_layer = curr_layer.__getattr__(temp_name) - if len(layer_infos) > 0: - temp_name = layer_infos.pop(0) - elif len(layer_infos) == 0: - break - except Exception: - if len(temp_name) > 0: - temp_name += "_" + layer_infos.pop(0) - else: - temp_name = layer_infos.pop(0) - - # get elements for this layer - weight_up = elems['lora_up.weight'].to(dtype) - weight_down = elems['lora_down.weight'].to(dtype) - alpha = elems['alpha'] - if alpha: - alpha = alpha.item() / weight_up.shape[1] - else: - alpha = 1.0 - - # update weight - if len(weight_up.shape) == 4: - curr_layer.weight.data += multiplier * alpha * torch.mm(weight_up.squeeze( - 3).squeeze(2), weight_down.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3) - else: - curr_layer.weight.data += multiplier * \ - alpha * torch.mm(weight_up, weight_down) - return pipeline - - -def make_inpaint_condition(image, image_mask): - # image = np.array(image.convert("RGB")).astype(np.float32) / 255.0 - image = image / 255.0 - print("img", image.max(), image.min(), image_mask.max(), image_mask.min()) - # image_mask = np.array(image_mask.convert("L")) - assert image.shape[0:1] == image_mask.shape[0: - 1], "image and image_mask must have the same image size" - image[image_mask > 128] = -1.0 # set as masked pixel - image = np.expand_dims(image, 0).transpose(0, 3, 1, 2) - image = torch.from_numpy(image) - return image - - -def obtain_generation_model(base_model_path, lora_model_path, controlnet_path, generation_only=False, extra_inpaint=True, lora_weight=1.0): - controlnet = [] - controlnet.append(ControlNetModel.from_pretrained( - controlnet_path, torch_dtype=torch.float16)) # sam control - if (not generation_only) and extra_inpaint: # inpainting control - print("Warning: ControlNet based inpainting model only support SD1.5 for now.") - controlnet.append( - ControlNetModel.from_pretrained( - 'lllyasviel/control_v11p_sd15_inpaint', torch_dtype=torch.float16) # inpainting controlnet - ) - - if generation_only and extra_inpaint: - pipe = StableDiffusionControlNetPipeline.from_pretrained( - base_model_path, controlnet=controlnet, torch_dtype=torch.float16, safety_checker=None - ) - else: - pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( - base_model_path, controlnet=controlnet, torch_dtype=torch.float16, safety_checker=None - ) - if lora_model_path is not None: - pipe = load_lora_weights( - pipe, [lora_model_path], lora_weight, 'cpu', torch.float32) - # speed up diffusion process with faster scheduler and memory optimization - pipe.scheduler = UniPCMultistepScheduler.from_config( - pipe.scheduler.config) - # remove following line if xformers is not installed - pipe.enable_xformers_memory_efficient_attention() - - pipe.enable_model_cpu_offload() - return pipe - - -def obtain_tile_model(base_model_path, lora_model_path, lora_weight=1.0): - controlnet = ControlNetModel.from_pretrained( - 'lllyasviel/control_v11f1e_sd15_tile', torch_dtype=torch.float16) # tile controlnet - if base_model_path == 'runwayml/stable-diffusion-v1-5' or base_model_path == 'stabilityai/stable-diffusion-2-inpainting': - print("base_model_path", base_model_path) - pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( - "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16, safety_checker=None - ) - else: - pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( - base_model_path, controlnet=controlnet, torch_dtype=torch.float16, safety_checker=None - ) - if lora_model_path is not None: - pipe = load_lora_weights( - pipe, [lora_model_path], lora_weight, 'cpu', torch.float32) - # speed up diffusion process with faster scheduler and memory optimization - pipe.scheduler = UniPCMultistepScheduler.from_config( - pipe.scheduler.config) - # remove following line if xformers is not installed - pipe.enable_xformers_memory_efficient_attention() - - pipe.enable_model_cpu_offload() - return pipe - - -def show_anns(anns): - if len(anns) == 0: - return - sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True) - full_img = None - - # for ann in sorted_anns: - for i in range(len(sorted_anns)): - ann = anns[i] - m = ann['segmentation'] - if full_img is None: - full_img = np.zeros((m.shape[0], m.shape[1], 3)) - map = np.zeros((m.shape[0], m.shape[1]), dtype=np.uint16) - map[m != 0] = i + 1 - color_mask = np.random.random((1, 3)).tolist()[0] - full_img[m != 0] = color_mask - full_img = full_img*255 - # anno encoding from https://github.com/LUSSeg/ImageNet-S - res = np.zeros((map.shape[0], map.shape[1], 3)) - res[:, :, 0] = map % 256 - res[:, :, 1] = map // 256 - res.astype(np.float32) - full_img = Image.fromarray(np.uint8(full_img)) - return full_img, res - - -class EditAnythingLoraModel: - def __init__(self, - base_model_path='../chilloutmix_NiPrunedFp32Fix', - lora_model_path='../40806/mix4', use_blip=True, - blip_processor=None, - blip_model=None, - sam_generator=None, - controlmodel_name='LAION Pretrained(v0-4)-SD15', - # used when the base model is not an inpainting model. - extra_inpaint=True, - tile_model=None, - lora_weight=1.0, - mask_predictor=None - ): - self.device = device - self.use_blip = use_blip - - # Diffusion init using diffusers. - self.default_controlnet_path = config_dict[controlmodel_name] - self.base_model_path = base_model_path - self.lora_model_path = lora_model_path - self.defalut_enable_all_generate = False - self.extra_inpaint = extra_inpaint - self.pipe = obtain_generation_model( - base_model_path, lora_model_path, self.default_controlnet_path, generation_only=False, extra_inpaint=extra_inpaint, lora_weight=lora_weight) - - # Segment-Anything init. - self.sam_generator, self.mask_predictor = init_sam_model( - sam_generator, mask_predictor) - # BLIP2 init. - if use_blip: - if blip_processor is not None: - self.blip_processor = blip_processor - else: - self.blip_processor = init_blip_processor() - - if blip_model is not None: - self.blip_model = blip_model - else: - self.blip_model = init_blip_model() - - # tile model init. - if tile_model is not None: - self.tile_pipe = tile_model - else: - self.tile_pipe = obtain_tile_model( - base_model_path, lora_model_path, lora_weight=lora_weight) - - def get_blip2_text(self, image): - inputs = self.blip_processor(image, return_tensors="pt").to( - self.device, torch.float16) - generated_ids = self.blip_model.generate(**inputs, max_new_tokens=50) - generated_text = self.blip_processor.batch_decode( - generated_ids, skip_special_tokens=True)[0].strip() - return generated_text - - def get_sam_control(self, image): - masks = self.sam_generator.generate(image) - full_img, res = show_anns(masks) - return full_img, res - - def get_click_mask(self, image, clicked_points): - self.mask_predictor.set_image(image) - # Separate the points and labels - points, labels = zip(*[(point[:2], point[2]) - for point in clicked_points]) - - # Convert the points and labels to numpy arrays - input_point = np.array(points) - input_label = np.array(labels) - - masks, _, _ = self.mask_predictor.predict( - point_coords=input_point, - point_labels=input_label, - multimask_output=False, - ) - - return masks - - @torch.inference_mode() - def process_image_click(self, original_image: gr.Image, - point_prompt: gr.Radio, - clicked_points: gr.State, - image_resolution, - evt: gr.SelectData): - # Get the clicked coordinates - clicked_coords = evt.index - x, y = clicked_coords - label = point_prompt - lab = 1 if label == "Foreground Point" else 0 - clicked_points.append((x, y, lab)) - - input_image = np.array(original_image, dtype=np.uint8) - H, W, C = input_image.shape - input_image = HWC3(input_image) - img = resize_image(input_image, image_resolution) - - # Update the clicked_points - resized_points = resize_points(clicked_points, - input_image.shape, - image_resolution) - mask_click_np = self.get_click_mask(img, resized_points) - - # Convert mask_click_np to HWC format - mask_click_np = np.transpose(mask_click_np, (1, 2, 0)) * 255.0 - - mask_image = HWC3(mask_click_np.astype(np.uint8)) - mask_image = cv2.resize( - mask_image, (W, H), interpolation=cv2.INTER_LINEAR) - # mask_image = Image.fromarray(mask_image_tmp) - - # Draw circles for all clicked points - edited_image = input_image - for x, y, lab in clicked_points: - # Set the circle color based on the label - color = (255, 0, 0) if lab == 1 else (0, 0, 255) - - # Draw the circle - edited_image = cv2.circle(edited_image, (x, y), 20, color, -1) - - # Set the opacity for the mask_image and edited_image - opacity_mask = 0.75 - opacity_edited = 1.0 - - # Combine the edited_image and the mask_image using cv2.addWeighted() - overlay_image = cv2.addWeighted( - edited_image, opacity_edited, - (mask_image * np.array([0/255, 255/255, 0/255])).astype(np.uint8), - opacity_mask, 0 - ) - - return Image.fromarray(overlay_image), clicked_points, Image.fromarray(mask_image) - - @torch.inference_mode() - def process(self, source_image, enable_all_generate, mask_image, - control_scale, - enable_auto_prompt, a_prompt, n_prompt, - num_samples, image_resolution, detect_resolution, - ddim_steps, guess_mode, scale, seed, eta, - enable_tile=True, refine_alignment_ratio=None, refine_image_resolution=None, condition_model=None): - - if condition_model is None: - this_controlnet_path = self.default_controlnet_path - else: - this_controlnet_path = config_dict[condition_model] - input_image = source_image["image"] if isinstance( - source_image, dict) else np.array(source_image, dtype=np.uint8) - if mask_image is None: - if enable_all_generate != self.defalut_enable_all_generate: - self.pipe = obtain_generation_model( - self.base_model_path, self.lora_model_path, this_controlnet_path, enable_all_generate, self.extra_inpaint) - - self.defalut_enable_all_generate = enable_all_generate - if enable_all_generate: - print("source_image", - source_image["mask"].shape, input_image.shape,) - mask_image = np.ones( - (input_image.shape[0], input_image.shape[1], 3))*255 - else: - mask_image = source_image["mask"] - else: - mask_image = np.array(mask_image, dtype=np.uint8) - if self.default_controlnet_path != this_controlnet_path: - print("To Use:", this_controlnet_path, - "Current:", self.default_controlnet_path) - print("Change condition model to:", this_controlnet_path) - self.pipe = obtain_generation_model( - self.base_model_path, self.lora_model_path, this_controlnet_path, enable_all_generate, self.extra_inpaint) - self.default_controlnet_path = this_controlnet_path - torch.cuda.empty_cache() - - with torch.no_grad(): - if self.use_blip and enable_auto_prompt: - print("Generating text:") - blip2_prompt = self.get_blip2_text(input_image) - print("Generated text:", blip2_prompt) - if len(a_prompt) > 0: - a_prompt = blip2_prompt + ',' + a_prompt - else: - a_prompt = blip2_prompt - - input_image = HWC3(input_image) - - img = resize_image(input_image, image_resolution) - H, W, C = img.shape - - print("Generating SAM seg:") - # the default SAM model is trained with 1024 size. - full_segmask, detected_map = self.get_sam_control( - resize_image(input_image, detect_resolution)) - - detected_map = HWC3(detected_map.astype(np.uint8)) - detected_map = cv2.resize( - detected_map, (W, H), interpolation=cv2.INTER_LINEAR) - - control = torch.from_numpy( - detected_map.copy()).float().cuda() - control = torch.stack([control for _ in range(num_samples)], dim=0) - control = einops.rearrange(control, 'b h w c -> b c h w').clone() - - mask_imag_ori = HWC3(mask_image.astype(np.uint8)) - mask_image_tmp = cv2.resize( - mask_imag_ori, (W, H), interpolation=cv2.INTER_LINEAR) - mask_image = Image.fromarray(mask_image_tmp) - - if seed == -1: - seed = random.randint(0, 65535) - seed_everything(seed) - generator = torch.manual_seed(seed) - postive_prompt = a_prompt - negative_prompt = n_prompt - prompt_embeds, negative_prompt_embeds = get_pipeline_embeds( - self.pipe, postive_prompt, negative_prompt, "cuda") - prompt_embeds = torch.cat([prompt_embeds] * num_samples, dim=0) - negative_prompt_embeds = torch.cat( - [negative_prompt_embeds] * num_samples, dim=0) - if enable_all_generate and self.extra_inpaint: - self.pipe.safety_checker = lambda images, clip_input: ( - images, False) - x_samples = self.pipe( - prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, - num_images_per_prompt=num_samples, - num_inference_steps=ddim_steps, - generator=generator, - height=H, - width=W, - image=[control.type(torch.float16)], - controlnet_conditioning_scale=[float(control_scale)], - guidance_scale=scale, - ).images - else: - multi_condition_image = [] - multi_condition_scale = [] - multi_condition_image.append(control.type(torch.float16)) - multi_condition_scale.append(float(control_scale)) - if self.extra_inpaint: - inpaint_image = make_inpaint_condition(img, mask_image_tmp) - print(inpaint_image.shape) - multi_condition_image.append( - inpaint_image.type(torch.float16)) - multi_condition_scale.append(1.0) - x_samples = self.pipe( - image=img, - mask_image=mask_image, - prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, - num_images_per_prompt=num_samples, - num_inference_steps=ddim_steps, - generator=generator, - controlnet_conditioning_image=multi_condition_image, - height=H, - width=W, - controlnet_conditioning_scale=multi_condition_scale, - guidance_scale=scale, - ).images - results = [x_samples[i] for i in range(num_samples)] - - results_tile = [] - if enable_tile: - prompt_embeds, negative_prompt_embeds = get_pipeline_embeds( - self.tile_pipe, postive_prompt, negative_prompt, "cuda") - for i in range(num_samples): - img_tile = PIL.Image.fromarray(resize_image( - np.array(x_samples[i]), refine_image_resolution)) - if i == 0: - mask_image_tile = cv2.resize( - mask_imag_ori, (img_tile.size[0], img_tile.size[1]), interpolation=cv2.INTER_LINEAR) - mask_image_tile = Image.fromarray(mask_image_tile) - x_samples_tile = self.tile_pipe( - image=img_tile, - mask_image=mask_image_tile, - prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, - num_images_per_prompt=1, - num_inference_steps=ddim_steps, - generator=generator, - controlnet_conditioning_image=img_tile, - height=img_tile.size[1], - width=img_tile.size[0], - controlnet_conditioning_scale=1.0, - alignment_ratio=refine_alignment_ratio, - guidance_scale=scale, - ).images - results_tile += x_samples_tile - - return results_tile, results, [full_segmask, mask_image], postive_prompt - - def download_image(url): - response = requests.get(url) - return Image.open(BytesIO(response.content)).convert("RGB") diff --git a/sam2groundingdino_edit.py b/sam2groundingdino_edit.py new file mode 100644 index 0000000000000000000000000000000000000000..5ad2dfae2c098c510e86add527c09cc314620048 --- /dev/null +++ b/sam2groundingdino_edit.py @@ -0,0 +1,351 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +from torchvision.utils import save_image +from PIL import Image +from pytorch_lightning import seed_everything + +import cv2 +import einops +import gradio as gr +import numpy as np +import torch +import random +import requests +from io import BytesIO +from annotator.util import resize_image, HWC3 + +device = "cuda" if torch.cuda.is_available() else "cpu" +use_blip = False +use_gradio = False + +# Diffusion init using diffusers. + +import groundingdino.datasets.transforms as T +from groundingdino.models import build_model +from groundingdino.util import box_ops +from groundingdino.util.slconfig import SLConfig +from groundingdino.util.utils import clean_state_dict, get_phrases_from_posmap +from groundingdino.util.inference import annotate, load_image, predict +from segment_anything import build_sam, SamPredictor +from segment_anything.utils.amg import remove_small_regions + +# diffusers==0.14.0 required. +from diffusers import ControlNetModel, UniPCMultistepScheduler +from utils.stable_diffusion_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline +import torch + +base_model_path = "stabilityai/stable-diffusion-2-inpainting" +controlnet_path = "shgao/edit-anything-v0-1-1" + +controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16) +pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( + base_model_path, controlnet=controlnet, torch_dtype=torch.float16 +) +# speed up diffusion process with faster scheduler and memory optimization +pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) +# remove following line if xformers is not installed +pipe.enable_xformers_memory_efficient_attention() + +# pipe.enable_model_cpu_offload() # disable for now because of unknow bug in accelerate +pipe.to(device) + +# Segment-Anything init. +# pip install git+https://github.com/facebookresearch/segment-anything.git +from segment_anything import sam_model_registry, SamAutomaticMaskGenerator + +sam_checkpoint = "./models/sam_vit_h_4b8939.pth" +groundingdino_checkpoint = "./models/groundingdino_swint_ogc.pth" +groundingdino_config_file = "./GroundingDINO_SwinT_OGC.py" +model_type = "default" +sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) +sam.to(device=device) +mask_generator = SamAutomaticMaskGenerator(sam) + + +def load_groundingdino_model(model_config_path, model_checkpoint_path): + args = SLConfig.fromfile(model_config_path) + args.device = device + model = build_model(args) + checkpoint = torch.load(model_checkpoint_path, map_location="cpu") + load_res = model.load_state_dict(clean_state_dict(checkpoint["model"]), strict=False) + print(load_res) + _ = model.eval() + return model + + +grounding_model = load_groundingdino_model(groundingdino_config_file, groundingdino_checkpoint).to(device) +sam_predictor = SamPredictor(build_sam(checkpoint=sam_checkpoint).to(device=device)) + +# BLIP2 init. +if use_blip: + # need the latest transformers + # pip install git+https://github.com/huggingface/transformers.git + from transformers import AutoProcessor, Blip2ForConditionalGeneration + + processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b") + blip_model = Blip2ForConditionalGeneration.from_pretrained( + "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16) + blip_model.to(device) + blip_model.to(device) + + +def get_blip2_text(image): + inputs = processor(image, return_tensors="pt").to(device, torch.float16) + generated_ids = blip_model.generate(**inputs, max_new_tokens=50) + generated_text = processor.batch_decode( + generated_ids, skip_special_tokens=True)[0].strip() + return generated_text + + +def show_anns(anns): + if len(anns) == 0: + return + sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True) + full_img = None + + # for ann in sorted_anns: + for i in range(len(sorted_anns)): + ann = anns[i] + m = ann['segmentation'] + if full_img is None: + full_img = np.zeros((m.shape[0], m.shape[1], 3)) + map = np.zeros((m.shape[0], m.shape[1]), dtype=np.uint16) + map[m != 0] = i + 1 + color_mask = np.random.random((1, 3)).tolist()[0] + full_img[m != 0] = color_mask + full_img = full_img * 255 + # anno encoding from https://github.com/LUSSeg/ImageNet-S + res = np.zeros((map.shape[0], map.shape[1], 3)) + res[:, :, 0] = map % 256 + res[:, :, 1] = map // 256 + res.astype(np.float32) + return full_img, res + + +def get_sam_control(image): + masks = mask_generator.generate(image) + full_img, res = show_anns(masks) + return full_img, res + + +def prompt2mask(original_image, caption, box_threshold=0.25, text_threshold=0.25, num_boxes=2): + def image_transform_grounding(init_image): + transform = T.Compose([ + T.RandomResize([800], max_size=1333), + T.ToTensor(), + T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) + ]) + image, _ = transform(init_image, None) # 3, h, w + return init_image, image + + image_np = np.array(original_image, dtype=np.uint8) + caption = caption.lower() + caption = caption.strip() + if not caption.endswith("."): + caption = caption + "." + _, image_tensor = image_transform_grounding(original_image) + boxes, logits, phrases = predict(grounding_model, + image_tensor, caption, box_threshold, text_threshold, device='cpu') + print(logits) + print('number of boxes: ', boxes.size(0)) + # exit(0) + # from PIL import Image, ImageDraw, ImageFont + H, W = original_image.size[1], original_image.size[0] + boxes = boxes * torch.Tensor([W, H, W, H]) + boxes[:, :2] = boxes[:, :2] - boxes[:, 2:] / 2 + boxes[:, 2:] = boxes[:, 2:] + boxes[:, :2] + # draw = ImageDraw.Draw(original_image) + # for box in boxes: + # # from 0..1 to 0..W, 0..H + # # box = box * torch.Tensor([W, H, W, H]) + # # # from xywh to xyxy + # # box[:2] -= box[2:] / 2 + # # box[2:] += box[:2] + # # random color + # color = tuple(np.random.randint(0, 255, size=3).tolist()) + # # draw + # x0, y0, x1, y1 = box + # x0, y0, x1, y1 = int(x0), int(y0), int(x1), int(y1) + # + # draw.rectangle([x0, y0, x1, y1], outline=color, width=6) + # original_image.save('debug.jpg') + # exit(0) + + final_m = torch.zeros((image_np.shape[0], image_np.shape[1])) + + if boxes.size(0) > 0: + sam_predictor.set_image(image_np) + + transformed_boxes = sam_predictor.transform.apply_boxes_torch(boxes, image_np.shape[:2]) + masks, _, _ = sam_predictor.predict_torch( + point_coords=None, + point_labels=None, + boxes=transformed_boxes.to(device), + multimask_output=False, + ) + + # remove small disconnected regions and holes + fine_masks = [] + for mask in masks.to('cpu').numpy(): # masks: [num_masks, 1, h, w] + fine_masks.append(remove_small_regions(mask[0], 400, mode="holes")[0]) + masks = np.stack(fine_masks, axis=0)[:, np.newaxis] + masks = torch.from_numpy(masks) + + num_obj = min(len(logits), num_boxes) + for obj_ind in range(num_obj): + # box = boxes[obj_ind] + + m = masks[obj_ind][0] + final_m += m + final_m = (final_m > 0).to('cpu').numpy() + # print(final_m.max(), final_m.min()) + return np.dstack((final_m, final_m, final_m)) * 255 + + +def process(input_image, mask_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, detect_resolution, + ddim_steps, guess_mode, strength, scale, seed, eta): + with torch.no_grad(): + mask_image = np.array(prompt2mask(input_image, mask_prompt), dtype=np.uint8) + input_image = np.array(input_image, dtype=np.uint8)[:, :, :3] + + if use_blip: + print("Generating text:") + blip2_prompt = get_blip2_text(input_image) + print("Generated text:", blip2_prompt) + if len(prompt) > 0: + prompt = blip2_prompt + ',' + prompt + else: + prompt = blip2_prompt + print("All text:", prompt) + + input_image = HWC3(input_image) + + img = resize_image(input_image, image_resolution) + H, W, C = img.shape + + print("Generating SAM seg:") + # the default SAM model is trained with 1024 size. + full_segmask, detected_map = get_sam_control( + resize_image(input_image, detect_resolution)) + + detected_map = HWC3(detected_map.astype(np.uint8)) + detected_map = cv2.resize( + detected_map, (W, H), interpolation=cv2.INTER_LINEAR) + + control = torch.from_numpy( + detected_map.copy()).float().cuda() + control = torch.stack([control for _ in range(num_samples)], dim=0) + control = einops.rearrange(control, 'b h w c -> b c h w').clone() + + mask_image = HWC3(mask_image.astype(np.uint8)) + mask_image = cv2.resize( + mask_image, (W, H), interpolation=cv2.INTER_LINEAR) + mask_image = Image.fromarray(mask_image) + + if seed == -1: + seed = random.randint(0, 65535) + seed_everything(seed) + generator = torch.manual_seed(seed) + x_samples = pipe( + image=img, + mask_image=mask_image, + prompt=[prompt + ', ' + a_prompt] * num_samples, + negative_prompt=[n_prompt] * num_samples, + num_images_per_prompt=num_samples, + num_inference_steps=ddim_steps, + generator=generator, + controlnet_conditioning_image=control.type(torch.float16), + height=H, + width=W, + ).images + + results = [x_samples[i] for i in range(num_samples)] + return [full_segmask, mask_image] + results + + +def download_image(url): + response = requests.get(url) + return Image.open(BytesIO(response.content)).convert("RGB") + + +# disable gradio when not using GUI. +if not use_gradio: + image_path = "assets/dog.png" + input_image_pil = Image.open(image_path).convert('RGB') + input_image = np.array(input_image_pil, dtype=np.uint8)[:, :, :3] + + mask_prompt = 'bench.' + # mask_image = np.array(prompt2mask(input_image, mask_prompt), dtype=np.uint8) + prompt = "cat" + a_prompt = 'best quality, extremely detailed' + n_prompt = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality' + num_samples = 3 + image_resolution = 512 + detect_resolution = 512 + ddim_steps = 30 + guess_mode = False + strength = 1.0 + scale = 9.0 + seed = -1 + eta = 0.0 + + outputs = process(input_image_pil, mask_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta) + + image_list = [] + input_image = resize_image(input_image, 512) + image_list.append(torch.tensor(input_image)) + for i in range(len(outputs)): + each = outputs[i] + if type(each) is not np.ndarray: + each = np.array(each, dtype=np.uint8) + each = resize_image(each, 512) + print(i, each.shape) + image_list.append(torch.tensor(each)) + + image_list = torch.stack(image_list).permute(0, 3, 1, 2) + + save_image(image_list, "sample.jpg", nrow=3, + normalize=True, value_range=(0, 255)) +else: + print("The GUI is not tested yet. Please open an issue if you find bugs.") + block = gr.Blocks().queue() + with block: + with gr.Row(): + gr.Markdown( + "## Edit Anything powered by ControlNet+SAM+BLIP2+Stable Diffusion") + with gr.Row(): + with gr.Column(): + input_image = gr.Image(source='upload', type="numpy") + prompt = gr.Textbox(label="Prompt") + run_button = gr.Button(label="Run") + with gr.Accordion("Advanced options", open=False): + num_samples = gr.Slider( + label="Images", minimum=1, maximum=12, value=1, step=1) + image_resolution = gr.Slider( + label="Image Resolution", minimum=256, maximum=768, value=512, step=64) + strength = gr.Slider( + label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01) + guess_mode = gr.Checkbox(label='Guess Mode', value=False) + detect_resolution = gr.Slider( + label="SAM Resolution", minimum=128, maximum=2048, value=1024, step=1) + ddim_steps = gr.Slider( + label="Steps", minimum=1, maximum=100, value=20, step=1) + scale = gr.Slider( + label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1) + seed = gr.Slider(label="Seed", minimum=-1, + maximum=2147483647, step=1, randomize=True) + eta = gr.Number(label="eta (DDIM)", value=0.0) + mask_prompt = gr.Textbox( + label="Mask Prompt", value='best quality, extremely detailed') + a_prompt = gr.Textbox( + label="Added Prompt", value='best quality, extremely detailed') + n_prompt = gr.Textbox(label="Negative Prompt", + value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality') + with gr.Column(): + result_gallery = gr.Gallery( + label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto') + ips = [input_image, mask_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta] + run_button.click(fn=process, inputs=ips, outputs=[result_gallery]) + + block.launch(server_name='0.0.0.0') diff --git a/sam2image.py b/sam2image.py index b1f669e8d48950cc092bc8be591e899492679d21..9bdf77395a0ca9bb994dfd4ed1e0262e5d425228 100644 --- a/sam2image.py +++ b/sam2image.py @@ -16,6 +16,7 @@ import random import os from annotator.util import resize_image, HWC3 + def create_demo(): device = "cuda" if torch.cuda.is_available() else "cpu" use_blip = True @@ -26,11 +27,10 @@ def create_demo(): base_model_path = "stabilityai/stable-diffusion-2-1" - config_dict = OrderedDict([('SAM Pretrained(v0-1)', 'shgao/edit-anything-v0-1-1'), - ('LAION Pretrained(v0-3)', 'shgao/edit-anything-v0-3'), - ]) - + ('LAION Pretrained(v0-3)', 'shgao/edit-anything-v0-3'), + ('LAION Pretrained(v0-4)', 'shgao/edit-anything-v0-4-sd21'), + ]) def obtain_generation_model(controlnet_path): controlnet = ControlNetModel.from_pretrained( @@ -42,34 +42,34 @@ def create_demo(): pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) # remove following line if xformers is not installed pipe.enable_xformers_memory_efficient_attention() - pipe.enable_model_cpu_offload() # disable for now because of unknow bug in accelerate - # pipe.to(device) + # pipe.enable_model_cpu_offload() # disable for now because of unknow bug in accelerate + pipe.to(device) return pipe global default_controlnet_path - default_controlnet_path = config_dict['LAION Pretrained(v0-3)'] + default_controlnet_path = config_dict['LAION Pretrained(v0-4)'] pipe = obtain_generation_model(default_controlnet_path) - # Segment-Anything init. # pip install git+https://github.com/facebookresearch/segment-anything.git try: from segment_anything import sam_model_registry, SamAutomaticMaskGenerator except ImportError: print('segment_anything not installed') - result = subprocess.run(['pip', 'install', 'git+https://github.com/facebookresearch/segment-anything.git'], check=True) + result = subprocess.run(['pip', 'install', 'git+https://github.com/facebookresearch/segment-anything.git'], + check=True) print(f'Install segment_anything {result}') - from segment_anything import sam_model_registry, SamAutomaticMaskGenerator if not os.path.exists('./models/sam_vit_h_4b8939.pth'): - result = subprocess.run(['wget', 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth', '-P', 'models'], check=True) - print(f'Download sam_vit_h_4b8939.pth {result}') + result = subprocess.run( + ['wget', 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth', '-P', 'models'], + check=True) + print(f'Download sam_vit_h_4b8939.pth {result}') sam_checkpoint = "models/sam_vit_h_4b8939.pth" model_type = "default" sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) sam.to(device=device) mask_generator = SamAutomaticMaskGenerator(sam) - # BLIP2 init. if use_blip: # need the latest transformers @@ -82,7 +82,6 @@ def create_demo(): blip_model.to(device) blip_model.to(device) - def get_blip2_text(image): inputs = processor(image, return_tensors="pt").to(device, torch.float16) generated_ids = blip_model.generate(**inputs, max_new_tokens=50) @@ -90,7 +89,6 @@ def create_demo(): generated_ids, skip_special_tokens=True)[0].strip() return generated_text - def show_anns(anns): if len(anns) == 0: return @@ -107,7 +105,7 @@ def create_demo(): map[m != 0] = i + 1 color_mask = np.random.random((1, 3)).tolist()[0] full_img[m != 0] = color_mask - full_img = full_img*255 + full_img = full_img * 255 # anno encoding from https://github.com/LUSSeg/ImageNet-S res = np.zeros((map.shape[0], map.shape[1], 3)) res[:, :, 0] = map % 256 @@ -116,19 +114,18 @@ def create_demo(): full_img = Image.fromarray(np.uint8(full_img)) return full_img, res - def get_sam_control(image): masks = mask_generator.generate(image) full_img, res = show_anns(masks) return full_img, res + def process(condition_model, input_image, enable_auto_prompt, prompt, a_prompt, n_prompt, num_samples, + image_resolution, detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta): - def process(condition_model, input_image, enable_auto_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta): - global default_controlnet_path global pipe print("To Use:", config_dict[condition_model], "Current:", default_controlnet_path) - if default_controlnet_path!=config_dict[condition_model]: + if default_controlnet_path != config_dict[condition_model]: print("Change condition model to:", config_dict[condition_model]) pipe = obtain_generation_model(config_dict[condition_model]) default_controlnet_path = config_dict[condition_model] @@ -182,7 +179,6 @@ def create_demo(): results = [x_samples[i] for i in range(num_samples)] return [full_segmask] + results, prompt - # disable gradio when not using GUI. if not use_gradio: # This part is not updated, it's just a example to use it without GUI. @@ -203,8 +199,9 @@ def create_demo(): seed = 10086 eta = 0.0 - outputs, full_text = process(condition_model, input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, - detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta) + outputs, full_text = process(condition_model, input_image, prompt, a_prompt, n_prompt, num_samples, + image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta) image_list = [] input_image = resize_image(input_image, 512) @@ -220,7 +217,7 @@ def create_demo(): image_list = torch.stack(image_list).permute(0, 3, 1, 2) save_image(image_list, "sample.jpg", nrow=3, - normalize=True, value_range=(0, 255)) + normalize=True, value_range=(0, 255)) else: block = gr.Blocks() with block as demo: @@ -233,12 +230,12 @@ def create_demo(): prompt = gr.Textbox(label="Prompt (Optional)") run_button = gr.Button(label="Run") condition_model = gr.Dropdown(choices=list(config_dict.keys()), - value=list(config_dict.keys())[0], - label='Model', - multiselect=False) + value=list(config_dict.keys())[0], + label='Model', + multiselect=False) num_samples = gr.Slider( - label="Images", minimum=1, maximum=12, value=1, step=1) - + label="Images", minimum=1, maximum=12, value=1, step=1) + enable_auto_prompt = gr.Checkbox(label='Auto generated BLIP2 prompt', value=True) with gr.Accordion("Advanced options", open=False): image_resolution = gr.Slider( @@ -253,22 +250,24 @@ def create_demo(): scale = gr.Slider( label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1) seed = gr.Slider(label="Seed", minimum=-1, - maximum=2147483647, step=1, randomize=True) + maximum=2147483647, step=1, randomize=True) eta = gr.Number(label="eta (DDIM)", value=0.0) a_prompt = gr.Textbox( label="Added Prompt", value='best quality, extremely detailed') n_prompt = gr.Textbox(label="Negative Prompt", - value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality') + value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality') with gr.Column(): result_gallery = gr.Gallery( label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto') result_text = gr.Text(label='BLIP2+Human Prompt Text') - ips = [condition_model, input_image, enable_auto_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, - detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta] + ips = [condition_model, input_image, enable_auto_prompt, prompt, a_prompt, n_prompt, num_samples, + image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta] run_button.click(fn=process, inputs=ips, outputs=[result_gallery, result_text]) return demo + if __name__ == '__main__': demo = create_demo() demo.queue().launch(server_name='0.0.0.0') diff --git a/sam2semantic.py b/sam2semantic.py new file mode 100644 index 0000000000000000000000000000000000000000..16658c04325f66502b6874fdfa68ef7e1645fc61 --- /dev/null +++ b/sam2semantic.py @@ -0,0 +1,174 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +# pip install mmcv + +from torchvision.utils import save_image +from PIL import Image +import subprocess +from collections import OrderedDict +import numpy as np +import cv2 +import textwrap +import torch +import os +from annotator.util import resize_image, HWC3 +import mmcv +import random + +# device = "cuda" if torch.cuda.is_available() else "cpu" # > 15GB GPU memory required +device = "cpu" +use_blip = True +use_gradio = True + +if device == 'cpu': + data_type = torch.float32 +else: + data_type = torch.float16 +# Diffusion init using diffusers. + +# diffusers==0.14.0 required. +from diffusers.utils import load_image + +base_model_path = "stabilityai/stable-diffusion-2-inpainting" +config_dict = OrderedDict([('SAM Pretrained(v0-1): Good Natural Sense', 'shgao/edit-anything-v0-1-1'), + ('LAION Pretrained(v0-3): Good Face', 'shgao/edit-anything-v0-3'), + ('SD Inpainting: Not keep position', 'stabilityai/stable-diffusion-2-inpainting') + ]) + +# Segment-Anything init. +# pip install git+https://github.com/facebookresearch/segment-anything.git +try: + from segment_anything import sam_model_registry, SamAutomaticMaskGenerator +except ImportError: + print('segment_anything not installed') + result = subprocess.run(['pip', 'install', 'git+https://github.com/facebookresearch/segment-anything.git'], check=True) + print(f'Install segment_anything {result}') + from segment_anything import sam_model_registry, SamAutomaticMaskGenerator +if not os.path.exists('./models/sam_vit_h_4b8939.pth'): + result = subprocess.run(['wget', 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth', '-P', 'models'], check=True) + print(f'Download sam_vit_h_4b8939.pth {result}') +sam_checkpoint = "models/sam_vit_h_4b8939.pth" +model_type = "default" +sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) +sam.to(device=device) +mask_generator = SamAutomaticMaskGenerator(sam) + + +# BLIP2 init. +if use_blip: + # need the latest transformers + # pip install git+https://github.com/huggingface/transformers.git + from transformers import AutoProcessor, Blip2ForConditionalGeneration + processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b") + blip_model = Blip2ForConditionalGeneration.from_pretrained( + "Salesforce/blip2-opt-2.7b", torch_dtype=data_type) + + +def region_classify_w_blip2(image): + inputs = processor(image, return_tensors="pt").to(device, data_type) + generated_ids = blip_model.generate(**inputs, max_new_tokens=15) + generated_text = processor.batch_decode( + generated_ids, skip_special_tokens=True)[0].strip() + return generated_text + +def region_level_semantic_api(image, topk=5): + """ + rank regions by area, and classify each region with blip2 + Args: + image: numpy array + topk: int + Returns: + topk_region_w_class_label: list of dict with key 'class_label' + """ + topk_region_w_class_label = [] + anns = mask_generator.generate(image) + if len(anns) == 0: + return [] + sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True) + for i in range(min(topk, len(sorted_anns))): + ann = anns[i] + m = ann['segmentation'] + m_3c = m[:,:, np.newaxis] + m_3c = np.concatenate((m_3c,m_3c,m_3c), axis=2) + bbox = ann['bbox'] + region = mmcv.imcrop(image*m_3c, np.array([bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3]]), scale=1) + region_class_label = region_classify_w_blip2(region) + ann['class_label'] = region_class_label + print(ann['class_label'], str(bbox)) + topk_region_w_class_label.append(ann) + return topk_region_w_class_label + +def show_semantic_image_label(anns): + """ + show semantic image label for each region + Args: + anns: list of dict with key 'class_label' + Returns: + full_img: numpy array + """ + full_img = None + # generate mask image + for i in range(len(anns)): + m = anns[i]['segmentation'] + if full_img is None: + full_img = np.zeros((m.shape[0], m.shape[1], 3)) + color_mask = np.random.random((1, 3)).tolist()[0] + full_img[m != 0] = color_mask + full_img = full_img*255 + # add text on this mask image + for i in range(len(anns)): + m = anns[i]['segmentation'] + class_label = anns[i]['class_label'] + # add text to region + # Calculate the centroid of the region to place the text + y, x = np.where(m != 0) + x_center, y_center = int(np.mean(x)), int(np.mean(y)) + + # Split the text into multiple lines + max_width = 20 # Adjust this value based on your preferred maximum width + wrapped_text = textwrap.wrap(class_label, width=max_width) + + # Add text to region + font = cv2.FONT_HERSHEY_SIMPLEX + font_scale = 1.2 + font_thickness = 2 + font_color = (random.randint(0, 255), random.randint(0, 255), random.randint(0, 255)) # red + line_spacing = 40 # Adjust this value based on your preferred line + + for idx, line in enumerate(wrapped_text): + y_offset = y_center - (len(wrapped_text) - 1) * line_spacing // 2 + idx * line_spacing + text_size = cv2.getTextSize(line, font, font_scale, font_thickness)[0] + x_offset = x_center - text_size[0] // 2 + # Draw the text multiple times with small offsets to create a bolder appearance + offsets = [(-1, -1), (-1, 0), (-1, 1), (0, -1), (0, 1), (1, -1), (1, 0), (1, 1)] + for off_x, off_y in offsets: + cv2.putText(full_img, line, (x_offset + off_x, y_offset + off_y), font, font_scale, font_color, font_thickness, cv2.LINE_AA) + + return full_img + + + +image_path = "images/sa_224577.jpg" +input_image = Image.open(image_path) +detect_resolution=1024 +input_image = resize_image(np.array(input_image, dtype=np.uint8), detect_resolution) +region_level_annots = region_level_semantic_api(input_image, topk=5) +output = show_semantic_image_label(region_level_annots) + +image_list = [] +input_image = resize_image(input_image, 512) +output = resize_image(output, 512) +input_image = np.array(input_image, dtype=np.uint8) +output = np.array(output, dtype=np.uint8) +image_list.append(torch.tensor(input_image).float()) +image_list.append(torch.tensor(output).float()) +for each in image_list: + print(each.shape, type(each)) + print(each.max(), each.min()) + + +image_list = torch.stack(image_list).permute(0, 3, 1, 2) +print(image_list.shape) + +save_image(image_list, "images/sample_semantic.jpg", nrow=2, + normalize=True) + diff --git a/sam2vlpart_edit.py b/sam2vlpart_edit.py new file mode 100644 index 0000000000000000000000000000000000000000..f4018329ec767883827e3fb6e006b7505cbe170d --- /dev/null +++ b/sam2vlpart_edit.py @@ -0,0 +1,328 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +from torchvision.utils import save_image +from PIL import Image +from cldm.ddim_hacked import DDIMSampler +from cldm.model import create_model, load_state_dict +from pytorch_lightning import seed_everything +from share import * +import config + +import cv2 +import einops +import gradio as gr +import numpy as np +import torch +import random +import os +import requests +from io import BytesIO +from annotator.util import resize_image, HWC3 + +device = "cuda" if torch.cuda.is_available() else "cpu" +use_blip = False +use_gradio = False + +# Diffusion init using diffusers. + +from vlpart.vlpart import build_vlpart +from segment_anything import build_sam, SamPredictor +from segment_anything.utils.amg import remove_small_regions +import detectron2.data.transforms as T + +# diffusers==0.14.0 required. +from diffusers import ControlNetModel, UniPCMultistepScheduler +from utils.stable_diffusion_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline +from diffusers.utils import load_image +import torch + +base_model_path = "stabilityai/stable-diffusion-2-inpainting" +controlnet_path = "shgao/edit-anything-v0-1-1" + +controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16) +pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( + base_model_path, controlnet=controlnet, torch_dtype=torch.float16 +) +# speed up diffusion process with faster scheduler and memory optimization +pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) +# remove following line if xformers is not installed +pipe.enable_xformers_memory_efficient_attention() + +# pipe.enable_model_cpu_offload() # disable for now because of unknow bug in accelerate +pipe.to(device) + +# Segment-Anything init. +# pip install git+https://github.com/facebookresearch/segment-anything.git +from segment_anything import sam_model_registry, SamAutomaticMaskGenerator + +sam_checkpoint = "./models/sam_vit_h_4b8939.pth" +vlpart_checkpoint = "./models/swinbase_part_0a0000.pth" +model_type = "default" +sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) +sam.to(device=device) +mask_generator = SamAutomaticMaskGenerator(sam) +vlpart = build_vlpart(checkpoint=vlpart_checkpoint) +vlpart.to(device=device) +sam_predictor = SamPredictor(build_sam(checkpoint=sam_checkpoint).to(device=device)) + +# BLIP2 init. +if use_blip: + # need the latest transformers + # pip install git+https://github.com/huggingface/transformers.git + from transformers import AutoProcessor, Blip2ForConditionalGeneration + + processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b") + blip_model = Blip2ForConditionalGeneration.from_pretrained( + "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16) + blip_model.to(device) + blip_model.to(device) + + +def get_blip2_text(image): + inputs = processor(image, return_tensors="pt").to(device, torch.float16) + generated_ids = blip_model.generate(**inputs, max_new_tokens=50) + generated_text = processor.batch_decode( + generated_ids, skip_special_tokens=True)[0].strip() + return generated_text + + +def show_anns(anns): + if len(anns) == 0: + return + sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True) + full_img = None + + # for ann in sorted_anns: + for i in range(len(sorted_anns)): + ann = anns[i] + m = ann['segmentation'] + if full_img is None: + full_img = np.zeros((m.shape[0], m.shape[1], 3)) + map = np.zeros((m.shape[0], m.shape[1]), dtype=np.uint16) + map[m != 0] = i + 1 + color_mask = np.random.random((1, 3)).tolist()[0] + full_img[m != 0] = color_mask + full_img = full_img * 255 + # anno encoding from https://github.com/LUSSeg/ImageNet-S + res = np.zeros((map.shape[0], map.shape[1], 3)) + res[:, :, 0] = map % 256 + res[:, :, 1] = map // 256 + res.astype(np.float32) + return full_img, res + + +def get_sam_control(image): + masks = mask_generator.generate(image) + full_img, res = show_anns(masks) + return full_img, res + + +def prompt2mask(original_image, text_prompt): + # original_image = original_image[:, :, :3] + preprocess = T.ResizeShortestEdge([800, 800], 1333) + height, width = original_image.shape[:2] + image = preprocess.get_transform(original_image).apply_image(original_image) + image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1)) + inputs = {"image": image, "height": height, "width": width} + with torch.no_grad(): + predictions = vlpart.inference([inputs], text_prompt=text_prompt)[0] + boxes, masks = None, None + filter_scores, filter_boxes, filter_classes = [], [], [] + + if "instances" in predictions: + instances = predictions['instances'].to('cpu') + boxes = instances.pred_boxes.tensor if instances.has("pred_boxes") else None + scores = instances.scores if instances.has("scores") else None + classes = instances.pred_classes.tolist() if instances.has("pred_classes") else None + + num_obj = len(scores) + for obj_ind in range(num_obj): + category_score = scores[obj_ind] + if category_score < 0.7: + continue + filter_scores.append(category_score) + filter_boxes.append(boxes[obj_ind]) + filter_classes.append(classes[obj_ind]) + + final_m = torch.zeros((original_image.shape[0], original_image.shape[1])) + + if len(filter_boxes) > 0: + # sam model inference + sam_predictor.set_image(original_image) + + boxes_filter = torch.stack(filter_boxes) + transformed_boxes = sam_predictor.transform.apply_boxes_torch(boxes_filter, original_image.shape[:2]) + masks, _, _ = sam_predictor.predict_torch( + point_coords=None, + point_labels=None, + boxes=transformed_boxes.to(device), + multimask_output=False, + ) + + # remove small disconnected regions and holes + fine_masks = [] + for mask in masks.to('cpu').numpy(): # masks: [num_masks, 1, h, w] + fine_masks.append(remove_small_regions(mask[0], 400, mode="holes")[0]) + masks = np.stack(fine_masks, axis=0)[:, np.newaxis] + masks = torch.from_numpy(masks) + + num_obj = len(scores) + for obj_ind in range(num_obj): + # box = boxes[obj_ind] + score = scores[obj_ind] + if score < 0.5: + continue + m = masks[obj_ind][0] + final_m += m + final_m = (final_m > 0).to('cpu').numpy() + # print(final_m.max(), final_m.min()) + return np.dstack((final_m, final_m, final_m)) * 255 + + +def process(input_image, mask_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, detect_resolution, + ddim_steps, guess_mode, strength, scale, seed, eta): + with torch.no_grad(): + mask_image = np.array(prompt2mask(input_image, mask_prompt), dtype=np.uint8) + + if use_blip: + print("Generating text:") + blip2_prompt = get_blip2_text(input_image) + print("Generated text:", blip2_prompt) + if len(prompt) > 0: + prompt = blip2_prompt + ',' + prompt + else: + prompt = blip2_prompt + print("All text:", prompt) + + input_image = HWC3(input_image) + + img = resize_image(input_image, image_resolution) + H, W, C = img.shape + + print("Generating SAM seg:") + # the default SAM model is trained with 1024 size. + full_segmask, detected_map = get_sam_control( + resize_image(input_image, detect_resolution)) + + detected_map = HWC3(detected_map.astype(np.uint8)) + detected_map = cv2.resize( + detected_map, (W, H), interpolation=cv2.INTER_LINEAR) + + control = torch.from_numpy( + detected_map.copy()).float().cuda() + control = torch.stack([control for _ in range(num_samples)], dim=0) + control = einops.rearrange(control, 'b h w c -> b c h w').clone() + + mask_image = HWC3(mask_image.astype(np.uint8)) + mask_image = cv2.resize( + mask_image, (W, H), interpolation=cv2.INTER_LINEAR) + mask_image = Image.fromarray(mask_image) + + if seed == -1: + seed = random.randint(0, 65535) + seed_everything(seed) + generator = torch.manual_seed(seed) + x_samples = pipe( + image=img, + mask_image=mask_image, + prompt=[prompt + ', ' + a_prompt] * num_samples, + negative_prompt=[n_prompt] * num_samples, + num_images_per_prompt=num_samples, + num_inference_steps=ddim_steps, + generator=generator, + controlnet_conditioning_image=control.type(torch.float16), + height=H, + width=W, + ).images + + results = [x_samples[i] for i in range(num_samples)] + return [full_segmask, mask_image] + results + + +def download_image(url): + response = requests.get(url) + return Image.open(BytesIO(response.content)).convert("RGB") + + +# disable gradio when not using GUI. +if not use_gradio: + image_path = "assets/dog.png" + input_image = Image.open(image_path) + + input_image = np.array(input_image, dtype=np.uint8)[:, :, :3] + + mask_prompt = 'dog head.' + # mask_image = np.array(prompt2mask(input_image, mask_prompt), dtype=np.uint8) + prompt = "cute dogs" + a_prompt = 'best quality, extremely detailed' + n_prompt = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality' + num_samples = 3 + image_resolution = 512 + detect_resolution = 512 + ddim_steps = 30 + guess_mode = False + strength = 1.0 + scale = 9.0 + seed = -1 + eta = 0.0 + + outputs = process(input_image, mask_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta) + + image_list = [] + input_image = resize_image(input_image, 512) + image_list.append(torch.tensor(input_image)) + for i in range(len(outputs)): + each = outputs[i] + if type(each) is not np.ndarray: + each = np.array(each, dtype=np.uint8) + each = resize_image(each, 512) + print(i, each.shape) + image_list.append(torch.tensor(each)) + + image_list = torch.stack(image_list).permute(0, 3, 1, 2) + + save_image(image_list, "sample.jpg", nrow=3, + normalize=True, value_range=(0, 255)) +else: + print("The GUI is not tested yet. Please open an issue if you find bugs.") + block = gr.Blocks().queue() + with block: + with gr.Row(): + gr.Markdown( + "## Edit Anything powered by ControlNet+SAM+BLIP2+Stable Diffusion") + with gr.Row(): + with gr.Column(): + input_image = gr.Image(source='upload', type="numpy") + prompt = gr.Textbox(label="Prompt") + run_button = gr.Button(label="Run") + with gr.Accordion("Advanced options", open=False): + num_samples = gr.Slider( + label="Images", minimum=1, maximum=12, value=1, step=1) + image_resolution = gr.Slider( + label="Image Resolution", minimum=256, maximum=768, value=512, step=64) + strength = gr.Slider( + label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01) + guess_mode = gr.Checkbox(label='Guess Mode', value=False) + detect_resolution = gr.Slider( + label="SAM Resolution", minimum=128, maximum=2048, value=1024, step=1) + ddim_steps = gr.Slider( + label="Steps", minimum=1, maximum=100, value=20, step=1) + scale = gr.Slider( + label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1) + seed = gr.Slider(label="Seed", minimum=-1, + maximum=2147483647, step=1, randomize=True) + eta = gr.Number(label="eta (DDIM)", value=0.0) + mask_prompt = gr.Textbox( + label="Mask Prompt", value='best quality, extremely detailed') + a_prompt = gr.Textbox( + label="Added Prompt", value='best quality, extremely detailed') + n_prompt = gr.Textbox(label="Negative Prompt", + value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality') + with gr.Column(): + result_gallery = gr.Gallery( + label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto') + ips = [input_image, mask_prompt, prompt, a_prompt, n_prompt, num_samples, image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta] + run_button.click(fn=process, inputs=ips, outputs=[result_gallery]) + + block.launch(server_name='0.0.0.0') diff --git a/sam_train_sd21.py b/sam_train_sd21.py new file mode 100644 index 0000000000000000000000000000000000000000..75d76a6371d20e643bb4de6fd4637ab895b60f4d --- /dev/null +++ b/sam_train_sd21.py @@ -0,0 +1,39 @@ +from share import * + +import pytorch_lightning as pl +from torch.utils.data import DataLoader +from utils.sam_dataset import SAMDataset +from cldm.logger import ImageLogger +from cldm.model import create_model, load_state_dict +import torch + + +# Configs +resume_path = './models/control_sd21_ini.ckpt' +batch_size = 4 +logger_freq = 300 +learning_rate = 1e-5 +sd_locked = True +only_mid_control = False +data_path = '../data/files' +txt_path = '../data/data_85616.txt' + + +# First use cpu to load models. Pytorch Lightning will automatically move it to GPUs. +model = create_model('./models/cldm_v21.yaml').cpu() +model.load_state_dict(load_state_dict(resume_path, location='cpu')) +model.learning_rate = learning_rate +model.sd_locked = sd_locked +model.only_mid_control = only_mid_control + + +# Misc +dataset = SAMDataset(data_path=data_path, txt_path=txt_path) +dataloader = DataLoader(dataset, num_workers=16, + batch_size=batch_size, shuffle=True) +logger = ImageLogger(batch_frequency=logger_freq) +trainer = pl.Trainer(gpus=8, strategy="ddp", precision=32, callbacks=[logger]) + + +# Train! +trainer.fit(model, dataloader) diff --git a/share.py b/share.py new file mode 100644 index 0000000000000000000000000000000000000000..463af08fb936d650b5dd2e66183661181c34a3d6 --- /dev/null +++ b/share.py @@ -0,0 +1,8 @@ +import config +from cldm.hack import disable_verbosity, enable_sliced_attention + + +disable_verbosity() + +if config.save_memory: + enable_sliced_attention() diff --git a/sketch2image.py b/sketch2image.py new file mode 100644 index 0000000000000000000000000000000000000000..bdebe48f4fa980ad991a44011524a237bd0682bd --- /dev/null +++ b/sketch2image.py @@ -0,0 +1,243 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +from diffusers.utils import load_image +from diffusers import UniPCMultistepScheduler +from torchvision.utils import save_image +from PIL import Image +from pytorch_lightning import seed_everything +import subprocess +from collections import OrderedDict + +import cv2 +import einops +import gradio as gr +import numpy as np +import torch +import random +import os +from annotator.util import resize_image, HWC3 +import base64 +from io import BytesIO + +from utils.stable_diffusion_controlnet import StableDiffusionControlNetPipeline2, ControlNetModel2 + + +def create_demo(): + MAX_COLORS = 12 + canvas_html = "
" + load_js = """ + async () => { + const url = "https://huggingface.co/datasets/radames/gradio-components/raw/main/sketch-canvas.js" + fetch(url) + .then(res => res.text()) + .then(text => { + const script = document.createElement('script'); + script.type = "module" + script.src = URL.createObjectURL(new Blob([text], { type: 'application/javascript' })); + document.head.appendChild(script); + }); + } + """ + + get_js_colors = """ + async (canvasData) => { + const canvasEl = document.getElementById("canvas-root"); + return [canvasEl._data] + } + """ + + set_canvas_size = """ + async (aspect) => { + if(aspect ==='square'){ + _updateCanvas(512,512) + } + if(aspect ==='horizontal'){ + _updateCanvas(768,512) + } + if(aspect ==='vertical'){ + _updateCanvas(512,768) + } + } + """ + + device = "cuda" if torch.cuda.is_available() else "cpu" + # aspect = gr.Radio(["square", "horizontal", "vertical"], value="square", label="Aspect Ratio", visible=False if is_shared_ui else True) + + # Diffusion init using diffusers. + # diffusers==0.14.0 required. + + base_model_path = "stabilityai/stable-diffusion-2-1" + + config_dict = OrderedDict([('SAM Pretrained(v0-1)', 'shgao/edit-anything-v0-1-1'), + ('LAION Pretrained(v0-3)', 'shgao/edit-anything-v0-3'), + ('LAION Pretrained(v0-4)', 'shgao/edit-anything-v0-4-sd21'), + ]) + + def obtain_generation_model(controlnet_path): + controlnet = ControlNetModel2.from_pretrained( + controlnet_path, torch_dtype=torch.float16) + pipe = StableDiffusionControlNetPipeline2.from_pretrained( + base_model_path, controlnet=controlnet, torch_dtype=torch.float16 + ) + # speed up diffusion process with faster scheduler and memory optimization + pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) + # remove following line if xformers is not installed + pipe.enable_xformers_memory_efficient_attention() + # pipe.enable_model_cpu_offload() # disable for now because of unknow bug in accelerate + pipe.to(device) + return pipe + + global default_controlnet_path + default_controlnet_path = config_dict['LAION Pretrained(v0-4)'] + pipe = obtain_generation_model(default_controlnet_path) + + def get_sam_control(image): + im2arr = np.array(image) + colors_map, res = None, None + ptr = 0 + for color in colors: + r, g, b = color + if any(c != 255 for c in (r, g, b)): + binary_matrix = np.all(im2arr == (r, g, b), axis=-1) + if colors_map is None: + colors_map = np.zeros((im2arr.shape[0], im2arr.shape[1]), dtype=np.uint16) + res = np.zeros((im2arr.shape[0], im2arr.shape[1], 3)) + colors_map[binary_matrix != 0] = ptr + 1 + ptr += 1 + white = np.all(im2arr == (255, 255, 255), axis=-1) + scale_map = (white != 1).astype(np.float32) + res[:, :, 0] = colors_map % 256 + res[:, :, 1] = colors_map // 256 + res.astype(np.float32) + return image, res, scale_map + + def process_sketch(canvas_data): + nonlocal colors + base64_img = canvas_data['image'] + image_data = base64.b64decode(base64_img.split(',')[1]) + image = Image.open(BytesIO(image_data)).convert("RGB") + colors = [tuple(map(int, rgb[4:-1].split(','))) for rgb in canvas_data['colors']] + print(colors) + # binary_matrixes['sketch'] = res + return image, "sketch loaded." + + def process(condition_model, input_image, control_scale, prompt, a_prompt, n_prompt, + num_samples, image_resolution, ddim_steps, guess_mode, use_scale_map, strength, scale, seed, eta): + + global default_controlnet_path + global pipe + print("To Use:", config_dict[condition_model], "Current:", default_controlnet_path) + if default_controlnet_path != config_dict[condition_model]: + print("Change condition model to:", config_dict[condition_model]) + pipe = obtain_generation_model(config_dict[condition_model]) + default_controlnet_path = config_dict[condition_model] + + with torch.no_grad(): + print("All text:", prompt) + + input_image = HWC3(input_image) + + img = resize_image(input_image, image_resolution) + H, W, C = img.shape + + # the default SAM model is trained with 1024 size. + fullseg, detected_map, scale_map = get_sam_control(input_image) + + detected_map = HWC3(detected_map.astype(np.uint8)) + detected_map = cv2.resize( + detected_map, (W, H), interpolation=cv2.INTER_LINEAR) + + control = torch.from_numpy( + detected_map.copy()).float().cuda() + control = torch.stack([control for _ in range(num_samples)], dim=0) + control = einops.rearrange(control, 'b h w c -> b c h w').clone() + + scale_map = torch.from_numpy(scale_map).float().cuda() if use_scale_map else None + + if seed == -1: + seed = random.randint(0, 65535) + seed_everything(seed) + print("control.shape", control.shape) + generator = torch.manual_seed(seed) + x_samples = pipe( + prompt=[prompt + ', ' + a_prompt] * num_samples, + negative_prompt=[n_prompt] * num_samples, + num_images_per_prompt=num_samples, + num_inference_steps=ddim_steps, + generator=generator, + height=H, + width=W, + controlnet_conditioning_scale=float(control_scale), + controlnet_conditioning_scale_map=scale_map, + image=control.type(torch.float16), + ).images + + results = [x_samples[i] for i in range(num_samples)] + return [fullseg] + results, prompt, "waiting for sketch..." + + # disable gradio when not using GUI. + block = gr.Blocks() + with block as demo: + colors = [] + with gr.Row(): + gr.Markdown( + "## Generate Anything") + with gr.Row(): + with gr.Column(): + canvas_data = gr.JSON(value={}, visible=False) + canvas = gr.HTML(canvas_html) + aspect = gr.Radio(["square", "horizontal", "vertical"], value="square", label="Aspect Ratio", + visible=False) + button_run = gr.Button("I've finished my sketch", elem_id="main_button", interactive=True) + result_text1 = gr.Text(label='sketch status:') + + with gr.Column(visible=True) as post_sketch: + input_image = gr.Image(type="numpy", visible=False) + prompt = gr.Textbox(label="Prompt (Optional)") + run_button = gr.Button(label="Run") + condition_model = gr.Dropdown(choices=list(config_dict.keys()), + value=list(config_dict.keys())[0], + label='Model', + multiselect=False) + control_scale = gr.Slider( + label="Mask Align strength", info="Large value -> strict alignment with SAM mask", minimum=0, + maximum=1, value=1, step=0.1) + num_samples = gr.Slider( + label="Images", minimum=1, maximum=12, value=1, step=1) + + # enable_auto_prompt = True + with gr.Accordion("Advanced options", open=False): + image_resolution = gr.Slider( + label="Image Resolution", minimum=256, maximum=768, value=512, step=64) + strength = gr.Slider( + label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01) + guess_mode = gr.Checkbox(label='Guess Mode', value=False) + use_scale_map = gr.Checkbox(label='Use scale map', value=False) + ddim_steps = gr.Slider( + label="Steps", minimum=1, maximum=100, value=20, step=1) + scale = gr.Slider( + label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1) + seed = gr.Slider(label="Seed", minimum=-1, + maximum=2147483647, step=1, randomize=True) + eta = gr.Number(label="eta (DDIM)", value=0.0) + a_prompt = gr.Textbox( + label="Added Prompt", value='best quality, extremely detailed') + n_prompt = gr.Textbox(label="Negative Prompt", + value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality') + + with gr.Column(): + result_gallery = gr.Gallery( + label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto') + result_text = gr.Text(label='BLIP2+Human Prompt Text') + aspect.change(None, inputs=[aspect], outputs=None, _js=set_canvas_size) + button_run.click(process_sketch, inputs=[canvas_data], + outputs=[input_image, result_text1], _js=get_js_colors, queue=False) + ips = [condition_model, input_image, control_scale, prompt, a_prompt, n_prompt, + num_samples, image_resolution, ddim_steps, guess_mode, use_scale_map, strength, scale, seed, eta] + run_button.click(fn=process, inputs=ips, outputs=[result_gallery, result_text, result_text1]) + demo.load(None, None, None, _js=load_js) + return demo + + +if __name__ == '__main__': + demo = create_demo() + demo.queue().launch(server_name='0.0.0.0', share=True) diff --git a/tool_transfer_control.py b/tool_transfer_control.py new file mode 100644 index 0000000000000000000000000000000000000000..b84442cc93f7f9c30cb7311b8675d9124a6e8ec9 --- /dev/null +++ b/tool_transfer_control.py @@ -0,0 +1,59 @@ +path_sd15 = './models/v1-5-pruned.ckpt' +path_sd15_with_control = './models/control_sd15_openpose.pth' +path_input = './models/anything-v3-full.safetensors' +path_output = './models/control_any3_openpose.pth' + + +import os + + +assert os.path.exists(path_sd15), 'Input path_sd15 does not exists!' +assert os.path.exists(path_sd15_with_control), 'Input path_sd15_with_control does not exists!' +assert os.path.exists(path_input), 'Input path_input does not exists!' +assert os.path.exists(os.path.dirname(path_output)), 'Output folder not exists!' + + +import torch +from share import * +from cldm.model import load_state_dict + + +sd15_state_dict = load_state_dict(path_sd15) +sd15_with_control_state_dict = load_state_dict(path_sd15_with_control) +input_state_dict = load_state_dict(path_input) + + +def get_node_name(name, parent_name): + if len(name) <= len(parent_name): + return False, '' + p = name[:len(parent_name)] + if p != parent_name: + return False, '' + return True, name[len(parent_name):] + + +keys = sd15_with_control_state_dict.keys() + +final_state_dict = {} +for key in keys: + is_first_stage, _ = get_node_name(key, 'first_stage_model') + is_cond_stage, _ = get_node_name(key, 'cond_stage_model') + if is_first_stage or is_cond_stage: + final_state_dict[key] = input_state_dict[key] + continue + p = sd15_with_control_state_dict[key] + is_control, node_name = get_node_name(key, 'control_') + if is_control: + sd15_key_name = 'model.diffusion_' + node_name + else: + sd15_key_name = key + if sd15_key_name in input_state_dict: + p_new = p + input_state_dict[sd15_key_name] - sd15_state_dict[sd15_key_name] + # print(f'Offset clone from [{sd15_key_name}] to [{key}]') + else: + p_new = p + # print(f'Direct clone to [{key}]') + final_state_dict[key] = p_new + +torch.save(final_state_dict, path_output) +print('Transferred model saved at ' + path_output) diff --git a/tools/convert_controlnet_to_diffusers.py b/tools/convert_controlnet_to_diffusers.py new file mode 100644 index 0000000000000000000000000000000000000000..d11bccea15b9e9abdaa80669a87ba67c92bc6179 --- /dev/null +++ b/tools/convert_controlnet_to_diffusers.py @@ -0,0 +1,91 @@ +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. +# +# 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. +""" Conversion script for stable diffusion checkpoints which _only_ contain a contrlnet. """ + +import argparse + +from diffusers.pipelines.stable_diffusion.convert_from_ckpt import download_controlnet_from_original_ckpt + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + + parser.add_argument( + "--checkpoint_path", default=None, type=str, required=True, help="Path to the checkpoint to convert." + ) + parser.add_argument( + "--original_config_file", + type=str, + required=True, + help="The YAML config file corresponding to the original architecture.", + ) + parser.add_argument( + "--num_in_channels", + default=None, + type=int, + help="The number of input channels. If `None` number of input channels will be automatically inferred.", + ) + parser.add_argument( + "--image_size", + default=512, + type=int, + help=( + "The image size that the model was trained on. Use 512 for Stable Diffusion v1.X and Stable Siffusion v2" + " Base. Use 768 for Stable Diffusion v2." + ), + ) + parser.add_argument( + "--extract_ema", + action="store_true", + help=( + "Only relevant for checkpoints that have both EMA and non-EMA weights. Whether to extract the EMA weights" + " or not. Defaults to `False`. Add `--extract_ema` to extract the EMA weights. EMA weights usually yield" + " higher quality images for inference. Non-EMA weights are usually better to continue fine-tuning." + ), + ) + parser.add_argument( + "--upcast_attention", + action="store_true", + help=( + "Whether the attention computation should always be upcasted. This is necessary when running stable" + " diffusion 2.1." + ), + ) + parser.add_argument( + "--from_safetensors", + action="store_true", + help="If `--checkpoint_path` is in `safetensors` format, load checkpoint with safetensors instead of PyTorch.", + ) + parser.add_argument( + "--to_safetensors", + action="store_true", + help="Whether to store pipeline in safetensors format or not.", + ) + parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.") + parser.add_argument("--device", type=str, help="Device to use (e.g. cpu, cuda:0, cuda:1, etc.)") + args = parser.parse_args() + + controlnet = download_controlnet_from_original_ckpt( + checkpoint_path=args.checkpoint_path, + original_config_file=args.original_config_file, + image_size=args.image_size, + extract_ema=args.extract_ema, + num_in_channels=args.num_in_channels, + upcast_attention=args.upcast_attention, + from_safetensors=args.from_safetensors, + device=args.device, + ) + + controlnet.save_pretrained(args.dump_path, safe_serialization=args.to_safetensors) \ No newline at end of file diff --git a/tools/sam2image_ori_version.py b/tools/sam2image_ori_version.py new file mode 100644 index 0000000000000000000000000000000000000000..a82b8a56a7b59cada5256e8a830226719d480704 --- /dev/null +++ b/tools/sam2image_ori_version.py @@ -0,0 +1,234 @@ +# Edit Anything trained with Stable Diffusion + ControlNet + SAM + BLIP2 +from torchvision.utils import save_image +from PIL import Image +from cldm.ddim_hacked import DDIMSampler +from cldm.model import create_model, load_state_dict +from pytorch_lightning import seed_everything +from share import * +import config + +import cv2 +import einops +import gradio as gr +import numpy as np +import torch +import random +import os +from annotator.util import resize_image, HWC3 + + +device = "cuda" if torch.cuda.is_available() else "cpu" +use_blip = True +use_gradio = False + +# Diffusion init. + +model = create_model('./models/cldm_v21.yaml').cpu() +model.load_state_dict(load_state_dict( + 'models/edit-anything-ckpt-v0-1.ckpt', location='cuda')) +model.to(device=device) +ddim_sampler = DDIMSampler(model) + + +# Segment-Anything init. +# pip install git+https://github.com/facebookresearch/segment-anything.git +from segment_anything import sam_model_registry, SamAutomaticMaskGenerator + +sam_checkpoint = "models/sam_vit_h_4b8939.pth" +model_type = "default" +sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) +sam.to(device=device) +mask_generator = SamAutomaticMaskGenerator(sam) + + +# BLIP2 init. +if use_blip: + # need the latest transformers + # pip install git+https://github.com/huggingface/transformers.git + from transformers import AutoProcessor, Blip2ForConditionalGeneration + + processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b") + blip_model = Blip2ForConditionalGeneration.from_pretrained( + "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16) + blip_model.to(device) + + +def get_blip2_text(image): + inputs = processor(image, return_tensors="pt").to(device, torch.float16) + generated_ids = blip_model.generate(**inputs, max_new_tokens=50) + generated_text = processor.batch_decode( + generated_ids, skip_special_tokens=True)[0].strip() + return generated_text + + +def show_anns(anns): + if len(anns) == 0: + return + sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True) + full_img = None + + # for ann in sorted_anns: + for i in range(len(sorted_anns)): + ann = anns[i] + m = ann['segmentation'] + if full_img is None: + full_img = np.zeros((m.shape[0], m.shape[1], 3)) + map = np.zeros((m.shape[0], m.shape[1]), dtype=np.uint16) + map[m != 0] = i + 1 + color_mask = np.random.random((1, 3)).tolist()[0] + full_img[m != 0] = color_mask + full_img = full_img*255 + # anno encoding from https://github.com/LUSSeg/ImageNet-S + res = np.zeros((map.shape[0], map.shape[1], 3)) + res[:, :, 0] = map % 256 + res[:, :, 1] = map // 256 + res.astype(np.float32) + return full_img, res + + +def get_sam_control(image): + masks = mask_generator.generate(image) + full_img, res = show_anns(masks) + return full_img, res + + +def process(input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta): + with torch.no_grad(): + if use_blip: + print("Generating text:") + blip2_prompt = get_blip2_text(input_image) + print("Generated text:", blip2_prompt) + if len(prompt)>0: + prompt = blip2_prompt + ',' + prompt + else: + prompt = blip2_prompt + print("All text:", prompt) + + input_image = HWC3(input_image) + + img = resize_image(input_image, image_resolution) + H, W, C = img.shape + + print("Generating SAM seg:") + # the default SAM model is trained with 1024 size. + full_segmask, detected_map = get_sam_control( + resize_image(input_image, detect_resolution)) + + detected_map = HWC3(detected_map.astype(np.uint8)) + detected_map = cv2.resize( + detected_map, (W, H), interpolation=cv2.INTER_LINEAR) + + control = torch.from_numpy( + detected_map.copy()).float().cuda() + control = torch.stack([control for _ in range(num_samples)], dim=0) + control = einops.rearrange(control, 'b h w c -> b c h w').clone() + + if seed == -1: + seed = random.randint(0, 65535) + seed_everything(seed) + + if config.save_memory: + model.low_vram_shift(is_diffusing=False) + + cond = {"c_concat": [control], "c_crossattn": [ + model.get_learned_conditioning([prompt + ', ' + a_prompt] * num_samples)]} + un_cond = {"c_concat": None if guess_mode else [control], "c_crossattn": [ + model.get_learned_conditioning([n_prompt] * num_samples)]} + shape = (4, H // 8, W // 8) + + if config.save_memory: + model.low_vram_shift(is_diffusing=True) + + model.control_scales = [strength * (0.825 ** float(12 - i)) for i in range(13)] if guess_mode else ( + [strength] * 13) # Magic number. IDK why. Perhaps because 0.825**12<0.01 but 0.826**12>0.01 + samples, intermediates = ddim_sampler.sample(ddim_steps, num_samples, + shape, cond, verbose=False, eta=eta, + unconditional_guidance_scale=scale, + unconditional_conditioning=un_cond) + + if config.save_memory: + model.low_vram_shift(is_diffusing=False) + + x_samples = model.decode_first_stage(samples) + x_samples = (einops.rearrange(x_samples, 'b c h w -> b h w c') + * 127.5 + 127.5).cpu().numpy().clip(0, 255).astype(np.uint8) + + results = [x_samples[i] for i in range(num_samples)] + return [full_segmask] + results + + +# disable gradio when not using GUI. +if not use_gradio: + image_path = "images/sa_309398.jpg" + input_image = Image.open(image_path) + input_image = np.array(input_image, dtype=np.uint8) + prompt = "" + a_prompt = 'best quality, extremely detailed' + n_prompt = 'longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality' + num_samples = 5 + image_resolution = 512 + detect_resolution = 512 + ddim_steps = 100 + guess_mode = False + strength = 1.0 + scale = 9.0 + seed = 10086 + eta = 0.0 + + outputs = process(input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta) + + image_list = [] + input_image = resize_image(input_image, 512) + image_list.append(torch.tensor(input_image)) + for i in range(len(outputs)): + each = outputs[i] + each = resize_image(each, 512) + print(i, each.shape) + image_list.append(torch.tensor(each)) + + image_list = torch.stack(image_list).permute(0, 3, 1, 2) + + save_image(image_list, "sample.jpg", nrow=4, + normalize=True, value_range=(0, 255)) + +else: + block = gr.Blocks().queue() + with block: + with gr.Row(): + gr.Markdown( + "## Edit Anything powered by ControlNet+SAM+BLIP2+Stable Diffusion") + with gr.Row(): + with gr.Column(): + input_image = gr.Image(source='upload', type="numpy") + prompt = gr.Textbox(label="Prompt") + run_button = gr.Button(label="Run") + with gr.Accordion("Advanced options", open=False): + num_samples = gr.Slider( + label="Images", minimum=1, maximum=12, value=1, step=1) + image_resolution = gr.Slider( + label="Image Resolution", minimum=256, maximum=768, value=512, step=64) + strength = gr.Slider( + label="Control Strength", minimum=0.0, maximum=2.0, value=1.0, step=0.01) + guess_mode = gr.Checkbox(label='Guess Mode', value=False) + detect_resolution = gr.Slider( + label="SAM Resolution", minimum=128, maximum=2048, value=1024, step=1) + ddim_steps = gr.Slider( + label="Steps", minimum=1, maximum=100, value=20, step=1) + scale = gr.Slider( + label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1) + seed = gr.Slider(label="Seed", minimum=-1, + maximum=2147483647, step=1, randomize=True) + eta = gr.Number(label="eta (DDIM)", value=0.0) + a_prompt = gr.Textbox( + label="Added Prompt", value='best quality, extremely detailed') + n_prompt = gr.Textbox(label="Negative Prompt", + value='longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality') + with gr.Column(): + result_gallery = gr.Gallery( + label='Output', show_label=False, elem_id="gallery").style(grid=2, height='auto') + ips = [input_image, prompt, a_prompt, n_prompt, num_samples, image_resolution, + detect_resolution, ddim_steps, guess_mode, strength, scale, seed, eta] + run_button.click(fn=process, inputs=ips, outputs=[result_gallery]) + + block.launch(server_name='0.0.0.0') diff --git a/tools/tool_add_control_sd15.py b/tools/tool_add_control_sd15.py new file mode 100644 index 0000000000000000000000000000000000000000..1c949b601c02184c835593d4deb4602f73b662ff --- /dev/null +++ b/tools/tool_add_control_sd15.py @@ -0,0 +1,50 @@ +import sys +import os + +assert len(sys.argv) == 3, 'Args are wrong.' + +input_path = sys.argv[1] +output_path = sys.argv[2] + +assert os.path.exists(input_path), 'Input model does not exist.' +assert not os.path.exists(output_path), 'Output filename already exists.' +assert os.path.exists(os.path.dirname(output_path)), 'Output path is not valid.' + +import torch +from share import * +from cldm.model import create_model + + +def get_node_name(name, parent_name): + if len(name) <= len(parent_name): + return False, '' + p = name[:len(parent_name)] + if p != parent_name: + return False, '' + return True, name[len(parent_name):] + + +model = create_model(config_path='./models/cldm_v15.yaml') + +pretrained_weights = torch.load(input_path) +if 'state_dict' in pretrained_weights: + pretrained_weights = pretrained_weights['state_dict'] + +scratch_dict = model.state_dict() + +target_dict = {} +for k in scratch_dict.keys(): + is_control, name = get_node_name(k, 'control_') + if is_control: + copy_k = 'model.diffusion_' + name + else: + copy_k = k + if copy_k in pretrained_weights: + target_dict[k] = pretrained_weights[copy_k].clone() + else: + target_dict[k] = scratch_dict[k].clone() + print(f'These weights are newly added: {k}') + +model.load_state_dict(target_dict, strict=True) +torch.save(model.state_dict(), output_path) +print('Done.') \ No newline at end of file diff --git a/tools/tool_add_control_sd21.py b/tools/tool_add_control_sd21.py new file mode 100644 index 0000000000000000000000000000000000000000..7c3ac5f9bb398b9d20cb7d750f1f7c0717670eae --- /dev/null +++ b/tools/tool_add_control_sd21.py @@ -0,0 +1,50 @@ +import sys +import os + +assert len(sys.argv) == 3, 'Args are wrong.' + +input_path = sys.argv[1] +output_path = sys.argv[2] + +assert os.path.exists(input_path), 'Input model does not exist.' +assert not os.path.exists(output_path), 'Output filename already exists.' +assert os.path.exists(os.path.dirname(output_path)), 'Output path is not valid.' + +import torch +from share import * +from cldm.model import create_model + + +def get_node_name(name, parent_name): + if len(name) <= len(parent_name): + return False, '' + p = name[:len(parent_name)] + if p != parent_name: + return False, '' + return True, name[len(parent_name):] + + +model = create_model(config_path='./models/cldm_v21.yaml') + +pretrained_weights = torch.load(input_path) +if 'state_dict' in pretrained_weights: + pretrained_weights = pretrained_weights['state_dict'] + +scratch_dict = model.state_dict() + +target_dict = {} +for k in scratch_dict.keys(): + is_control, name = get_node_name(k, 'control_') + if is_control: + copy_k = 'model.diffusion_' + name + else: + copy_k = k + if copy_k in pretrained_weights: + target_dict[k] = pretrained_weights[copy_k].clone() + else: + target_dict[k] = scratch_dict[k].clone() + print(f'These weights are newly added: {k}') + +model.load_state_dict(target_dict, strict=True) +torch.save(model.state_dict(), output_path) +print('Done.') diff --git a/tools/train_dreambooth_inpaint.py b/tools/train_dreambooth_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..c06cb7ce8904b800ec15c9732d14f69bb727e0ee --- /dev/null +++ b/tools/train_dreambooth_inpaint.py @@ -0,0 +1,812 @@ +# ported from https://github.com/huggingface/diffusers/tree/a6fb9407fd45e76ccd47d13f08f0dd835967d620/examples/research_projects/dreambooth_inpaint +import argparse +import hashlib +import itertools +import math +import os +import random +from pathlib import Path + +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.utils import ProjectConfiguration, set_seed +from huggingface_hub import create_repo, upload_folder +from PIL import Image, ImageDraw +from torch.utils.data import Dataset +from torchvision import transforms +from tqdm.auto import tqdm +from transformers import CLIPTextModel, CLIPTokenizer + +from diffusers import ( + AutoencoderKL, + DDPMScheduler, + StableDiffusionInpaintPipeline, + StableDiffusionPipeline, + UNet2DConditionModel, +) +from diffusers.optimization import get_scheduler +from diffusers.utils import check_min_version + + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.13.0.dev0") + +logger = get_logger(__name__) + + +def prepare_mask_and_masked_image(image, mask): + image = np.array(image.convert("RGB")) + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + mask = np.array(mask.convert("L")) + mask = mask.astype(np.float32) / 255.0 + mask = mask[None, None] + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + masked_image = image * (mask < 0.5) + + return mask, masked_image + + +# generate random masks +def random_mask(im_shape, ratio=1, mask_full_image=False): + mask = Image.new("L", im_shape, 0) + draw = ImageDraw.Draw(mask) + size = (random.randint(0, int(im_shape[0] * ratio)), random.randint(0, int(im_shape[1] * ratio))) + # use this to always mask the whole image + if mask_full_image: + size = (int(im_shape[0] * ratio), int(im_shape[1] * ratio)) + limits = (im_shape[0] - size[0] // 2, im_shape[1] - size[1] // 2) + center = (random.randint(size[0] // 2, limits[0]), random.randint(size[1] // 2, limits[1])) + draw_type = random.randint(0, 1) + if draw_type == 0 or mask_full_image: + draw.rectangle( + (center[0] - size[0] // 2, center[1] - size[1] // 2, center[0] + size[0] // 2, center[1] + size[1] // 2), + fill=255, + ) + else: + draw.ellipse( + (center[0] - size[0] // 2, center[1] - size[1] // 2, center[0] + size[0] // 2, center[1] + size[1] // 2), + fill=255, + ) + + return mask + + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + required=True, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--tokenizer_name", + type=str, + default=None, + help="Pretrained tokenizer name or path if not the same as model_name", + ) + parser.add_argument( + "--instance_data_dir", + type=str, + default=None, + required=True, + help="A folder containing the training data of instance images.", + ) + parser.add_argument( + "--class_data_dir", + type=str, + default=None, + required=False, + help="A folder containing the training data of class images.", + ) + parser.add_argument( + "--instance_prompt", + type=str, + default=None, + help="The prompt with identifier specifying the instance", + ) + parser.add_argument( + "--class_prompt", + type=str, + default=None, + help="The prompt to specify images in the same class as provided instance images.", + ) + parser.add_argument( + "--with_prior_preservation", + default=False, + action="store_true", + help="Flag to add prior preservation loss.", + ) + parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.") + parser.add_argument( + "--num_class_images", + type=int, + default=100, + help=( + "Minimal class images for prior preservation loss. If not have enough images, additional images will be" + " sampled with class_prompt." + ), + ) + parser.add_argument( + "--output_dir", + type=str, + default="text-inversion-model", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--resolution", + type=int, + default=512, + help=( + "The resolution for input images, all the images in the train/validation dataset will be resized to this" + " resolution" + ), + ) + parser.add_argument( + "--center_crop", + default=False, + action="store_true", + help=( + "Whether to center crop the input images to the resolution. If not set, the images will be randomly" + " cropped. The images will be resized to the resolution first before cropping." + ), + ) + parser.add_argument("--train_text_encoder", action="store_true", help="Whether to train the text encoder") + parser.add_argument( + "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument( + "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images." + ) + parser.add_argument("--num_train_epochs", type=int, default=1) + parser.add_argument( + "--max_train_steps", + type=int, + default=None, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=5e-6, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--mixed_precision", + type=str, + default="no", + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose" + "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10." + "and an Nvidia Ampere GPU." + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final" + " checkpoints in case they are better than the last checkpoint and are suitable for resuming training" + " using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=( + "Max number of checkpoints to store. Passed as `total_limit` to the `Accelerator` `ProjectConfiguration`." + " See Accelerator::save_state https://huggingface.co/docs/accelerate/package_reference/accelerator#accelerate.Accelerator.save_state" + " for more docs" + ), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + if args.instance_data_dir is None: + raise ValueError("You must specify a train data directory.") + + if args.with_prior_preservation: + if args.class_data_dir is None: + raise ValueError("You must specify a data directory for class images.") + if args.class_prompt is None: + raise ValueError("You must specify prompt for class images.") + + return args + + +class DreamBoothDataset(Dataset): + """ + A dataset to prepare the instance and class images with the prompts for fine-tuning the model. + It pre-processes the images and the tokenizes prompts. + """ + + def __init__( + self, + instance_data_root, + instance_prompt, + tokenizer, + class_data_root=None, + class_prompt=None, + size=512, + center_crop=False, + ): + self.size = size + self.center_crop = center_crop + self.tokenizer = tokenizer + + self.instance_data_root = Path(instance_data_root) + if not self.instance_data_root.exists(): + raise ValueError("Instance images root doesn't exists.") + + self.instance_images_path = list(Path(instance_data_root).iterdir()) + self.num_instance_images = len(self.instance_images_path) + self.instance_prompt = instance_prompt + self._length = self.num_instance_images + + if class_data_root is not None: + self.class_data_root = Path(class_data_root) + self.class_data_root.mkdir(parents=True, exist_ok=True) + self.class_images_path = list(self.class_data_root.iterdir()) + self.num_class_images = len(self.class_images_path) + self._length = max(self.num_class_images, self.num_instance_images) + self.class_prompt = class_prompt + else: + self.class_data_root = None + + self.image_transforms_resize_and_crop = transforms.Compose( + [ + transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR), + transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size), + ] + ) + + self.image_transforms = transforms.Compose( + [ + transforms.ToTensor(), + transforms.Normalize([0.5], [0.5]), + ] + ) + + def __len__(self): + return self._length + + def __getitem__(self, index): + example = {} + instance_image = Image.open(self.instance_images_path[index % self.num_instance_images]) + if not instance_image.mode == "RGB": + instance_image = instance_image.convert("RGB") + instance_image = self.image_transforms_resize_and_crop(instance_image) + + example["PIL_images"] = instance_image + example["instance_images"] = self.image_transforms(instance_image) + + example["instance_prompt_ids"] = self.tokenizer( + self.instance_prompt, + padding="do_not_pad", + truncation=True, + max_length=self.tokenizer.model_max_length, + ).input_ids + + if self.class_data_root: + class_image = Image.open(self.class_images_path[index % self.num_class_images]) + if not class_image.mode == "RGB": + class_image = class_image.convert("RGB") + class_image = self.image_transforms_resize_and_crop(class_image) + example["class_images"] = self.image_transforms(class_image) + example["class_PIL_images"] = class_image + example["class_prompt_ids"] = self.tokenizer( + self.class_prompt, + padding="do_not_pad", + truncation=True, + max_length=self.tokenizer.model_max_length, + ).input_ids + + return example + + +class PromptDataset(Dataset): + "A simple dataset to prepare the prompts to generate class images on multiple GPUs." + + def __init__(self, prompt, num_samples): + self.prompt = prompt + self.num_samples = num_samples + + def __len__(self): + return self.num_samples + + def __getitem__(self, index): + example = {} + example["prompt"] = self.prompt + example["index"] = index + return example + + +def main(): + args = parse_args() + logging_dir = Path(args.output_dir, args.logging_dir) + + project_config = ProjectConfiguration(total_limit=args.checkpoints_total_limit) + + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with="tensorboard", + logging_dir=logging_dir, + project_config=project_config, + ) + + # Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate + # This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models. + # TODO (patil-suraj): Remove this check when gradient accumulation with two models is enabled in accelerate. + if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1: + raise ValueError( + "Gradient accumulation is not supported when training the text encoder in distributed training. " + "Please set gradient_accumulation_steps to 1. This feature will be supported in the future." + ) + + if args.seed is not None: + set_seed(args.seed) + + if args.with_prior_preservation: + class_images_dir = Path(args.class_data_dir) + if not class_images_dir.exists(): + class_images_dir.mkdir(parents=True) + cur_class_images = len(list(class_images_dir.iterdir())) + + if cur_class_images < args.num_class_images: + torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32 + pipeline = StableDiffusionInpaintPipeline.from_pretrained( + args.pretrained_model_name_or_path, torch_dtype=torch_dtype, safety_checker=None + ) + pipeline.set_progress_bar_config(disable=True) + + num_new_images = args.num_class_images - cur_class_images + logger.info(f"Number of class images to sample: {num_new_images}.") + + sample_dataset = PromptDataset(args.class_prompt, num_new_images) + sample_dataloader = torch.utils.data.DataLoader( + sample_dataset, batch_size=args.sample_batch_size, num_workers=1 + ) + + sample_dataloader = accelerator.prepare(sample_dataloader) + pipeline.to(accelerator.device) + transform_to_pil = transforms.ToPILImage() + for example in tqdm( + sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process + ): + bsz = len(example["prompt"]) + fake_images = torch.rand((3, args.resolution, args.resolution)) + transform_to_pil = transforms.ToPILImage() + fake_pil_images = transform_to_pil(fake_images) + + fake_mask = random_mask((args.resolution, args.resolution), ratio=1, mask_full_image=True) + + images = pipeline(prompt=example["prompt"], mask_image=fake_mask, image=fake_pil_images).images + + for i, image in enumerate(images): + hash_image = hashlib.sha1(image.tobytes()).hexdigest() + image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg" + image.save(image_filename) + + del pipeline + if torch.cuda.is_available(): + torch.cuda.empty_cache() + + # Handle the repository creation + if accelerator.is_main_process: + if args.output_dir is not None: + os.makedirs(args.output_dir, exist_ok=True) + + if args.push_to_hub: + repo_id = create_repo( + repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token + ).repo_id + + # Load the tokenizer + if args.tokenizer_name: + tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name) + elif args.pretrained_model_name_or_path: + tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer") + + # Load models and create wrapper for stable diffusion + text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder") + vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae") + unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet") + + vae.requires_grad_(False) + if not args.train_text_encoder: + text_encoder.requires_grad_(False) + + if args.gradient_checkpointing: + unet.enable_gradient_checkpointing() + if args.train_text_encoder: + text_encoder.gradient_checkpointing_enable() + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`." + ) + + optimizer_class = bnb.optim.AdamW8bit + else: + optimizer_class = torch.optim.AdamW + + params_to_optimize = ( + itertools.chain(unet.parameters(), text_encoder.parameters()) if args.train_text_encoder else unet.parameters() + ) + optimizer = optimizer_class( + params_to_optimize, + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler") + + train_dataset = DreamBoothDataset( + instance_data_root=args.instance_data_dir, + instance_prompt=args.instance_prompt, + class_data_root=args.class_data_dir if args.with_prior_preservation else None, + class_prompt=args.class_prompt, + tokenizer=tokenizer, + size=args.resolution, + center_crop=args.center_crop, + ) + + def collate_fn(examples): + input_ids = [example["instance_prompt_ids"] for example in examples] + pixel_values = [example["instance_images"] for example in examples] + + # Concat class and instance examples for prior preservation. + # We do this to avoid doing two forward passes. + if args.with_prior_preservation: + input_ids += [example["class_prompt_ids"] for example in examples] + pixel_values += [example["class_images"] for example in examples] + pior_pil = [example["class_PIL_images"] for example in examples] + + masks = [] + masked_images = [] + for example in examples: + pil_image = example["PIL_images"] + # generate a random mask + mask = random_mask(pil_image.size, 1, False) + # prepare mask and masked image + mask, masked_image = prepare_mask_and_masked_image(pil_image, mask) + + masks.append(mask) + masked_images.append(masked_image) + + if args.with_prior_preservation: + for pil_image in pior_pil: + # generate a random mask + mask = random_mask(pil_image.size, 1, False) + # prepare mask and masked image + mask, masked_image = prepare_mask_and_masked_image(pil_image, mask) + + masks.append(mask) + masked_images.append(masked_image) + + pixel_values = torch.stack(pixel_values) + pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() + + input_ids = tokenizer.pad({"input_ids": input_ids}, padding=True, return_tensors="pt").input_ids + masks = torch.stack(masks) + masked_images = torch.stack(masked_images) + batch = {"input_ids": input_ids, "pixel_values": pixel_values, "masks": masks, "masked_images": masked_images} + return batch + + train_dataloader = torch.utils.data.DataLoader( + train_dataset, batch_size=args.train_batch_size, shuffle=True, collate_fn=collate_fn + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + ) + + if args.train_text_encoder: + unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, text_encoder, optimizer, train_dataloader, lr_scheduler + ) + else: + unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, optimizer, train_dataloader, lr_scheduler + ) + accelerator.register_for_checkpointing(lr_scheduler) + + weight_dtype = torch.float32 + if args.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif args.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + # Move text_encode and vae to gpu. + # For mixed precision training we cast the text_encoder and vae weights to half-precision + # as these models are only used for inference, keeping weights in full precision is not required. + vae.to(accelerator.device, dtype=weight_dtype) + if not args.train_text_encoder: + text_encoder.to(accelerator.device, dtype=weight_dtype) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + accelerator.init_trackers("dreambooth", config=vars(args)) + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num batches each epoch = {len(train_dataloader)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + # Only show the progress bar once on each machine. + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + + for epoch in range(first_epoch, args.num_train_epochs): + unet.train() + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(unet): + # Convert images to latent space + + latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample() + latents = latents * vae.config.scaling_factor + + # Convert masked images to latent space + masked_latents = vae.encode( + batch["masked_images"].reshape(batch["pixel_values"].shape).to(dtype=weight_dtype) + ).latent_dist.sample() + masked_latents = masked_latents * vae.config.scaling_factor + + masks = batch["masks"] + # resize the mask to latents shape as we concatenate the mask to the latents + mask = torch.stack( + [ + torch.nn.functional.interpolate(mask, size=(args.resolution // 8, args.resolution // 8)) + for mask in masks + ] + ) + mask = mask.reshape(-1, 1, args.resolution // 8, args.resolution // 8) + + # Sample noise that we'll add to the latents + noise = torch.randn_like(latents) + bsz = latents.shape[0] + # Sample a random timestep for each image + timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device) + timesteps = timesteps.long() + + # Add noise to the latents according to the noise magnitude at each timestep + # (this is the forward diffusion process) + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + # concatenate the noised latents with the mask and the masked latents + latent_model_input = torch.cat([noisy_latents, mask, masked_latents], dim=1) + + # Get the text embedding for conditioning + encoder_hidden_states = text_encoder(batch["input_ids"])[0] + + # Predict the noise residual + noise_pred = unet(latent_model_input, timesteps, encoder_hidden_states).sample + + # Get the target for loss depending on the prediction type + if noise_scheduler.config.prediction_type == "epsilon": + target = noise + elif noise_scheduler.config.prediction_type == "v_prediction": + target = noise_scheduler.get_velocity(latents, noise, timesteps) + else: + raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") + + if args.with_prior_preservation: + # Chunk the noise and noise_pred into two parts and compute the loss on each part separately. + noise_pred, noise_pred_prior = torch.chunk(noise_pred, 2, dim=0) + target, target_prior = torch.chunk(target, 2, dim=0) + + # Compute instance loss + loss = F.mse_loss(noise_pred.float(), target.float(), reduction="none").mean([1, 2, 3]).mean() + + # Compute prior loss + prior_loss = F.mse_loss(noise_pred_prior.float(), target_prior.float(), reduction="mean") + + # Add the prior loss to the instance loss. + loss = loss + args.prior_loss_weight * prior_loss + else: + loss = F.mse_loss(noise_pred.float(), target.float(), reduction="mean") + + accelerator.backward(loss) + if accelerator.sync_gradients: + params_to_clip = ( + itertools.chain(unet.parameters(), text_encoder.parameters()) + if args.train_text_encoder + else unet.parameters() + ) + accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + progress_bar.update(1) + global_step += 1 + + if global_step % args.checkpointing_steps == 0: + if accelerator.is_main_process: + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + accelerator.log(logs, step=global_step) + + if global_step >= args.max_train_steps: + break + + accelerator.wait_for_everyone() + + # Create the pipeline using using the trained modules and save it. + if accelerator.is_main_process: + pipeline = StableDiffusionPipeline.from_pretrained( + args.pretrained_model_name_or_path, + unet=accelerator.unwrap_model(unet), + text_encoder=accelerator.unwrap_model(text_encoder), + ) + pipeline.save_pretrained(args.output_dir) + + if args.push_to_hub: + upload_folder( + repo_id=repo_id, + folder_path=args.output_dir, + commit_message="End of training", + ignore_patterns=["step_*", "epoch_*"], + ) + + accelerator.end_training() + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/utils/GroundingDINO_SwinT_OGC.py b/utils/GroundingDINO_SwinT_OGC.py new file mode 100644 index 0000000000000000000000000000000000000000..c2e64680e46c241d99cf32b6a2122fd7aa9aa545 --- /dev/null +++ b/utils/GroundingDINO_SwinT_OGC.py @@ -0,0 +1,43 @@ +batch_size = 1 +modelname = "groundingdino" +backbone = "swin_T_224_1k" +position_embedding = "sine" +pe_temperatureH = 20 +pe_temperatureW = 20 +return_interm_indices = [1, 2, 3] +backbone_freeze_keywords = None +enc_layers = 6 +dec_layers = 6 +pre_norm = False +dim_feedforward = 2048 +hidden_dim = 256 +dropout = 0.0 +nheads = 8 +num_queries = 900 +query_dim = 4 +num_patterns = 0 +num_feature_levels = 4 +enc_n_points = 4 +dec_n_points = 4 +two_stage_type = "standard" +two_stage_bbox_embed_share = False +two_stage_class_embed_share = False +transformer_activation = "relu" +dec_pred_bbox_embed_share = True +dn_box_noise_scale = 1.0 +dn_label_noise_ratio = 0.5 +dn_label_coef = 1.0 +dn_bbox_coef = 1.0 +embed_init_tgt = True +dn_labelbook_size = 2000 +max_text_len = 256 +text_encoder_type = "bert-base-uncased" +use_text_enhancer = True +use_fusion_layer = True +use_checkpoint = True +use_transformer_ckpt = True +use_text_cross_attention = True +text_dropout = 0.0 +fusion_dropout = 0.0 +fusion_droppath = 0.1 +sub_sentence_present = True \ No newline at end of file diff --git a/utils/run_texutal_inversion.sh b/utils/run_texutal_inversion.sh new file mode 100644 index 0000000000000000000000000000000000000000..6646e871d6d751bceb7c9b9dc8ad80565ee65e42 --- /dev/null +++ b/utils/run_texutal_inversion.sh @@ -0,0 +1,18 @@ +export MODEL_NAME="runwayml/stable-diffusion-v1-5" +export DATA_DIR="./tmp/textinv/img" +export OUTPUT_DIR="./tmp/textinv/model" + +CUDA_VISIBLE_DEVICES=0 accelerate launch --main_process_port 1111 texutal_inversion.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --train_data_dir=$DATA_DIR \ + --learnable_property="object" \ + --placeholder_token="" --initializer_token="mark" \ + --resolution=512 \ + --train_batch_size=4 \ + --gradient_accumulation_steps=1 \ + --max_train_steps=3000 \ + --learning_rate=5.0e-04 --scale_lr \ + --lr_scheduler="constant" \ + --lr_warmup_steps=0 \ + --output_dir=$OUTPUT_DIR \ + --num_vectors 10 \ No newline at end of file diff --git a/utils/sam_dataset.py b/utils/sam_dataset.py new file mode 100644 index 0000000000000000000000000000000000000000..2ff359342f9e3849f3ea22f02849348113cce47d --- /dev/null +++ b/utils/sam_dataset.py @@ -0,0 +1,72 @@ +import json +import cv2 +import numpy as np +import os + +from torch.utils.data import Dataset +import pycocotools.mask as maskUtils +from torchvision import transforms +import utils.transforms as custom_transforms +from PIL import Image +class SAMDataset(Dataset): + def __init__(self, data_path='../data/files', txt_path='../data/data_85616.txt'): + self.data = [] + with open(txt_path, 'rt') as f: + for line in f: + self.data.append(eval(line)) + self.data_path = data_path + randomresizedcrop = custom_transforms.RandomResizedCrop( + 512, + scale=(0.9, 1), + ) + self.transform = custom_transforms.Compose([ + randomresizedcrop, + custom_transforms.RandomHorizontalFlip(p=0.5), + custom_transforms.ToTensor(), + transforms.Normalize(mean=0.5, std=0.5) + ]) + + + + def __len__(self): + return len(self.data) + + def load_rle_annotations_from_json(self, json_file_path, return_pil=True): + with open(json_file_path, 'r', encoding='utf-8') as f: + anno_data = json.load(f) + annotations = anno_data['annotations'] + height = int(anno_data['image']['height']) + width = int(anno_data['image']['width']) + + map = np.zeros((height,width), dtype=np.uint16) + for i in range(len(annotations)): + ann = annotations[i] + mask = maskUtils.decode(ann['segmentation']) + map[mask != 0] = i + 1 + if return_pil: + res = np.zeros((map.shape[0], map.shape[1], 3)) + res[:, :, 0] = map % 256 + res[:, :, 1] = map // 256 + res = Image.fromarray(res.astype(np.uint8)) + return res + return map + + def __getitem__(self, idx): + item = self.data[idx] + + source_filename = item['source'] + target_filename = item['target'] + prompt = item['prompt'] + + + source = self.load_rle_annotations_from_json(os.path.join(self.data_path, source_filename)) + target = Image.open(os.path.join(self.data_path, target_filename)) + + + target, source = self.transform(target, source) + + print(source.max(), source.min()) + target = target.permute(1,2,0) + + return dict(jpg=target, txt=prompt, hint=source) + diff --git a/utils/sam_dataset_test.py b/utils/sam_dataset_test.py new file mode 100644 index 0000000000000000000000000000000000000000..5d523d1309762970b851bc792524c4622aec563d --- /dev/null +++ b/utils/sam_dataset_test.py @@ -0,0 +1,19 @@ +from sam_dataset import SAMDataset + +dataset = SAMDataset() +print(len(dataset)) + +item = dataset[1234] +jpg = item['jpg'] +txt = item['txt'] +hint = item['hint'] +print(txt) +print(jpg.shape) +print(hint.shape) + +for each in dataset: + jpg = item['jpg'] + txt = item['txt'] + hint = item['hint'] + print(txt) + print(jpg.shape,hint.shape) \ No newline at end of file diff --git a/utils/sketch_helpers.py b/utils/sketch_helpers.py new file mode 100644 index 0000000000000000000000000000000000000000..e71551158b3ae3d3f7f2bff1a3976c50ccb16a24 --- /dev/null +++ b/utils/sketch_helpers.py @@ -0,0 +1,84 @@ +import numpy as np +import cv2 +from PIL import Image +from skimage.color import rgb2lab +from skimage.color import lab2rgb +from sklearn.cluster import KMeans + + +def count_high_freq_colors(image): + im = image.getcolors(maxcolors=1024 * 1024) + sorted_colors = sorted(im, key=lambda x: x[0], reverse=True) + + freqs = [c[0] for c in sorted_colors] + mean_freq = sum(freqs) / len(freqs) + + high_freq_colors = [c for c in sorted_colors if c[0] > max(2, mean_freq * 1.25)] + return high_freq_colors + + +def get_high_freq_colors(image, similarity_threshold=30): + image_copy = image.copy() + high_freq_colors = count_high_freq_colors(image) + # Check for similar colors and replace the lower frequency color with the higher frequency color in the image + for i, (freq1, color1) in enumerate(high_freq_colors): + for j, (freq2, color2) in enumerate(high_freq_colors): + if (color_distance(color1, color2) < similarity_threshold) or ( + color_distance(color1, opaque_color_on_white(color2, 0.5)) < 5): + if (freq2 > freq1): + replace_color(image_copy, color1, color2) + + high_freq_colors = count_high_freq_colors(image_copy) + print(high_freq_colors) + return [high_freq_colors, image_copy] + + +def color_quantization(image, color_frequency_list): + # Convert the color frequency list to a set of unique colors + unique_colors = set([color for _, color in color_frequency_list]) + + # Create a mask for the image with True where the color is in the unique colors set + mask = np.any(np.all(image.reshape(-1, 1, 3) == np.array(list(unique_colors)), axis=2), axis=1).reshape( + image.shape[:2]) + + # Create a new image with all pixels set to white + new_image = np.full_like(image, 255) + + # Copy the pixels from the original image that have a color in the color frequency list + new_image[mask] = image[mask] + return new_image + + +def create_binary_matrix(img_arr, target_color): + # Create mask of pixels with target color + mask = np.all(img_arr == target_color, axis=-1) + + # Convert mask to binary matrix + binary_matrix = mask.astype(int) + from datetime import datetime + binary_file_name = f'mask-{datetime.now().timestamp()}.png' + cv2.imwrite(binary_file_name, binary_matrix * 255) + + # binary_matrix = torch.from_numpy(binary_matrix).unsqueeze(0).unsqueeze(0) + return binary_file_name + + +def color_distance(color1, color2): + return sum((a - b) ** 2 for a, b in zip(color1, color2)) ** 0.5 + + +def replace_color(image, old_color, new_color): + pixels = image.load() + width, height = image.size + for x in range(width): + for y in range(height): + if pixels[x, y] == old_color: + pixels[x, y] = new_color + + +def opaque_color_on_white(color, a): + r, g, b = color + opaque_red = int((1 - a) * 255 + a * r) + opaque_green = int((1 - a) * 255 + a * g) + opaque_blue = int((1 - a) * 255 + a * b) + return (opaque_red, opaque_green, opaque_blue) diff --git a/utils/stable_diffusion_controlnet.py b/utils/stable_diffusion_controlnet.py new file mode 100644 index 0000000000000000000000000000000000000000..cf970d3bef6ba17293608c16deed56f39425985c --- /dev/null +++ b/utils/stable_diffusion_controlnet.py @@ -0,0 +1,815 @@ +# Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/ +# From https://raw.githubusercontent.com/huggingface/diffusers/53377ef83c6446033f3ee506e3ef718db817b293/examples/community/stable_diffusion_controlnet_inpaint.py +import inspect +from typing import Any, Callable, Dict, List, Optional, Union, Tuple + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, \ + UNet2DConditionModel, logging, StableDiffusionControlNetPipeline +from diffusers.models.controlnet import ControlNetOutput +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker +from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_controlnet import MultiControlNetModel +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + PIL_INTERPOLATION, + is_accelerate_available, + is_accelerate_version, + is_compiled_module, + randn_tensor, + replace_example_docstring, +) +from diffusers.loaders import LoraLoaderMixin + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import numpy as np + >>> import torch + >>> from PIL import Image + >>> from stable_diffusion_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline + + >>> from transformers import AutoImageProcessor, UperNetForSemanticSegmentation + >>> from diffusers import ControlNetModel, UniPCMultistepScheduler + >>> from diffusers.utils import load_image + + >>> def ade_palette(): + return [[120, 120, 120], [180, 120, 120], [6, 230, 230], [80, 50, 50], + [4, 200, 3], [120, 120, 80], [140, 140, 140], [204, 5, 255], + [230, 230, 230], [4, 250, 7], [224, 5, 255], [235, 255, 7], + [150, 5, 61], [120, 120, 70], [8, 255, 51], [255, 6, 82], + [143, 255, 140], [204, 255, 4], [255, 51, 7], [204, 70, 3], + [0, 102, 200], [61, 230, 250], [255, 6, 51], [11, 102, 255], + [255, 7, 71], [255, 9, 224], [9, 7, 230], [220, 220, 220], + [255, 9, 92], [112, 9, 255], [8, 255, 214], [7, 255, 224], + [255, 184, 6], [10, 255, 71], [255, 41, 10], [7, 255, 255], + [224, 255, 8], [102, 8, 255], [255, 61, 6], [255, 194, 7], + [255, 122, 8], [0, 255, 20], [255, 8, 41], [255, 5, 153], + [6, 51, 255], [235, 12, 255], [160, 150, 20], [0, 163, 255], + [140, 140, 140], [250, 10, 15], [20, 255, 0], [31, 255, 0], + [255, 31, 0], [255, 224, 0], [153, 255, 0], [0, 0, 255], + [255, 71, 0], [0, 235, 255], [0, 173, 255], [31, 0, 255], + [11, 200, 200], [255, 82, 0], [0, 255, 245], [0, 61, 255], + [0, 255, 112], [0, 255, 133], [255, 0, 0], [255, 163, 0], + [255, 102, 0], [194, 255, 0], [0, 143, 255], [51, 255, 0], + [0, 82, 255], [0, 255, 41], [0, 255, 173], [10, 0, 255], + [173, 255, 0], [0, 255, 153], [255, 92, 0], [255, 0, 255], + [255, 0, 245], [255, 0, 102], [255, 173, 0], [255, 0, 20], + [255, 184, 184], [0, 31, 255], [0, 255, 61], [0, 71, 255], + [255, 0, 204], [0, 255, 194], [0, 255, 82], [0, 10, 255], + [0, 112, 255], [51, 0, 255], [0, 194, 255], [0, 122, 255], + [0, 255, 163], [255, 153, 0], [0, 255, 10], [255, 112, 0], + [143, 255, 0], [82, 0, 255], [163, 255, 0], [255, 235, 0], + [8, 184, 170], [133, 0, 255], [0, 255, 92], [184, 0, 255], + [255, 0, 31], [0, 184, 255], [0, 214, 255], [255, 0, 112], + [92, 255, 0], [0, 224, 255], [112, 224, 255], [70, 184, 160], + [163, 0, 255], [153, 0, 255], [71, 255, 0], [255, 0, 163], + [255, 204, 0], [255, 0, 143], [0, 255, 235], [133, 255, 0], + [255, 0, 235], [245, 0, 255], [255, 0, 122], [255, 245, 0], + [10, 190, 212], [214, 255, 0], [0, 204, 255], [20, 0, 255], + [255, 255, 0], [0, 153, 255], [0, 41, 255], [0, 255, 204], + [41, 0, 255], [41, 255, 0], [173, 0, 255], [0, 245, 255], + [71, 0, 255], [122, 0, 255], [0, 255, 184], [0, 92, 255], + [184, 255, 0], [0, 133, 255], [255, 214, 0], [25, 194, 194], + [102, 255, 0], [92, 0, 255]] + + >>> image_processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-convnext-small") + >>> image_segmentor = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-convnext-small") + + >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-seg", torch_dtype=torch.float16) + + >>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( + "runwayml/stable-diffusion-inpainting", controlnet=controlnet, safety_checker=None, torch_dtype=torch.float16 + ) + + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) + >>> pipe.enable_xformers_memory_efficient_attention() + >>> pipe.enable_model_cpu_offload() + + >>> def image_to_seg(image): + pixel_values = image_processor(image, return_tensors="pt").pixel_values + with torch.no_grad(): + outputs = image_segmentor(pixel_values) + seg = image_processor.post_process_semantic_segmentation(outputs, target_sizes=[image.size[::-1]])[0] + color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8) # height, width, 3 + palette = np.array(ade_palette()) + for label, color in enumerate(palette): + color_seg[seg == label, :] = color + color_seg = color_seg.astype(np.uint8) + seg_image = Image.fromarray(color_seg) + return seg_image + + >>> image = load_image( + "https://github.com/CompVis/latent-diffusion/raw/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + ) + + >>> mask_image = load_image( + "https://github.com/CompVis/latent-diffusion/raw/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + ) + + >>> controlnet_conditioning_image = image_to_seg(image) + + >>> image = pipe( + "Face of a yellow cat, high resolution, sitting on a park bench", + image, + mask_image, + controlnet_conditioning_image, + num_inference_steps=20, + ).images[0] + + >>> image.save("out.png") + ``` +""" + + +def prepare_image(image): + if isinstance(image, torch.Tensor): + # Batch single image + if image.ndim == 3: + image = image.unsqueeze(0) + + image = image.to(dtype=torch.float32) + else: + # preprocess image + if isinstance(image, (PIL.Image.Image, np.ndarray)): + image = [image] + + if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): + image = [np.array(i.convert("RGB"))[None, :] for i in image] + image = np.concatenate(image, axis=0) + elif isinstance(image, list) and isinstance(image[0], np.ndarray): + image = np.concatenate([i[None, :] for i in image], axis=0) + + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.prepare_mask_and_masked_image +def prepare_mask_and_masked_image(image, mask, height, width, return_image=False): + """ + Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be + converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the + ``image`` and ``1`` for the ``mask``. + + The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be + binarized (``mask > 0.5``) and cast to ``torch.float32`` too. + + Args: + image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint. + It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width`` + ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``. + mask (_type_): The mask to apply to the image, i.e. regions to inpaint. + It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width`` + ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``. + + + Raises: + ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask + should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions. + TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not + (ot the other way around). + + Returns: + tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4 + dimensions: ``batch x channels x height x width``. + """ + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + if mask is None: + raise ValueError("`mask_image` input cannot be undefined.") + + if isinstance(image, torch.Tensor): + if not isinstance(mask, torch.Tensor): + raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not") + + # Batch single image + if image.ndim == 3: + assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)" + image = image.unsqueeze(0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions" + assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions" + assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size" + + # Check image is in [-1, 1] + if image.min() < -1 or image.max() > 1: + raise ValueError("Image should be in [-1, 1] range") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("Mask should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + # Image as float32 + image = image.to(dtype=torch.float32) + elif isinstance(mask, torch.Tensor): + raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not") + else: + # preprocess image + if isinstance(image, (PIL.Image.Image, np.ndarray)): + image = [image] + if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): + # resize all images w.r.t passed height an width + image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image] + image = [np.array(i.convert("RGB"))[None, :] for i in image] + image = np.concatenate(image, axis=0) + elif isinstance(image, list) and isinstance(image[0], np.ndarray): + image = np.concatenate([i[None, :] for i in image], axis=0) + + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): + mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask] + mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) + mask = mask.astype(np.float32) / 255.0 + elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): + mask = np.concatenate([m[None, None, :] for m in mask], axis=0) + + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + masked_image = image * (mask < 0.5) + + # n.b. ensure backwards compatibility as old function does not return image + if return_image: + return mask, masked_image, image + + return mask, masked_image + + +def prepare_mask_image(mask_image): + if isinstance(mask_image, torch.Tensor): + if mask_image.ndim == 2: + # Batch and add channel dim for single mask + mask_image = mask_image.unsqueeze(0).unsqueeze(0) + elif mask_image.ndim == 3 and mask_image.shape[0] == 1: + # Single mask, the 0'th dimension is considered to be + # the existing batch size of 1 + mask_image = mask_image.unsqueeze(0) + elif mask_image.ndim == 3 and mask_image.shape[0] != 1: + # Batch of mask, the 0'th dimension is considered to be + # the batching dimension + mask_image = mask_image.unsqueeze(1) + + # Binarize mask + mask_image[mask_image < 0.5] = 0 + mask_image[mask_image >= 0.5] = 1 + else: + # preprocess mask + if isinstance(mask_image, (PIL.Image.Image, np.ndarray)): + mask_image = [mask_image] + + if isinstance(mask_image, list) and isinstance(mask_image[0], PIL.Image.Image): + mask_image = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask_image], axis=0) + mask_image = mask_image.astype(np.float32) / 255.0 + elif isinstance(mask_image, list) and isinstance(mask_image[0], np.ndarray): + mask_image = np.concatenate([m[None, None, :] for m in mask_image], axis=0) + + mask_image[mask_image < 0.5] = 0 + mask_image[mask_image >= 0.5] = 1 + mask_image = torch.from_numpy(mask_image) + + return mask_image + + +def prepare_controlnet_conditioning_image( + controlnet_conditioning_image, width, height, batch_size, num_images_per_prompt, device, dtype, + do_classifier_free_guidance, +): + if not isinstance(controlnet_conditioning_image, torch.Tensor): + if isinstance(controlnet_conditioning_image, PIL.Image.Image): + controlnet_conditioning_image = [controlnet_conditioning_image] + + if isinstance(controlnet_conditioning_image[0], PIL.Image.Image): + controlnet_conditioning_image = [ + np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"]))[None, :] + for i in controlnet_conditioning_image + ] + controlnet_conditioning_image = np.concatenate(controlnet_conditioning_image, axis=0) + controlnet_conditioning_image = np.array(controlnet_conditioning_image).astype(np.float32) / 255.0 + controlnet_conditioning_image = controlnet_conditioning_image.transpose(0, 3, 1, 2) + controlnet_conditioning_image = torch.from_numpy(controlnet_conditioning_image) + elif isinstance(controlnet_conditioning_image[0], torch.Tensor): + controlnet_conditioning_image = torch.cat(controlnet_conditioning_image, dim=0) + + image_batch_size = controlnet_conditioning_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 + + controlnet_conditioning_image = controlnet_conditioning_image.repeat_interleave(repeat_by, dim=0) + + controlnet_conditioning_image = controlnet_conditioning_image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance: + controlnet_conditioning_image = torch.cat([controlnet_conditioning_image] * 2) + + return controlnet_conditioning_image + + +class StableDiffusionControlNetPipeline2(StableDiffusionControlNetPipeline): + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + controlnet_conditioning_scale_map=None, + guess_mode: bool = False, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, + `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`): + The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If + the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can + also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If + height and/or width are passed, `image` is resized according to them. If multiple ControlNets are + specified in init, images must be passed as a list such that each element of the list can be correctly + batched for input to a single controlnet. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + 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. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + 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` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](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`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + 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)`. + 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. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py). + controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0): + The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original unet. If multiple ControlNets are specified in init, you can set the + corresponding scale as a list. + guess_mode (`bool`, *optional*, defaults to `False`): + In this mode, the ControlNet encoder will try best to recognize the content of the input image even if + you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended. + + Examples: + + 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`. + """ + # 0. Default height and width to unet + height, width = self._default_height_width(height, width, image) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + image, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + controlnet_conditioning_scale, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + 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 + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) + + if controlnet_conditioning_scale_map is not None: + if isinstance(controlnet_conditioning_scale, list): + controlnet_conditioning_scale = [scale * controlnet_conditioning_scale_map for scale in + controlnet_conditioning_scale] + else: + controlnet_conditioning_scale = controlnet_conditioning_scale * controlnet_conditioning_scale_map + + global_pool_conditions = ( + controlnet.config.global_pool_conditions + if isinstance(controlnet, ControlNetModel) + else controlnet.nets[0].config.global_pool_conditions + ) + guess_mode = guess_mode or global_pool_conditions + + # 3. Encode input prompt + prompt_embeds = self._encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # 4. Prepare image + if isinstance(controlnet, ControlNetModel): + image = self.prepare_image( + image=image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + guess_mode=guess_mode, + ) + elif isinstance(controlnet, MultiControlNetModel): + images = [] + + for image_ in image: + image_ = self.prepare_image( + image=image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + guess_mode=guess_mode, + ) + + images.append(image_) + + image = images + else: + assert False + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.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 + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(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) + + # controlnet(s) inference + if guess_mode and do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + controlnet_latent_model_input = latents + controlnet_prompt_embeds = prompt_embeds.chunk(2)[1] + else: + controlnet_latent_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds + + down_block_res_samples, mid_block_res_sample = self.controlnet( + controlnet_latent_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + conditioning_scale=controlnet_conditioning_scale, + guess_mode=guess_mode, + return_dict=False, + ) + + if guess_mode and do_classifier_free_guidance: + # Infered ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + return_dict=False, + )[0] + + # 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, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # If we do sequential model offloading, let's offload unet and controlnet + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + self.controlnet.to("cpu") + torch.cuda.empty_cache() + + if output_type == "latent": + image = latents + has_nsfw_concept = None + elif output_type == "pil": + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # 10. Convert to PIL + image = self.numpy_to_pil(image) + else: + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # Offload last model to CPU + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.final_offload_hook.offload() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + +class ControlNetModel2(ControlNetModel): + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + controlnet_cond: torch.FloatTensor, + conditioning_scale: float = 1.0, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guess_mode: bool = False, + return_dict: bool = True, + ) -> Union[ControlNetOutput, Tuple]: + # check channel order + channel_order = self.config.controlnet_conditioning_channel_order + + if channel_order == "rgb": + # in rgb order by default + ... + elif channel_order == "bgr": + controlnet_cond = torch.flip(controlnet_cond, dims=[1]) + else: + raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}") + + # prepare attention_mask + if attention_mask is not None: + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # 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 = 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. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + class_emb = self.class_embedding(class_labels).to(dtype=self.dtype) + emb = emb + class_emb + + # 2. pre-process + sample = self.conv_in(sample) + + controlnet_cond = self.controlnet_cond_embedding(controlnet_cond) + + sample = sample + controlnet_cond + + # 3. down + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + + down_block_res_samples += res_samples + + # 4. mid + if self.mid_block is not None: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + ) + + # 5. Control net blocks + + controlnet_down_block_res_samples = () + + for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks): + down_block_res_sample = controlnet_block(down_block_res_sample) + controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = controlnet_down_block_res_samples + + mid_block_res_sample = self.controlnet_mid_block(sample) + + # 6. scaling + if guess_mode and not self.config.global_pool_conditions: + scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device) # 0.1 to 1.0 + + scales = scales * conditioning_scale + down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)] + mid_block_res_sample = mid_block_res_sample * scales[-1] # last one + else: + if isinstance(conditioning_scale, float): + down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples] + mid_block_res_sample = mid_block_res_sample * conditioning_scale + else: + assert isinstance(conditioning_scale, torch.Tensor) + if len(conditioning_scale.shape) == 2: + conditioning_scale = conditioning_scale[None, None] + elif len(conditioning_scale.shape) == 3: + conditioning_scale = conditioning_scale[None] + down_block_res_samples = [ + sample * F.interpolate(conditioning_scale, sample.shape[-2:], + mode='bilinear', align_corners=True).type(sample.dtype) + for sample in down_block_res_samples + ] + mid_block_res_sample = mid_block_res_sample * F.interpolate( + conditioning_scale, mid_block_res_sample.shape[-2:], + mode='bilinear', align_corners=True + ).type(sample.dtype) + + if self.config.global_pool_conditions: + down_block_res_samples = [ + torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples + ] + mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True) + + if not return_dict: + return (down_block_res_samples, mid_block_res_sample) + + return ControlNetOutput( + down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample + ) diff --git a/utils/stable_diffusion_controlnet_inpaint.py b/utils/stable_diffusion_controlnet_inpaint.py index 3864aea1cb8c310dda0d4ee72db253569e7c5c87..536d37555a261ec0e45034014d0d9c00e992db50 100644 --- a/utils/stable_diffusion_controlnet_inpaint.py +++ b/utils/stable_diffusion_controlnet_inpaint.py @@ -9,9 +9,22 @@ import torch import torch.nn.functional as F from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer -from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging -from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker -from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_controlnet import MultiControlNetModel +from diffusers import ( + AutoencoderKL, + ControlNetModel, + DiffusionPipeline, + UNet2DConditionModel, + logging, +) +from diffusers.models.controlnet import ControlNetOutput +from diffusers.pipelines.stable_diffusion import ( + StableDiffusionPipelineOutput, + StableDiffusionSafetyChecker, +) +from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_controlnet import ( + MultiControlNetModel, +) + from diffusers.schedulers import KarrasDiffusionSchedulers from diffusers.utils import ( PIL_INTERPOLATION, @@ -20,7 +33,8 @@ from diffusers.utils import ( randn_tensor, replace_example_docstring, ) -from diffusers.loaders import LoraLoaderMixin +from diffusers.loaders import LoraLoaderMixin, TextualInversionLoaderMixin +from utils.stable_diffusion_reference import StableDiffusionReferencePipeline logger = logging.get_logger(__name__) # pylint: disable=invalid-name @@ -149,6 +163,130 @@ def prepare_image(image): return image +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.prepare_mask_and_masked_image +def prepare_mask_and_masked_image(image, mask, height, width, return_image=False): + """ + Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be + converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the + ``image`` and ``1`` for the ``mask``. + + The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be + binarized (``mask > 0.5``) and cast to ``torch.float32`` too. + + Args: + image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint. + It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width`` + ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``. + mask (_type_): The mask to apply to the image, i.e. regions to inpaint. + It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width`` + ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``. + + + Raises: + ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask + should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions. + TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not + (ot the other way around). + + Returns: + tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4 + dimensions: ``batch x channels x height x width``. + """ + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + if mask is None: + raise ValueError("`mask_image` input cannot be undefined.") + + if isinstance(image, torch.Tensor): + if not isinstance(mask, torch.Tensor): + raise TypeError( + f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not") + + # Batch single image + if image.ndim == 3: + assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)" + image = image.unsqueeze(0) + + # Batch and add channel dim for single mask + if mask.ndim == 2: + mask = mask.unsqueeze(0).unsqueeze(0) + + # Batch single mask or add channel dim + if mask.ndim == 3: + # Single batched mask, no channel dim or single mask not batched but channel dim + if mask.shape[0] == 1: + mask = mask.unsqueeze(0) + + # Batched masks no channel dim + else: + mask = mask.unsqueeze(1) + + assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions" + assert image.shape[-2:] == mask.shape[-2: + ], "Image and Mask must have the same spatial dimensions" + assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size" + + # Check image is in [-1, 1] + if image.min() < -1 or image.max() > 1: + raise ValueError("Image should be in [-1, 1] range") + + # Check mask is in [0, 1] + if mask.min() < 0 or mask.max() > 1: + raise ValueError("Mask should be in [0, 1] range") + + # Binarize mask + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + # Image as float32 + image = image.to(dtype=torch.float32) + elif isinstance(mask, torch.Tensor): + raise TypeError( + f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not") + else: + # preprocess image + if isinstance(image, (PIL.Image.Image, np.ndarray)): + image = [image] + if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): + # resize all images w.r.t passed height an width + image = [i.resize((width, height), resample=PIL.Image.LANCZOS) + for i in image] + image = [np.array(i.convert("RGB"))[None, :] for i in image] + image = np.concatenate(image, axis=0) + elif isinstance(image, list) and isinstance(image[0], np.ndarray): + image = np.concatenate([i[None, :] for i in image], axis=0) + + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): + mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) + for i in mask] + mask = np.concatenate( + [np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) + mask = mask.astype(np.float32) / 255.0 + elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): + mask = np.concatenate([m[None, None, :] for m in mask], axis=0) + + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + masked_image = image * (mask < 0.5) + + # n.b. ensure backwards compatibility as old function does not return image + if return_image: + return mask, masked_image, image + + return mask, masked_image + + def prepare_mask_image(mask_image): if isinstance(mask_image, torch.Tensor): if mask_image.ndim == 2: @@ -172,10 +310,13 @@ def prepare_mask_image(mask_image): mask_image = [mask_image] if isinstance(mask_image, list) and isinstance(mask_image[0], PIL.Image.Image): - mask_image = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask_image], axis=0) + mask_image = np.concatenate( + [np.array(m.convert("L"))[None, None, :] for m in mask_image], axis=0 + ) mask_image = mask_image.astype(np.float32) / 255.0 elif isinstance(mask_image, list) and isinstance(mask_image[0], np.ndarray): - mask_image = np.concatenate([m[None, None, :] for m in mask_image], axis=0) + mask_image = np.concatenate( + [m[None, None, :] for m in mask_image], axis=0) mask_image[mask_image < 0.5] = 0 mask_image[mask_image >= 0.5] = 1 @@ -185,7 +326,14 @@ def prepare_mask_image(mask_image): def prepare_controlnet_conditioning_image( - controlnet_conditioning_image, width, height, batch_size, num_images_per_prompt, device, dtype, do_classifier_free_guidance, + controlnet_conditioning_image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance, ): if not isinstance(controlnet_conditioning_image, torch.Tensor): if isinstance(controlnet_conditioning_image, PIL.Image.Image): @@ -193,15 +341,29 @@ def prepare_controlnet_conditioning_image( if isinstance(controlnet_conditioning_image[0], PIL.Image.Image): controlnet_conditioning_image = [ - np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"]))[None, :] + np.array( + i.resize((width, height), + resample=PIL_INTERPOLATION["lanczos"]) + )[None, :] for i in controlnet_conditioning_image ] - controlnet_conditioning_image = np.concatenate(controlnet_conditioning_image, axis=0) - controlnet_conditioning_image = np.array(controlnet_conditioning_image).astype(np.float32) / 255.0 - controlnet_conditioning_image = controlnet_conditioning_image.transpose(0, 3, 1, 2) - controlnet_conditioning_image = torch.from_numpy(controlnet_conditioning_image) + controlnet_conditioning_image = np.concatenate( + controlnet_conditioning_image, axis=0 + ) + controlnet_conditioning_image = ( + np.array(controlnet_conditioning_image).astype( + np.float32) / 255.0 + ) + controlnet_conditioning_image = controlnet_conditioning_image.transpose( + 0, 3, 1, 2 + ) + controlnet_conditioning_image = torch.from_numpy( + controlnet_conditioning_image + ) elif isinstance(controlnet_conditioning_image[0], torch.Tensor): - controlnet_conditioning_image = torch.cat(controlnet_conditioning_image, dim=0) + controlnet_conditioning_image = torch.cat( + controlnet_conditioning_image, dim=0 + ) image_batch_size = controlnet_conditioning_image.shape[0] @@ -211,17 +373,27 @@ def prepare_controlnet_conditioning_image( # image batch size is the same as prompt batch size repeat_by = num_images_per_prompt - controlnet_conditioning_image = controlnet_conditioning_image.repeat_interleave(repeat_by, dim=0) + controlnet_conditioning_image = controlnet_conditioning_image.repeat_interleave( + repeat_by, dim=0 + ) - controlnet_conditioning_image = controlnet_conditioning_image.to(device=device, dtype=dtype) + controlnet_conditioning_image = controlnet_conditioning_image.to( + device=device, dtype=dtype + ) if do_classifier_free_guidance: - controlnet_conditioning_image = torch.cat([controlnet_conditioning_image] * 2) + controlnet_conditioning_image = torch.cat( + [controlnet_conditioning_image] * 2) return controlnet_conditioning_image -class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixin): +class StableDiffusionControlNetInpaintPipeline( + DiffusionPipeline, + LoraLoaderMixin, + StableDiffusionReferencePipeline, + TextualInversionLoaderMixin, +): """ Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/ """ @@ -234,7 +406,12 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi text_encoder: CLIPTextModel, tokenizer: CLIPTokenizer, unet: UNet2DConditionModel, - controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel], + controlnet: Union[ + ControlNetModel, + List[ControlNetModel], + Tuple[ControlNetModel], + MultiControlNetModel, + ], scheduler: KarrasDiffusionSchedulers, safety_checker: StableDiffusionSafetyChecker, feature_extractor: CLIPImageProcessor, @@ -269,8 +446,10 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi safety_checker=safety_checker, feature_extractor=feature_extractor, ) - self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) - self.register_to_config(requires_safety_checker=requires_safety_checker) + self.vae_scale_factor = 2 ** ( + len(self.vae.config.block_out_channels) - 1) + self.register_to_config( + requires_safety_checker=requires_safety_checker) def enable_vae_slicing(self): r""" @@ -299,15 +478,23 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi if is_accelerate_available(): from accelerate import cpu_offload else: - raise ImportError("Please install accelerate via `pip install accelerate`") + raise ImportError( + "Please install accelerate via `pip install accelerate`") device = torch.device(f"cuda:{gpu_id}") - for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]: + for cpu_offloaded_model in [ + self.unet, + self.text_encoder, + self.vae, + self.controlnet, + ]: cpu_offload(cpu_offloaded_model, device) if self.safety_checker is not None: - cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True) + cpu_offload( + self.safety_checker, execution_device=device, offload_buffers=True + ) def enable_model_cpu_offload(self, gpu_id=0): r""" @@ -319,17 +506,23 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): from accelerate import cpu_offload_with_hook else: - raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") + raise ImportError( + "`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher." + ) device = torch.device(f"cuda:{gpu_id}") hook = None for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]: - _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook) + _, hook = cpu_offload_with_hook( + cpu_offloaded_model, device, prev_module_hook=hook + ) if self.safety_checker is not None: # the safety checker can offload the vae again - _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook) + _, hook = cpu_offload_with_hook( + self.safety_checker, device, prev_module_hook=hook + ) # control net hook has be manually offloaded as it alternates with unet cpu_offload_with_hook(self.controlnet, device) @@ -348,22 +541,22 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi 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 + 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=None, - prompt_embeds: Optional[torch.FloatTensor] = None, - negative_prompt_embeds: Optional[torch.FloatTensor] = None, + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, ): r""" Encodes the prompt into text encoder hidden states. @@ -404,20 +597,25 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi return_tensors="pt", ) text_input_ids = text_inputs.input_ids - untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + untruncated_ids = self.tokenizer( + prompt, padding="longest", return_tensors="pt" + ).input_ids - if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( - text_input_ids, untruncated_ids - ): + if untruncated_ids.shape[-1] >= text_input_ids.shape[ + -1 + ] and not torch.equal(text_input_ids, untruncated_ids): removed_text = self.tokenizer.batch_decode( - untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + untruncated_ids[:, self.tokenizer.model_max_length - 1: -1] ) logger.warning( "The following part of your input was truncated because CLIP can only handle sequences up to" f" {self.tokenizer.model_max_length} tokens: {removed_text}" ) - if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + if ( + hasattr(self.text_encoder.config, "use_attention_mask") + and self.text_encoder.config.use_attention_mask + ): attention_mask = text_inputs.attention_mask.to(device) else: attention_mask = None @@ -428,12 +626,15 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi ) prompt_embeds = prompt_embeds[0] - prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + prompt_embeds = prompt_embeds.to( + dtype=self.text_encoder.dtype, device=device) bs_embed, seq_len, _ = prompt_embeds.shape # duplicate text embeddings for each generation per prompt, using mps friendly method prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) - prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_embeds = prompt_embeds.view( + bs_embed * num_images_per_prompt, seq_len, -1 + ) # get unconditional embeddings for classifier free guidance if do_classifier_free_guidance and negative_prompt_embeds is None: @@ -465,7 +666,10 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi return_tensors="pt", ) - if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + if ( + hasattr(self.text_encoder.config, "use_attention_mask") + and self.text_encoder.config.use_attention_mask + ): attention_mask = uncond_input.attention_mask.to(device) else: attention_mask = None @@ -480,10 +684,16 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi # duplicate unconditional embeddings for each generation per prompt, using mps friendly method seq_len = negative_prompt_embeds.shape[1] - negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.to( + dtype=self.text_encoder.dtype, device=device + ) - negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) - negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + negative_prompt_embeds = negative_prompt_embeds.repeat( + 1, num_images_per_prompt, 1 + ) + negative_prompt_embeds = negative_prompt_embeds.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) # For classifier free guidance, we need to do two forward passes. # Here we concatenate the unconditional and text embeddings into a single batch @@ -494,9 +704,12 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi 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) + 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) + images=image, clip_input=safety_checker_input.pixel_values.to( + dtype) ) else: has_nsfw_concept = None @@ -516,13 +729,17 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi # 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()) + 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()) + accepts_generator = "generator" in set( + inspect.signature(self.scheduler.step).parameters.keys() + ) if accepts_generator: extra_step_kwargs["generator"] = generator return extra_step_kwargs @@ -530,10 +747,19 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi def check_controlnet_conditioning_image(self, image, prompt, prompt_embeds): image_is_pil = isinstance(image, PIL.Image.Image) image_is_tensor = isinstance(image, torch.Tensor) - image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) - image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_pil_list = isinstance(image, list) and isinstance( + image[0], PIL.Image.Image + ) + image_is_tensor_list = isinstance(image, list) and isinstance( + image[0], torch.Tensor + ) - if not image_is_pil and not image_is_tensor and not image_is_pil_list and not image_is_tensor_list: + if ( + not image_is_pil + and not image_is_tensor + and not image_is_pil_list + and not image_is_tensor_list + ): raise TypeError( "image must be passed and be one of PIL image, torch tensor, list of PIL images, or list of torch tensors" ) @@ -564,24 +790,27 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi ) def check_inputs( - self, - prompt, - image, - mask_image, - controlnet_conditioning_image, - height, - width, - callback_steps, - negative_prompt=None, - prompt_embeds=None, - negative_prompt_embeds=None, - controlnet_conditioning_scale=None, + self, + prompt, + image, + mask_image, + controlnet_conditioning_image, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + controlnet_conditioning_scale=None, ): 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}.") + 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) + 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" @@ -597,8 +826,12 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi raise ValueError( "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." ) - elif prompt is not None and (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)}") + elif prompt is not None and ( + 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 negative_prompt is not None and negative_prompt_embeds is not None: raise ValueError( @@ -616,27 +849,33 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi # check controlnet condition image if isinstance(self.controlnet, ControlNetModel): - self.check_controlnet_conditioning_image(controlnet_conditioning_image, prompt, prompt_embeds) + self.check_controlnet_conditioning_image( + controlnet_conditioning_image, prompt, prompt_embeds + ) elif isinstance(self.controlnet, MultiControlNetModel): if not isinstance(controlnet_conditioning_image, list): - raise TypeError("For multiple controlnets: `image` must be type `list`") + raise TypeError( + "For multiple controlnets: `image` must be type `list`") if len(controlnet_conditioning_image) != len(self.controlnet.nets): raise ValueError( "For multiple controlnets: `image` must have the same length as the number of controlnets." ) for image_ in controlnet_conditioning_image: - self.check_controlnet_conditioning_image(image_, prompt, prompt_embeds) + self.check_controlnet_conditioning_image( + image_, prompt, prompt_embeds) else: assert False # Check `controlnet_conditioning_scale` if isinstance(self.controlnet, ControlNetModel): if not isinstance(controlnet_conditioning_scale, float): - raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + raise TypeError( + "For single controlnet: `controlnet_conditioning_scale` must be type `float`." + ) elif isinstance(self.controlnet, MultiControlNetModel): - if isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len( - self.controlnet.nets - ): + if isinstance(controlnet_conditioning_scale, list) and len( + controlnet_conditioning_scale + ) != len(self.controlnet.nets): raise ValueError( "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" " the same length as the number of controlnets" @@ -645,23 +884,35 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi assert False if isinstance(image, torch.Tensor) and not isinstance(mask_image, torch.Tensor): - raise TypeError("if `image` is a tensor, `mask_image` must also be a tensor") + raise TypeError( + "if `image` is a tensor, `mask_image` must also be a tensor" + ) - if isinstance(image, PIL.Image.Image) and not isinstance(mask_image, PIL.Image.Image): - raise TypeError("if `image` is a PIL image, `mask_image` must also be a PIL image") + if isinstance(image, PIL.Image.Image) and not isinstance( + mask_image, PIL.Image.Image + ): + raise TypeError( + "if `image` is a PIL image, `mask_image` must also be a PIL image" + ) if isinstance(image, torch.Tensor): if image.ndim != 3 and image.ndim != 4: raise ValueError("`image` must have 3 or 4 dimensions") if mask_image.ndim != 2 and mask_image.ndim != 3 and mask_image.ndim != 4: - raise ValueError("`mask_image` must have 2, 3, or 4 dimensions") + raise ValueError( + "`mask_image` must have 2, 3, or 4 dimensions") if image.ndim == 3: image_batch_size = 1 image_channels, image_height, image_width = image.shape elif image.ndim == 4: - image_batch_size, image_channels, image_height, image_width = image.shape + ( + image_batch_size, + image_channels, + image_height, + image_width, + ) = image.shape else: assert False @@ -671,9 +922,18 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi mask_image_height, mask_image_width = mask_image.shape elif mask_image.ndim == 3: mask_image_channels = 1 - mask_image_batch_size, mask_image_height, mask_image_width = mask_image.shape + ( + mask_image_batch_size, + mask_image_height, + mask_image_width, + ) = mask_image.shape elif mask_image.ndim == 4: - mask_image_batch_size, mask_image_channels, mask_image_height, mask_image_width = mask_image.shape + ( + mask_image_batch_size, + mask_image_channels, + mask_image_height, + mask_image_width, + ) = mask_image.shape if image_channels != 3: raise ValueError("`image` must have 3 channels") @@ -682,10 +942,14 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi raise ValueError("`mask_image` must have 1 channel") if image_batch_size != mask_image_batch_size: - raise ValueError("`image` and `mask_image` mush have the same batch sizes") + raise ValueError( + "`image` and `mask_image` mush have the same batch sizes" + ) if image_height != mask_image_height or image_width != mask_image_width: - raise ValueError("`image` and `mask_image` must have the same height and width dimensions") + raise ValueError( + "`image` and `mask_image` must have the same height and width dimensions" + ) if image.min() < -1 or image.max() > 1: raise ValueError("`image` should be in range [-1, 1]") @@ -698,11 +962,14 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi single_image_latent_channels = self.vae.config.latent_channels - if self.unet.config.in_channels==4: - total_latent_channels = single_image_latent_channels # support base model without inpainting ability. + if self.unet.config.in_channels == 4: + # support base model without inpainting ability. + total_latent_channels = single_image_latent_channels else: - total_latent_channels = single_image_latent_channels * 2 + mask_image_channels - + total_latent_channels = ( + single_image_latent_channels * 2 + mask_image_channels + ) + if total_latent_channels != self.unet.config.in_channels: raise ValueError( f"The config of `pipeline.unet` expects {self.unet.config.in_channels} but received" @@ -711,8 +978,23 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi f" Please verify the config of `pipeline.unet` and the `mask_image` and `image` inputs." ) - def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): - shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + shape = ( + batch_size, + num_channels_latents, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) if isinstance(generator, list) and len(generator) != batch_size: raise ValueError( f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" @@ -720,7 +1002,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi ) if latents is None: - latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = randn_tensor( + shape, generator=generator, device=device, dtype=dtype + ) else: latents = latents.to(device) @@ -729,11 +1013,24 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi return latents - def prepare_mask_latents(self, mask_image, batch_size, height, width, dtype, device, do_classifier_free_guidance): + def prepare_mask_latents( + self, + mask_image, + batch_size, + height, + width, + dtype, + device, + do_classifier_free_guidance, + ): # resize the mask to latents shape as we concatenate the mask to the latents # we do that before converting to dtype to avoid breaking in case we're using cpu_offload # and half precision - mask_image = F.interpolate(mask_image, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)) + mask_image = F.interpolate( + mask_image, + size=(height // self.vae_scale_factor, + width // self.vae_scale_factor), + ) mask_image = mask_image.to(device=device, dtype=dtype) # duplicate mask for each generation per prompt, using mps friendly method @@ -744,28 +1041,44 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi f" a total batch size of {batch_size}, but {mask_image.shape[0]} masks were passed. Make sure the number" " of masks that you pass is divisible by the total requested batch size." ) - mask_image = mask_image.repeat(batch_size // mask_image.shape[0], 1, 1, 1) + mask_image = mask_image.repeat( + batch_size // mask_image.shape[0], 1, 1, 1) - mask_image = torch.cat([mask_image] * 2) if do_classifier_free_guidance else mask_image + mask_image = ( + torch.cat([mask_image] * + 2) if do_classifier_free_guidance else mask_image + ) mask_image_latents = mask_image return mask_image_latents def prepare_masked_image_latents( - self, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + self, + masked_image, + batch_size, + height, + width, + dtype, + device, + generator, + do_classifier_free_guidance, ): masked_image = masked_image.to(device=device, dtype=dtype) # encode the mask image into latents space so we can concatenate it to the latents if isinstance(generator, list): masked_image_latents = [ - self.vae.encode(masked_image[i : i + 1]).latent_dist.sample(generator=generator[i]) + self.vae.encode(masked_image[i: i + 1]).latent_dist.sample( + generator=generator[i] + ) for i in range(batch_size) ] masked_image_latents = torch.cat(masked_image_latents, dim=0) else: - masked_image_latents = self.vae.encode(masked_image).latent_dist.sample(generator=generator) + masked_image_latents = self.vae.encode(masked_image).latent_dist.sample( + generator=generator + ) masked_image_latents = self.vae.config.scaling_factor * masked_image_latents # duplicate masked_image_latents for each generation per prompt, using mps friendly method @@ -776,14 +1089,19 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." " Make sure the number of images that you pass is divisible by the total requested batch size." ) - masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + masked_image_latents = masked_image_latents.repeat( + batch_size // masked_image_latents.shape[0], 1, 1, 1 + ) masked_image_latents = ( - torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + torch.cat([masked_image_latents] * 2) + if do_classifier_free_guidance + else masked_image_latents ) # aligning device to prevent device errors when concating it with the latent model input - masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + masked_image_latents = masked_image_latents.to( + device=device, dtype=dtype) return masked_image_latents def _default_height_width(self, height, width, image): @@ -816,7 +1134,10 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi image: Union[torch.Tensor, PIL.Image.Image] = None, mask_image: Union[torch.Tensor, PIL.Image.Image] = None, controlnet_conditioning_image: Union[ - torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image] + torch.FloatTensor, + PIL.Image.Image, + List[torch.FloatTensor], + List[PIL.Image.Image], ] = None, height: Optional[int] = None, width: Optional[int] = None, @@ -825,17 +1146,39 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi negative_prompt: Optional[Union[str, List[str]]] = None, num_images_per_prompt: Optional[int] = 1, eta: float = 0.0, - generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + generator: Optional[Union[torch.Generator, + List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, - callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback: Optional[Callable[[ + int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, cross_attention_kwargs: Optional[Dict[str, Any]] = None, controlnet_conditioning_scale: Union[float, List[float]] = 1.0, - alignment_ratio = None, + alignment_ratio=None, + guess_mode: bool = False, + ref_image: Union[ + torch.FloatTensor, + PIL.Image.Image, + List[torch.FloatTensor], + List[PIL.Image.Image], + ] = None, + ref_mask: Union[ + torch.FloatTensor, + PIL.Image.Image, + List[torch.FloatTensor], + List[PIL.Image.Image], + ] = None, + ref_controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + ref_prompt: Union[str, List[str]] = None, + attention_auto_machine_weight: float = 1.0, + gn_auto_machine_weight: float = 1.0, + style_fidelity: float = 0.5, + reference_attn: bool = True, + reference_adain: bool = True, ): r""" Function invoked when calling the pipeline for generation. @@ -910,6 +1253,25 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi controlnet_conditioning_scale (`float`, *optional*, defaults to 1.0): The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added to the residual in the original unet. + guess_mode (`bool`, *optional*, defaults to `False`): + In this mode, the ControlNet encoder will try best to recognize the content of the input image even if + you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended. + ref_image (`torch.FloatTensor`, `PIL.Image.Image`): + The Reference Control input condition. Reference Control uses this input condition to generate guidance to Unet. If + the type is specified as `Torch.FloatTensor`, it is passed to Reference Control as is. `PIL.Image.Image` can + also be accepted as an image. + attention_auto_machine_weight (`float`): + Weight of using reference query for self attention's context. + If attention_auto_machine_weight=1.0, use reference query for all self attention's context. + gn_auto_machine_weight (`float`): + Weight of using reference adain. If gn_auto_machine_weight=2.0, use all reference adain plugins. + style_fidelity (`float`): + style fidelity of ref_uncond_xt. If style_fidelity=1.0, control more important, + elif style_fidelity=0.0, prompt more important, else balanced. + reference_attn (`bool`): + Whether to use reference query for self attention's context. + reference_adain (`bool`): + Whether to use reference adain. Examples: @@ -921,7 +1283,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi (nsfw) content, according to the `safety_checker`. """ # 0. Default height and width to unet - height, width = self._default_height_width(height, width, controlnet_conditioning_image) + height, width = self._default_height_width( + height, width, controlnet_conditioning_image + ) # 1. Check inputs. Raise error if not correct self.check_inputs( @@ -937,6 +1301,15 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi negative_prompt_embeds, controlnet_conditioning_scale, ) + if ref_image is not None: # for ref_only mode + self.check_ref_input(reference_attn, reference_adain) + if ref_mask is not None: + ref_mask = prepare_mask_image(ref_mask) + ref_mask = F.interpolate( + ref_mask, + size=(height // self.vae_scale_factor, + width // self.vae_scale_factor), + ) # 2. Define call parameters if prompt is not None and isinstance(prompt, str): @@ -952,8 +1325,12 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi # corresponds to doing no classifier free guidance. do_classifier_free_guidance = guidance_scale > 1.0 - if isinstance(self.controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): - controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(self.controlnet.nets) + if isinstance(self.controlnet, MultiControlNetModel) and isinstance( + controlnet_conditioning_scale, float + ): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len( + self.controlnet.nets + ) # 3. Encode input prompt prompt_embeds = self._encode_prompt( @@ -965,8 +1342,17 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, ) + if ref_image is not None: + ref_prompt_embeds = self._encode_prompt( + ref_prompt, + device, + num_images_per_prompt * 2, + do_classifier_free_guidance, + negative_prompt="longbody, lowres, bad anatomy, bad hands, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality", + prompt_embeds=None, + ) - # 4. Prepare mask, image, and controlnet_conditioning_image + # 4. Prepare mask, image, and controlnet_conditioning_image + ref_img image = prepare_image(image) mask_image = prepare_mask_image(mask_image) @@ -1005,6 +1391,37 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi masked_image = image * (mask_image < 0.5) + if ref_image is not None: # for ref_only mode + # Preprocess reference image + # from controlnet_aux import LineartDetector + # processor = LineartDetector.from_pretrained("lllyasviel/Annotators") + ref_ori = ref_image + ref_image = self.prepare_ref_image( + image=ref_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=prompt_embeds.dtype, + ) + + ref_control_image = prepare_controlnet_conditioning_image( + controlnet_conditioning_image=ref_ori, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + ref_controlnet_conditioning_image = controlnet_conditioning_image.copy() + ref_controlnet_conditioning_image[-1] = ref_control_image + # ref_controlnet_conditioning_scale = controlnet_conditioning_scale.copy() + # ref_controlnet_conditioning_scale[0] = 1.0 # disable the first sam controlnet + # ref_controlnet_conditioning_scale[-1] = 0.2 + # 5. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.scheduler.timesteps @@ -1024,7 +1441,7 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi noise = latents - if self.unet.config.in_channels!=4: + if self.unet.config.in_channels != 4: # inpainting base model mask_image_latents = self.prepare_mask_latents( mask_image, batch_size * num_images_per_prompt, @@ -1045,7 +1462,7 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi generator, do_classifier_free_guidance, ) - if self.unet.config.in_channels==4: + if self.unet.config.in_channels == 4: # non-inpainting base model init_masked_image_latents = self.prepare_masked_image_latents( image, batch_size * num_images_per_prompt, @@ -1057,44 +1474,137 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi do_classifier_free_guidance, ) if do_classifier_free_guidance: - init_masked_image_latents, _ = init_masked_image_latents.chunk(2) + init_masked_image_latents, _ = init_masked_image_latents.chunk( + 2) # print(type(mask_image), mask_image.shape) _, _, w, h = mask_image.shape - mask_image = torch.nn.functional.interpolate(mask_image, ((w // 8, h // 8)), mode='nearest') + mask_image = torch.nn.functional.interpolate( + mask_image, ((w // 8, h // 8)), mode="nearest" + ) mask_image = mask_image.to(latents.device).type_as(latents) mask_image = 1 - mask_image + if ref_image is not None: # for ref_only mode + ref_image_latents = self.prepare_ref_latents( + ref_image, + batch_size * num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + do_classifier_free_guidance, + ) + # 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) + if ref_image is not None: # for ref_only mode + # Modify self attention and group norm + self.uc_mask = ( + torch.Tensor( + [1] * batch_size * num_images_per_prompt + + [0] * batch_size * num_images_per_prompt + ) + .type_as(ref_image_latents) + .bool() + ) + self.attention_auto_machine_weight = attention_auto_machine_weight + self.gn_auto_machine_weight = gn_auto_machine_weight + self.do_classifier_free_guidance = do_classifier_free_guidance + self.style_fidelity = style_fidelity + self.ref_mask = ref_mask + self.inpaint_mask = mask_image + attn_modules, gn_modules = self.redefine_ref_model( + self.unet, reference_attn, reference_adain, model_type="unet" + ) + + control_attn_modules, control_gn_modules = self.redefine_ref_model( + self.controlnet, reference_attn, False, model_type="controlnet" + ) + # 8. Denoising loop - num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + num_warmup_steps = len(timesteps) - \ + num_inference_steps * self.scheduler.order with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): # expand the latents if we are doing classifier free guidance non_inpainting_latent_model_input = ( - torch.cat([latents] * 2) if do_classifier_free_guidance else latents + torch.cat( + [latents] * 2) if do_classifier_free_guidance else latents ) non_inpainting_latent_model_input = self.scheduler.scale_model_input( non_inpainting_latent_model_input, t ) - if self.unet.config.in_channels!=4: + if self.unet.config.in_channels != 4: # inpainting base model inpainting_latent_model_input = torch.cat( - [non_inpainting_latent_model_input, mask_image_latents, masked_image_latents], dim=1 + [ + non_inpainting_latent_model_input, + mask_image_latents, + masked_image_latents, + ], + dim=1, ) else: inpainting_latent_model_input = non_inpainting_latent_model_input + if ref_image is not None: # for ref_only mode + # ref only part + noise = randn_tensor( + ref_image_latents.shape, + generator=generator, + device=ref_image_latents.device, + dtype=ref_image_latents.dtype, + ) + ref_xt = self.scheduler.add_noise( + ref_image_latents, + noise, + t.reshape( + 1, + ), + ) + ref_xt = self.scheduler.scale_model_input(ref_xt, t) + + MODE = "write" + self.change_module_mode( + MODE, control_attn_modules, control_gn_modules + ) + + ( + ref_down_block_res_samples, + ref_mid_block_res_sample, + ) = self.controlnet( + ref_xt, + t, + encoder_hidden_states=ref_prompt_embeds, + controlnet_cond=ref_controlnet_conditioning_image, + conditioning_scale=ref_controlnet_conditioning_scale, + guess_mode=guess_mode, + return_dict=False, + ) + + self.change_module_mode(MODE, attn_modules, gn_modules) + self.unet( + ref_xt, + t, + encoder_hidden_states=ref_prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + down_block_additional_residuals=ref_down_block_res_samples, + mid_block_additional_residual=ref_mid_block_res_sample, + return_dict=False, + ) + + # predict the noise residual + MODE = "read" # change to read mode for following noise_pred + self.change_module_mode(MODE, attn_modules, gn_modules) down_block_res_samples, mid_block_res_sample = self.controlnet( non_inpainting_latent_model_input, t, encoder_hidden_states=prompt_embeds, controlnet_cond=controlnet_conditioning_image, conditioning_scale=controlnet_conditioning_scale, + guess_mode=guess_mode, return_dict=False, ) - # predict the noise residual noise_pred = self.unet( inpainting_latent_model_input, @@ -1108,27 +1618,429 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi # 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) + 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 + latents = self.scheduler.step( + noise_pred, t, latents, **extra_step_kwargs + ).prev_sample # call the callback, if provided - if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + if i == len(timesteps) - 1 or ( + (i + 1) > num_warmup_steps and (i + + 1) % self.scheduler.order == 0 + ): progress_bar.update() if callback is not None and i % callback_steps == 0: callback(i, t, latents) - if self.unet.config.in_channels==4 and alignment_ratio is not None: + if self.unet.config.in_channels == 4 and alignment_ratio is not None: if i < len(timesteps) * alignment_ratio: # print(i, len(timesteps)) # masking for non-inpainting models - init_latents_proper = self.scheduler.add_noise(init_masked_image_latents, noise, t) - latents = (init_latents_proper * mask_image) + (latents * (1 - mask_image)) - - if self.unet.config.in_channels==4 and (alignment_ratio==1.0 or alignment_ratio is None): + init_latents_proper = self.scheduler.add_noise( + init_masked_image_latents, noise, timesteps[i + 1] + ) + latents = (init_latents_proper * mask_image) + ( + latents * (1 - mask_image) + ) + + if self.unet.config.in_channels == 4 and ( + alignment_ratio == 1.0 or alignment_ratio is None + ): # fill the unmasked part with original image - latents = (init_masked_image_latents * mask_image) + (latents * (1 - mask_image)) + latents = (init_masked_image_latents * mask_image) + ( + latents * (1 - mask_image) + ) + + # If we do sequential model offloading, let's offload unet and controlnet + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + self.controlnet.to("cpu") + torch.cuda.empty_cache() + + if output_type == "latent": + image = latents + has_nsfw_concept = None + elif output_type == "pil": + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker( + image, device, prompt_embeds.dtype + ) + + # 10. Convert to PIL + image = self.numpy_to_pil(image) + else: + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker( + image, device, prompt_embeds.dtype + ) + + # Offload last model to CPU + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.final_offload_hook.offload() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + +class StableDiffusionControlNetInpaintMixingPipeline(StableDiffusionControlNetInpaintPipeline): + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: Union[torch.Tensor, PIL.Image.Image] = None, + mask_image: Union[torch.Tensor, PIL.Image.Image] = None, + controlnet_conditioning_image: Union[ + torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image] + ] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, + List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[ + int, int, torch.FloatTensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + controlnet_conditioning_scale_map=None, + alignment_ratio=0.95, + alpha_weight=0.5, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`torch.Tensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + 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)`. + controlnet_conditioning_image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]` or `List[PIL.Image.Image]`): + The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If + the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. PIL.Image.Image` can + also be accepted as an image. The control image is automatically resized to fit the output image. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + 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. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` instead. + Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). + 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` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](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`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + 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)`. + 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. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py). + controlnet_conditioning_scale (`float`, *optional*, defaults to 1.0): + The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original unet. + + Examples: + + 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`. + """ + # 0. Default height and width to unet + height, width = self._default_height_width( + height, width, controlnet_conditioning_image) + print('Running with alpha = {}...'.format(alpha_weight)) + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + image, + mask_image, + controlnet_conditioning_image, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + controlnet_conditioning_scale, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + 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 + + if isinstance(self.controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [ + controlnet_conditioning_scale] * len(self.controlnet.nets) + + # 3. Encode input prompt + prompt_embeds = self._encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # 4. Prepare mask, image, and controlnet_conditioning_image + image = prepare_image(image) + + mask_image = prepare_mask_image(mask_image) + + if controlnet_conditioning_scale_map is not None: + if isinstance(controlnet_conditioning_scale, list): + controlnet_conditioning_scale = [scale * controlnet_conditioning_scale_map for scale in + controlnet_conditioning_scale] + else: + controlnet_conditioning_scale = controlnet_conditioning_scale * \ + controlnet_conditioning_scale_map + + # condition image(s) + if isinstance(self.controlnet, ControlNetModel): + controlnet_conditioning_image = prepare_controlnet_conditioning_image( + controlnet_conditioning_image=controlnet_conditioning_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + elif isinstance(self.controlnet, MultiControlNetModel): + controlnet_conditioning_images = [] + + for image_ in controlnet_conditioning_image: + image_ = prepare_controlnet_conditioning_image( + controlnet_conditioning_image=image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + controlnet_conditioning_images.append(image_) + + controlnet_conditioning_image = controlnet_conditioning_images + else: + assert False + + masked_image = image * (mask_image < 0.5) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + if self.unet.config.in_channels != 4: + mask_image_latents = self.prepare_mask_latents( + mask_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + do_classifier_free_guidance, + ) + + masked_image_latents = self.prepare_masked_image_latents( + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + do_classifier_free_guidance, + ) + elif self.unet.config.in_channels == 4: + init_masked_image_latents = self.prepare_masked_image_latents( + image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + do_classifier_free_guidance, + ) + if do_classifier_free_guidance: + init_masked_image_latents, _ = init_masked_image_latents.chunk( + 2) + # print(type(mask_image), mask_image.shape) + _, _, w, h = mask_image.shape + mask_image = torch.nn.functional.interpolate( + mask_image, ((w // 8, h // 8)), mode='nearest') + mask_image = mask_image.to(latents.device).type_as(latents) + mask_image = 1 - mask_image + latents = mask_image * self.scheduler.add_noise( + init_masked_image_latents, torch.randn_like( + init_masked_image_latents), timesteps[0] + ) + (1 - mask_image) * 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 + num_warmup_steps = len(timesteps) - \ + num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + non_inpainting_latent_model_input = ( + torch.cat( + [latents] * 2) if do_classifier_free_guidance else latents + ) + + non_inpainting_latent_model_input = self.scheduler.scale_model_input( + non_inpainting_latent_model_input, t + ) + if self.unet.config.in_channels != 4: + inpainting_latent_model_input = torch.cat( + [non_inpainting_latent_model_input, + mask_image_latents, masked_image_latents], dim=1 + ) + else: + inpainting_latent_model_input = non_inpainting_latent_model_input + + down_block_res_samples, mid_block_res_sample = self.controlnet( + non_inpainting_latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + controlnet_cond=controlnet_conditioning_image, + conditioning_scale=controlnet_conditioning_scale, + return_dict=False, + ) + + # predict the noise residual + noise_pred = self.unet( + inpainting_latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + ).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 + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + if self.unet.config.in_channels == 4 and i < len(timesteps) - 1: + init_latents_proper = self.scheduler.add_noise(init_masked_image_latents, + torch.randn_like( + init_masked_image_latents), + timesteps[i + 1]) + if i < len(timesteps) * alignment_ratio: + latents = init_latents_proper * mask_image \ + + ((1 - alpha_weight) * latents + alpha_weight * init_latents_proper) * ( + 1 - mask_image) + else: + latents = latents * mask_image \ + + ((1 - alpha_weight) * latents + alpha_weight * init_latents_proper) * ( + 1 - mask_image) # If we do sequential model offloading, let's offload unet and controlnet # manually for max memory savings @@ -1145,7 +2057,8 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi image = self.decode_latents(latents) # 9. Run safety checker - image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + image, has_nsfw_concept = self.run_safety_checker( + image, device, prompt_embeds.dtype) # 10. Convert to PIL image = self.numpy_to_pil(image) @@ -1154,7 +2067,8 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi image = self.decode_latents(latents) # 9. Run safety checker - image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + image, has_nsfw_concept = self.run_safety_checker( + image, device, prompt_embeds.dtype) # Offload last model to CPU if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: @@ -1163,4 +2077,4 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixi if not return_dict: return (image, has_nsfw_concept) - return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) \ No newline at end of file + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/stable_diffusion_controlnet_inpaint.py b/utils/stable_diffusion_controlnet_inpaint_v1.py similarity index 72% rename from stable_diffusion_controlnet_inpaint.py rename to utils/stable_diffusion_controlnet_inpaint_v1.py index c8231a78e40f5e2cb60b80cf422d2f741745ce46..2e5a9064542e330791ccc70e0e93aea0377d37ac 100644 --- a/stable_diffusion_controlnet_inpaint.py +++ b/utils/stable_diffusion_controlnet_inpaint_v1.py @@ -1,7 +1,7 @@ # Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/ # From https://raw.githubusercontent.com/huggingface/diffusers/53377ef83c6446033f3ee506e3ef718db817b293/examples/community/stable_diffusion_controlnet_inpaint.py import inspect -from typing import Any, Callable, Dict, List, Optional, Union +from typing import Any, Callable, Dict, List, Optional, Union, Tuple import numpy as np import PIL.Image @@ -9,8 +9,20 @@ import torch import torch.nn.functional as F from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer -from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, UNet2DConditionModel, logging -from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker +from diffusers import ( + AutoencoderKL, + ControlNetModel, + DiffusionPipeline, + UNet2DConditionModel, + logging, +) +from diffusers.pipelines.stable_diffusion import ( + StableDiffusionPipelineOutput, + StableDiffusionSafetyChecker, +) +from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_controlnet import ( + MultiControlNetModel, +) from diffusers.schedulers import KarrasDiffusionSchedulers from diffusers.utils import ( PIL_INTERPOLATION, @@ -19,7 +31,7 @@ from diffusers.utils import ( randn_tensor, replace_example_docstring, ) - +from diffusers.loaders import LoraLoaderMixin logger = logging.get_logger(__name__) # pylint: disable=invalid-name @@ -171,7 +183,9 @@ def prepare_mask_image(mask_image): mask_image = [mask_image] if isinstance(mask_image, list) and isinstance(mask_image[0], PIL.Image.Image): - mask_image = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask_image], axis=0) + mask_image = np.concatenate( + [np.array(m.convert("L"))[None, None, :] for m in mask_image], axis=0 + ) mask_image = mask_image.astype(np.float32) / 255.0 elif isinstance(mask_image, list) and isinstance(mask_image[0], np.ndarray): mask_image = np.concatenate([m[None, None, :] for m in mask_image], axis=0) @@ -184,7 +198,14 @@ def prepare_mask_image(mask_image): def prepare_controlnet_conditioning_image( - controlnet_conditioning_image, width, height, batch_size, num_images_per_prompt, device, dtype + controlnet_conditioning_image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance, ): if not isinstance(controlnet_conditioning_image, torch.Tensor): if isinstance(controlnet_conditioning_image, PIL.Image.Image): @@ -192,15 +213,27 @@ def prepare_controlnet_conditioning_image( if isinstance(controlnet_conditioning_image[0], PIL.Image.Image): controlnet_conditioning_image = [ - np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"]))[None, :] + np.array( + i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"]) + )[None, :] for i in controlnet_conditioning_image ] - controlnet_conditioning_image = np.concatenate(controlnet_conditioning_image, axis=0) - controlnet_conditioning_image = np.array(controlnet_conditioning_image).astype(np.float32) / 255.0 - controlnet_conditioning_image = controlnet_conditioning_image.transpose(0, 3, 1, 2) - controlnet_conditioning_image = torch.from_numpy(controlnet_conditioning_image) + controlnet_conditioning_image = np.concatenate( + controlnet_conditioning_image, axis=0 + ) + controlnet_conditioning_image = ( + np.array(controlnet_conditioning_image).astype(np.float32) / 255.0 + ) + controlnet_conditioning_image = controlnet_conditioning_image.transpose( + 0, 3, 1, 2 + ) + controlnet_conditioning_image = torch.from_numpy( + controlnet_conditioning_image + ) elif isinstance(controlnet_conditioning_image[0], torch.Tensor): - controlnet_conditioning_image = torch.cat(controlnet_conditioning_image, dim=0) + controlnet_conditioning_image = torch.cat( + controlnet_conditioning_image, dim=0 + ) image_batch_size = controlnet_conditioning_image.shape[0] @@ -210,14 +243,21 @@ def prepare_controlnet_conditioning_image( # image batch size is the same as prompt batch size repeat_by = num_images_per_prompt - controlnet_conditioning_image = controlnet_conditioning_image.repeat_interleave(repeat_by, dim=0) + controlnet_conditioning_image = controlnet_conditioning_image.repeat_interleave( + repeat_by, dim=0 + ) + + controlnet_conditioning_image = controlnet_conditioning_image.to( + device=device, dtype=dtype + ) - controlnet_conditioning_image = controlnet_conditioning_image.to(device=device, dtype=dtype) + if do_classifier_free_guidance: + controlnet_conditioning_image = torch.cat([controlnet_conditioning_image] * 2) return controlnet_conditioning_image -class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): +class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline, LoraLoaderMixin): """ Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/ """ @@ -230,7 +270,12 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): text_encoder: CLIPTextModel, tokenizer: CLIPTokenizer, unet: UNet2DConditionModel, - controlnet: ControlNetModel, + controlnet: Union[ + ControlNetModel, + List[ControlNetModel], + Tuple[ControlNetModel], + MultiControlNetModel, + ], scheduler: KarrasDiffusionSchedulers, safety_checker: StableDiffusionSafetyChecker, feature_extractor: CLIPImageProcessor, @@ -253,7 +298,8 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." ) - + if isinstance(controlnet, (list, tuple)): + controlnet = MultiControlNetModel(controlnet) self.register_modules( vae=vae, text_encoder=text_encoder, @@ -298,11 +344,18 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): device = torch.device(f"cuda:{gpu_id}") - for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.controlnet]: + for cpu_offloaded_model in [ + self.unet, + self.text_encoder, + self.vae, + self.controlnet, + ]: cpu_offload(cpu_offloaded_model, device) if self.safety_checker is not None: - cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True) + cpu_offload( + self.safety_checker, execution_device=device, offload_buffers=True + ) def enable_model_cpu_offload(self, gpu_id=0): r""" @@ -314,17 +367,23 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): from accelerate import cpu_offload_with_hook else: - raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") + raise ImportError( + "`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher." + ) device = torch.device(f"cuda:{gpu_id}") hook = None for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]: - _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook) + _, hook = cpu_offload_with_hook( + cpu_offloaded_model, device, prev_module_hook=hook + ) if self.safety_checker is not None: # the safety checker can offload the vae again - _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook) + _, hook = cpu_offload_with_hook( + self.safety_checker, device, prev_module_hook=hook + ) # control net hook has be manually offloaded as it alternates with unet cpu_offload_with_hook(self.controlnet, device) @@ -399,11 +458,13 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): return_tensors="pt", ) text_input_ids = text_inputs.input_ids - untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + untruncated_ids = self.tokenizer( + prompt, padding="longest", return_tensors="pt" + ).input_ids - if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( - text_input_ids, untruncated_ids - ): + if untruncated_ids.shape[-1] >= text_input_ids.shape[ + -1 + ] and not torch.equal(text_input_ids, untruncated_ids): removed_text = self.tokenizer.batch_decode( untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] ) @@ -412,7 +473,10 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): f" {self.tokenizer.model_max_length} tokens: {removed_text}" ) - if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + if ( + hasattr(self.text_encoder.config, "use_attention_mask") + and self.text_encoder.config.use_attention_mask + ): attention_mask = text_inputs.attention_mask.to(device) else: attention_mask = None @@ -428,7 +492,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): bs_embed, seq_len, _ = prompt_embeds.shape # duplicate text embeddings for each generation per prompt, using mps friendly method prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) - prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_embeds = prompt_embeds.view( + bs_embed * num_images_per_prompt, seq_len, -1 + ) # get unconditional embeddings for classifier free guidance if do_classifier_free_guidance and negative_prompt_embeds is None: @@ -460,7 +526,10 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): return_tensors="pt", ) - if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + if ( + hasattr(self.text_encoder.config, "use_attention_mask") + and self.text_encoder.config.use_attention_mask + ): attention_mask = uncond_input.attention_mask.to(device) else: attention_mask = None @@ -475,10 +544,16 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): # duplicate unconditional embeddings for each generation per prompt, using mps friendly method seq_len = negative_prompt_embeds.shape[1] - negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.to( + dtype=self.text_encoder.dtype, device=device + ) - negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) - negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + negative_prompt_embeds = negative_prompt_embeds.repeat( + 1, num_images_per_prompt, 1 + ) + negative_prompt_embeds = negative_prompt_embeds.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) # For classifier free guidance, we need to do two forward passes. # Here we concatenate the unconditional and text embeddings into a single batch @@ -489,7 +564,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): 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) + 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) ) @@ -511,17 +588,66 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): # 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()) + 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()) + 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 check_controlnet_conditioning_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_pil_list = isinstance(image, list) and isinstance( + image[0], PIL.Image.Image + ) + image_is_tensor_list = isinstance(image, list) and isinstance( + image[0], torch.Tensor + ) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_pil_list + and not image_is_tensor_list + ): + raise TypeError( + "image must be passed and be one of PIL image, torch tensor, list of PIL images, or list of torch tensors" + ) + + if image_is_pil: + image_batch_size = 1 + elif image_is_tensor: + image_batch_size = image.shape[0] + elif image_is_pil_list: + image_batch_size = len(image) + elif image_is_tensor_list: + image_batch_size = len(image) + else: + raise ValueError("controlnet condition image is not valid") + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + else: + raise ValueError("prompt or prompt_embeds are not valid") + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + def check_inputs( self, prompt, @@ -534,12 +660,16 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None, + controlnet_conditioning_scale=None, ): 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}.") + 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) + 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" @@ -555,8 +685,12 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): raise ValueError( "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." ) - elif prompt is not None and (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)}") + elif prompt is not None and ( + 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 negative_prompt is not None and negative_prompt_embeds is not None: raise ValueError( @@ -572,51 +706,51 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): f" {negative_prompt_embeds.shape}." ) - controlnet_cond_image_is_pil = isinstance(controlnet_conditioning_image, PIL.Image.Image) - controlnet_cond_image_is_tensor = isinstance(controlnet_conditioning_image, torch.Tensor) - controlnet_cond_image_is_pil_list = isinstance(controlnet_conditioning_image, list) and isinstance( - controlnet_conditioning_image[0], PIL.Image.Image - ) - controlnet_cond_image_is_tensor_list = isinstance(controlnet_conditioning_image, list) and isinstance( - controlnet_conditioning_image[0], torch.Tensor - ) - - if ( - not controlnet_cond_image_is_pil - and not controlnet_cond_image_is_tensor - and not controlnet_cond_image_is_pil_list - and not controlnet_cond_image_is_tensor_list - ): - raise TypeError( - "image must be passed and be one of PIL image, torch tensor, list of PIL images, or list of torch tensors" + # check controlnet condition image + if isinstance(self.controlnet, ControlNetModel): + self.check_controlnet_conditioning_image( + controlnet_conditioning_image, prompt, prompt_embeds ) + elif isinstance(self.controlnet, MultiControlNetModel): + if not isinstance(controlnet_conditioning_image, list): + raise TypeError("For multiple controlnets: `image` must be type `list`") + if len(controlnet_conditioning_image) != len(self.controlnet.nets): + raise ValueError( + "For multiple controlnets: `image` must have the same length as the number of controlnets." + ) + for image_ in controlnet_conditioning_image: + self.check_controlnet_conditioning_image(image_, prompt, prompt_embeds) + else: + assert False - if controlnet_cond_image_is_pil: - controlnet_cond_image_batch_size = 1 - elif controlnet_cond_image_is_tensor: - controlnet_cond_image_batch_size = controlnet_conditioning_image.shape[0] - elif controlnet_cond_image_is_pil_list: - controlnet_cond_image_batch_size = len(controlnet_conditioning_image) - elif controlnet_cond_image_is_tensor_list: - controlnet_cond_image_batch_size = len(controlnet_conditioning_image) - - if prompt is not None and isinstance(prompt, str): - prompt_batch_size = 1 - elif prompt is not None and isinstance(prompt, list): - prompt_batch_size = len(prompt) - elif prompt_embeds is not None: - prompt_batch_size = prompt_embeds.shape[0] - - if controlnet_cond_image_batch_size != 1 and controlnet_cond_image_batch_size != prompt_batch_size: - raise ValueError( - f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {controlnet_cond_image_batch_size}, prompt batch size: {prompt_batch_size}" - ) + # Check `controlnet_conditioning_scale` + if isinstance(self.controlnet, ControlNetModel): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError( + "For single controlnet: `controlnet_conditioning_scale` must be type `float`." + ) + elif isinstance(self.controlnet, MultiControlNetModel): + if isinstance(controlnet_conditioning_scale, list) and len( + controlnet_conditioning_scale + ) != len(self.controlnet.nets): + raise ValueError( + "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" + " the same length as the number of controlnets" + ) + else: + assert False if isinstance(image, torch.Tensor) and not isinstance(mask_image, torch.Tensor): - raise TypeError("if `image` is a tensor, `mask_image` must also be a tensor") + raise TypeError( + "if `image` is a tensor, `mask_image` must also be a tensor" + ) - if isinstance(image, PIL.Image.Image) and not isinstance(mask_image, PIL.Image.Image): - raise TypeError("if `image` is a PIL image, `mask_image` must also be a PIL image") + if isinstance(image, PIL.Image.Image) and not isinstance( + mask_image, PIL.Image.Image + ): + raise TypeError( + "if `image` is a PIL image, `mask_image` must also be a PIL image" + ) if isinstance(image, torch.Tensor): if image.ndim != 3 and image.ndim != 4: @@ -629,7 +763,14 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): image_batch_size = 1 image_channels, image_height, image_width = image.shape elif image.ndim == 4: - image_batch_size, image_channels, image_height, image_width = image.shape + ( + image_batch_size, + image_channels, + image_height, + image_width, + ) = image.shape + else: + assert False if mask_image.ndim == 2: mask_image_batch_size = 1 @@ -637,9 +778,18 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): mask_image_height, mask_image_width = mask_image.shape elif mask_image.ndim == 3: mask_image_channels = 1 - mask_image_batch_size, mask_image_height, mask_image_width = mask_image.shape + ( + mask_image_batch_size, + mask_image_height, + mask_image_width, + ) = mask_image.shape elif mask_image.ndim == 4: - mask_image_batch_size, mask_image_channels, mask_image_height, mask_image_width = mask_image.shape + ( + mask_image_batch_size, + mask_image_channels, + mask_image_height, + mask_image_width, + ) = mask_image.shape if image_channels != 3: raise ValueError("`image` must have 3 channels") @@ -648,10 +798,14 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): raise ValueError("`mask_image` must have 1 channel") if image_batch_size != mask_image_batch_size: - raise ValueError("`image` and `mask_image` mush have the same batch sizes") + raise ValueError( + "`image` and `mask_image` mush have the same batch sizes" + ) if image_height != mask_image_height or image_width != mask_image_width: - raise ValueError("`image` and `mask_image` must have the same height and width dimensions") + raise ValueError( + "`image` and `mask_image` must have the same height and width dimensions" + ) if image.min() < -1 or image.max() > 1: raise ValueError("`image` should be in range [-1, 1]") @@ -664,7 +818,12 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): single_image_latent_channels = self.vae.config.latent_channels - total_latent_channels = single_image_latent_channels * 2 + mask_image_channels + if self.unet.config.in_channels == 4: + total_latent_channels = single_image_latent_channels # support base model without inpainting ability. + else: + total_latent_channels = ( + single_image_latent_channels * 2 + mask_image_channels + ) if total_latent_channels != self.unet.config.in_channels: raise ValueError( @@ -674,8 +833,23 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): f" Please verify the config of `pipeline.unet` and the `mask_image` and `image` inputs." ) - def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): - shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + shape = ( + batch_size, + num_channels_latents, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) if isinstance(generator, list) and len(generator) != batch_size: raise ValueError( f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" @@ -683,7 +857,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): ) if latents is None: - latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = randn_tensor( + shape, generator=generator, device=device, dtype=dtype + ) else: latents = latents.to(device) @@ -692,11 +868,23 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): return latents - def prepare_mask_latents(self, mask_image, batch_size, height, width, dtype, device, do_classifier_free_guidance): + def prepare_mask_latents( + self, + mask_image, + batch_size, + height, + width, + dtype, + device, + do_classifier_free_guidance, + ): # resize the mask to latents shape as we concatenate the mask to the latents # we do that before converting to dtype to avoid breaking in case we're using cpu_offload # and half precision - mask_image = F.interpolate(mask_image, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)) + mask_image = F.interpolate( + mask_image, + size=(height // self.vae_scale_factor, width // self.vae_scale_factor), + ) mask_image = mask_image.to(device=device, dtype=dtype) # duplicate mask for each generation per prompt, using mps friendly method @@ -709,26 +897,40 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): ) mask_image = mask_image.repeat(batch_size // mask_image.shape[0], 1, 1, 1) - mask_image = torch.cat([mask_image] * 2) if do_classifier_free_guidance else mask_image + mask_image = ( + torch.cat([mask_image] * 2) if do_classifier_free_guidance else mask_image + ) mask_image_latents = mask_image return mask_image_latents def prepare_masked_image_latents( - self, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + self, + masked_image, + batch_size, + height, + width, + dtype, + device, + generator, + do_classifier_free_guidance, ): masked_image = masked_image.to(device=device, dtype=dtype) # encode the mask image into latents space so we can concatenate it to the latents if isinstance(generator, list): masked_image_latents = [ - self.vae.encode(masked_image[i : i + 1]).latent_dist.sample(generator=generator[i]) + self.vae.encode(masked_image[i : i + 1]).latent_dist.sample( + generator=generator[i] + ) for i in range(batch_size) ] masked_image_latents = torch.cat(masked_image_latents, dim=0) else: - masked_image_latents = self.vae.encode(masked_image).latent_dist.sample(generator=generator) + masked_image_latents = self.vae.encode(masked_image).latent_dist.sample( + generator=generator + ) masked_image_latents = self.vae.config.scaling_factor * masked_image_latents # duplicate masked_image_latents for each generation per prompt, using mps friendly method @@ -739,10 +941,14 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." " Make sure the number of images that you pass is divisible by the total requested batch size." ) - masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + masked_image_latents = masked_image_latents.repeat( + batch_size // masked_image_latents.shape[0], 1, 1, 1 + ) masked_image_latents = ( - torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + torch.cat([masked_image_latents] * 2) + if do_classifier_free_guidance + else masked_image_latents ) # aligning device to prevent device errors when concating it with the latent model input @@ -779,7 +985,10 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): image: Union[torch.Tensor, PIL.Image.Image] = None, mask_image: Union[torch.Tensor, PIL.Image.Image] = None, controlnet_conditioning_image: Union[ - torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image] + torch.FloatTensor, + PIL.Image.Image, + List[torch.FloatTensor], + List[PIL.Image.Image], ] = None, height: Optional[int] = None, width: Optional[int] = None, @@ -797,7 +1006,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, cross_attention_kwargs: Optional[Dict[str, Any]] = None, - controlnet_conditioning_scale: float = 1.0, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + alignment_ratio=None, + guess_mode: bool = False, ): r""" Function invoked when calling the pipeline for generation. @@ -872,6 +1083,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): controlnet_conditioning_scale (`float`, *optional*, defaults to 1.0): The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added to the residual in the original unet. + guess_mode (`bool`, *optional*, defaults to `False`): + In this mode, the ControlNet encoder will try best to recognize the content of the input image even if + you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended. Examples: @@ -883,7 +1097,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): (nsfw) content, according to the `safety_checker`. """ # 0. Default height and width to unet - height, width = self._default_height_width(height, width, controlnet_conditioning_image) + height, width = self._default_height_width( + height, width, controlnet_conditioning_image + ) # 1. Check inputs. Raise error if not correct self.check_inputs( @@ -897,6 +1113,7 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): negative_prompt, prompt_embeds, negative_prompt_embeds, + controlnet_conditioning_scale, ) # 2. Define call parameters @@ -913,6 +1130,13 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): # corresponds to doing no classifier free guidance. do_classifier_free_guidance = guidance_scale > 1.0 + if isinstance(self.controlnet, MultiControlNetModel) and isinstance( + controlnet_conditioning_scale, float + ): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len( + self.controlnet.nets + ) + # 3. Encode input prompt prompt_embeds = self._encode_prompt( prompt, @@ -929,15 +1153,37 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): mask_image = prepare_mask_image(mask_image) - controlnet_conditioning_image = prepare_controlnet_conditioning_image( - controlnet_conditioning_image, - width, - height, - batch_size * num_images_per_prompt, - num_images_per_prompt, - device, - self.controlnet.dtype, - ) + # condition image(s) + if isinstance(self.controlnet, ControlNetModel): + controlnet_conditioning_image = prepare_controlnet_conditioning_image( + controlnet_conditioning_image=controlnet_conditioning_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + elif isinstance(self.controlnet, MultiControlNetModel): + controlnet_conditioning_images = [] + + for image_ in controlnet_conditioning_image: + image_ = prepare_controlnet_conditioning_image( + controlnet_conditioning_image=image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + controlnet_conditioning_images.append(image_) + + controlnet_conditioning_image = controlnet_conditioning_images + else: + assert False masked_image = image * (mask_image < 0.5) @@ -958,29 +1204,49 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): latents, ) - mask_image_latents = self.prepare_mask_latents( - mask_image, - batch_size * num_images_per_prompt, - height, - width, - prompt_embeds.dtype, - device, - do_classifier_free_guidance, - ) + noise = latents - masked_image_latents = self.prepare_masked_image_latents( - masked_image, - batch_size * num_images_per_prompt, - height, - width, - prompt_embeds.dtype, - device, - generator, - do_classifier_free_guidance, - ) + if self.unet.config.in_channels != 4: + mask_image_latents = self.prepare_mask_latents( + mask_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + do_classifier_free_guidance, + ) - if do_classifier_free_guidance: - controlnet_conditioning_image = torch.cat([controlnet_conditioning_image] * 2) + masked_image_latents = self.prepare_masked_image_latents( + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + do_classifier_free_guidance, + ) + if self.unet.config.in_channels == 4: + init_masked_image_latents = self.prepare_masked_image_latents( + image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + do_classifier_free_guidance, + ) + if do_classifier_free_guidance: + init_masked_image_latents, _ = init_masked_image_latents.chunk(2) + # print(type(mask_image), mask_image.shape) + _, _, w, h = mask_image.shape + mask_image = torch.nn.functional.interpolate( + mask_image, ((w // 8, h // 8)), mode="nearest" + ) + mask_image = mask_image.to(latents.device).type_as(latents) + mask_image = 1 - mask_image # 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) @@ -997,25 +1263,28 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): non_inpainting_latent_model_input = self.scheduler.scale_model_input( non_inpainting_latent_model_input, t ) - - inpainting_latent_model_input = torch.cat( - [non_inpainting_latent_model_input, mask_image_latents, masked_image_latents], dim=1 - ) + if self.unet.config.in_channels != 4: + inpainting_latent_model_input = torch.cat( + [ + non_inpainting_latent_model_input, + mask_image_latents, + masked_image_latents, + ], + dim=1, + ) + else: + inpainting_latent_model_input = non_inpainting_latent_model_input down_block_res_samples, mid_block_res_sample = self.controlnet( non_inpainting_latent_model_input, t, encoder_hidden_states=prompt_embeds, controlnet_cond=controlnet_conditioning_image, + conditioning_scale=controlnet_conditioning_scale, + guess_mode=guess_mode, return_dict=False, ) - down_block_res_samples = [ - down_block_res_sample * controlnet_conditioning_scale - for down_block_res_sample in down_block_res_samples - ] - mid_block_res_sample *= controlnet_conditioning_scale - # predict the noise residual noise_pred = self.unet( inpainting_latent_model_input, @@ -1029,17 +1298,42 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): # 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) + 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 + latents = self.scheduler.step( + noise_pred, t, latents, **extra_step_kwargs + ).prev_sample # call the callback, if provided - if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + if i == len(timesteps) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0 + ): progress_bar.update() if callback is not None and i % callback_steps == 0: callback(i, t, latents) + if self.unet.config.in_channels == 4 and alignment_ratio is not None: + if i < len(timesteps) * alignment_ratio: + # print(i, len(timesteps)) + # masking for non-inpainting models + init_latents_proper = self.scheduler.add_noise( + init_masked_image_latents, noise, t + ) + latents = (init_latents_proper * mask_image) + ( + latents * (1 - mask_image) + ) + + if self.unet.config.in_channels == 4 and ( + alignment_ratio == 1.0 or alignment_ratio is None + ): + # fill the unmasked part with original image + latents = (init_masked_image_latents * mask_image) + ( + latents * (1 - mask_image) + ) + # If we do sequential model offloading, let's offload unet and controlnet # manually for max memory savings if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: @@ -1055,7 +1349,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): image = self.decode_latents(latents) # 9. Run safety checker - image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + image, has_nsfw_concept = self.run_safety_checker( + image, device, prompt_embeds.dtype + ) # 10. Convert to PIL image = self.numpy_to_pil(image) @@ -1064,7 +1360,9 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): image = self.decode_latents(latents) # 9. Run safety checker - image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + image, has_nsfw_concept = self.run_safety_checker( + image, device, prompt_embeds.dtype + ) # Offload last model to CPU if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: @@ -1073,4 +1371,6 @@ class StableDiffusionControlNetInpaintPipeline(DiffusionPipeline): if not return_dict: return (image, has_nsfw_concept) - return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) \ No newline at end of file + return StableDiffusionPipelineOutput( + images=image, nsfw_content_detected=has_nsfw_concept + ) diff --git a/utils/stable_diffusion_reference.py b/utils/stable_diffusion_reference.py new file mode 100644 index 0000000000000000000000000000000000000000..6a2e3a943a1d1dd7c16e7ebf07c851c20120eca7 --- /dev/null +++ b/utils/stable_diffusion_reference.py @@ -0,0 +1,1128 @@ +# Based on https://raw.githubusercontent.com/okotaku/diffusers/feature/reference_only_control/examples/community/stable_diffusion_reference.py +# Inspired by: https://github.com/Mikubill/sd-webui-controlnet/discussions/1236 +from typing import Any, Callable, Dict, List, Optional, Union, Tuple + +import numpy as np +import PIL.Image +import torch + +from diffusers import StableDiffusionPipeline +from diffusers.models.attention import BasicTransformerBlock +from diffusers.models.unet_2d_blocks import ( + CrossAttnDownBlock2D, + CrossAttnUpBlock2D, + DownBlock2D, + UpBlock2D, +) +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput +from diffusers.utils import PIL_INTERPOLATION, logging +import torch.nn.functional as F + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import UniPCMultistepScheduler + >>> from diffusers.utils import load_image + + >>> input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png") + + >>> pipe = StableDiffusionReferencePipeline.from_pretrained( + "runwayml/stable-diffusion-v1-5", + safety_checker=None, + torch_dtype=torch.float16 + ).to('cuda:0') + + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe_controlnet.scheduler.config) + + >>> result_img = pipe(ref_image=input_image, + prompt="1girl", + num_inference_steps=20, + reference_attn=True, + reference_adain=True).images[0] + + >>> result_img.show() + ``` +""" + + +def torch_dfs(model: torch.nn.Module): + result = [model] + for child in model.children(): + result += torch_dfs(child) + return result + + +class StableDiffusionReferencePipeline: + def prepare_ref_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=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 - 0.5) / 0.5 + 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 + + def prepare_ref_latents( + self, + refimage, + batch_size, + dtype, + device, + generator, + do_classifier_free_guidance, + ): + refimage = refimage.to(device=device, dtype=dtype) + + # encode the mask image into latents space so we can concatenate it to the latents + if isinstance(generator, list): + ref_image_latents = [ + self.vae.encode(refimage[i: i + 1]).latent_dist.sample( + generator=generator[i] + ) + for i in range(batch_size) + ] + ref_image_latents = torch.cat(ref_image_latents, dim=0) + else: + ref_image_latents = self.vae.encode(refimage).latent_dist.sample( + generator=generator + ) + ref_image_latents = self.vae.config.scaling_factor * ref_image_latents + + # duplicate mask and ref_image_latents for each generation per prompt, using mps friendly method + if ref_image_latents.shape[0] < batch_size: + if not batch_size % ref_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {ref_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + ref_image_latents = ref_image_latents.repeat( + batch_size // ref_image_latents.shape[0], 1, 1, 1 + ) + + ref_image_latents = ( + torch.cat([ref_image_latents] * 2) + if do_classifier_free_guidance + else ref_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + ref_image_latents = ref_image_latents.to(device=device, dtype=dtype) + return ref_image_latents + + def check_ref_input(self, reference_attn, reference_adain): + assert ( + reference_attn or reference_adain + ), "`reference_attn` or `reference_adain` must be True." + + def redefine_ref_model( + self, model, reference_attn, reference_adain, model_type="unet" + ): + def hacked_basic_transformer_inner_forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + timestep: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + class_labels: Optional[torch.LongTensor] = None, + ): + if self.use_ada_layer_norm: + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.use_ada_layer_norm_zero: + ( + norm_hidden_states, + gate_msa, + shift_mlp, + scale_mlp, + gate_mlp, + ) = self.norm1( + hidden_states, + timestep, + class_labels, + hidden_dtype=hidden_states.dtype, + ) + else: + norm_hidden_states = self.norm1(hidden_states) + + # 1. Self-Attention + cross_attention_kwargs = ( + cross_attention_kwargs if cross_attention_kwargs is not None else {} + ) + if self.only_cross_attention: + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states + if self.only_cross_attention + else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + else: + if self.MODE == "write": + if self.attention_auto_machine_weight > self.attn_weight: + # print("hacked_basic_transformer_inner_forward") + scale_ratio = ( + (self.ref_mask.shape[2] * self.ref_mask.shape[3]) + / norm_hidden_states.shape[1] + ) ** 0.5 + this_ref_mask = F.interpolate( + self.ref_mask.to(norm_hidden_states.device), + scale_factor=1 / scale_ratio, + ) + resize_norm_hidden_states = norm_hidden_states.view( + norm_hidden_states.shape[0], + this_ref_mask.shape[2], + this_ref_mask.shape[3], + -1, + ).permute(0, 3, 1, 2) + + ref_scale = 1.0 + resize_norm_hidden_states = F.interpolate( + resize_norm_hidden_states, + scale_factor=ref_scale, + mode="bilinear", + ) + this_ref_mask = F.interpolate( + this_ref_mask, scale_factor=ref_scale + ) + # print("this_ref_mask",this_ref_mask.shape) + + # this_ref_mask = this_ref_mask.view(1,-1,1) + this_ref_mask = this_ref_mask.repeat( + resize_norm_hidden_states.shape[0], + resize_norm_hidden_states.shape[1], + 1, + 1, + ).bool() + masked_norm_hidden_states = ( + resize_norm_hidden_states[this_ref_mask] + .detach() + .clone() + .view( + resize_norm_hidden_states.shape[0], + resize_norm_hidden_states.shape[1], + -1, + ) + ) + masked_norm_hidden_states = masked_norm_hidden_states.permute( + 0, 2, 1 + ) + self.bank.append(masked_norm_hidden_states) + # self.bank.append(norm_hidden_states.detach().clone()) + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states + if self.only_cross_attention + else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.MODE == "read": + if self.attention_auto_machine_weight > self.attn_weight: + # scale_ratio = ((self.ref_mask.shape[2] * self.ref_mask.shape[3])/norm_hidden_states.shape[1])**0.5 + # print(scale_ratio) + # this_ref_mask = F.interpolate(self.ref_mask.to(norm_hidden_states.device), scale_factor=1/scale_ratio).view(1,1,-1) + # print("resized mask", this_ref_mask.shape, this_ref_mask.max(), this_ref_mask.min(), this_ref_mask.sum()) + # ref_hidden_states = torch.cat([norm_hidden_states] + self.bank, dim=1) + # if attention_mask is None: + # attention_mask = torch.ones( + # norm_hidden_states.shape[0], norm_hidden_states.shape[1], ref_hidden_states.shape[1], dtype=norm_hidden_states.dtype, device=norm_hidden_states.device + # ) + # this_ref_mask = this_ref_mask.repeat(norm_hidden_states.shape[0], norm_hidden_states.shape[1], 1) + # this_ref_mask = torch.zeros( + # norm_hidden_states.shape[0], norm_hidden_states.shape[1], this_ref_mask.shape[1], dtype=norm_hidden_states.dtype, device=norm_hidden_states.device + # ) + # print(attention_mask.shape, this_ref_mask.shape) + # attention_mask = torch.cat((attention_mask, this_ref_mask), dim=-1) + # print("merge", attention_mask.shape) + ref_hidden_states = torch.cat( + [norm_hidden_states] + self.bank, dim=1 + ) + attn_output_uc = self.attn1( + norm_hidden_states, + encoder_hidden_states=ref_hidden_states, + # attention_mask=attention_mask, + **cross_attention_kwargs, + ) + attn_output_c = attn_output_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + attn_output_c[self.uc_mask] = self.attn1( + norm_hidden_states[self.uc_mask], + encoder_hidden_states=norm_hidden_states[self.uc_mask], + **cross_attention_kwargs, + ) + attn_output = ( + self.style_fidelity * attn_output_c + + (1.0 - self.style_fidelity) * attn_output_uc + ) + self.bank.clear() + else: + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states + if self.only_cross_attention + else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.use_ada_layer_norm_zero: + attn_output = gate_msa.unsqueeze(1) * attn_output + hidden_states = attn_output + hidden_states + + if self.attn2 is not None: + norm_hidden_states = ( + self.norm2(hidden_states, timestep) + if self.use_ada_layer_norm + else self.norm2(hidden_states) + ) + + # 2. Cross-Attention + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output + hidden_states + + # 3. Feed-forward + norm_hidden_states = self.norm3(hidden_states) + + if self.use_ada_layer_norm_zero: + norm_hidden_states = ( + norm_hidden_states * + (1 + scale_mlp[:, None]) + shift_mlp[:, None] + ) + + ff_output = self.ff(norm_hidden_states) + + if self.use_ada_layer_norm_zero: + ff_output = gate_mlp.unsqueeze(1) * ff_output + + hidden_states = ff_output + hidden_states + + return hidden_states + + def hacked_mid_forward(self, *args, **kwargs): + eps = 1e-6 + x = self.original_forward(*args, **kwargs) + if self.MODE == "write": + if self.gn_auto_machine_weight >= self.gn_weight: + # mask var mean + scale_ratio = self.ref_mask.shape[2] / x.shape[2] + this_ref_mask = F.interpolate( + self.ref_mask.to(x.device), scale_factor=1 / scale_ratio + ) + this_ref_mask = this_ref_mask.repeat( + x.shape[0], x.shape[1], 1, 1 + ).bool() + masked_x = ( + x[this_ref_mask] + .detach() + .clone() + .view(x.shape[0], x.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_x, dim=(2, 3), keepdim=True, correction=0 + ) + + self.mean_bank.append(mean) + self.var_bank.append(var) + if self.MODE == "read": + if ( + self.gn_auto_machine_weight >= self.gn_weight + and len(self.mean_bank) > 0 + and len(self.var_bank) > 0 + ): + # print("hacked_mid_forward") + scale_ratio = self.inpaint_mask.shape[2] / x.shape[2] + this_inpaint_mask = F.interpolate( + self.inpaint_mask.to(x.device), scale_factor=1 / scale_ratio + ) + this_inpaint_mask = this_inpaint_mask.repeat( + x.shape[0], x.shape[1], 1, 1 + ).bool() + masked_x = ( + x[this_inpaint_mask] + .detach() + .clone() + .view(x.shape[0], x.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_x, dim=(2, 3), keepdim=True, correction=0 + ) + std = torch.maximum( + var, torch.zeros_like(var) + eps) ** 0.5 + mean_acc = sum(self.mean_bank) / float(len(self.mean_bank)) + var_acc = sum(self.var_bank) / float(len(self.var_bank)) + std_acc = ( + torch.maximum(var_acc, torch.zeros_like( + var_acc) + eps) ** 0.5 + ) + x_uc = (((masked_x - mean) / std) * std_acc) + mean_acc + x_c = x_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + x_c[self.uc_mask] = masked_x[self.uc_mask] + masked_x = self.style_fidelity * x_c + \ + (1.0 - self.style_fidelity) * x_uc + x[this_inpaint_mask] = masked_x.view(-1) + self.mean_bank = [] + self.var_bank = [] + return x + + def hack_CrossAttnDownBlock2D_forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ): + eps = 1e-6 + # TODO(Patrick, William) - attention mask is not used + output_states = () + + for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)): + hidden_states = resnet(hidden_states, temb) + + if self.MODE == "write": + if self.gn_auto_machine_weight >= self.gn_weight: + # var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0) + # mask var mean + scale_ratio = self.ref_mask.shape[2] / \ + hidden_states.shape[2] + this_ref_mask = F.interpolate( + self.ref_mask.to(hidden_states.device), + scale_factor=1 / scale_ratio, + ) + this_ref_mask = this_ref_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_ref_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + self.mean_bank0.append(mean) + self.var_bank0.append(var) + if self.MODE == "read": + if ( + self.gn_auto_machine_weight >= self.gn_weight + and len(self.mean_bank0) > 0 + and len(self.var_bank0) > 0 + ): + # print("hacked_CrossAttnDownBlock2D_forward0") + scale_ratio = self.inpaint_mask.shape[2] / \ + hidden_states.shape[2] + this_inpaint_mask = F.interpolate( + self.inpaint_mask.to(hidden_states.device), scale_factor=1 / scale_ratio + ) + this_inpaint_mask = this_inpaint_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_inpaint_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + std = torch.maximum( + var, torch.zeros_like(var) + eps) ** 0.5 + mean_acc = sum(self.mean_bank0[i]) / float( + len(self.mean_bank0[i]) + ) + var_acc = sum( + self.var_bank0[i]) / float(len(self.var_bank0[i])) + std_acc = ( + torch.maximum( + var_acc, torch.zeros_like(var_acc) + eps) + ** 0.5 + ) + hidden_states_uc = ( + ((masked_hidden_states - mean) / std) * std_acc + ) + mean_acc + hidden_states_c = hidden_states_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask] + masked_hidden_states = ( + self.style_fidelity * hidden_states_c + + (1.0 - self.style_fidelity) * hidden_states_uc + ) + hidden_states[this_inpaint_mask] = masked_hidden_states.view( + -1) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + # attention_mask=attention_mask, + # encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + if self.MODE == "write": + if self.gn_auto_machine_weight >= self.gn_weight: + # mask var mean + scale_ratio = self.ref_mask.shape[2] / \ + hidden_states.shape[2] + this_ref_mask = F.interpolate( + self.ref_mask.to(hidden_states.device), + scale_factor=1 / scale_ratio, + ) + this_ref_mask = this_ref_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_ref_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + self.mean_bank.append(mean) + self.var_bank.append(var) + if self.MODE == "read": + if ( + self.gn_auto_machine_weight >= self.gn_weight + and len(self.mean_bank) > 0 + and len(self.var_bank) > 0 + ): + # print("hack_CrossAttnDownBlock2D_forward") + scale_ratio = self.inpaint_mask.shape[2] / \ + hidden_states.shape[2] + this_inpaint_mask = F.interpolate( + self.inpaint_mask.to(hidden_states.device), scale_factor=1 / scale_ratio + ) + this_inpaint_mask = this_inpaint_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_inpaint_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + std = torch.maximum( + var, torch.zeros_like(var) + eps) ** 0.5 + mean_acc = sum(self.mean_bank[i]) / float( + len(self.mean_bank[i]) + ) + var_acc = sum( + self.var_bank[i]) / float(len(self.var_bank[i])) + std_acc = ( + torch.maximum( + var_acc, torch.zeros_like(var_acc) + eps) + ** 0.5 + ) + hidden_states_uc = ( + ((masked_hidden_states - mean) / std) * std_acc + ) + mean_acc + hidden_states_c = hidden_states_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask] + masked_hidden_states = ( + self.style_fidelity * hidden_states_c + + (1.0 - self.style_fidelity) * hidden_states_uc + ) + hidden_states[this_inpaint_mask] = masked_hidden_states.view( + -1) + + output_states = output_states + (hidden_states,) + + if self.MODE == "read": + self.mean_bank0 = [] + self.var_bank0 = [] + self.mean_bank = [] + self.var_bank = [] + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + def hacked_DownBlock2D_forward(self, hidden_states, temb=None): + eps = 1e-6 + + output_states = () + + for i, resnet in enumerate(self.resnets): + hidden_states = resnet(hidden_states, temb) + + if self.MODE == "write": + if self.gn_auto_machine_weight >= self.gn_weight: + # var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0) + # mask var mean + scale_ratio = self.ref_mask.shape[2] / \ + hidden_states.shape[2] + this_ref_mask = F.interpolate( + self.ref_mask.to(hidden_states.device), + scale_factor=1 / scale_ratio, + ) + this_ref_mask = this_ref_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_ref_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + self.mean_bank.append(mean) + self.var_bank.append(var) + if self.MODE == "read": + if ( + self.gn_auto_machine_weight >= self.gn_weight + and len(self.mean_bank) > 0 + and len(self.var_bank) > 0 + ): + # print("hacked_DownBlock2D_forward") + scale_ratio = self.inpaint_mask.shape[2] / \ + hidden_states.shape[2] + this_inpaint_mask = F.interpolate( + self.inpaint_mask.to(hidden_states.device), scale_factor=1 / scale_ratio + ) + this_inpaint_mask = this_inpaint_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_inpaint_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + std = torch.maximum( + var, torch.zeros_like(var) + eps) ** 0.5 + mean_acc = sum(self.mean_bank[i]) / float( + len(self.mean_bank[i]) + ) + var_acc = sum( + self.var_bank[i]) / float(len(self.var_bank[i])) + std_acc = ( + torch.maximum( + var_acc, torch.zeros_like(var_acc) + eps) + ** 0.5 + ) + hidden_states_uc = ( + ((masked_hidden_states - mean) / std) * std_acc + ) + mean_acc + hidden_states_c = hidden_states_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask] + masked_hidden_states = ( + self.style_fidelity * hidden_states_c + + (1.0 - self.style_fidelity) * hidden_states_uc + ) + hidden_states[this_inpaint_mask] = masked_hidden_states.view( + -1) + + output_states = output_states + (hidden_states,) + + if self.MODE == "read": + self.mean_bank = [] + self.var_bank = [] + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + def hacked_CrossAttnUpBlock2D_forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ): + eps = 1e-6 + # TODO(Patrick, William) - attention mask is not used + for i, (resnet, attn) in enumerate(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) + hidden_states = resnet(hidden_states, temb) + + if self.MODE == "write": + if self.gn_auto_machine_weight >= self.gn_weight: + # var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0) + # mask var mean + scale_ratio = self.ref_mask.shape[2] / \ + hidden_states.shape[2] + this_ref_mask = F.interpolate( + self.ref_mask.to(hidden_states.device), + scale_factor=1 / scale_ratio, + ) + this_ref_mask = this_ref_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_ref_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + self.mean_bank0.append(mean) + self.var_bank0.append(var) + if self.MODE == "read": + if ( + self.gn_auto_machine_weight >= self.gn_weight + and len(self.mean_bank0) > 0 + and len(self.var_bank0) > 0 + ): + # print("hacked_CrossAttnUpBlock2D_forward1") + scale_ratio = self.inpaint_mask.shape[2] / \ + hidden_states.shape[2] + this_inpaint_mask = F.interpolate( + self.inpaint_mask.to(hidden_states.device), scale_factor=1 / scale_ratio + ) + this_inpaint_mask = this_inpaint_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_inpaint_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + std = torch.maximum( + var, torch.zeros_like(var) + eps) ** 0.5 + mean_acc = sum(self.mean_bank0[i]) / float( + len(self.mean_bank0[i]) + ) + var_acc = sum( + self.var_bank0[i]) / float(len(self.var_bank0[i])) + std_acc = ( + torch.maximum( + var_acc, torch.zeros_like(var_acc) + eps) + ** 0.5 + ) + hidden_states_uc = ( + ((masked_hidden_states - mean) / std) * std_acc + ) + mean_acc + hidden_states_c = hidden_states_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask] + masked_hidden_states = ( + self.style_fidelity * hidden_states_c + + (1.0 - self.style_fidelity) * hidden_states_uc + ) + hidden_states[this_inpaint_mask] = masked_hidden_states.view( + -1) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + # attention_mask=attention_mask, + # encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + if self.MODE == "write": + if self.gn_auto_machine_weight >= self.gn_weight: + # var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0) + # mask var mean + scale_ratio = self.ref_mask.shape[2] / \ + hidden_states.shape[2] + this_ref_mask = F.interpolate( + self.ref_mask.to(hidden_states.device), + scale_factor=1 / scale_ratio, + ) + this_ref_mask = this_ref_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_ref_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + self.mean_bank.append(mean) + self.var_bank.append(var) + if self.MODE == "read": + if ( + self.gn_auto_machine_weight >= self.gn_weight + and len(self.mean_bank) > 0 + and len(self.var_bank) > 0 + ): + # print("hacked_CrossAttnUpBlock2D_forward") + scale_ratio = self.inpaint_mask.shape[2] / \ + hidden_states.shape[2] + this_inpaint_mask = F.interpolate( + self.inpaint_mask.to(hidden_states.device), scale_factor=1 / scale_ratio + ) + this_inpaint_mask = this_inpaint_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_inpaint_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + std = torch.maximum( + var, torch.zeros_like(var) + eps) ** 0.5 + mean_acc = sum(self.mean_bank[i]) / float( + len(self.mean_bank[i]) + ) + var_acc = sum( + self.var_bank[i]) / float(len(self.var_bank[i])) + std_acc = ( + torch.maximum( + var_acc, torch.zeros_like(var_acc) + eps) + ** 0.5 + ) + hidden_states_uc = ( + ((masked_hidden_states - mean) / std) * std_acc + ) + mean_acc + hidden_states_c = hidden_states_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask] + masked_hidden_states = ( + self.style_fidelity * hidden_states_c + + (1.0 - self.style_fidelity) * hidden_states_uc + ) + hidden_states[this_inpaint_mask] = masked_hidden_states.view( + -1) + + if self.MODE == "read": + self.mean_bank0 = [] + self.var_bank0 = [] + self.mean_bank = [] + self.var_bank = [] + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + def hacked_UpBlock2D_forward( + self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None + ): + eps = 1e-6 + for i, resnet in enumerate(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) + hidden_states = resnet(hidden_states, temb) + + if self.MODE == "write": + if self.gn_auto_machine_weight >= self.gn_weight: + # var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0) + # mask var mean + scale_ratio = self.ref_mask.shape[2] / \ + hidden_states.shape[2] + this_ref_mask = F.interpolate( + self.ref_mask.to(hidden_states.device), + scale_factor=1 / scale_ratio, + ) + this_ref_mask = this_ref_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_ref_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + self.mean_bank.append(mean) + self.var_bank.append(var) + if self.MODE == "read": + if ( + self.gn_auto_machine_weight >= self.gn_weight + and len(self.mean_bank) > 0 + and len(self.var_bank) > 0 + ): + # print("hacked_UpBlock2D_forward") + scale_ratio = self.inpaint_mask.shape[2] / \ + hidden_states.shape[2] + this_inpaint_mask = F.interpolate( + self.inpaint_mask.to(hidden_states.device), scale_factor=1 / scale_ratio + ) + this_inpaint_mask = this_inpaint_mask.repeat( + hidden_states.shape[0], hidden_states.shape[1], 1, 1 + ).bool() + masked_hidden_states = ( + hidden_states[this_inpaint_mask] + .detach() + .clone() + .view(hidden_states.shape[0], hidden_states.shape[1], -1, 1) + ) + var, mean = torch.var_mean( + masked_hidden_states, dim=(2, 3), keepdim=True, correction=0 + ) + std = torch.maximum( + var, torch.zeros_like(var) + eps) ** 0.5 + mean_acc = sum(self.mean_bank[i]) / float( + len(self.mean_bank[i]) + ) + var_acc = sum( + self.var_bank[i]) / float(len(self.var_bank[i])) + std_acc = ( + torch.maximum( + var_acc, torch.zeros_like(var_acc) + eps) + ** 0.5 + ) + hidden_states_uc = ( + ((masked_hidden_states - mean) / std) * std_acc + ) + mean_acc + hidden_states_c = hidden_states_uc.clone() + if self.do_classifier_free_guidance and self.style_fidelity > 0: + hidden_states_c[self.uc_mask] = masked_hidden_states[self.uc_mask] + masked_hidden_states = ( + self.style_fidelity * hidden_states_c + + (1.0 - self.style_fidelity) * hidden_states_uc + ) + hidden_states[this_inpaint_mask] = masked_hidden_states.view( + -1) + + if self.MODE == "read": + self.mean_bank = [] + self.var_bank = [] + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size) + + return hidden_states + + if model_type == "unet": + + if reference_attn: + attn_modules = [ + module + for module in torch_dfs(model) + if isinstance(module, BasicTransformerBlock) + ] + attn_modules = sorted( + attn_modules, key=lambda x: -x.norm1.normalized_shape[0] + ) + + for i, module in enumerate(attn_modules): + module._original_inner_forward = module.forward + module.forward = hacked_basic_transformer_inner_forward.__get__( + module, BasicTransformerBlock + ) + module.bank = [] + module.attn_weight = float(i) / float(len(attn_modules)) + module.attention_auto_machine_weight = ( + self.attention_auto_machine_weight + ) + module.gn_auto_machine_weight = self.gn_auto_machine_weight + module.do_classifier_free_guidance = ( + self.do_classifier_free_guidance + ) + module.do_classifier_free_guidance = ( + self.do_classifier_free_guidance + ) + module.uc_mask = self.uc_mask + module.style_fidelity = self.style_fidelity + module.ref_mask = self.ref_mask + else: + attn_modules = None + if reference_adain: + gn_modules = [model.mid_block] + model.mid_block.gn_weight = 0 + + down_blocks = model.down_blocks + for w, module in enumerate(down_blocks): + module.gn_weight = 1.0 - float(w) / float(len(down_blocks)) + gn_modules.append(module) + # print(module.__class__.__name__,module.gn_weight) + + up_blocks = model.up_blocks + for w, module in enumerate(up_blocks): + module.gn_weight = float(w) / float(len(up_blocks)) + gn_modules.append(module) + # print(module.__class__.__name__,module.gn_weight) + + for i, module in enumerate(gn_modules): + if getattr(module, "original_forward", None) is None: + module.original_forward = module.forward + if i == 0: + # mid_block + module.forward = hacked_mid_forward.__get__( + module, torch.nn.Module + ) + elif isinstance(module, CrossAttnDownBlock2D): + module.forward = hack_CrossAttnDownBlock2D_forward.__get__( + module, CrossAttnDownBlock2D + ) + module.mean_bank0 = [] + module.var_bank0 = [] + elif isinstance(module, DownBlock2D): + module.forward = hacked_DownBlock2D_forward.__get__( + module, DownBlock2D + ) + # elif isinstance(module, CrossAttnUpBlock2D): + # module.forward = hacked_CrossAttnUpBlock2D_forward.__get__(module, CrossAttnUpBlock2D) + # module.mean_bank0 = [] + # module.var_bank0 = [] + elif isinstance(module, UpBlock2D): + module.forward = hacked_UpBlock2D_forward.__get__( + module, UpBlock2D + ) + module.mean_bank0 = [] + module.var_bank0 = [] + module.mean_bank = [] + module.var_bank = [] + module.attention_auto_machine_weight = ( + self.attention_auto_machine_weight + ) + module.gn_auto_machine_weight = self.gn_auto_machine_weight + module.do_classifier_free_guidance = ( + self.do_classifier_free_guidance + ) + module.do_classifier_free_guidance = ( + self.do_classifier_free_guidance + ) + module.uc_mask = self.uc_mask + module.style_fidelity = self.style_fidelity + module.ref_mask = self.ref_mask + module.inpaint_mask = self.inpaint_mask + else: + gn_modules = None + elif model_type == "controlnet": + model = model.nets[-1] # only hack the inpainting controlnet + if reference_attn: + attn_modules = [ + module + for module in torch_dfs(model) + if isinstance(module, BasicTransformerBlock) + ] + attn_modules = sorted( + attn_modules, key=lambda x: -x.norm1.normalized_shape[0] + ) + for i, module in enumerate(attn_modules): + module._original_inner_forward = module.forward + module.forward = hacked_basic_transformer_inner_forward.__get__( + module, BasicTransformerBlock + ) + module.bank = [] + # float(i) / float(len(attn_modules)) + module.attn_weight = 0.0 + module.attention_auto_machine_weight = ( + self.attention_auto_machine_weight + ) + module.gn_auto_machine_weight = self.gn_auto_machine_weight + module.do_classifier_free_guidance = ( + self.do_classifier_free_guidance + ) + module.do_classifier_free_guidance = ( + self.do_classifier_free_guidance + ) + module.uc_mask = self.uc_mask + module.style_fidelity = self.style_fidelity + module.ref_mask = self.ref_mask + else: + attn_modules = None + gn_modules = None + + return attn_modules, gn_modules + + def change_module_mode(self, mode, attn_modules, gn_modules): + if attn_modules is not None: + for i, module in enumerate(attn_modules): + module.MODE = mode + if gn_modules is not None: + for i, module in enumerate(gn_modules): + module.MODE = mode diff --git a/utils/texutal_inversion.py b/utils/texutal_inversion.py new file mode 100644 index 0000000000000000000000000000000000000000..f232ed6b1e20d6d3419232960397f2066a1e8838 --- /dev/null +++ b/utils/texutal_inversion.py @@ -0,0 +1,959 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. 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 + +import argparse +import logging +import math +import os +import random +import shutil +import warnings +from pathlib import Path + +import numpy as np +import PIL +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +import transformers +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.utils import ProjectConfiguration, set_seed +from huggingface_hub import create_repo, upload_folder + +# TODO: remove and import from diffusers.utils when the new version of diffusers is released +from packaging import version +from PIL import Image +from torch.utils.data import Dataset +from torchvision import transforms +from tqdm.auto import tqdm +from transformers import CLIPTextModel, CLIPTokenizer + +import diffusers +from diffusers import ( + AutoencoderKL, + DDPMScheduler, + DiffusionPipeline, + DPMSolverMultistepScheduler, + StableDiffusionPipeline, + UNet2DConditionModel, +) +from diffusers.optimization import get_scheduler +from diffusers.utils import check_min_version, is_wandb_available +from diffusers.utils.import_utils import is_xformers_available + + +if is_wandb_available(): + import wandb + +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, + } +# ------------------------------------------------------------------------------ + + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +# check_min_version("0.18.0.dev0") + +logger = get_logger(__name__) + + +def save_model_card(repo_id: str, images=None, base_model=str, repo_folder=None): + img_str = "" + for i, image in enumerate(images): + image.save(os.path.join(repo_folder, f"image_{i}.png")) + img_str += f"![img_{i}](./image_{i}.png)\n" + + yaml = f""" +--- +license: creativeml-openrail-m +base_model: {base_model} +tags: +- stable-diffusion +- stable-diffusion-diffusers +- text-to-image +- diffusers +- textual_inversion +inference: true +--- + """ + model_card = f""" +# Textual inversion text2image fine-tuning - {repo_id} +These are textual inversion adaption weights for {base_model}. You can find some example images in the following. \n +{img_str} +""" + with open(os.path.join(repo_folder, "README.md"), "w") as f: + f.write(yaml + model_card) + + +def log_validation(text_encoder, tokenizer, unet, vae, args, accelerator, weight_dtype, epoch): + logger.info( + f"Running validation... \n Generating {args.num_validation_images} images with prompt:" + f" {args.validation_prompt}." + ) + # create pipeline (note: unet and vae are loaded again in float32) + pipeline = DiffusionPipeline.from_pretrained( + args.pretrained_model_name_or_path, + text_encoder=accelerator.unwrap_model(text_encoder), + tokenizer=tokenizer, + unet=unet, + vae=vae, + safety_checker=None, + revision=args.revision, + torch_dtype=weight_dtype, + ) + pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config) + pipeline = pipeline.to(accelerator.device) + pipeline.set_progress_bar_config(disable=True) + + # run inference + generator = None if args.seed is None else torch.Generator(device=accelerator.device).manual_seed(args.seed) + images = [] + for _ in range(args.num_validation_images): + with torch.autocast("cuda"): + image = pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0] + images.append(image) + + for tracker in accelerator.trackers: + if tracker.name == "tensorboard": + np_images = np.stack([np.asarray(img) for img in images]) + tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC") + if tracker.name == "wandb": + tracker.log( + { + "validation": [ + wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images) + ] + } + ) + + del pipeline + torch.cuda.empty_cache() + return images + + +def save_progress(text_encoder, placeholder_token_ids, accelerator, args, save_path): + logger.info("Saving embeddings") + learned_embeds = ( + accelerator.unwrap_model(text_encoder) + .get_input_embeddings() + .weight[min(placeholder_token_ids) : max(placeholder_token_ids) + 1] + ) + learned_embeds_dict = {args.placeholder_token: learned_embeds.detach().cpu()} + torch.save(learned_embeds_dict, save_path) + + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--save_steps", + type=int, + default=500, + help="Save learned_embeds.bin every X updates steps.", + ) + parser.add_argument( + "--save_as_full_pipeline", + action="store_true", + help="Save the complete stable diffusion pipeline.", + ) + parser.add_argument( + "--num_vectors", + type=int, + default=1, + help="How many textual inversion vectors shall be used to learn the concept.", + ) + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + required=True, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--revision", + type=str, + default=None, + required=False, + help="Revision of pretrained model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--tokenizer_name", + type=str, + default=None, + help="Pretrained tokenizer name or path if not the same as model_name", + ) + parser.add_argument( + "--train_data_dir", type=str, default=None, required=True, help="A folder containing the training data." + ) + parser.add_argument( + "--placeholder_token", + type=str, + default=None, + required=True, + help="A token to use as a placeholder for the concept.", + ) + parser.add_argument( + "--initializer_token", type=str, default=None, required=True, help="A token to use as initializer word." + ) + parser.add_argument("--learnable_property", type=str, default="object", help="Choose between 'object' and 'style'") + parser.add_argument("--repeats", type=int, default=100, help="How many times to repeat the training data.") + parser.add_argument( + "--output_dir", + type=str, + default="text-inversion-model", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--resolution", + type=int, + default=512, + help=( + "The resolution for input images, all the images in the train/validation dataset will be resized to this" + " resolution" + ), + ) + parser.add_argument( + "--center_crop", action="store_true", help="Whether to center crop images before resizing to resolution." + ) + parser.add_argument( + "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument("--num_train_epochs", type=int, default=100) + parser.add_argument( + "--max_train_steps", + type=int, + default=5000, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=1e-4, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--lr_num_cycles", + type=int, + default=1, + help="Number of hard resets of the lr in cosine_with_restarts scheduler.", + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--mixed_precision", + type=str, + default="no", + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose" + "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10." + "and an Nvidia Ampere GPU." + ), + ) + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument( + "--report_to", + type=str, + default="tensorboard", + help=( + 'The integration to report the results and logs to. Supported platforms are `"tensorboard"`' + ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.' + ), + ) + parser.add_argument( + "--validation_prompt", + type=str, + default=None, + help="A prompt that is used during validation to verify that the model is learning.", + ) + parser.add_argument( + "--num_validation_images", + type=int, + default=4, + help="Number of images that should be generated during validation with `validation_prompt`.", + ) + parser.add_argument( + "--validation_steps", + type=int, + default=100, + help=( + "Run validation every X steps. Validation consists of running the prompt" + " `args.validation_prompt` multiple times: `args.num_validation_images`" + " and logging the images." + ), + ) + parser.add_argument( + "--validation_epochs", + type=int, + default=None, + help=( + "Deprecated in favor of validation_steps. Run validation every X epochs. Validation consists of running the prompt" + " `args.validation_prompt` multiple times: `args.num_validation_images`" + " and logging the images." + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming" + " training using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=("Max number of checkpoints to store."), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + parser.add_argument( + "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers." + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + if args.train_data_dir is None: + raise ValueError("You must specify a train data directory.") + + return args + + +imagenet_templates_small = [ + "a photo of a {}", + "a rendering of a {}", + "a cropped photo of the {}", + "the photo of a {}", + "a photo of a clean {}", + "a photo of a dirty {}", + "a dark photo of the {}", + "a photo of my {}", + "a photo of the cool {}", + "a close-up photo of a {}", + "a bright photo of the {}", + "a cropped photo of a {}", + "a photo of the {}", + "a good photo of the {}", + "a photo of one {}", + "a close-up photo of the {}", + "a rendition of the {}", + "a photo of the clean {}", + "a rendition of a {}", + "a photo of a nice {}", + "a good photo of a {}", + "a photo of the nice {}", + "a photo of the small {}", + "a photo of the weird {}", + "a photo of the large {}", + "a photo of a cool {}", + "a photo of a small {}", +] + +imagenet_style_templates_small = [ + "a painting in the style of {}", + "a rendering in the style of {}", + "a cropped painting in the style of {}", + "the painting in the style of {}", + "a clean painting in the style of {}", + "a dirty painting in the style of {}", + "a dark painting in the style of {}", + "a picture in the style of {}", + "a cool painting in the style of {}", + "a close-up painting in the style of {}", + "a bright painting in the style of {}", + "a cropped painting in the style of {}", + "a good painting in the style of {}", + "a close-up painting in the style of {}", + "a rendition in the style of {}", + "a nice painting in the style of {}", + "a small painting in the style of {}", + "a weird painting in the style of {}", + "a large painting in the style of {}", +] + + +class TextualInversionDataset(Dataset): + def __init__( + self, + data_root, + tokenizer, + learnable_property="object", # [object, style] + size=512, + repeats=100, + interpolation="bicubic", + flip_p=0.5, + set="train", + placeholder_token="*", + center_crop=False, + ): + self.data_root = data_root + self.tokenizer = tokenizer + self.learnable_property = learnable_property + self.size = size + self.placeholder_token = placeholder_token + self.center_crop = center_crop + self.flip_p = flip_p + + self.image_paths = [os.path.join(self.data_root, file_path) for file_path in os.listdir(self.data_root)] + + self.num_images = len(self.image_paths) + self._length = self.num_images + + if set == "train": + self._length = self.num_images * repeats + + self.interpolation = { + "linear": PIL_INTERPOLATION["linear"], + "bilinear": PIL_INTERPOLATION["bilinear"], + "bicubic": PIL_INTERPOLATION["bicubic"], + "lanczos": PIL_INTERPOLATION["lanczos"], + }[interpolation] + + self.templates = imagenet_style_templates_small if learnable_property == "style" else imagenet_templates_small + self.flip_transform = transforms.RandomHorizontalFlip(p=self.flip_p) + + def __len__(self): + return self._length + + def __getitem__(self, i): + example = {} + image = Image.open(self.image_paths[i % self.num_images]) + + if not image.mode == "RGB": + image = image.convert("RGB") + + placeholder_string = self.placeholder_token + text = random.choice(self.templates).format(placeholder_string) + + example["input_ids"] = self.tokenizer( + text, + padding="max_length", + truncation=True, + max_length=self.tokenizer.model_max_length, + return_tensors="pt", + ).input_ids[0] + + # default to score-sde preprocessing + img = np.array(image).astype(np.uint8) + + if self.center_crop: + crop = min(img.shape[0], img.shape[1]) + ( + h, + w, + ) = ( + img.shape[0], + img.shape[1], + ) + img = img[(h - crop) // 2 : (h + crop) // 2, (w - crop) // 2 : (w + crop) // 2] + + image = Image.fromarray(img) + image = image.resize((self.size, self.size), resample=self.interpolation) + + image = self.flip_transform(image) + image = np.array(image).astype(np.uint8) + image = (image / 127.5 - 1.0).astype(np.float32) + + example["pixel_values"] = torch.from_numpy(image).permute(2, 0, 1) + return example + + +def main(): + args = parse_args() + logging_dir = os.path.join(args.output_dir, args.logging_dir) + accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with=args.report_to, + project_config=accelerator_project_config, + ) + + if args.report_to == "wandb": + if not is_wandb_available(): + raise ImportError("Make sure to install wandb if you want to use it for logging during training.") + + # Make one log on every process with the configuration for debugging. + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + level=logging.INFO, + ) + logger.info(accelerator.state, main_process_only=False) + if accelerator.is_local_main_process: + transformers.utils.logging.set_verbosity_warning() + diffusers.utils.logging.set_verbosity_info() + else: + transformers.utils.logging.set_verbosity_error() + diffusers.utils.logging.set_verbosity_error() + + # If passed along, set the training seed now. + if args.seed is not None: + set_seed(args.seed) + + # Handle the repository creation + if accelerator.is_main_process: + if args.output_dir is not None: + os.makedirs(args.output_dir, exist_ok=True) + + if args.push_to_hub: + repo_id = create_repo( + repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token + ).repo_id + + # Load tokenizer + if args.tokenizer_name: + tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name) + elif args.pretrained_model_name_or_path: + tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer") + + # Load scheduler and models + noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler") + text_encoder = CLIPTextModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision + ) + vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision) + unet = UNet2DConditionModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision + ) + + # Add the placeholder token in tokenizer + placeholder_tokens = [args.placeholder_token] + + if args.num_vectors < 1: + raise ValueError(f"--num_vectors has to be larger or equal to 1, but is {args.num_vectors}") + + # add dummy tokens for multi-vector + additional_tokens = [] + for i in range(1, args.num_vectors): + additional_tokens.append(f"{args.placeholder_token}_{i}") + placeholder_tokens += additional_tokens + + num_added_tokens = tokenizer.add_tokens(placeholder_tokens) + if num_added_tokens != args.num_vectors: + raise ValueError( + f"The tokenizer already contains the token {args.placeholder_token}. Please pass a different" + " `placeholder_token` that is not already in the tokenizer." + ) + + # Convert the initializer_token, placeholder_token to ids + token_ids = tokenizer.encode(args.initializer_token, add_special_tokens=False) + # Check if initializer_token is a single token or a sequence of tokens + if len(token_ids) > 1: + raise ValueError("The initializer token must be a single token.") + + initializer_token_id = token_ids[0] + placeholder_token_ids = tokenizer.convert_tokens_to_ids(placeholder_tokens) + + # Resize the token embeddings as we are adding new special tokens to the tokenizer + text_encoder.resize_token_embeddings(len(tokenizer)) + + # Initialise the newly added placeholder token with the embeddings of the initializer token + token_embeds = text_encoder.get_input_embeddings().weight.data + with torch.no_grad(): + for token_id in placeholder_token_ids: + token_embeds[token_id] = token_embeds[initializer_token_id].clone() + + # Freeze vae and unet + vae.requires_grad_(False) + unet.requires_grad_(False) + # Freeze all parameters except for the token embeddings in text encoder + text_encoder.text_model.encoder.requires_grad_(False) + text_encoder.text_model.final_layer_norm.requires_grad_(False) + text_encoder.text_model.embeddings.position_embedding.requires_grad_(False) + + if args.gradient_checkpointing: + # Keep unet in train mode if we are using gradient checkpointing to save memory. + # The dropout cannot be != 0 so it doesn't matter if we are in eval or train mode. + unet.train() + text_encoder.gradient_checkpointing_enable() + unet.enable_gradient_checkpointing() + + if args.enable_xformers_memory_efficient_attention: + if is_xformers_available(): + import xformers + + xformers_version = version.parse(xformers.__version__) + if xformers_version == version.parse("0.0.16"): + logger.warn( + "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details." + ) + unet.enable_xformers_memory_efficient_attention() + else: + raise ValueError("xformers is not available. Make sure it is installed correctly") + + # Enable TF32 for faster training on Ampere GPUs, + # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Initialize the optimizer + optimizer = torch.optim.AdamW( + text_encoder.get_input_embeddings().parameters(), # only optimize the embeddings + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + # Dataset and DataLoaders creation: + train_dataset = TextualInversionDataset( + data_root=args.train_data_dir, + tokenizer=tokenizer, + size=args.resolution, + placeholder_token=args.placeholder_token, + repeats=args.repeats, + learnable_property=args.learnable_property, + center_crop=args.center_crop, + set="train", + ) + train_dataloader = torch.utils.data.DataLoader( + train_dataset, batch_size=args.train_batch_size, shuffle=True, num_workers=args.dataloader_num_workers + ) + if args.validation_epochs is not None: + warnings.warn( + f"FutureWarning: You are doing logging with validation_epochs={args.validation_epochs}." + " Deprecated validation_epochs in favor of `validation_steps`" + f"Setting `args.validation_steps` to {args.validation_epochs * len(train_dataset)}", + FutureWarning, + stacklevel=2, + ) + args.validation_steps = args.validation_epochs * len(train_dataset) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + num_cycles=args.lr_num_cycles * args.gradient_accumulation_steps, + ) + + # Prepare everything with our `accelerator`. + text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + text_encoder, optimizer, train_dataloader, lr_scheduler + ) + + # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + # Move vae and unet to device and cast to weight_dtype + unet.to(accelerator.device, dtype=weight_dtype) + vae.to(accelerator.device, dtype=weight_dtype) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + accelerator.init_trackers("textual_inversion", config=vars(args)) + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + # Only show the progress bar once on each machine. + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + + # keep original embeddings as reference + orig_embeds_params = accelerator.unwrap_model(text_encoder).get_input_embeddings().weight.data.clone() + + for epoch in range(first_epoch, args.num_train_epochs): + text_encoder.train() + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(text_encoder): + # Convert images to latent space + latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample().detach() + latents = latents * vae.config.scaling_factor + + # Sample noise that we'll add to the latents + noise = torch.randn_like(latents) + bsz = latents.shape[0] + # Sample a random timestep for each image + timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device) + timesteps = timesteps.long() + + # Add noise to the latents according to the noise magnitude at each timestep + # (this is the forward diffusion process) + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + # Get the text embedding for conditioning + encoder_hidden_states = text_encoder(batch["input_ids"])[0].to(dtype=weight_dtype) + + # Predict the noise residual + model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample + + # Get the target for loss depending on the prediction type + if noise_scheduler.config.prediction_type == "epsilon": + target = noise + elif noise_scheduler.config.prediction_type == "v_prediction": + target = noise_scheduler.get_velocity(latents, noise, timesteps) + else: + raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") + + loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") + + accelerator.backward(loss) + + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Let's make sure we don't update any embedding weights besides the newly added token + index_no_updates = torch.ones((len(tokenizer),), dtype=torch.bool) + index_no_updates[min(placeholder_token_ids) : max(placeholder_token_ids) + 1] = False + + with torch.no_grad(): + accelerator.unwrap_model(text_encoder).get_input_embeddings().weight[ + index_no_updates + ] = orig_embeds_params[index_no_updates] + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + images = [] + progress_bar.update(1) + global_step += 1 + if global_step % args.save_steps == 0: + save_path = os.path.join(args.output_dir, f"learned_embeds-steps-{global_step}.bin") + save_progress(text_encoder, placeholder_token_ids, accelerator, args, save_path) + + if accelerator.is_main_process: + if global_step % args.checkpointing_steps == 0: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + if args.validation_prompt is not None and global_step % args.validation_steps == 0: + images = log_validation( + text_encoder, tokenizer, unet, vae, args, accelerator, weight_dtype, epoch + ) + + logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + accelerator.log(logs, step=global_step) + + if global_step >= args.max_train_steps: + break + # Create the pipeline using the trained modules and save it. + accelerator.wait_for_everyone() + if accelerator.is_main_process: + if args.push_to_hub and not args.save_as_full_pipeline: + logger.warn("Enabling full model saving because --push_to_hub=True was specified.") + save_full_model = True + else: + save_full_model = args.save_as_full_pipeline + if save_full_model: + pipeline = StableDiffusionPipeline.from_pretrained( + args.pretrained_model_name_or_path, + text_encoder=accelerator.unwrap_model(text_encoder), + vae=vae, + unet=unet, + tokenizer=tokenizer, + ) + pipeline.save_pretrained(args.output_dir) + # Save the newly trained embeddings + save_path = os.path.join(args.output_dir, "learned_embeds.bin") + save_progress(text_encoder, placeholder_token_ids, accelerator, args, save_path) + + if args.push_to_hub: + save_model_card( + repo_id, + images=images, + base_model=args.pretrained_model_name_or_path, + repo_folder=args.output_dir, + ) + upload_folder( + repo_id=repo_id, + folder_path=args.output_dir, + commit_message="End of training", + ignore_patterns=["step_*", "epoch_*"], + ) + + accelerator.end_training() + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/utils/train_dreambooth_lora_inpaint.py b/utils/train_dreambooth_lora_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..351521df7c7beddb12511d502cadeb88b5219fa2 --- /dev/null +++ b/utils/train_dreambooth_lora_inpaint.py @@ -0,0 +1,881 @@ +import argparse +import hashlib +import itertools +import math +import os +import random +from pathlib import Path + +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.utils import ProjectConfiguration, set_seed +from huggingface_hub import create_repo, upload_folder +from PIL import Image, ImageDraw +from torch.utils.data import Dataset +from torchvision import transforms +from tqdm.auto import tqdm +from transformers import CLIPTextModel, CLIPTokenizer + +from diffusers import ( + AutoencoderKL, + DDPMScheduler, + StableDiffusionInpaintPipeline, + StableDiffusionPipeline, + UNet2DConditionModel, +) +from diffusers.optimization import get_scheduler +from diffusers.utils import check_min_version +from diffusers.loaders import AttnProcsLayers, LoraLoaderMixin +from diffusers.models.attention_processor import LoRAAttnProcessor + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.13.0.dev0") + +logger = get_logger(__name__) + + +def prepare_mask_and_masked_image(image, mask): + image = np.array(image.convert("RGB")) + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + mask = np.array(mask.convert("L")) + mask = mask.astype(np.float32) / 255.0 + mask = mask[None, None] + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + masked_image = image * (mask < 0.5) + + return mask, masked_image + + +# generate random masks +def random_mask(im_shape, ratio=1, mask_full_image=False): + mask = Image.new("L", im_shape, 0) + draw = ImageDraw.Draw(mask) + size = (random.randint(0, int(im_shape[0] * ratio)), random.randint(0, int(im_shape[1] * ratio))) + # use this to always mask the whole image + if mask_full_image: + size = (int(im_shape[0] * ratio), int(im_shape[1] * ratio)) + limits = (im_shape[0] - size[0] // 2, im_shape[1] - size[1] // 2) + center = (random.randint(size[0] // 2, limits[0]), random.randint(size[1] // 2, limits[1])) + draw_type = random.randint(0, 1) + if draw_type == 0 or mask_full_image: + draw.rectangle( + (center[0] - size[0] // 2, center[1] - size[1] // 2, center[0] + size[0] // 2, center[1] + size[1] // 2), + fill=255, + ) + else: + draw.ellipse( + (center[0] - size[0] // 2, center[1] - size[1] // 2, center[0] + size[0] // 2, center[1] + size[1] // 2), + fill=255, + ) + + return mask + + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--pretrained_model_name_or_path", + type=str, + default=None, + required=True, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--tokenizer_name", + type=str, + default=None, + help="Pretrained tokenizer name or path if not the same as model_name", + ) + parser.add_argument( + "--instance_data_dir", + type=str, + default=None, + required=True, + help="A folder containing the training data of instance images.", + ) + parser.add_argument( + "--class_data_dir", + type=str, + default=None, + required=False, + help="A folder containing the training data of class images.", + ) + parser.add_argument( + "--instance_prompt", + type=str, + default=None, + help="The prompt with identifier specifying the instance", + ) + parser.add_argument( + "--class_prompt", + type=str, + default=None, + help="The prompt to specify images in the same class as provided instance images.", + ) + parser.add_argument( + "--with_prior_preservation", + default=False, + action="store_true", + help="Flag to add prior preservation loss.", + ) + parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.") + parser.add_argument( + "--num_class_images", + type=int, + default=100, + help=( + "Minimal class images for prior preservation loss. If not have enough images, additional images will be" + " sampled with class_prompt." + ), + ) + parser.add_argument( + "--output_dir", + type=str, + default="text-inversion-model", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--resolution", + type=int, + default=512, + help=( + "The resolution for input images, all the images in the train/validation dataset will be resized to this" + " resolution" + ), + ) + parser.add_argument( + "--center_crop", + default=False, + action="store_true", + help=( + "Whether to center crop the input images to the resolution. If not set, the images will be randomly" + " cropped. The images will be resized to the resolution first before cropping." + ), + ) + parser.add_argument("--train_text_encoder", action="store_true", help="Whether to train the text encoder") + parser.add_argument( + "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument( + "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images." + ) + parser.add_argument("--num_train_epochs", type=int, default=1) + parser.add_argument( + "--max_train_steps", + type=int, + default=None, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=5e-6, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--mixed_precision", + type=str, + default="no", + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose" + "between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >= 1.10." + "and an Nvidia Ampere GPU." + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final" + " checkpoints in case they are better than the last checkpoint and are suitable for resuming training" + " using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=( + "Max number of checkpoints to store. Passed as `total_limit` to the `Accelerator` `ProjectConfiguration`." + " See Accelerator::save_state https://huggingface.co/docs/accelerate/package_reference/accelerator#accelerate.Accelerator.save_state" + " for more docs" + ), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + if args.instance_data_dir is None: + raise ValueError("You must specify a train data directory.") + + if args.with_prior_preservation: + if args.class_data_dir is None: + raise ValueError("You must specify a data directory for class images.") + if args.class_prompt is None: + raise ValueError("You must specify prompt for class images.") + + return args + + +class DreamBoothDataset(Dataset): + """ + A dataset to prepare the instance and class images with the prompts for fine-tuning the model. + It pre-processes the images and the tokenizes prompts. + """ + + def __init__( + self, + instance_data_root, + instance_prompt, + tokenizer, + class_data_root=None, + class_prompt=None, + size=512, + center_crop=False, + ): + self.size = size + self.center_crop = center_crop + self.tokenizer = tokenizer + + self.instance_data_root = Path(instance_data_root) + if not self.instance_data_root.exists(): + raise ValueError("Instance images root doesn't exists.") + + self.instance_images_path = list(Path(instance_data_root).iterdir()) + self.num_instance_images = len(self.instance_images_path) + self.instance_prompt = instance_prompt + self._length = self.num_instance_images + + if class_data_root is not None: + self.class_data_root = Path(class_data_root) + self.class_data_root.mkdir(parents=True, exist_ok=True) + self.class_images_path = list(self.class_data_root.iterdir()) + self.num_class_images = len(self.class_images_path) + self._length = max(self.num_class_images, self.num_instance_images) + self.class_prompt = class_prompt + else: + self.class_data_root = None + + self.image_transforms_resize_and_crop = transforms.Compose( + [ + transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR), + transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size), + ] + ) + + self.image_transforms = transforms.Compose( + [ + transforms.ToTensor(), + transforms.Normalize([0.5], [0.5]), + ] + ) + + def __len__(self): + return self._length + + def __getitem__(self, index): + example = {} + instance_image = Image.open(self.instance_images_path[index % self.num_instance_images]) + if not instance_image.mode == "RGB": + instance_image = instance_image.convert("RGB") + instance_image = self.image_transforms_resize_and_crop(instance_image) + + example["PIL_images"] = instance_image + example["instance_images"] = self.image_transforms(instance_image) + + example["instance_prompt_ids"] = self.tokenizer( + self.instance_prompt, + padding="do_not_pad", + truncation=True, + max_length=self.tokenizer.model_max_length, + ).input_ids + + if self.class_data_root: + class_image = Image.open(self.class_images_path[index % self.num_class_images]) + if not class_image.mode == "RGB": + class_image = class_image.convert("RGB") + class_image = self.image_transforms_resize_and_crop(class_image) + example["class_images"] = self.image_transforms(class_image) + example["class_PIL_images"] = class_image + example["class_prompt_ids"] = self.tokenizer( + self.class_prompt, + padding="do_not_pad", + truncation=True, + max_length=self.tokenizer.model_max_length, + ).input_ids + + return example + + +class PromptDataset(Dataset): + "A simple dataset to prepare the prompts to generate class images on multiple GPUs." + + def __init__(self, prompt, num_samples): + self.prompt = prompt + self.num_samples = num_samples + + def __len__(self): + return self.num_samples + + def __getitem__(self, index): + example = {} + example["prompt"] = self.prompt + example["index"] = index + return example + + +def main(): + args = parse_args() + logging_dir = Path(args.output_dir, args.logging_dir) + + project_config = ProjectConfiguration(total_limit=args.checkpoints_total_limit) + + accelerator = Accelerator( + gradient_accumulation_steps=args.gradient_accumulation_steps, + mixed_precision=args.mixed_precision, + log_with="tensorboard", + logging_dir=logging_dir, + project_config=project_config, + ) + + # Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate + # This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models. + # TODO (patil-suraj): Remove this check when gradient accumulation with two models is enabled in accelerate. + if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1: + raise ValueError( + "Gradient accumulation is not supported when training the text encoder in distributed training. " + "Please set gradient_accumulation_steps to 1. This feature will be supported in the future." + ) + + if args.seed is not None: + set_seed(args.seed) + + if args.with_prior_preservation: + class_images_dir = Path(args.class_data_dir) + if not class_images_dir.exists(): + class_images_dir.mkdir(parents=True) + cur_class_images = len(list(class_images_dir.iterdir())) + + if cur_class_images < args.num_class_images: + torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32 + pipeline = StableDiffusionInpaintPipeline.from_pretrained( + args.pretrained_model_name_or_path, torch_dtype=torch_dtype, safety_checker=None + ) + pipeline.set_progress_bar_config(disable=True) + + num_new_images = args.num_class_images - cur_class_images + logger.info(f"Number of class images to sample: {num_new_images}.") + + sample_dataset = PromptDataset(args.class_prompt, num_new_images) + sample_dataloader = torch.utils.data.DataLoader( + sample_dataset, batch_size=args.sample_batch_size, num_workers=1 + ) + + sample_dataloader = accelerator.prepare(sample_dataloader) + pipeline.to(accelerator.device) + transform_to_pil = transforms.ToPILImage() + for example in tqdm( + sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process + ): + bsz = len(example["prompt"]) + fake_images = torch.rand((3, args.resolution, args.resolution)) + transform_to_pil = transforms.ToPILImage() + fake_pil_images = transform_to_pil(fake_images) + + fake_mask = random_mask((args.resolution, args.resolution), ratio=1, mask_full_image=True) + + images = pipeline(prompt=example["prompt"], mask_image=fake_mask, image=fake_pil_images).images + + for i, image in enumerate(images): + hash_image = hashlib.sha1(image.tobytes()).hexdigest() + image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg" + image.save(image_filename) + + del pipeline + if torch.cuda.is_available(): + torch.cuda.empty_cache() + + # Handle the repository creation + if accelerator.is_main_process: + if args.output_dir is not None: + os.makedirs(args.output_dir, exist_ok=True) + + if args.push_to_hub: + repo_id = create_repo( + repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token + ).repo_id + + # Load the tokenizer + if args.tokenizer_name: + tokenizer = CLIPTokenizer.from_pretrained(args.tokenizer_name) + elif args.pretrained_model_name_or_path: + tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer") + + # Load models and create wrapper for stable diffusion + text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder") + vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae") + unet = UNet2DConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet") + + vae.requires_grad_(False) + unet.requires_grad_(False) + text_encoder.requires_grad_(False) + # if not args.train_text_encoder: + # text_encoder.requires_grad_(False) + + if args.gradient_checkpointing: + unet.enable_gradient_checkpointing() + if args.train_text_encoder: + text_encoder.gradient_checkpointing_enable() + + # Set correct lora layers + unet_lora_attn_procs = {} + for name in unet.attn_processors.keys(): + cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim + if name.startswith("mid_block"): + hidden_size = unet.config.block_out_channels[-1] + elif name.startswith("up_blocks"): + block_id = int(name[len("up_blocks.")]) + hidden_size = list(reversed(unet.config.block_out_channels))[block_id] + elif name.startswith("down_blocks"): + block_id = int(name[len("down_blocks.")]) + hidden_size = unet.config.block_out_channels[block_id] + + unet_lora_attn_procs[name] = LoRAAttnProcessor( + hidden_size=hidden_size, cross_attention_dim=cross_attention_dim + ) + + unet.set_attn_processor(unet_lora_attn_procs) + unet_lora_layers = AttnProcsLayers(unet.attn_processors) + accelerator.register_for_checkpointing(unet_lora_layers) + + # The text encoder comes from 🤗 transformers, so we cannot directly modify it. + # So, instead, we monkey-patch the forward calls of its attention-blocks. For this, + # we first load a dummy pipeline with the text encoder and then do the monkey-patching. + text_encoder_lora_layers = None + if args.train_text_encoder: + text_lora_attn_procs = {} + for name, module in text_encoder.named_modules(): + if any(x in name for x in TEXT_ENCODER_TARGET_MODULES): + text_lora_attn_procs[name] = LoRAAttnProcessor( + hidden_size=module.out_features, cross_attention_dim=None + ) + text_encoder_lora_layers = AttnProcsLayers(text_lora_attn_procs) + temp_pipeline = StableDiffusionPipeline.from_pretrained( + args.pretrained_model_name_or_path, text_encoder=text_encoder + ) + temp_pipeline._modify_text_encoder(text_lora_attn_procs) + text_encoder = temp_pipeline.text_encoder + accelerator.register_for_checkpointing(text_encoder_lora_layers) + del temp_pipeline + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`." + ) + + optimizer_class = bnb.optim.AdamW8bit + else: + optimizer_class = torch.optim.AdamW + + params_to_optimize = ( + itertools.chain(unet_lora_layers.parameters(), text_encoder_lora_layers.parameters()) if args.train_text_encoder else unet_lora_layers.parameters() + ) + optimizer = optimizer_class( + params_to_optimize, + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler") + + train_dataset = DreamBoothDataset( + instance_data_root=args.instance_data_dir, + instance_prompt=args.instance_prompt, + class_data_root=args.class_data_dir if args.with_prior_preservation else None, + class_prompt=args.class_prompt, + tokenizer=tokenizer, + size=args.resolution, + center_crop=args.center_crop, + ) + + def collate_fn(examples): + input_ids = [example["instance_prompt_ids"] for example in examples] + pixel_values = [example["instance_images"] for example in examples] + + # Concat class and instance examples for prior preservation. + # We do this to avoid doing two forward passes. + if args.with_prior_preservation: + input_ids += [example["class_prompt_ids"] for example in examples] + pixel_values += [example["class_images"] for example in examples] + pior_pil = [example["class_PIL_images"] for example in examples] + + masks = [] + masked_images = [] + for example in examples: + pil_image = example["PIL_images"] + # generate a random mask + mask = random_mask(pil_image.size, 1, False) + # prepare mask and masked image + mask, masked_image = prepare_mask_and_masked_image(pil_image, mask) + + masks.append(mask) + masked_images.append(masked_image) + + if args.with_prior_preservation: + for pil_image in pior_pil: + # generate a random mask + mask = random_mask(pil_image.size, 1, False) + # prepare mask and masked image + mask, masked_image = prepare_mask_and_masked_image(pil_image, mask) + + masks.append(mask) + masked_images.append(masked_image) + + pixel_values = torch.stack(pixel_values) + pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() + + input_ids = tokenizer.pad({"input_ids": input_ids}, padding=True, return_tensors="pt").input_ids + masks = torch.stack(masks) + masked_images = torch.stack(masked_images) + batch = {"input_ids": input_ids, "pixel_values": pixel_values, "masks": masks, "masked_images": masked_images} + return batch + + train_dataloader = torch.utils.data.DataLoader( + train_dataset, batch_size=args.train_batch_size, shuffle=True, collate_fn=collate_fn + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + ) + + # if args.train_text_encoder: + # unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + # unet, text_encoder, optimizer, train_dataloader, lr_scheduler + # ) + # else: + # unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + # unet, optimizer, train_dataloader, lr_scheduler + # ) + if args.train_text_encoder: + unet_lora_layers, text_encoder_lora_layers, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet_lora_layers, text_encoder_lora_layers, optimizer, train_dataloader, lr_scheduler + ) + else: + unet_lora_layers, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet_lora_layers, optimizer, train_dataloader, lr_scheduler + ) + + accelerator.register_for_checkpointing(lr_scheduler) + + weight_dtype = torch.float32 + if args.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif args.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + # Move text_encode and vae to gpu. + # For mixed precision training we cast the text_encoder and vae weights to half-precision + # as these models are only used for inference, keeping weights in full precision is not required. + unet.to(accelerator.device, dtype=weight_dtype) + vae.to(accelerator.device, dtype=weight_dtype) + if not args.train_text_encoder: + text_encoder.to(accelerator.device, dtype=weight_dtype) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + accelerator.init_trackers("dreambooth-lora", config=vars(args)) + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num batches each epoch = {len(train_dataloader)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + # Only show the progress bar once on each machine. + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + + for epoch in range(first_epoch, args.num_train_epochs): + unet.train() + if args.train_text_encoder: + text_encoder.train() + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(unet): + # Convert images to latent space + + latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample() + latents = latents * vae.config.scaling_factor + + # Convert masked images to latent space + masked_latents = vae.encode( + batch["masked_images"].reshape(batch["pixel_values"].shape).to(dtype=weight_dtype) + ).latent_dist.sample() + masked_latents = masked_latents * vae.config.scaling_factor + + masks = batch["masks"] + # resize the mask to latents shape as we concatenate the mask to the latents + mask = torch.stack( + [ + torch.nn.functional.interpolate(mask, size=(args.resolution // 8, args.resolution // 8)) + for mask in masks + ] + ) + mask = mask.reshape(-1, 1, args.resolution // 8, args.resolution // 8) + + # Sample noise that we'll add to the latents + noise = torch.randn_like(latents) + bsz = latents.shape[0] + # Sample a random timestep for each image + timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device) + timesteps = timesteps.long() + + # Add noise to the latents according to the noise magnitude at each timestep + # (this is the forward diffusion process) + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + # concatenate the noised latents with the mask and the masked latents + latent_model_input = torch.cat([noisy_latents, mask, masked_latents], dim=1) + + # Get the text embedding for conditioning + encoder_hidden_states = text_encoder(batch["input_ids"])[0] + + # Predict the noise residual + noise_pred = unet(latent_model_input, timesteps, encoder_hidden_states).sample + + # Get the target for loss depending on the prediction type + if noise_scheduler.config.prediction_type == "epsilon": + target = noise + elif noise_scheduler.config.prediction_type == "v_prediction": + target = noise_scheduler.get_velocity(latents, noise, timesteps) + else: + raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") + + if args.with_prior_preservation: + # Chunk the noise and noise_pred into two parts and compute the loss on each part separately. + noise_pred, noise_pred_prior = torch.chunk(noise_pred, 2, dim=0) + target, target_prior = torch.chunk(target, 2, dim=0) + + # Compute instance loss + loss = F.mse_loss(noise_pred.float(), target.float(), reduction="none").mean([1, 2, 3]).mean() + + # Compute prior loss + prior_loss = F.mse_loss(noise_pred_prior.float(), target_prior.float(), reduction="mean") + + # Add the prior loss to the instance loss. + loss = loss + args.prior_loss_weight * prior_loss + else: + loss = F.mse_loss(noise_pred.float(), target.float(), reduction="mean") + + accelerator.backward(loss) + # if accelerator.sync_gradients: + # params_to_clip = ( + # itertools.chain(unet.parameters(), text_encoder.parameters()) + # if args.train_text_encoder + # else unet.parameters() + # ) + if accelerator.sync_gradients: + params_to_clip = ( + itertools.chain(unet_lora_layers.parameters(), text_encoder_lora_layers.parameters()) + if args.train_text_encoder + else unet_lora_layers.parameters() + ) + accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + progress_bar.update(1) + global_step += 1 + + if global_step % args.checkpointing_steps == 0: + if accelerator.is_main_process: + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + accelerator.log(logs, step=global_step) + + if global_step >= args.max_train_steps: + break + + accelerator.wait_for_everyone() + + # Create the pipeline using using the trained modules and save it. + if accelerator.is_main_process: + # pipeline = StableDiffusionPipeline.from_pretrained( + # args.pretrained_model_name_or_path, + # unet=accelerator.unwrap_model(unet), + # text_encoder=accelerator.unwrap_model(text_encoder), + # ) + # pipeline.save_pretrained(args.output_dir) + unet = unet.to(torch.float32) + unet.save_attn_procs(args.output_dir) + # text_encoder = text_encoder.to(torch.float32) + # LoraLoaderMixin.save_lora_weights( + # save_directory=args.output_dir, + # unet_lora_layers=unet_lora_layers, + # text_encoder_lora_layers=text_encoder_lora_layers, + # ) + + if args.push_to_hub: + upload_folder( + repo_id=repo_id, + folder_path=args.output_dir, + commit_message="End of training", + ignore_patterns=["step_*", "epoch_*"], + ) + + accelerator.end_training() + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/utils/transforms.py b/utils/transforms.py new file mode 100644 index 0000000000000000000000000000000000000000..12f6ba4274fa1f91d2844568ecbe41af586ba48c --- /dev/null +++ b/utils/transforms.py @@ -0,0 +1,247 @@ +# Code from https://github.com/LUSSeg/ImageNetSegModel/blob/main/util/transforms.py +from __future__ import division + +import math +import random +import warnings +from collections import Iterable + +import numpy as np +import torch +from torchvision.transforms import functional as F + +try: + from torchvision.transforms import InterpolationMode + + NEAREST = InterpolationMode.NEAREST + BILINEAR = InterpolationMode.BILINEAR + BICUBIC = InterpolationMode.BICUBIC + LANCZOS = InterpolationMode.LANCZOS + HAMMING = InterpolationMode.HAMMING + HAMMING = InterpolationMode.HAMMING + + _pil_interpolation_to_str = { + InterpolationMode.NEAREST: 'InterpolationMode.NEAREST', + InterpolationMode.BILINEAR: 'InterpolationMode.BILINEAR', + InterpolationMode.BICUBIC: 'InterpolationMode.BICUBIC', + InterpolationMode.LANCZOS: 'InterpolationMode.LANCZOS', + InterpolationMode.HAMMING: 'InterpolationMode.HAMMING', + InterpolationMode.BOX: 'InterpolationMode.BOX', + } + +except: + from PIL import Image + + NEAREST = Image.NEAREST + BILINEAR = Image.BILINEAR + BICUBIC = Image.BICUBIC + LANCZOS = Image.LANCZOS + HAMMING = Image.HAMMING + HAMMING = Image.HAMMING + + _pil_interpolation_to_str = { + Image.NEAREST: 'PIL.Image.NEAREST', + Image.BILINEAR: 'PIL.Image.BILINEAR', + Image.BICUBIC: 'PIL.Image.BICUBIC', + Image.LANCZOS: 'PIL.Image.LANCZOS', + Image.HAMMING: 'PIL.Image.HAMMING', + Image.BOX: 'PIL.Image.BOX', + } + +def _get_image_size(img): + if F._is_pil_image(img): + return img.size + elif isinstance(img, torch.Tensor) and img.dim() > 2: + return img.shape[-2:][::-1] + else: + raise TypeError('Unexpected type {}'.format(type(img))) + + +class Compose(object): + """Composes several transforms together. + + Args: + transforms (list of ``Transform`` objects): + list of transforms to compose. + + Example: + >>> transforms.Compose([ + >>> transforms.CenterCrop(10), + >>> transforms.ToTensor(), + >>> ]) + """ + def __init__(self, transforms): + self.transforms = transforms + + def __call__(self, img, gt): + for t in self.transforms: + if 'RandomResizedCrop' in t.__class__.__name__: + img, gt = t(img, gt) + elif 'Flip' in t.__class__.__name__: + img, gt = t(img, gt) + elif 'ToTensor' in t.__class__.__name__: + img, gt = t(img, gt) + else: + img = t(img) + gt = gt.float() + + return img, gt + + def __repr__(self): + format_string = self.__class__.__name__ + '(' + for t in self.transforms: + format_string += '\n' + format_string += ' {0}'.format(t) + format_string += '\n)' + return format_string + + +class RandomHorizontalFlip(object): + """Horizontally flip the given PIL Image randomly with a given probability. + + Args: + p (float): probability of the image being flipped. Default value is 0.5 + """ + def __init__(self, p=0.5): + self.p = p + + def __call__(self, img, gt): + """ + Args: + img (PIL Image): Image to be flipped. + + Returns: + PIL Image: Randomly flipped image. + """ + if random.random() < self.p: + return F.hflip(img), F.hflip(gt) + return img, gt + + def __repr__(self): + return self.__class__.__name__ + '(p={})'.format(self.p) + + +class RandomResizedCrop(object): + """Crop the given PIL Image to random size and aspect ratio. + + A crop of random size (default: of 0.08 to 1.0) of the original size + and a random aspect ratio (default: of 3/4 to 4/3) of the original + aspect ratio is made. This crop is finally resized to given size. + This is popularly used to train the Inception networks. + + Args: + size: expected output size of each edge + scale: range of size of the origin size cropped + ratio: range of aspect ratio of the origin aspect ratio cropped + interpolation: Default: PIL.Image.BILINEAR + """ + def __init__(self, + size, + scale=(0.08, 1.0), + ratio=(3. / 4., 4. / 3.), + interpolation=BILINEAR): + if isinstance(size, (tuple, list)): + self.size = size + else: + self.size = (size, size) + if (scale[0] > scale[1]) or (ratio[0] > ratio[1]): + warnings.warn('range should be of kind (min, max)') + + self.interpolation = interpolation + self.scale = scale + self.ratio = ratio + + @staticmethod + def get_params(img, scale, ratio): + """Get parameters for ``crop`` for a random sized crop. + + Args: + img (PIL Image): Image to be cropped. + scale (tuple): + range of size of the origin size cropped + ratio (tuple): + range of aspect ratio of the origin aspect ratio cropped + + Returns: + tuple: params (i, j, h, w) to be passed to ``crop`` for a random + sized crop. + """ + width, height = _get_image_size(img) + area = height * width + + for attempt in range(10): + target_area = random.uniform(*scale) * area + log_ratio = (math.log(ratio[0]), math.log(ratio[1])) + aspect_ratio = math.exp(random.uniform(*log_ratio)) + + w = int(round(math.sqrt(target_area * aspect_ratio))) + h = int(round(math.sqrt(target_area / aspect_ratio))) + + if 0 < w <= width and 0 < h <= height: + i = random.randint(0, height - h) + j = random.randint(0, width - w) + return i, j, h, w + + # Fallback to central crop + in_ratio = float(width) / float(height) + if (in_ratio < min(ratio)): + w = width + h = int(round(w / min(ratio))) + elif (in_ratio > max(ratio)): + h = height + w = int(round(h * max(ratio))) + else: # whole image + w = width + h = height + i = (height - h) // 2 + j = (width - w) // 2 + return i, j, h, w + + def __call__(self, img, gt): + """ + Args: + img (PIL Image): Image to be cropped and resized. + + Returns: + PIL Image: Randomly cropped and resized image. + """ + i, j, h, w = self.get_params(img, self.scale, self.ratio) + return F.resized_crop( + img, i, j, h, w, self.size, self.interpolation), \ + F.resized_crop( + gt, i, j, h, w, self.size, NEAREST) + + def __repr__(self): + interpolate_str = _pil_interpolation_to_str[self.interpolation] + format_string = self.__class__.__name__ + '(size={0}'.format(self.size) + format_string += ', scale={0}'.format( + tuple(round(s, 4) for s in self.scale)) + format_string += ', ratio={0}'.format( + tuple(round(r, 4) for r in self.ratio)) + format_string += ', interpolation={0})'.format(interpolate_str) + return format_string + +class ToTensor(object): + """Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor. + + Converts a PIL Image or numpy.ndarray (H x W x C) in the range + [0, 255] to a torch.FloatTensor of + shape (C x H x W) in the range [0.0, 1.0] + if the PIL Image belongs to one of the + modes (L, LA, P, I, F, RGB, YCbCr, RGBA, CMYK, 1) + or if the numpy.ndarray has dtype = np.uint8 + + In the other cases, tensors are returned without scaling. + """ + def __call__(self, pic, gt): + """ + Args: + pic (PIL Image or numpy.ndarray): Image to be converted to tensor. + + Returns: + Tensor: Converted image. + """ + return F.to_tensor(pic), torch.from_numpy(np.array(gt)) + + def __repr__(self): + return self.__class__.__name__ + '()' \ No newline at end of file diff --git a/vlpart/swintransformer.py b/vlpart/swintransformer.py new file mode 100644 index 0000000000000000000000000000000000000000..ac115f1e6204ac402c6feed1888119dd516145cc --- /dev/null +++ b/vlpart/swintransformer.py @@ -0,0 +1,733 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# Modified by Xingyi Zhou from https://github.com/SwinTransformer/Swin-Transformer-Object-Detection/blob/master/mmdet/models/backbones/swin_transformer.py +# -------------------------------------------------------- +# Swin Transformer +# Copyright (c) 2021 Microsoft +# Licensed under The MIT License [see LICENSE for details] +# Written by Ze Liu, Yutong Lin, Yixuan Wei +# -------------------------------------------------------- + +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.checkpoint as checkpoint +import numpy as np +from timm.models.layers import DropPath, to_2tuple, trunc_normal_ + +import fvcore.nn.weight_init as weight_init + +from detectron2.layers import ShapeSpec +from detectron2.modeling.backbone.backbone import Backbone +from detectron2.modeling.backbone.build import BACKBONE_REGISTRY +from detectron2.modeling.backbone.fpn import FPN, LastLevelMaxPool + + +class LastLevelP6P7_P5(nn.Module): + """ + This module is used in RetinaNet to generate extra layers, P6 and P7 from + C5 feature. + """ + + def __init__(self, in_channels, out_channels): + super().__init__() + self.num_levels = 2 + self.in_feature = "p5" + self.p6 = nn.Conv2d(in_channels, out_channels, 3, 2, 1) + self.p7 = nn.Conv2d(out_channels, out_channels, 3, 2, 1) + for module in [self.p6, self.p7]: + weight_init.c2_xavier_fill(module) + + def forward(self, c5): + p6 = self.p6(c5) + p7 = self.p7(F.relu(p6)) + return [p6, p7] + + +class Mlp(nn.Module): + """ Multilayer perceptron.""" + + 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 + + +def window_partition(x, window_size): + """ + Args: + x: (B, H, W, C) + window_size (int): window size + Returns: + windows: (num_windows*B, window_size, window_size, C) + """ + B, H, W, C = x.shape + x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) + windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) + return windows + + +def window_reverse(windows, window_size, H, W): + """ + Args: + windows: (num_windows*B, window_size, window_size, C) + window_size (int): Window size + H (int): Height of image + W (int): Width of image + Returns: + x: (B, H, W, C) + """ + B = int(windows.shape[0] / (H * W / window_size / window_size)) + x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) + x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) + return x + + +class WindowAttention(nn.Module): + """ Window based multi-head self attention (W-MSA) module with relative position bias. + It supports both of shifted and non-shifted window. + Args: + dim (int): Number of input channels. + window_size (tuple[int]): The height and width of the window. + num_heads (int): Number of attention heads. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set + attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 + proj_drop (float, optional): Dropout ratio of output. Default: 0.0 + """ + + def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): + + super().__init__() + self.dim = dim + self.window_size = window_size # Wh, Ww + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + # define a parameter table of relative position bias + self.relative_position_bias_table = nn.Parameter( + torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH + + # get pair-wise relative position index for each token inside the window + coords_h = torch.arange(self.window_size[0]) + coords_w = torch.arange(self.window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 + relative_coords[:, :, 1] += self.window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 + relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + self.register_buffer("relative_position_index", relative_position_index) + + 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) + + trunc_normal_(self.relative_position_bias_table, std=.02) + self.softmax = nn.Softmax(dim=-1) + + def forward(self, x, mask=None): + """ Forward function. + Args: + x: input features with shape of (num_windows*B, N, C) + mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None + """ + 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) + + q = q * self.scale + attn = (q @ k.transpose(-2, -1)) + + relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( + self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH + relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + attn = attn + relative_position_bias.unsqueeze(0) + + if mask is not None: + nW = mask.shape[0] + attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) + attn = attn.view(-1, self.num_heads, N, N) + attn = self.softmax(attn) + else: + attn = self.softmax(attn) + + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B_, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SwinTransformerBlock(nn.Module): + """ Swin Transformer Block. + Args: + dim (int): Number of input channels. + num_heads (int): Number of attention heads. + window_size (int): Window size. + shift_size (int): Shift size for SW-MSA. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float, optional): Stochastic depth rate. Default: 0.0 + act_layer (nn.Module, optional): Activation layer. Default: nn.GELU + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + + def __init__(self, dim, num_heads, window_size=7, shift_size=0, + mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0., + act_layer=nn.GELU, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.num_heads = num_heads + self.window_size = window_size + self.shift_size = shift_size + self.mlp_ratio = mlp_ratio + assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size" + + self.norm1 = norm_layer(dim) + self.attn = WindowAttention( + dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, + qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) + + 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) + + self.H = None + self.W = None + + def forward(self, x, mask_matrix): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + mask_matrix: Attention mask for cyclic shift. + """ + B, L, C = x.shape + H, W = self.H, self.W + assert L == H * W, "input feature has wrong size" + + shortcut = x + x = self.norm1(x) + x = x.view(B, H, W, C) + + # pad feature maps to multiples of window size + pad_l = pad_t = 0 + pad_r = (self.window_size - W % self.window_size) % self.window_size + pad_b = (self.window_size - H % self.window_size) % self.window_size + x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + _, Hp, Wp, _ = x.shape + + # cyclic shift + if self.shift_size > 0: + shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) + attn_mask = mask_matrix + else: + shifted_x = x + attn_mask = None + + # partition windows + x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C + x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C + + # W-MSA/SW-MSA + attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C + + # merge windows + attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) + shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C + + # reverse cyclic shift + if self.shift_size > 0: + x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) + else: + x = shifted_x + + if pad_r > 0 or pad_b > 0: + x = x[:, :H, :W, :].contiguous() + + x = x.view(B, H * W, C) + + # FFN + x = shortcut + self.drop_path(x) + x = x + self.drop_path(self.mlp(self.norm2(x))) + + return x + + +class PatchMerging(nn.Module): + """ Patch Merging Layer + Args: + dim (int): Number of input channels. + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + def __init__(self, dim, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) + self.norm = norm_layer(4 * dim) + + def forward(self, x, H, W): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + B, L, C = x.shape + assert L == H * W, "input feature has wrong size" + + x = x.view(B, H, W, C) + + # padding + pad_input = (H % 2 == 1) or (W % 2 == 1) + if pad_input: + x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2)) + + x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C + x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C + x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C + x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C + x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C + x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C + + x = self.norm(x) + x = self.reduction(x) + + return x + + +class BasicLayer(nn.Module): + """ A basic Swin Transformer layer for one stage. + Args: + dim (int): Number of feature channels + depth (int): Depths of this stage. + num_heads (int): Number of attention head. + window_size (int): Local window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + dim, + depth, + num_heads, + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop=0., + attn_drop=0., + drop_path=0., + norm_layer=nn.LayerNorm, + downsample=None, + use_checkpoint=False): + super().__init__() + self.window_size = window_size + self.shift_size = window_size // 2 + self.depth = depth + self.use_checkpoint = use_checkpoint + + # build blocks + self.blocks = nn.ModuleList([ + SwinTransformerBlock( + dim=dim, + num_heads=num_heads, + window_size=window_size, + shift_size=0 if (i % 2 == 0) else window_size // 2, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop, + attn_drop=attn_drop, + drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, + norm_layer=norm_layer) + for i in range(depth)]) + + # patch merging layer + if downsample is not None: + self.downsample = downsample(dim=dim, norm_layer=norm_layer) + else: + self.downsample = None + + def forward(self, x, H, W): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + + # calculate attention mask for SW-MSA + Hp = int(np.ceil(H / self.window_size)) * self.window_size + Wp = int(np.ceil(W / self.window_size)) * self.window_size + img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1 + h_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + w_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + cnt = 0 + for h in h_slices: + for w in w_slices: + img_mask[:, h, w, :] = cnt + cnt += 1 + + mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 + mask_windows = mask_windows.view(-1, self.window_size * self.window_size) + attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) + + for blk in self.blocks: + blk.H, blk.W = H, W + if self.use_checkpoint: + x = checkpoint.checkpoint(blk, x, attn_mask) + else: + x = blk(x, attn_mask) + if self.downsample is not None: + x_down = self.downsample(x, H, W) + Wh, Ww = (H + 1) // 2, (W + 1) // 2 + return x, H, W, x_down, Wh, Ww + else: + return x, H, W, x, H, W + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding + Args: + patch_size (int): Patch token size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + norm_layer (nn.Module, optional): Normalization layer. Default: None + """ + + def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): + super().__init__() + patch_size = to_2tuple(patch_size) + self.patch_size = patch_size + + self.in_chans = in_chans + self.embed_dim = embed_dim + + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + if norm_layer is not None: + self.norm = norm_layer(embed_dim) + else: + self.norm = None + + def forward(self, x): + """Forward function.""" + # padding + _, _, H, W = x.size() + if W % self.patch_size[1] != 0: + x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) + if H % self.patch_size[0] != 0: + x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) + + x = self.proj(x) # B C Wh Ww + if self.norm is not None: + Wh, Ww = x.size(2), x.size(3) + x = x.flatten(2).transpose(1, 2) + x = self.norm(x) + x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) + + return x + + +class SwinTransformer(Backbone): + """ Swin Transformer backbone. + A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - + https://arxiv.org/pdf/2103.14030 + Args: + pretrain_img_size (int): Input image size for training the pretrained model, + used in absolute postion embedding. Default 224. + patch_size (int | tuple(int)): Patch size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + depths (tuple[int]): Depths of each Swin Transformer stage. + num_heads (tuple[int]): Number of attention head of each stage. + window_size (int): Window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. + drop_rate (float): Dropout rate. + attn_drop_rate (float): Attention dropout rate. Default: 0. + drop_path_rate (float): Stochastic depth rate. Default: 0.2. + norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. + ape (bool): If True, add absolute position embedding to the patch embedding. Default: False. + patch_norm (bool): If True, add normalization after patch embedding. Default: True. + out_indices (Sequence[int]): Output from which stages. + frozen_stages (int): Stages to be frozen (stop grad and set eval mode). + -1 means not freezing any parameters. + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + pretrain_img_size=224, + patch_size=4, + in_chans=3, + embed_dim=96, + depths=(2, 2, 6, 2), + num_heads=(3, 6, 12, 24), + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop_rate=0., + attn_drop_rate=0., + drop_path_rate=0.2, + norm_layer=nn.LayerNorm, + ape=False, + patch_norm=True, + out_indices=(0, 1, 2, 3), + frozen_stages=-1, + use_checkpoint=False): + super().__init__() + + self.pretrain_img_size = pretrain_img_size + self.num_layers = len(depths) + self.embed_dim = embed_dim + self.ape = ape + self.patch_norm = patch_norm + self.out_indices = out_indices + self.frozen_stages = frozen_stages + + # split image into non-overlapping patches + self.patch_embed = PatchEmbed( + patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, + norm_layer=norm_layer if self.patch_norm else None) + + # absolute position embedding + if self.ape: + pretrain_img_size = to_2tuple(pretrain_img_size) + patch_size = to_2tuple(patch_size) + patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]] + + self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])) + trunc_normal_(self.absolute_pos_embed, std=.02) + + self.pos_drop = nn.Dropout(p=drop_rate) + + # stochastic depth + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule + + # build layers + self.layers = nn.ModuleList() + for i_layer in range(self.num_layers): + layer = BasicLayer( + dim=int(embed_dim * 2 ** i_layer), + depth=depths[i_layer], + num_heads=num_heads[i_layer], + window_size=window_size, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop_rate, + attn_drop=attn_drop_rate, + drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], + norm_layer=norm_layer, + downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, + use_checkpoint=use_checkpoint) + self.layers.append(layer) + + num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)] + self.num_features = num_features + + # add a norm layer for each output + for i_layer in out_indices: + layer = norm_layer(num_features[i_layer]) + layer_name = f'norm{i_layer}' + self.add_module(layer_name, layer) + + self._freeze_stages() + self._out_features = ['swin{}'.format(i) for i in self.out_indices] + self._out_feature_channels = { + 'swin{}'.format(i): self.embed_dim * 2 ** i for i in self.out_indices + } + self._out_feature_strides = { + 'swin{}'.format(i): 2 ** (i + 2) for i in self.out_indices + } + self._size_devisibility = 32 + + + def _freeze_stages(self): + if self.frozen_stages >= 0: + self.patch_embed.eval() + for param in self.patch_embed.parameters(): + param.requires_grad = False + + if self.frozen_stages >= 1 and self.ape: + self.absolute_pos_embed.requires_grad = False + + if self.frozen_stages >= 2: + self.pos_drop.eval() + for i in range(0, self.frozen_stages - 1): + m = self.layers[i] + m.eval() + for param in m.parameters(): + param.requires_grad = False + + def init_weights(self, pretrained=None): + """Initialize the weights in backbone. + Args: + pretrained (str, optional): Path to pre-trained weights. + Defaults to None. + """ + + def _init_weights(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) + + if isinstance(pretrained, str): + self.apply(_init_weights) + # load_checkpoint(self, pretrained, strict=False) + elif pretrained is None: + self.apply(_init_weights) + else: + raise TypeError('pretrained must be a str or None') + + def forward(self, x): + """Forward function.""" + x = self.patch_embed(x) + + Wh, Ww = x.size(2), x.size(3) + if self.ape: + # interpolate the position embedding to the corresponding size + absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic') + x = (x + absolute_pos_embed).flatten(2).transpose(1, 2) # B Wh*Ww C + else: + x = x.flatten(2).transpose(1, 2) + x = self.pos_drop(x) + + # outs = [] + outs = {} + for i in range(self.num_layers): + layer = self.layers[i] + x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww) + + if i in self.out_indices: + norm_layer = getattr(self, f'norm{i}') + x_out = norm_layer(x_out) + + out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous() + # outs.append(out) + outs['swin{}'.format(i)] = out + + return outs + + def train(self, mode=True): + """Convert the model into training mode while keep layers freezed.""" + super(SwinTransformer, self).train(mode) + self._freeze_stages() + +size2config = { + 'T': { + 'window_size': 7, + 'embed_dim': 96, + 'depth': [2, 2, 6, 2], + 'num_heads': [3, 6, 12, 24], + 'drop_path_rate': 0.2, + 'pretrained': 'models/swin_tiny_patch4_window7_224.pth' + }, + 'S': { + 'window_size': 7, + 'embed_dim': 96, + 'depth': [2, 2, 18, 2], + 'num_heads': [3, 6, 12, 24], + 'drop_path_rate': 0.2, + 'pretrained': 'models/swin_small_patch4_window7_224.pth' + }, + 'B': { + 'window_size': 7, + 'embed_dim': 128, + 'depth': [2, 2, 18, 2], + 'num_heads': [4, 8, 16, 32], + 'drop_path_rate': 0.3, + 'pretrained': 'models/swin_base_patch4_window7_224.pth' + }, + 'B-22k': { + 'window_size': 7, + 'embed_dim': 128, + 'depth': [2, 2, 18, 2], + 'num_heads': [4, 8, 16, 32], + 'drop_path_rate': 0.3, + 'pretrained': 'models/swin_base_patch4_window7_224_22k.pth' + }, + 'B-22k-384': { + 'window_size': 12, + 'embed_dim': 128, + 'depth': [2, 2, 18, 2], + 'num_heads': [4, 8, 16, 32], + 'drop_path_rate': 0.3, + 'pretrained': 'models/swin_base_patch4_window12_384_22k.pth' + }, + 'L-22k': { + 'window_size': 7, + 'embed_dim': 192, + 'depth': [2, 2, 18, 2], + 'num_heads': [6, 12, 24, 48], + 'drop_path_rate': 0.3, # TODO (xingyi): this is unclear + 'pretrained': 'models/swin_large_patch4_window7_224_22k.pth' + }, + 'L-22k-384': { + 'window_size': 12, + 'embed_dim': 192, + 'depth': [2, 2, 18, 2], + 'num_heads': [6, 12, 24, 48], + 'drop_path_rate': 0.3, # TODO (xingyi): this is unclear + 'pretrained': 'models/swin_large_patch4_window12_384_22k.pth' + } +} + +def build_swinbase_fpn_backbone(): + config = size2config['B-22k'] + bottom_up = SwinTransformer( + embed_dim=config['embed_dim'], + window_size=config['window_size'], + depths=config['depth'], + num_heads=config['num_heads'], + drop_path_rate=config['drop_path_rate'], + out_indices=[0, 1, 2, 3], + frozen_stages=-1, + use_checkpoint=False, + ) + backbone = FPN( + bottom_up=bottom_up, + in_features=["swin0", "swin1", "swin2", "swin3"], + out_channels=256, + norm="", + top_block=LastLevelMaxPool(), + fuse_type="sum", + ) + return backbone diff --git a/vlpart/text_encoder.py b/vlpart/text_encoder.py new file mode 100644 index 0000000000000000000000000000000000000000..ede4644276d783bcbaa6f87cb54712202f10b0e6 --- /dev/null +++ b/vlpart/text_encoder.py @@ -0,0 +1,196 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +from typing import Union, List +from collections import OrderedDict +import torch +from torch import nn +import torch + +from clip.simple_tokenizer import SimpleTokenizer as _Tokenizer + +__all__ = ["tokenize"] + + +class LayerNorm(nn.LayerNorm): + """Subclass torch's LayerNorm to handle fp16.""" + + def forward(self, x: torch.Tensor): + orig_type = x.dtype + ret = super().forward(x.type(torch.float32)) + return ret.type(orig_type) + + +class QuickGELU(nn.Module): + def forward(self, x: torch.Tensor): + return x * torch.sigmoid(1.702 * x) + + +class ResidualAttentionBlock(nn.Module): + def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None): + super().__init__() + + self.attn = nn.MultiheadAttention(d_model, n_head) + self.ln_1 = LayerNorm(d_model) + self.mlp = nn.Sequential(OrderedDict([ + ("c_fc", nn.Linear(d_model, d_model * 4)), + ("gelu", QuickGELU()), + ("c_proj", nn.Linear(d_model * 4, d_model)) + ])) + self.ln_2 = LayerNorm(d_model) + self.attn_mask = attn_mask + + def attention(self, x: torch.Tensor): + self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None + return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0] + + def forward(self, x: torch.Tensor): + x = x + self.attention(self.ln_1(x)) + x = x + self.mlp(self.ln_2(x)) + return x + + +class Transformer(nn.Module): + def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None): + super().__init__() + self.width = width + self.layers = layers + self.resblocks = nn.Sequential( + *[ResidualAttentionBlock(width, heads, attn_mask) \ + for _ in range(layers)]) + + def forward(self, x: torch.Tensor): + return self.resblocks(x) + +class CLIPTEXT(nn.Module): + def __init__(self, + embed_dim=512, + # text + context_length=77, + vocab_size=49408, + transformer_width=512, + transformer_heads=8, + transformer_layers=12 + ): + super().__init__() + + self._tokenizer = _Tokenizer() + self.context_length = context_length + + self.transformer = Transformer( + width=transformer_width, + layers=transformer_layers, + heads=transformer_heads, + attn_mask=self.build_attention_mask() + ) + + self.vocab_size = vocab_size + self.token_embedding = nn.Embedding(vocab_size, transformer_width) + self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width)) + self.ln_final = LayerNorm(transformer_width) + + self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim)) + # self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07)) + + self.initialize_parameters() + + def initialize_parameters(self): + nn.init.normal_(self.token_embedding.weight, std=0.02) + nn.init.normal_(self.positional_embedding, std=0.01) + + proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5) + attn_std = self.transformer.width ** -0.5 + fc_std = (2 * self.transformer.width) ** -0.5 + for block in self.transformer.resblocks: + nn.init.normal_(block.attn.in_proj_weight, std=attn_std) + nn.init.normal_(block.attn.out_proj.weight, std=proj_std) + nn.init.normal_(block.mlp.c_fc.weight, std=fc_std) + nn.init.normal_(block.mlp.c_proj.weight, std=proj_std) + + if self.text_projection is not None: + nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5) + + def build_attention_mask(self): + # lazily create causal attention mask, with full attention between the vision tokens + # pytorch uses additive attention mask; fill with -inf + mask = torch.empty(self.context_length, self.context_length) + mask.fill_(float("-inf")) + mask.triu_(1) # zero out the lower diagonal + return mask + + @property + def device(self): + return self.text_projection.device + + @property + def dtype(self): + return self.text_projection.dtype + + def tokenize(self, + texts: Union[str, List[str]], \ + context_length: int = 77) -> torch.LongTensor: + """ + """ + if isinstance(texts, str): + texts = [texts] + + sot_token = self._tokenizer.encoder["<|startoftext|>"] + eot_token = self._tokenizer.encoder["<|endoftext|>"] + all_tokens = [[sot_token] + self._tokenizer.encode(text) + [eot_token] for text in texts] + result = torch.zeros(len(all_tokens), context_length, dtype=torch.long) + + for i, tokens in enumerate(all_tokens): + if len(tokens) > context_length: + st = torch.randint( + len(tokens) - context_length + 1, (1,))[0].item() + tokens = tokens[st: st + context_length] + # raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}") + result[i, :len(tokens)] = torch.tensor(tokens) + + return result + + def encode_text(self, text): + x = self.token_embedding(text).type(self.dtype) # [batch_size, n_ctx, d_model] + x = x + self.positional_embedding.type(self.dtype) + x = x.permute(1, 0, 2) # NLD -> LND + x = self.transformer(x) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.ln_final(x).type(self.dtype) + # take features from the eot embedding (eot_token is the highest number in each sequence) + x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection + return x + + def forward(self, captions): + ''' + captions: list of strings + ''' + text = self.tokenize(captions).to(self.device) # B x L x D + features = self.encode_text(text) # B x D + return features + + +def build_text_encoder(pretrain=True, visual_type="RN50"): + clip_dict = { + "visual_type": ["embed_dim", "context_length", "vocab_size", + "transformer_width", "transformer_heads", "transformer_layers"], + "RN50": [1024, 77, 49408, 512, 8, 12], + "RN50x4": [640, 77, 49408, 640, 10, 12], + "RN50x16": [768, 77, 49408, 768, 12, 12], + "RN50x64": [1024, 77, 49408, 1024, 16, 12], + } + text_encoder = CLIPTEXT(**{k: v for k, v in zip(clip_dict['visual_type'], clip_dict[visual_type])}) + if pretrain: + import clip + if visual_type in clip_dict: + pretrained_model, _ = clip.load(visual_type, device='cpu') + else: + raise NotImplementedError + + state_dict = pretrained_model.state_dict() + to_delete_keys = ["logit_scale", "input_resolution", \ + "context_length", "vocab_size"] + \ + [k for k in state_dict.keys() if k.startswith('visual.')] + for k in to_delete_keys: + if k in state_dict: + del state_dict[k] + # print('Loading pretrained CLIP') + text_encoder.load_state_dict(state_dict) + return text_encoder diff --git a/vlpart/vlpart.py b/vlpart/vlpart.py new file mode 100644 index 0000000000000000000000000000000000000000..1c15ae7b591d908153e6a751b64babaca9fe96de --- /dev/null +++ b/vlpart/vlpart.py @@ -0,0 +1,185 @@ +import logging +import numpy as np +import itertools +from typing import Dict, List, Optional, Tuple +import torch +import torch.nn.functional as F +from torch import nn + +from detectron2.config import configurable +from detectron2.data.detection_utils import convert_image_to_rgb +from detectron2.layers import move_device_like, batched_nms +from detectron2.structures import ImageList, Boxes, Instances, BitMasks, ROIMasks + +from detectron2.modeling.backbone import Backbone, build_backbone +from detectron2.modeling.proposal_generator import build_proposal_generator +from detectron2.config import get_cfg + + +import clip +from vlpart.text_encoder import build_text_encoder +from vlpart.swintransformer import build_swinbase_fpn_backbone +from vlpart.vlpart_roi_heads import build_vlpart_roi_heads + + +def build_vlpart(checkpoint=None): + cfg = get_cfg() + cfg.merge_from_list(['MODEL.RPN.IN_FEATURES', ["p2", "p3", "p4", "p5", "p6"], + 'MODEL.ROI_HEADS.IN_FEATURES', ["p2", "p3", "p4", "p5"], + 'MODEL.ROI_BOX_CASCADE_HEAD.IOUS', [0.5, 0.6, 0.7], + 'MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG', True, + 'MODEL.ROI_BOX_HEAD.NAME', "FastRCNNConvFCHead", + 'MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION', 7, + 'MODEL.ROI_BOX_HEAD.NUM_FC', 2, + 'MODEL.ANCHOR_GENERATOR.SIZES', [[32], [64], [128], [256], [512]], + 'MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS', [[0.5, 1.0, 2.0]], + ]) + backbone = build_swinbase_fpn_backbone() + vlpart = VLPart( + backbone=backbone, + proposal_generator=build_proposal_generator(cfg, backbone.output_shape()), + roi_heads=build_vlpart_roi_heads(cfg, backbone.output_shape()), + ) + vlpart.eval() + if checkpoint is not None: + with open(checkpoint, "rb") as f: + state_dict = torch.load(f) + vlpart.load_state_dict(state_dict['model'], strict=False) + + return vlpart + + +class VLPart(nn.Module): + def __init__( + self, + backbone: Backbone, + proposal_generator: nn.Module, + roi_heads: nn.Module, + ): + super().__init__() + + self.backbone = backbone + self.proposal_generator = proposal_generator + self.roi_heads = roi_heads + self.text_encoder = build_text_encoder(pretrain=True, visual_type='RN50') + + self.register_buffer("pixel_mean", + torch.tensor([123.675, 116.280, 103.530]).view(-1, 1, 1), False) + self.register_buffer("pixel_std", + torch.tensor([58.395, 57.120, 57.375]).view(-1, 1, 1), False) + + @property + def device(self): + return self.pixel_mean.device + + def _move_to_current_device(self, x): + return move_device_like(x, self.pixel_mean) + + def get_text_embeddings(self, vocabulary, prefix_prompt='a '): + vocabulary = vocabulary.split('.') + texts = [prefix_prompt + x.lower().replace(':', ' ') for x in vocabulary] + texts_aug = texts + ['background'] + emb = self.text_encoder(texts_aug).permute(1, 0) + emb = F.normalize(emb, p=2, dim=0) + return emb + + def inference( + self, + batched_inputs: List[Dict[str, torch.Tensor]], + do_postprocess: bool = True, + text_prompt: str = 'dog', + ): + assert not self.training + + images = self.preprocess_image(batched_inputs) + features = self.backbone(images.tensor) + proposals, _ = self.proposal_generator(images, features) + text_embed = self.get_text_embeddings(text_prompt) + results, _ = self.roi_heads(images, features, proposals, text_embed) + if do_postprocess: + assert not torch.jit.is_scripting(), "Scripting is not supported for postprocess." + max_shape = images.tensor.shape[2:] + return VLPart._postprocess(results, batched_inputs, images.image_sizes, max_shape) + else: + return results + + + def preprocess_image(self, batched_inputs: List[Dict[str, torch.Tensor]]): + """ + Normalize, pad and batch the input images. + """ + original_images = [self._move_to_current_device(x["image"]) for x in batched_inputs] + images = [(x - self.pixel_mean) / self.pixel_std for x in original_images] + images = ImageList.from_tensors( + images, + self.backbone.size_divisibility, + padding_constraints=self.backbone.padding_constraints, + ) + return images + + @staticmethod + def _postprocess(instances, batched_inputs: List[Dict[str, torch.Tensor]], image_sizes, max_shape): + """ + Rescale the output instances to the target size. + """ + # note: private function; subject to changes + processed_results = [] + for results_per_image, input_per_image, image_size in zip( + instances, batched_inputs, image_sizes + ): + height = input_per_image.get("height", image_size[0]) + width = input_per_image.get("width", image_size[1]) + r = custom_detector_postprocess(results_per_image, height, width, max_shape) + processed_results.append({"instances": r}) + return processed_results + + +def custom_detector_postprocess( + results: Instances, output_height: int, output_width: int, + max_shape, mask_threshold: float = 0.5 +): + """ + detector_postprocess with support on global_masks + """ + if isinstance(output_width, torch.Tensor): + # This shape might (but not necessarily) be tensors during tracing. + # Converts integer tensors to float temporaries to ensure true + # division is performed when computing scale_x and scale_y. + output_width_tmp = output_width.float() + output_height_tmp = output_height.float() + new_size = torch.stack([output_height, output_width]) + else: + new_size = (output_height, output_width) + output_width_tmp = output_width + output_height_tmp = output_height + + scale_x, scale_y = ( + output_width_tmp / results.image_size[1], + output_height_tmp / results.image_size[0], + ) + + resized_h, resized_w = results.image_size + results = Instances(new_size, **results.get_fields()) + + if results.has("pred_boxes"): + output_boxes = results.pred_boxes + else: + output_boxes = None + assert output_boxes is not None, "Predictions must contain boxes!" + + output_boxes.scale(scale_x, scale_y) + output_boxes.clip(results.image_size) + + results = results[output_boxes.nonempty()] + + if results.has("pred_masks"): + if isinstance(results.pred_masks, ROIMasks): + roi_masks = results.pred_masks + else: + # pred_masks is a tensor of shape (N, 1, M, M) + roi_masks = ROIMasks(results.pred_masks[:, 0, :, :]) + results.pred_masks = roi_masks.to_bitmasks( + results.pred_boxes, output_height, output_width, mask_threshold + ).tensor # TODO return ROIMasks/BitMask object in the future + + return results diff --git a/vlpart/vlpart_fast_rcnn.py b/vlpart/vlpart_fast_rcnn.py new file mode 100644 index 0000000000000000000000000000000000000000..cc1ed58760521fb9877da8ce9ce78ecff866badd --- /dev/null +++ b/vlpart/vlpart_fast_rcnn.py @@ -0,0 +1,112 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import logging +from typing import Callable, Dict, List, Optional, Tuple, Union +import math +import copy +import numpy as np +import torch +import torch.distributed as dist +from torch import nn +from torch.nn import functional as F +from fvcore.nn import sigmoid_focal_loss_jit, giou_loss, smooth_l1_loss +import fvcore.nn.weight_init as weight_init + +from detectron2.config import configurable +from detectron2.data.detection_utils import get_fed_loss_cls_weights +from detectron2.layers import ShapeSpec, batched_nms, cat, cross_entropy, nonzero_tuple +from detectron2.modeling.box_regression import Box2BoxTransform, _dense_box_regression_loss +from detectron2.structures import Boxes, Instances, BitMasks, pairwise_iou, pairwise_ioa +from detectron2.utils.events import get_event_storage +from detectron2.modeling.roi_heads.fast_rcnn import FastRCNNOutputLayers + + +class TexeEmbedClassifier(nn.Module): + def __init__( + self, + input_shape: ShapeSpec, + zs_weight_dim: int = 1024, + norm_weight: bool = True, + norm_temperature: float = 50.0, + ): + super().__init__() + if isinstance(input_shape, int): # some backward compatibility + input_shape = ShapeSpec(channels=input_shape) + input_size = input_shape.channels * (input_shape.width or 1) * (input_shape.height or 1) + self.norm_weight = norm_weight + self.norm_temperature = norm_temperature + + self.linear = nn.Linear(input_size, zs_weight_dim) + + def forward(self, x, text_embed): + + x = self.linear(x) + if self.norm_weight: + x = self.norm_temperature * F.normalize(x, p=2, dim=1) + x = torch.mm(x, text_embed) + return x + + +class VLMFastRCNNOutputLayers(nn.Module): + def __init__( + self, + input_shape: ShapeSpec, + box2box_transform, + use_sigmoid_ce: bool = True, + test_score_thresh: float = 0.0, + test_nms_thresh: float = 0.5, + test_topk_per_image: int = 100, + ): + super().__init__() + if isinstance(input_shape, int): # some backward compatibility + input_shape = ShapeSpec(channels=input_shape) + + self.box2box_transform = box2box_transform + self.use_sigmoid_ce = use_sigmoid_ce + self.test_score_thresh = test_score_thresh + self.test_nms_thresh = test_nms_thresh + self.test_topk_per_image = test_topk_per_image + + input_size = input_shape.channels * \ + (input_shape.width or 1) * (input_shape.height or 1) + + # bbox_pred + self.bbox_pred = nn.Sequential( + nn.Linear(input_size, input_size), + nn.ReLU(inplace=True), + nn.Linear(input_size, 4) + ) + # cls_score + self.cls_score = TexeEmbedClassifier(input_shape) + + def forward(self, x, text_embed): + if x.dim() > 2: + x = torch.flatten(x, start_dim=1) + cls_scores = self.cls_score(x, text_embed) + proposal_deltas = self.bbox_pred(x) + + return cls_scores, proposal_deltas + + def predict_boxes(self, predictions, proposals): + if not len(proposals): + return [] + _, proposal_deltas = predictions + num_prop_per_image = [len(p) for p in proposals] + proposal_boxes = cat([p.proposal_boxes.tensor for p in proposals], dim=0) + + predict_boxes = self.box2box_transform.apply_deltas( + proposal_deltas, + proposal_boxes, + ) # Nx(KxB) + return predict_boxes.split(num_prop_per_image) + + def predict_probs(self, predictions, proposals): + cls_scores, _ = predictions + num_inst_per_image = [len(p) for p in proposals] + cls_scores = cls_scores.split(num_inst_per_image, dim=0) + + final_scores = [] + for cls_score in cls_scores: + final_score = cls_score.sigmoid() if self.use_sigmoid_ce else F.softmax(cls_score, dim=-1) + final_scores.append(final_score) + return final_scores + diff --git a/vlpart/vlpart_roi_heads.py b/vlpart/vlpart_roi_heads.py new file mode 100644 index 0000000000000000000000000000000000000000..ef2523658f93ec18a2920a46593abd4b24333f3e --- /dev/null +++ b/vlpart/vlpart_roi_heads.py @@ -0,0 +1,152 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +import copy +import inspect +import logging + +import numpy as np +from typing import Dict, List, Optional, Tuple, Union +import torch +from torch import nn +from torch.nn import functional as F + +from detectron2.config import configurable +from detectron2.layers import ShapeSpec, nonzero_tuple, batched_nms, cat +from detectron2.structures import Boxes, ImageList, Instances, pairwise_iou, pairwise_ioa +from detectron2.utils.events import get_event_storage +from detectron2.utils.registry import Registry + +from detectron2.modeling.backbone.resnet import BottleneckBlock, ResNet +from detectron2.modeling.matcher import Matcher +from detectron2.modeling.poolers import ROIPooler +from detectron2.modeling.sampling import subsample_labels + +from detectron2.modeling.box_regression import Box2BoxTransform +from detectron2.modeling.roi_heads.fast_rcnn import fast_rcnn_inference +from detectron2.modeling.roi_heads.box_head import build_box_head +from detectron2.modeling import ROI_HEADS_REGISTRY, StandardROIHeads +from detectron2.modeling.roi_heads.roi_heads import Res5ROIHeads +from detectron2.modeling.roi_heads.cascade_rcnn import CascadeROIHeads, _ScaleGradient + +from .vlpart_fast_rcnn import VLMFastRCNNOutputLayers + + +def build_vlpart_roi_heads(cfg, input_shape): + return CascadeVLMROIHeads(cfg, input_shape) + + +class CascadeVLMROIHeads(CascadeROIHeads): + + @classmethod + def _init_box_head(self, cfg, input_shape): + # fmt: off + in_features = cfg.MODEL.ROI_HEADS.IN_FEATURES + pooler_resolution = cfg.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION + pooler_scales = tuple(1.0 / input_shape[k].stride for k in in_features) + sampling_ratio = cfg.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO + pooler_type = cfg.MODEL.ROI_BOX_HEAD.POOLER_TYPE + cascade_bbox_reg_weights = cfg.MODEL.ROI_BOX_CASCADE_HEAD.BBOX_REG_WEIGHTS + cascade_ious = cfg.MODEL.ROI_BOX_CASCADE_HEAD.IOUS + assert len(cascade_bbox_reg_weights) == len(cascade_ious) + assert cfg.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG, \ + "CascadeROIHeads only support class-agnostic regression now!" + assert cascade_ious[0] == cfg.MODEL.ROI_HEADS.IOU_THRESHOLDS[0] + # fmt: on + + # If StandardROIHeads is applied on multiple feature maps (as in FPN), + # then we share the same predictors and therefore the channel counts must be the same + in_channels = [input_shape[f].channels for f in in_features] + # Check all channel counts are equal + assert len(set(in_channels)) == 1, in_channels + in_channels = in_channels[0] + + box_pooler = ROIPooler( + output_size=pooler_resolution, + scales=pooler_scales, + sampling_ratio=sampling_ratio, + pooler_type=pooler_type, + ) + pooled_shape = ShapeSpec( + channels=in_channels, height=pooler_resolution, width=pooler_resolution + ) + + box_heads, box_predictors, proposal_matchers = [], [], [] + for match_iou, bbox_reg_weights in zip(cascade_ious, cascade_bbox_reg_weights): + box_head = build_box_head(cfg, pooled_shape) + box_heads.append(box_head) + box_predictors.append( + VLMFastRCNNOutputLayers( + box_head.output_shape, + box2box_transform=Box2BoxTransform(weights=bbox_reg_weights), + ) + ) + proposal_matchers.append(Matcher([match_iou], [0, 1], allow_low_quality_matches=False)) + return { + "box_in_features": in_features, + "box_pooler": box_pooler, + "box_heads": box_heads, + "box_predictors": box_predictors, + "proposal_matchers": proposal_matchers, + } + + def forward(self, images, features, proposals, text_embed): + del images + assert not self.training, 'only support inference now' + pred_instances = self._forward_box( + features, proposals, text_embed=text_embed) + pred_instances = self.forward_with_given_boxes(features, pred_instances) + return pred_instances, {} + + def _forward_box(self, features, proposals, text_embed): + + features = [features[f] for f in self.box_in_features] + head_outputs = [] # (predictor, predictions, proposals) + prev_pred_boxes = None + image_sizes = [x.image_size for x in proposals] + + for k in range(self.num_cascade_stages): + if k > 0: + proposals = self._create_proposals_from_boxes( + prev_pred_boxes, image_sizes) + if self.training and ann_type in ['box', 'part']: + proposals = self._match_and_label_boxes( + proposals, k, targets) + predictions = self._run_stage(features, proposals, k, text_embed) + prev_pred_boxes = self.box_predictor[k].predict_boxes( + (predictions[0], predictions[1]), proposals) + head_outputs.append((self.box_predictor[k], predictions, proposals)) + + assert not self.training, 'only support inference now' + # Each is a list[Tensor] of length #image. Each tensor is Ri x (K+1) + scores_per_stage = [h[0].predict_probs(h[1], h[2]) for h in head_outputs] + scores = [ + sum(list(scores_per_image)) * (1.0 / self.num_cascade_stages) + for scores_per_image in zip(*scores_per_stage) + ] + predictor, predictions, proposals = head_outputs[-1] + boxes = predictor.predict_boxes((predictions[0], predictions[1]), proposals) + pred_instances, _ = fast_rcnn_inference( + boxes, + scores, + image_sizes, + predictor.test_score_thresh, + predictor.test_nms_thresh, + predictor.test_topk_per_image, + ) + return pred_instances + + def _create_proposals_from_boxes(self, boxes, image_sizes): + boxes = [Boxes(b.detach()) for b in boxes] + proposals = [] + for boxes_per_image, image_size in zip(boxes, image_sizes): + boxes_per_image.clip(image_size) + prop = Instances(image_size) + prop.proposal_boxes = boxes_per_image + proposals.append(prop) + return proposals + + def _run_stage(self, features, proposals, stage, text_embed): + pool_boxes = [x.proposal_boxes for x in proposals] + box_features = self.box_pooler(features, pool_boxes) + box_features = _ScaleGradient.apply(box_features, 1.0 / self.num_cascade_stages) + box_features = self.box_head[stage](box_features) + return self.box_predictor[stage](box_features, text_embed)