# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py from dataclasses import dataclass from typing import List, Optional, Tuple, Union import os import sys sys.path.append(os.path.split(sys.path[0])[0]) import math import json import torch import einops import torch.nn as nn import torch.utils.checkpoint from diffusers.configuration_utils import ConfigMixin, register_to_config from diffusers.utils import BaseOutput, logging from diffusers.models.embeddings import TimestepEmbedding, Timesteps from einops import rearrange try: from diffusers.models.modeling_utils import ModelMixin except: from diffusers.modeling_utils import ModelMixin # 0.11.1 try: from .unet_blocks import ( CrossAttnDownBlock3D, CrossAttnUpBlock3D, DownBlock3D, UNetMidBlock3DCrossAttn, UpBlock3D, get_down_block, get_up_block, ) from .resnet import InflatedConv3d except: from unet_blocks import ( CrossAttnDownBlock3D, CrossAttnUpBlock3D, DownBlock3D, UNetMidBlock3DCrossAttn, UpBlock3D, get_down_block, get_up_block, ) from resnet import InflatedConv3d from rotary_embedding_torch import RotaryEmbedding logger = logging.get_logger(__name__) # pylint: disable=invalid-name class RelativePositionBias(nn.Module): def __init__( self, heads=8, num_buckets=32, max_distance=128, ): super().__init__() self.num_buckets = num_buckets self.max_distance = max_distance self.relative_attention_bias = nn.Embedding(num_buckets, heads) @staticmethod def _relative_position_bucket(relative_position, num_buckets=32, max_distance=128): ret = 0 n = -relative_position num_buckets //= 2 ret += (n < 0).long() * num_buckets n = torch.abs(n) max_exact = num_buckets // 2 is_small = n < max_exact val_if_large = max_exact + ( torch.log(n.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact) ).long() val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1)) ret += torch.where(is_small, n, val_if_large) return ret def forward(self, n, device): q_pos = torch.arange(n, dtype = torch.long, device = device) k_pos = torch.arange(n, dtype = torch.long, device = device) rel_pos = einops.rearrange(k_pos, 'j -> 1 j') - einops.rearrange(q_pos, 'i -> i 1') rp_bucket = self._relative_position_bucket(rel_pos, num_buckets = self.num_buckets, max_distance = self.max_distance) values = self.relative_attention_bias(rp_bucket) return einops.rearrange(values, 'i j h -> h i j') # num_heads, num_frames, num_frames @dataclass class UNet3DConditionOutput(BaseOutput): sample: torch.FloatTensor class UNet3DConditionModel(ModelMixin, ConfigMixin): _supports_gradient_checkpointing = True @register_to_config def __init__( self, sample_size: Optional[int] = None, # 64 in_channels: int = 4, out_channels: int = 4, center_input_sample: bool = False, flip_sin_to_cos: bool = True, freq_shift: int = 0, down_block_types: Tuple[str] = ( "CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "DownBlock3D", ), mid_block_type: str = "UNetMidBlock3DCrossAttn", up_block_types: Tuple[str] = ( "UpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D" ), only_cross_attention: Union[bool, Tuple[bool]] = False, block_out_channels: Tuple[int] = (320, 640, 1280, 1280), layers_per_block: int = 2, downsample_padding: int = 1, mid_block_scale_factor: float = 1, act_fn: str = "silu", norm_num_groups: int = 32, norm_eps: float = 1e-5, cross_attention_dim: int = 1280, attention_head_dim: Union[int, Tuple[int]] = 8, dual_cross_attention: bool = False, use_linear_projection: bool = False, class_embed_type: Optional[str] = None, num_class_embeds: Optional[int] = None, upcast_attention: bool = False, resnet_time_scale_shift: str = "default", use_first_frame: bool = False, use_relative_position: bool = False, ): super().__init__() # print(use_first_frame) self.sample_size = sample_size time_embed_dim = block_out_channels[0] * 4 # input self.conv_in = InflatedConv3d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)) # time self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) timestep_input_dim = block_out_channels[0] self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim) # class embedding if class_embed_type is None and num_class_embeds is not None: self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) elif class_embed_type == "timestep": self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim) elif class_embed_type == "identity": self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) else: self.class_embedding = None self.down_blocks = nn.ModuleList([]) self.mid_block = None self.up_blocks = nn.ModuleList([]) if isinstance(only_cross_attention, bool): only_cross_attention = [only_cross_attention] * len(down_block_types) if isinstance(attention_head_dim, int): attention_head_dim = (attention_head_dim,) * len(down_block_types) rotary_emb = RotaryEmbedding(32) # down output_channel = block_out_channels[0] for i, down_block_type in enumerate(down_block_types): input_channel = output_channel output_channel = block_out_channels[i] is_final_block = i == len(block_out_channels) - 1 down_block = get_down_block( down_block_type, num_layers=layers_per_block, in_channels=input_channel, out_channels=output_channel, temb_channels=time_embed_dim, add_downsample=not is_final_block, resnet_eps=norm_eps, resnet_act_fn=act_fn, resnet_groups=norm_num_groups, cross_attention_dim=cross_attention_dim, attn_num_head_channels=attention_head_dim[i], downsample_padding=downsample_padding, dual_cross_attention=dual_cross_attention, use_linear_projection=use_linear_projection, only_cross_attention=only_cross_attention[i], upcast_attention=upcast_attention, resnet_time_scale_shift=resnet_time_scale_shift, use_first_frame=use_first_frame, use_relative_position=use_relative_position, rotary_emb=rotary_emb, ) self.down_blocks.append(down_block) # mid if mid_block_type == "UNetMidBlock3DCrossAttn": self.mid_block = UNetMidBlock3DCrossAttn( in_channels=block_out_channels[-1], temb_channels=time_embed_dim, resnet_eps=norm_eps, resnet_act_fn=act_fn, output_scale_factor=mid_block_scale_factor, resnet_time_scale_shift=resnet_time_scale_shift, cross_attention_dim=cross_attention_dim, attn_num_head_channels=attention_head_dim[-1], resnet_groups=norm_num_groups, dual_cross_attention=dual_cross_attention, use_linear_projection=use_linear_projection, upcast_attention=upcast_attention, use_first_frame=use_first_frame, use_relative_position=use_relative_position, rotary_emb=rotary_emb, ) else: raise ValueError(f"unknown mid_block_type : {mid_block_type}") # count how many layers upsample the videos self.num_upsamplers = 0 # up reversed_block_out_channels = list(reversed(block_out_channels)) reversed_attention_head_dim = list(reversed(attention_head_dim)) only_cross_attention = list(reversed(only_cross_attention)) output_channel = reversed_block_out_channels[0] for i, up_block_type in enumerate(up_block_types): is_final_block = i == len(block_out_channels) - 1 prev_output_channel = output_channel output_channel = reversed_block_out_channels[i] input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] # add upsample block for all BUT final layer if not is_final_block: add_upsample = True self.num_upsamplers += 1 else: add_upsample = False up_block = get_up_block( up_block_type, num_layers=layers_per_block + 1, in_channels=input_channel, out_channels=output_channel, prev_output_channel=prev_output_channel, temb_channels=time_embed_dim, add_upsample=add_upsample, resnet_eps=norm_eps, resnet_act_fn=act_fn, resnet_groups=norm_num_groups, cross_attention_dim=cross_attention_dim, attn_num_head_channels=reversed_attention_head_dim[i], dual_cross_attention=dual_cross_attention, use_linear_projection=use_linear_projection, only_cross_attention=only_cross_attention[i], upcast_attention=upcast_attention, resnet_time_scale_shift=resnet_time_scale_shift, use_first_frame=use_first_frame, use_relative_position=use_relative_position, rotary_emb=rotary_emb, ) self.up_blocks.append(up_block) prev_output_channel = output_channel # out self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps) self.conv_act = nn.SiLU() self.conv_out = InflatedConv3d(block_out_channels[0], out_channels, kernel_size=3, padding=1) # relative time positional embeddings self.use_relative_position = use_relative_position if self.use_relative_position: self.time_rel_pos_bias = RelativePositionBias(heads=8, max_distance=32) # realistically will not be able to generate that many frames of video... yet def set_attention_slice(self, slice_size): r""" Enable sliced attention computation. When this option is enabled, the attention module will split the input tensor in slices, to compute attention in several steps. This is useful to save some memory in exchange for a small speed decrease. Args: slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` must be a multiple of `slice_size`. """ sliceable_head_dims = [] def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module): if hasattr(module, "set_attention_slice"): sliceable_head_dims.append(module.sliceable_head_dim) for child in module.children(): fn_recursive_retrieve_slicable_dims(child) # retrieve number of attention layers for module in self.children(): fn_recursive_retrieve_slicable_dims(module) num_slicable_layers = len(sliceable_head_dims) if slice_size == "auto": # half the attention head size is usually a good trade-off between # speed and memory slice_size = [dim // 2 for dim in sliceable_head_dims] elif slice_size == "max": # make smallest slice possible slice_size = num_slicable_layers * [1] slice_size = num_slicable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size if len(slice_size) != len(sliceable_head_dims): raise ValueError( f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." ) for i in range(len(slice_size)): size = slice_size[i] dim = sliceable_head_dims[i] if size is not None and size > dim: raise ValueError(f"size {size} has to be smaller or equal to {dim}.") # Recursively walk through all the children. # Any children which exposes the set_attention_slice method # gets the message def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): if hasattr(module, "set_attention_slice"): module.set_attention_slice(slice_size.pop()) for child in module.children(): fn_recursive_set_attention_slice(child, slice_size) reversed_slice_size = list(reversed(slice_size)) for module in self.children(): fn_recursive_set_attention_slice(module, reversed_slice_size) def _set_gradient_checkpointing(self, module, value=False): if isinstance(module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)): module.gradient_checkpointing = value def forward( self, sample: torch.FloatTensor, timestep: Union[torch.Tensor, float, int], encoder_hidden_states: torch.Tensor = None, class_labels: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, use_image_num: int = 0, return_dict: bool = True, ip_hidden_states = None, encoder_temporal_hidden_states = None ) -> Union[UNet3DConditionOutput, Tuple]: r""" Args: sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. Returns: [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor. """ # By default samples have to be AT least a multiple of the overall upsampling factor. # The overall upsampling factor is equal to 2 ** (# num of upsampling layears). # However, the upsampling interpolation output size can be forced to fit any upsampling size # on the fly if necessary. if ip_hidden_states is not None: b = ip_hidden_states.shape[0] ip_hidden_states = rearrange(ip_hidden_states, 'b n c -> (b n) c') ip_hidden_states = self.image_proj_model(ip_hidden_states) ip_hidden_states = rearrange(ip_hidden_states, '(b n) m c -> b n m c', b=b) default_overall_up_factor = 2**self.num_upsamplers # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` forward_upsample_size = False upsample_size = None if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]): logger.info("Forward upsample size to force interpolation output size.") forward_upsample_size = True # 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) # center input if necessary if self.config.center_input_sample: sample = 2 * sample - 1.0 # time timesteps = timestep if not torch.is_tensor(timesteps): # 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=self.dtype) emb = self.time_embedding(t_emb) 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) # print(emb.shape) # torch.Size([3, 1280]) # print(class_emb.shape) # torch.Size([3, 1280]) emb = emb + class_emb if self.use_relative_position: frame_rel_pos_bias = self.time_rel_pos_bias(sample.shape[2], device=sample.device) else: frame_rel_pos_bias = None # pre-process sample = self.conv_in(sample) # 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, use_image_num=use_image_num, ip_hidden_states=ip_hidden_states, encoder_temporal_hidden_states=encoder_temporal_hidden_states ) else: sample, res_samples = downsample_block(hidden_states=sample, temb=emb) down_block_res_samples += res_samples # mid sample = self.mid_block( sample, emb, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, use_image_num=use_image_num, ip_hidden_states=ip_hidden_states, encoder_temporal_hidden_states=encoder_temporal_hidden_states ) # up for i, upsample_block in enumerate(self.up_blocks): is_final_block = i == len(self.up_blocks) - 1 res_samples = down_block_res_samples[-len(upsample_block.resnets) :] down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] # if we have not reached the final block and need to forward the # upsample size, we do it here if not is_final_block and forward_upsample_size: upsample_size = down_block_res_samples[-1].shape[2:] if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: sample = upsample_block( hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, encoder_hidden_states=encoder_hidden_states, upsample_size=upsample_size, attention_mask=attention_mask, use_image_num=use_image_num, ip_hidden_states=ip_hidden_states, encoder_temporal_hidden_states=encoder_temporal_hidden_states ) else: sample = upsample_block( hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size ) # post-process sample = self.conv_norm_out(sample) sample = self.conv_act(sample) sample = self.conv_out(sample) # print(sample.shape) if not return_dict: return (sample,) sample = UNet3DConditionOutput(sample=sample) return sample def forward_with_cfg(self, x, t, encoder_hidden_states = None, class_labels: Optional[torch.Tensor] = None, cfg_scale=4.0, use_fp16=False, ip_hidden_states = None): """ Forward pass of DiT, but also batches the unconditional forward pass for classifier-free guidance. """ # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb half = x[: len(x) // 2] combined = torch.cat([half, half], dim=0) if use_fp16: combined = combined.to(dtype=torch.float16) model_out = self.forward(combined, t, encoder_hidden_states, class_labels, ip_hidden_states=ip_hidden_states).sample # For exact reproducibility reasons, we apply classifier-free guidance on only # three channels by default. The standard approach to cfg applies it to all channels. # This can be done by uncommenting the following line and commenting-out the line following that. eps, rest = model_out[:, :4], model_out[:, 4:] # eps, rest = model_out[:, :3], model_out[:, 3:] # b c f h w cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0) half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps) eps = torch.cat([half_eps, half_eps], dim=0) return torch.cat([eps, rest], dim=1) @classmethod def from_pretrained_2d(cls, pretrained_model_path, subfolder=None, use_concat=False): if subfolder is not None: pretrained_model_path = os.path.join(pretrained_model_path, subfolder) # the content of the config file # { # "_class_name": "UNet2DConditionModel", # "_diffusers_version": "0.2.2", # "act_fn": "silu", # "attention_head_dim": 8, # "block_out_channels": [ # 320, # 640, # 1280, # 1280 # ], # "center_input_sample": false, # "cross_attention_dim": 768, # "down_block_types": [ # "CrossAttnDownBlock2D", # "CrossAttnDownBlock2D", # "CrossAttnDownBlock2D", # "DownBlock2D" # ], # "downsample_padding": 1, # "flip_sin_to_cos": true, # "freq_shift": 0, # "in_channels": 4, # "layers_per_block": 2, # "mid_block_scale_factor": 1, # "norm_eps": 1e-05, # "norm_num_groups": 32, # "out_channels": 4, # "sample_size": 64, # "up_block_types": [ # "UpBlock2D", # "CrossAttnUpBlock2D", # "CrossAttnUpBlock2D", # "CrossAttnUpBlock2D" # ] # } config_file = os.path.join(pretrained_model_path, 'config.json') if not os.path.isfile(config_file): raise RuntimeError(f"{config_file} does not exist") with open(config_file, "r") as f: config = json.load(f) config["_class_name"] = cls.__name__ config["down_block_types"] = [ "CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "DownBlock3D" ] config["up_block_types"] = [ "UpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D" ] # config["use_first_frame"] = True config["use_first_frame"] = False if use_concat: config["in_channels"] = 9 # config["use_relative_position"] = True # # tmp # config["class_embed_type"] = "timestep" # config["num_class_embeds"] = 100 from diffusers.utils import WEIGHTS_NAME # diffusion_pytorch_model.bin # {'_class_name': 'UNet3DConditionModel', # '_diffusers_version': '0.2.2', # 'act_fn': 'silu', # 'attention_head_dim': 8, # 'block_out_channels': [320, 640, 1280, 1280], # 'center_input_sample': False, # 'cross_attention_dim': 768, # 'down_block_types': # ['CrossAttnDownBlock3D', # 'CrossAttnDownBlock3D', # 'CrossAttnDownBlock3D', # 'DownBlock3D'], # 'downsample_padding': 1, # 'flip_sin_to_cos': True, # 'freq_shift': 0, # 'in_channels': 4, # 'layers_per_block': 2, # 'mid_block_scale_factor': 1, # 'norm_eps': 1e-05, # 'norm_num_groups': 32, # 'out_channels': 4, # 'sample_size': 64, # 'up_block_types': # ['UpBlock3D', # 'CrossAttnUpBlock3D', # 'CrossAttnUpBlock3D', # 'CrossAttnUpBlock3D']} model = cls.from_config(config) model_file = os.path.join(pretrained_model_path, WEIGHTS_NAME) if not os.path.isfile(model_file): raise RuntimeError(f"{model_file} does not exist") state_dict = torch.load(model_file, map_location="cpu") if use_concat: new_state_dict = {} conv_in_weight = state_dict["conv_in.weight"] new_conv_weight = torch.zeros((conv_in_weight.shape[0], 9, *conv_in_weight.shape[2:]), dtype=conv_in_weight.dtype) for i, j in zip([0, 1, 2, 3], [0, 1, 2, 3, 4, 5, 6, 7, 8]): new_conv_weight[:, j] = conv_in_weight[:, i] new_state_dict["conv_in.weight"] = new_conv_weight new_state_dict["conv_in.bias"] = state_dict["conv_in.bias"] for k, v in model.state_dict().items(): # print(k) if '_temp.' in k: new_state_dict.update({k: v}) if 'attn_fcross' in k: # conpy parms of attn1 to attn_fcross k = k.replace('attn_fcross', 'attn1') state_dict.update({k: state_dict[k]}) if 'norm_fcross' in k: k = k.replace('norm_fcross', 'norm1') state_dict.update({k: state_dict[k]}) if 'conv_in' in k: continue else: new_state_dict[k] = v # # tmp # if 'class_embedding' in k: # state_dict.update({k: v}) # breakpoint() model.load_state_dict(new_state_dict) else: for k, v in model.state_dict().items(): # print(k) if '_temp' in k: state_dict.update({k: v}) if 'attn_fcross' in k: # conpy parms of attn1 to attn_fcross k = k.replace('attn_fcross', 'attn1') state_dict.update({k: state_dict[k]}) if 'norm_fcross' in k: k = k.replace('norm_fcross', 'norm1') state_dict.update({k: state_dict[k]}) model.load_state_dict(state_dict) return model