from dataclasses import dataclass from typing import List, Optional, Tuple, Union import torch import numpy as np import torch.nn.functional as F from torch import nn import torchvision # from torch_utils.ops import grid_sample_gradfix from diffusers.configuration_utils import ConfigMixin, register_to_config from diffusers.models.modeling_utils import ModelMixin from diffusers.utils import BaseOutput from diffusers.utils.import_utils import is_xformers_available from diffusers.models.attention import FeedForward # from diffusers.models.attention_processor import Attention try: from .diffusers_attention import CrossAttention from .resnet import Downsample3D, Upsample3D, InflatedConv3d, ResnetBlock3D, ResnetBlock3DCNN except: from diffusers_attention import CrossAttention from resnet import Downsample3D, Upsample3D, InflatedConv3d, ResnetBlock3D, ResnetBlock3DCNN from einops import rearrange, repeat import math import pdb 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 grid_sample_align(input, grid): return torch.nn.functional.grid_sample(input=input, grid=grid, mode='bilinear', padding_mode='zeros', align_corners=True) @dataclass class TemporalTransformer3DModelOutput(BaseOutput): sample: torch.FloatTensor if is_xformers_available(): import xformers import xformers.ops else: xformers = None class EmptyTemporalModule3D(nn.Module): def __init__(self): super().__init__() def forward(self, hidden_states, condition_video=None, encoder_hidden_states=None, timesteps=None, temb=None, attention_mask=None): return hidden_states class TemporalModule3D(nn.Module): def __init__( self, in_channels=None, out_channels=None, num_attention_layers=None, num_attention_head=8, attention_head_dim=None, cross_attention_dim=768, temb_channels=512, dropout=0., attention_bias=False, activation_fn="geglu", only_cross_attention=False, upcast_attention=False, norm_num_groups=8, use_linear_projection=True, use_scale_shift=False, # set True always produce nan loss, I don't know why attention_block_types: Tuple[str]=None, cross_frame_attention_mode=None, temporal_shift_fold_div=None, temporal_shift_direction=None, use_dcn_warpping=None, use_deformable_conv=None, attention_dim_div: int = None, video_condition=False, ): super().__init__() assert len(attention_block_types) == 2 self.use_scale_shift = use_scale_shift self.video_condition = video_condition self.non_linearity = nn.SiLU() # 1. 3d cnn if self.video_condition: video_condition_dim = int(out_channels//4) self.v_cond_conv = ResnetBlock3D(in_channels=3, out_channels=video_condition_dim, temb_channels=temb_channels, groups=3, groups_out=32) self.resblocks_3d_t = ResnetBlock3DCNN(in_channels=in_channels+video_condition_dim, out_channels=in_channels, kernel=(5,1,1), temb_channels=temb_channels) else: self.resblocks_3d_t = ResnetBlock3DCNN(in_channels=in_channels, out_channels=in_channels, kernel=(5,1,1), temb_channels=temb_channels) self.resblocks_3d_s = ResnetBlock3D(in_channels=in_channels, out_channels=in_channels, temb_channels=temb_channels, groups=32, groups_out=32) # 2. transformer blocks if not (attention_block_types[0]=='' and attention_block_types[1]==''): attentions = TemporalTransformer3DModel( num_attention_heads=num_attention_head, attention_head_dim=attention_head_dim if attention_head_dim is not None else in_channels // num_attention_head // attention_dim_div, in_channels=in_channels, num_layers=num_attention_layers, dropout=dropout, norm_num_groups=norm_num_groups, cross_attention_dim=cross_attention_dim, attention_bias=attention_bias, activation_fn=activation_fn, num_embeds_ada_norm=1000, # adaptive norm for timestep embedding injection use_linear_projection=use_linear_projection, only_cross_attention=only_cross_attention, upcast_attention=upcast_attention, attention_block_types=attention_block_types, cross_frame_attention_mode=cross_frame_attention_mode, temporal_shift_fold_div=temporal_shift_fold_div, temporal_shift_direction=temporal_shift_direction, use_dcn_warpping=use_dcn_warpping, use_deformable_conv=use_deformable_conv, ) self.attentions = nn.ModuleList([attentions]) if use_scale_shift: self.scale_shift_conv = zero_module(InflatedConv3d(in_channels=in_channels, out_channels=in_channels * 2, kernel_size=1, stride=1, padding=0)) else: self.shift_conv = zero_module(InflatedConv3d(in_channels=in_channels, out_channels=in_channels, kernel_size=1, stride=1, padding=0)) def forward(self, hidden_states, condition_video=None, encoder_hidden_states=None, timesteps=None, temb=None, attention_mask=None): input_tensor = hidden_states if self.video_condition: # obtain video attention assert condition_video is not None if isinstance(condition_video, dict): condition_video = condition_video[hidden_states.shape[-1]] hidden_condition = self.v_cond_conv(condition_video, temb) hidden_states = torch.cat([hidden_states, hidden_condition], dim=1) # 3DCNN hidden_states = self.resblocks_3d_t(hidden_states, temb) hidden_states = self.resblocks_3d_s(hidden_states, temb) if hasattr(self, "attentions"): for attn in self.attentions: hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states, timestep=timesteps).sample if self.use_scale_shift: hidden_states = self.scale_shift_conv(hidden_states) scale, shift = torch.chunk(hidden_states, chunks=2, dim=1) hidden_states = (1 + scale) * input_tensor + shift else: hidden_states = self.shift_conv(hidden_states) hidden_states = input_tensor + hidden_states return hidden_states class TemporalTransformer3DModel(ModelMixin, ConfigMixin): @register_to_config def __init__( self, num_attention_heads=None, attention_head_dim=None, in_channels=None, num_layers=None, dropout=None, norm_num_groups=None, cross_attention_dim=None, attention_bias=None, activation_fn=None, num_embeds_ada_norm=None, use_linear_projection=None, only_cross_attention=None, upcast_attention=None, attention_block_types=None, cross_frame_attention_mode=None, temporal_shift_fold_div=None, temporal_shift_direction=None, use_dcn_warpping=None, use_deformable_conv=None, ): super().__init__() self.use_linear_projection = use_linear_projection self.num_attention_heads = num_attention_heads self.attention_head_dim = attention_head_dim inner_dim = num_attention_heads * attention_head_dim # Define input layers self.in_channels = in_channels self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) if use_linear_projection: self.proj_in = nn.Linear(in_channels, inner_dim) else: self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) # Define transformers blocks self.transformer_blocks = nn.ModuleList( [ TemporalTransformerBlock( inner_dim, num_attention_heads, attention_head_dim, dropout=dropout, cross_attention_dim=cross_attention_dim, activation_fn=activation_fn, num_embeds_ada_norm=num_embeds_ada_norm, attention_bias=attention_bias, only_cross_attention=only_cross_attention, upcast_attention=upcast_attention, attention_block_types=attention_block_types, cross_frame_attention_mode=cross_frame_attention_mode, temporal_shift_fold_div=temporal_shift_fold_div, temporal_shift_direction=temporal_shift_direction, use_dcn_warpping=use_dcn_warpping, use_deformable_conv=use_deformable_conv, ) for d in range(num_layers) ] ) # 4. Define output layers if use_linear_projection: self.proj_out = nn.Linear(inner_dim, in_channels) else: self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True): # Input assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}." video_length = hidden_states.shape[2] hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w") if encoder_hidden_states is not None: encoder_hidden_states = repeat(encoder_hidden_states, 'b n c -> (b f) n c', f=video_length) batch, channel, height, weight = hidden_states.shape residual = hidden_states hidden_states = self.norm(hidden_states) if not self.use_linear_projection: hidden_states = self.proj_in(hidden_states) inner_dim = hidden_states.shape[1] hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) else: inner_dim = hidden_states.shape[1] hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim) hidden_states = self.proj_in(hidden_states) # Blocks for block in self.transformer_blocks: hidden_states = block( hidden_states, encoder_hidden_states=encoder_hidden_states, timestep=timestep, video_length=video_length ) # Output if not self.use_linear_projection: hidden_states = ( hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous() ) hidden_states = self.proj_out(hidden_states) else: hidden_states = self.proj_out(hidden_states) hidden_states = ( hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous() ) output = hidden_states + residual output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length) if not return_dict: return (output,) return TemporalTransformer3DModelOutput(sample=output) class TemporalTransformerBlock(nn.Module): def __init__( self, dim=None, num_attention_heads=None, attention_head_dim=None, dropout=None, cross_attention_dim=None, activation_fn=None, num_embeds_ada_norm=None, attention_bias=None, only_cross_attention=None, upcast_attention=None, attention_block_types=None, cross_frame_attention_mode=None, temporal_shift_fold_div=None, temporal_shift_direction=None, use_dcn_warpping=None, use_deformable_conv=None, ): super().__init__() assert len(attention_block_types) == 2 self.only_cross_attention = only_cross_attention self.use_ada_layer_norm = num_embeds_ada_norm is not None self.use_dcn_warpping = use_dcn_warpping # 1. Spatial-Attn (self) if not attention_block_types[0] == '': self.attn_spatial = VersatileSelfAttention( attention_mode=attention_block_types[0], query_dim=dim, heads=num_attention_heads, dim_head=attention_head_dim, dropout=dropout, bias=attention_bias, upcast_attention=upcast_attention, cross_frame_attention_mode=cross_frame_attention_mode, temporal_shift_fold_div=temporal_shift_fold_div, temporal_shift_direction=temporal_shift_direction, ) nn.init.zeros_(self.attn_spatial.to_out[0].weight.data) self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim) # 2. Temporal-Attn (self) self.attn_temporal = VersatileSelfAttention( attention_mode=attention_block_types[1], query_dim=dim, heads=num_attention_heads, dim_head=attention_head_dim, dropout=dropout, bias=attention_bias, upcast_attention=upcast_attention, cross_frame_attention_mode=cross_frame_attention_mode, temporal_shift_fold_div=temporal_shift_fold_div, temporal_shift_direction=temporal_shift_direction, ) nn.init.zeros_(self.attn_temporal.to_out[0].weight.data) self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim) self.dcn_module = WarpModule( in_channels=dim, use_deformable_conv=use_deformable_conv, ) if use_dcn_warpping else None # 3. Feed-forward self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn) self.norm3 = nn.LayerNorm(dim) def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool, attention_op: None): if not is_xformers_available(): print("Here is how to install it") raise ModuleNotFoundError( "Refer to https://github.com/facebookresearch/xformers for more information on how to install" " xformers", name="xformers", ) elif not torch.cuda.is_available(): raise ValueError( "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only" " available for GPU " ) else: try: # Make sure we can run the memory efficient attention _ = xformers.ops.memory_efficient_attention( torch.randn((1, 2, 40), device="cuda"), torch.randn((1, 2, 40), device="cuda"), torch.randn((1, 2, 40), device="cuda"), ) except Exception as e: raise e if hasattr(self, "attn_spatial"): self.attn_spatial._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None, video_length=None): # 1. Spatial-Attention if hasattr(self, "attn_spatial") and hasattr(self, "norm1"): norm_hidden_states = self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states) hidden_states = self.attn_spatial(norm_hidden_states, attention_mask=attention_mask, video_length=video_length) + hidden_states # 2. Temporal-Attention norm_hidden_states = self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states) if not self.use_dcn_warpping: hidden_states = self.attn_temporal(norm_hidden_states, attention_mask=attention_mask, video_length=video_length) + hidden_states else: hidden_states = self.dcn_module( hidden_states, offset_hidden_states=self.attn_temporal(norm_hidden_states, attention_mask=attention_mask, video_length=video_length), ) # 3. Feed-forward hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states return hidden_states class VersatileSelfAttention(CrossAttention): def __init__( self, attention_mode=None, cross_frame_attention_mode=None, temporal_shift_fold_div=None, temporal_shift_direction=None, temporal_position_encoding=False, temporal_position_encoding_max_len=24, *args, **kwargs ): super().__init__(*args, **kwargs) assert attention_mode in ("Temporal", "Spatial", "CrossFrame", "SpatialTemporalShift", None) assert cross_frame_attention_mode in ("0_i-1", "i-1_i", "0_i-1_i", "i-1_i_i+1", None) self.attention_mode = attention_mode self.cross_frame_attention_mode = cross_frame_attention_mode self.temporal_shift_fold_div = temporal_shift_fold_div self.temporal_shift_direction = temporal_shift_direction self.pos_encoder = PositionalEncoding( kwargs["query_dim"], dropout=0., max_len=temporal_position_encoding_max_len ) if temporal_position_encoding else None def temporal_token_concat(self, tensor, video_length): # print("### temporal token concat") current_frame_index = torch.arange(video_length) former_frame_index = current_frame_index - 1 former_frame_index[0] = 0 later_frame_index = current_frame_index + 1 later_frame_index[-1] = -1 # (b f) d c tensor = rearrange(tensor, "(b f) d c -> b f d c", f=video_length) if self.cross_frame_attention_mode == "0_i-1": tensor = torch.cat([tensor[:, [0] * video_length], tensor[:, former_frame_index]], dim=2) elif self.cross_frame_attention_mode == "i-1_i": tensor = torch.cat([tensor[:, former_frame_index], tensor[:, current_frame_index]], dim=2) elif self.cross_frame_attention_mode == "0_i-1_i": tensor = torch.cat([tensor[:, [0] * video_length], tensor[:, former_frame_index], tensor[:, current_frame_index]], dim=2) elif self.cross_frame_attention_mode == "i-1_i_i+1": tensor = torch.cat([tensor[:, former_frame_index], tensor[:, current_frame_index], tensor[:, later_frame_index]], dim=2) else: raise NotImplementedError tensor = rearrange(tensor, "b f d c -> (b f) d c") return tensor def temporal_shift(self, tensor, video_length): # print("### temporal shift") # (b f) d c tensor = rearrange(tensor, "(b f) d c -> b f d c", f=video_length) n_channels = tensor.shape[-1] fold = n_channels // self.temporal_shift_fold_div if self.temporal_shift_direction != "right": raise NotImplementedError tensor_out = torch.zeros_like(tensor) tensor_out[:, 1:, :, :fold] = tensor[:, :-1, :, :fold] tensor_out[:, :, :, fold:] = tensor[:, :, :, fold:] tensor_out = rearrange(tensor_out, "b f d c -> (b f) d c") return tensor_out def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, video_length=None): # pdb.set_trace() batch_size, sequence_length, _ = hidden_states.shape assert encoder_hidden_states is None # (b f) d c if self.attention_mode == "Temporal": # print("### temporal reshape") d = hidden_states.shape[1] hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) f c", f=video_length) if self.pos_encoder is not None: hidden_states = self.pos_encoder(hidden_states) encoder_hidden_states = encoder_hidden_states if self.group_norm is not None: hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = self.to_q(hidden_states) dim = query.shape[-1] query = self.reshape_heads_to_batch_dim(query) if self.added_kv_proj_dim is not None: raise NotImplementedError encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states key = self.to_k(encoder_hidden_states) value = self.to_v(encoder_hidden_states) if self.attention_mode == "SpatialTemporalShift": key = self.temporal_shift(key, video_length=video_length) value = self.temporal_shift(value, video_length=video_length) elif self.attention_mode == "CrossFrame": key = self.temporal_token_concat(key, video_length=video_length) value = self.temporal_token_concat(value, video_length=video_length) key = self.reshape_heads_to_batch_dim(key) value = self.reshape_heads_to_batch_dim(value) if attention_mask is not None: if attention_mask.shape[-1] != query.shape[1]: target_length = query.shape[1] attention_mask = F.pad(attention_mask, (0, target_length), value=0.0) attention_mask = attention_mask.repeat_interleave(self.heads, dim=0) # attention, what we cannot get enough of if self._use_memory_efficient_attention_xformers: hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask) # Some versions of xformers return output in fp32, cast it back to the dtype of the input hidden_states = hidden_states.to(query.dtype) else: if self._slice_size is None or query.shape[0] // self._slice_size == 1: hidden_states = self._attention(query, key, value, attention_mask) else: hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask) # linear proj hidden_states = self.to_out[0](hidden_states) # dropout hidden_states = self.to_out[1](hidden_states) if self.attention_mode == "Temporal": hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d) return hidden_states class WarpModule(nn.Module): def __init__( self, in_channels=None, use_deformable_conv=None, ): super().__init__() self.use_deformable_conv = use_deformable_conv self.conv = None self.dcn_weight = None if use_deformable_conv: self.conv = nn.Conv2d(in_channels*2, 27, kernel_size=3, stride=1, padding=1) self.dcn_weight = nn.Parameter(torch.randn(in_channels, in_channels, 3, 3) / np.sqrt(in_channels * 3 * 3)) self.alpha = nn.Parameter(torch.zeros(1, in_channels, 1, 1)) else: self.conv = zero_module(nn.Conv2d(in_channels, 2, kernel_size=3, stride=1, padding=1)) def forward(self, hidden_states, offset_hidden_states): # (b f) d c spatial_dim = hidden_states.shape[1] size = int(spatial_dim ** 0.5) assert size ** 2 == spatial_dim hidden_states = rearrange(hidden_states, "b (h w) c -> b c h w", h=size) offset_hidden_states = rearrange(offset_hidden_states, "b (h w) c -> b c h w", h=size) concat_hidden_states = torch.cat([hidden_states, offset_hidden_states], dim=1) input_tensor = hidden_states if self.use_deformable_conv: offset_x, offset_y, offsets_mask = torch.chunk(self.conv(concat_hidden_states), chunks=3, dim=1) offsets_mask = offsets_mask.sigmoid() * 2 offsets = torch.cat([offset_x, offset_y], dim=1) hidden_states = torchvision.ops.deform_conv2d( hidden_states, offset=offsets, weight=self.dcn_weight, mask=offsets_mask, padding=1, ) hidden_states = self.alpha * hidden_states + input_tensor else: offsets = self.conv(concat_hidden_states) hidden_states = self.optical_flow_warping(hidden_states, offsets) hidden_states = rearrange(hidden_states, "b c h w -> b (h w) c") return hidden_states @staticmethod def optical_flow_warping(x, flo): """ warp an image/tensor (im2) back to im1, according to the optical flow x: [B, C, H, W] (im2) flo: [B, 2, H, W] flow pad_mode (optional): ref to https://pytorch.org/docs/stable/nn.functional.html#grid-sample "zeros": use 0 for out-of-bound grid locations, "border": use border values for out-of-bound grid locations """ dtype = x.dtype if dtype != torch.float32: x = x.to(torch.float32) B, C, H, W = x.size() # mesh grid xx = torch.arange(0, W).view(1, -1).repeat(H, 1) yy = torch.arange(0, H).view(-1, 1).repeat(1, W) xx = xx.view(1, 1, H, W).repeat(B, 1, 1, 1) yy = yy.view(1, 1, H, W).repeat(B, 1, 1, 1) grid = torch.cat((xx, yy), 1).float().to(flo.device) vgrid = grid + flo # scale grid to [-1,1] vgrid[:, 0, :, :] = 2.0 * vgrid[:, 0, :, :].clone() / max(W - 1, 1) - 1.0 vgrid[:, 1, :, :] = 2.0 * vgrid[:, 1, :, :].clone() / max(H - 1, 1) - 1.0 vgrid = vgrid.permute(0, 2, 3, 1) # output = grid_sample_gradfix.grid_sample_align(x, vgrid) output = grid_sample_align(x, vgrid) #output = torch.nn.functional.grid_sample(x, vgrid, padding_mode='zeros', mode='bilinear', align_corners=True) mask = torch.ones_like(x) # mask = grid_sample_gradfix.grid_sample_align(mask, vgrid) mask = grid_sample_align(x, vgrid) #mask = torch.nn.functional.grid_sample(mask, vgrid, padding_mode='zeros', mode='bilinear', align_corners=True) mask[mask < 0.9999] = 0 mask[mask > 0] = 1 results = output * mask if dtype != torch.float32: results = results.to(dtype) return results class AdaLayerNorm(nn.Module): """ Norm layer modified to incorporate timestep embeddings. """ def __init__(self, embedding_dim, num_embeddings): super().__init__() self.emb = nn.Embedding(num_embeddings, embedding_dim) self.silu = nn.SiLU() self.linear = nn.Linear(embedding_dim, embedding_dim * 2) self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False) def forward(self, x, timestep): timestep = repeat(timestep, "b -> (b r)", r=x.shape[0] // timestep.shape[0]) emb = self.linear(self.silu(self.emb(timestep))).unsqueeze(1) # (b f) 1 2d scale, shift = torch.chunk(emb, 2, dim=-1) x = self.norm(x) * (1 + scale) + shift return x