Spaces:
Build error
Build error
| from __future__ import annotations | |
| from typing import Callable, Iterable, Union | |
| import torch | |
| from einops import rearrange | |
| from ...util import partialclass | |
| from ..diffusionmodules.model import Decoder, ResnetBlock | |
| from ..diffusionmodules.openaimodel import ResBlock | |
| class VideoResBlock(ResnetBlock): | |
| def __init__( | |
| self, | |
| out_channels, | |
| *args, | |
| dropout=0.0, | |
| video_kernel_size=3, | |
| alpha=0.0, | |
| merge_strategy="learned", | |
| **kwargs | |
| ): | |
| super().__init__(out_channels=out_channels, dropout=dropout, *args, **kwargs) | |
| if video_kernel_size is None: | |
| video_kernel_size = [3, 1, 1] | |
| self.time_stack = ResBlock( | |
| channels=out_channels, | |
| emb_channels=0, | |
| dropout=dropout, | |
| dims=3, | |
| use_scale_shift_norm=False, | |
| use_conv=False, | |
| up=False, | |
| down=False, | |
| kernel_size=video_kernel_size, | |
| use_checkpoint=False, | |
| skip_t_emb=True | |
| ) | |
| self.merge_strategy = merge_strategy | |
| if self.merge_strategy == "fixed": | |
| self.register_buffer("mix_factor", torch.Tensor([alpha])) | |
| elif self.merge_strategy == "learned": | |
| self.register_parameter("mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))) | |
| else: | |
| raise ValueError(f"Unknown merge strategy {self.merge_strategy}") | |
| def get_alpha(self): | |
| if self.merge_strategy == "fixed": | |
| return self.mix_factor | |
| elif self.merge_strategy == "learned": | |
| return torch.sigmoid(self.mix_factor) | |
| else: | |
| raise NotImplementedError | |
| def forward(self, x, temb, skip_video=False, timesteps=None): | |
| if timesteps is None: | |
| timesteps = self.timesteps | |
| x = super().forward(x, temb) | |
| if not skip_video: | |
| x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) | |
| x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) | |
| x = self.time_stack(x, temb) | |
| alpha = self.get_alpha() | |
| x = alpha * x + (1.0 - alpha) * x_mix | |
| x = rearrange(x, "b c t h w -> (b t) c h w") | |
| return x | |
| class AE3DConv(torch.nn.Conv2d): | |
| def __init__(self, in_channels, out_channels, video_kernel_size=3, *args, **kwargs): | |
| super().__init__(in_channels, out_channels, *args, **kwargs) | |
| if isinstance(video_kernel_size, Iterable): | |
| padding = [int(k // 2) for k in video_kernel_size] | |
| else: | |
| padding = int(video_kernel_size // 2) | |
| self.time_mix_conv = torch.nn.Conv3d( | |
| in_channels=out_channels, | |
| out_channels=out_channels, | |
| kernel_size=video_kernel_size, | |
| padding=padding | |
| ) | |
| def forward(self, input, timesteps, skip_video=False): | |
| x = super().forward(input) | |
| if skip_video: | |
| return x | |
| else: | |
| x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps) | |
| x = self.time_mix_conv(x) | |
| return rearrange(x, "b c t h w -> (b t) c h w") | |
| class Conv2DWrapper(torch.nn.Conv2d): | |
| def forward(self, input: torch.Tensor, **kwargs) -> torch.Tensor: | |
| return super().forward(input) | |
| class VideoDecoder(Decoder): | |
| available_time_modes = ["all", "conv-only", "attn-only"] | |
| def __init__( | |
| self, | |
| *args, | |
| video_kernel_size: Union[int, list] = 3, | |
| alpha: float = 0.0, | |
| merge_strategy: str = "learned", | |
| time_mode: str = "conv-only", | |
| **kwargs | |
| ): | |
| self.video_kernel_size = video_kernel_size | |
| self.alpha = alpha | |
| self.merge_strategy = merge_strategy | |
| self.time_mode = time_mode | |
| assert ( | |
| self.time_mode in self.available_time_modes | |
| ), f"time_mode parameter has to be in {self.available_time_modes}" | |
| super().__init__(*args, **kwargs) | |
| def get_last_layer(self, skip_time_mix=False, **kwargs): | |
| if self.time_mode == "attn-only": | |
| raise NotImplementedError | |
| else: | |
| return ( | |
| self.conv_out.time_mix_conv.weight | |
| if not skip_time_mix | |
| else self.conv_out.weight | |
| ) | |
| def _make_conv(self) -> Callable: | |
| if self.time_mode != "attn-only": | |
| return partialclass(AE3DConv, video_kernel_size=self.video_kernel_size) | |
| else: | |
| return Conv2DWrapper | |
| def _make_resblock(self) -> Callable: | |
| if self.time_mode not in ["attn-only", "only-last-conv"]: | |
| return partialclass( | |
| VideoResBlock, | |
| video_kernel_size=self.video_kernel_size, | |
| alpha=self.alpha, | |
| merge_strategy=self.merge_strategy | |
| ) | |
| else: | |
| return super()._make_resblock() | |