File size: 13,013 Bytes
07c6a04 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 |
# Adapted from CogVideo
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# CogVideo: https://github.com/THUDM/CogVideo
# diffusers: https://github.com/huggingface/diffusers
# --------------------------------------------------------
from typing import Optional, Tuple, Union
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusers.models.embeddings import get_1d_sincos_pos_embed_from_grid, get_2d_sincos_pos_embed_from_grid
class CogVideoXDownsample3D(nn.Module):
# Todo: Wait for paper relase.
r"""
A 3D Downsampling layer using in [CogVideoX]() by Tsinghua University & ZhipuAI
Args:
in_channels (`int`):
Number of channels in the input image.
out_channels (`int`):
Number of channels produced by the convolution.
kernel_size (`int`, defaults to `3`):
Size of the convolving kernel.
stride (`int`, defaults to `2`):
Stride of the convolution.
padding (`int`, defaults to `0`):
Padding added to all four sides of the input.
compress_time (`bool`, defaults to `False`):
Whether or not to compress the time dimension.
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
stride: int = 2,
padding: int = 0,
compress_time: bool = False,
):
super().__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
self.compress_time = compress_time
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.compress_time:
batch_size, channels, frames, height, width = x.shape
# (batch_size, channels, frames, height, width) -> (batch_size, height, width, channels, frames) -> (batch_size * height * width, channels, frames)
x = x.permute(0, 3, 4, 1, 2).reshape(batch_size * height * width, channels, frames)
if x.shape[-1] % 2 == 1:
x_first, x_rest = x[..., 0], x[..., 1:]
if x_rest.shape[-1] > 0:
# (batch_size * height * width, channels, frames - 1) -> (batch_size * height * width, channels, (frames - 1) // 2)
x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
x = torch.cat([x_first[..., None], x_rest], dim=-1)
# (batch_size * height * width, channels, (frames // 2) + 1) -> (batch_size, height, width, channels, (frames // 2) + 1) -> (batch_size, channels, (frames // 2) + 1, height, width)
x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
else:
# (batch_size * height * width, channels, frames) -> (batch_size * height * width, channels, frames // 2)
x = F.avg_pool1d(x, kernel_size=2, stride=2)
# (batch_size * height * width, channels, frames // 2) -> (batch_size, height, width, channels, frames // 2) -> (batch_size, channels, frames // 2, height, width)
x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
# Pad the tensor
pad = (0, 1, 0, 1)
x = F.pad(x, pad, mode="constant", value=0)
batch_size, channels, frames, height, width = x.shape
# (batch_size, channels, frames, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size * frames, channels, height, width)
x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channels, height, width)
x = self.conv(x)
# (batch_size * frames, channels, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size, channels, frames, height, width)
x = x.reshape(batch_size, frames, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
return x
class CogVideoXUpsample3D(nn.Module):
r"""
A 3D Upsample layer using in CogVideoX by Tsinghua University & ZhipuAI # Todo: Wait for paper relase.
Args:
in_channels (`int`):
Number of channels in the input image.
out_channels (`int`):
Number of channels produced by the convolution.
kernel_size (`int`, defaults to `3`):
Size of the convolving kernel.
stride (`int`, defaults to `1`):
Stride of the convolution.
padding (`int`, defaults to `1`):
Padding added to all four sides of the input.
compress_time (`bool`, defaults to `False`):
Whether or not to compress the time dimension.
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
stride: int = 1,
padding: int = 1,
compress_time: bool = False,
) -> None:
super().__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
self.compress_time = compress_time
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
if self.compress_time:
if inputs.shape[2] > 1 and inputs.shape[2] % 2 == 1:
# split first frame
x_first, x_rest = inputs[:, :, 0], inputs[:, :, 1:]
x_first = F.interpolate(x_first, scale_factor=2.0)
x_rest = F.interpolate(x_rest, scale_factor=2.0)
x_first = x_first[:, :, None, :, :]
inputs = torch.cat([x_first, x_rest], dim=2)
elif inputs.shape[2] > 1:
inputs = F.interpolate(inputs, scale_factor=2.0)
else:
inputs = inputs.squeeze(2)
inputs = F.interpolate(inputs, scale_factor=2.0)
inputs = inputs[:, :, None, :, :]
else:
# only interpolate 2D
b, c, t, h, w = inputs.shape
inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
inputs = F.interpolate(inputs, scale_factor=2.0)
inputs = inputs.reshape(b, t, c, *inputs.shape[2:]).permute(0, 2, 1, 3, 4)
b, c, t, h, w = inputs.shape
inputs = inputs.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
inputs = self.conv(inputs)
inputs = inputs.reshape(b, t, *inputs.shape[1:]).permute(0, 2, 1, 3, 4)
return inputs
def get_3d_sincos_pos_embed(
embed_dim: int,
spatial_size: Union[int, Tuple[int, int]],
temporal_size: int,
spatial_interpolation_scale: float = 1.0,
temporal_interpolation_scale: float = 1.0,
) -> np.ndarray:
r"""
Args:
embed_dim (`int`):
spatial_size (`int` or `Tuple[int, int]`):
temporal_size (`int`):
spatial_interpolation_scale (`float`, defaults to 1.0):
temporal_interpolation_scale (`float`, defaults to 1.0):
"""
if embed_dim % 4 != 0:
raise ValueError("`embed_dim` must be divisible by 4")
if isinstance(spatial_size, int):
spatial_size = (spatial_size, spatial_size)
embed_dim_spatial = 3 * embed_dim // 4
embed_dim_temporal = embed_dim // 4
# 1. Spatial
grid_h = np.arange(spatial_size[1], dtype=np.float32) / spatial_interpolation_scale
grid_w = np.arange(spatial_size[0], dtype=np.float32) / spatial_interpolation_scale
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, spatial_size[1], spatial_size[0]])
pos_embed_spatial = get_2d_sincos_pos_embed_from_grid(embed_dim_spatial, grid)
# 2. Temporal
grid_t = np.arange(temporal_size, dtype=np.float32) / temporal_interpolation_scale
pos_embed_temporal = get_1d_sincos_pos_embed_from_grid(embed_dim_temporal, grid_t)
# 3. Concat
pos_embed_spatial = pos_embed_spatial[np.newaxis, :, :]
pos_embed_spatial = np.repeat(pos_embed_spatial, temporal_size, axis=0) # [T, H*W, D // 4 * 3]
pos_embed_temporal = pos_embed_temporal[:, np.newaxis, :]
pos_embed_temporal = np.repeat(pos_embed_temporal, spatial_size[0] * spatial_size[1], axis=1) # [T, H*W, D // 4]
pos_embed = np.concatenate([pos_embed_temporal, pos_embed_spatial], axis=-1) # [T, H*W, D]
return pos_embed
class CogVideoXPatchEmbed(nn.Module):
def __init__(
self,
patch_size: int = 2,
in_channels: int = 16,
embed_dim: int = 1920,
text_embed_dim: int = 4096,
bias: bool = True,
) -> None:
super().__init__()
self.patch_size = patch_size
self.proj = nn.Conv2d(
in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
)
self.text_proj = nn.Linear(text_embed_dim, embed_dim)
def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
r"""
Args:
text_embeds (`torch.Tensor`):
Input text embeddings. Expected shape: (batch_size, seq_length, embedding_dim).
image_embeds (`torch.Tensor`):
Input image embeddings. Expected shape: (batch_size, num_frames, channels, height, width).
"""
text_embeds = self.text_proj(text_embeds)
batch, num_frames, channels, height, width = image_embeds.shape
image_embeds = image_embeds.reshape(-1, channels, height, width)
image_embeds = self.proj(image_embeds)
image_embeds = image_embeds.view(batch, num_frames, *image_embeds.shape[1:])
image_embeds = image_embeds.flatten(3).transpose(2, 3) # [batch, num_frames, height x width, channels]
image_embeds = image_embeds.flatten(1, 2) # [batch, num_frames x height x width, channels]
embeds = torch.cat(
[text_embeds, image_embeds], dim=1
).contiguous() # [batch, seq_length + num_frames x height x width, channels]
return embeds
class CogVideoXLayerNormZero(nn.Module):
def __init__(
self,
conditioning_dim: int,
embedding_dim: int,
elementwise_affine: bool = True,
eps: float = 1e-5,
bias: bool = True,
) -> None:
super().__init__()
self.silu = nn.SiLU()
self.linear = nn.Linear(conditioning_dim, 6 * embedding_dim, bias=bias)
self.norm = nn.LayerNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
def forward(
self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, temb: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
class AdaLayerNorm(nn.Module):
r"""
Norm layer modified to incorporate timestep embeddings.
Parameters:
embedding_dim (`int`): The size of each embedding vector.
num_embeddings (`int`, *optional*): The size of the embeddings dictionary.
output_dim (`int`, *optional*):
norm_elementwise_affine (`bool`, defaults to `False):
norm_eps (`bool`, defaults to `False`):
chunk_dim (`int`, defaults to `0`):
"""
def __init__(
self,
embedding_dim: int,
num_embeddings: Optional[int] = None,
output_dim: Optional[int] = None,
norm_elementwise_affine: bool = False,
norm_eps: float = 1e-5,
chunk_dim: int = 0,
):
super().__init__()
self.chunk_dim = chunk_dim
output_dim = output_dim or embedding_dim * 2
if num_embeddings is not None:
self.emb = nn.Embedding(num_embeddings, embedding_dim)
else:
self.emb = None
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, output_dim)
self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
def forward(
self, x: torch.Tensor, timestep: Optional[torch.Tensor] = None, temb: Optional[torch.Tensor] = None
) -> torch.Tensor:
if self.emb is not None:
temb = self.emb(timestep)
temb = self.linear(self.silu(temb))
if self.chunk_dim == 1:
# This is a bit weird why we have the order of "shift, scale" here and "scale, shift" in the
# other if-branch. This branch is specific to CogVideoX for now.
shift, scale = temb.chunk(2, dim=1)
shift = shift[:, None, :]
scale = scale[:, None, :]
else:
scale, shift = temb.chunk(2, dim=0)
x = self.norm(x) * (1 + scale) + shift
return x
|