autoencoder_kl_3d.py
| 1 | # Licensed under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT (the "License"); |
| 2 | # you may not use this file except in compliance with the License. |
| 3 | # You may obtain a copy of the License at |
| 4 | # |
| 5 | # https://github.com/Tencent-Hunyuan/HunyuanImage-3.0/blob/main/LICENSE |
| 6 | # |
| 7 | # Unless required by applicable law or agreed to in writing, software |
| 8 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 9 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 10 | # See the License for the specific language governing permissions and |
| 11 | # limitations under the License. |
| 12 | # ============================================================================== |
| 13 | |
| 14 | from dataclasses import dataclass |
| 15 | from typing import Tuple, Optional |
| 16 | import math |
| 17 | import random |
| 18 | import numpy as np |
| 19 | from einops import rearrange |
| 20 | import torch |
| 21 | from torch import Tensor, nn |
| 22 | import torch.nn.functional as F |
| 23 | |
| 24 | from diffusers.configuration_utils import ConfigMixin, register_to_config |
| 25 | from diffusers.models.modeling_outputs import AutoencoderKLOutput |
| 26 | from diffusers.models.modeling_utils import ModelMixin |
| 27 | from diffusers.utils.torch_utils import randn_tensor |
| 28 | from diffusers.utils import BaseOutput |
| 29 | |
| 30 | |
| 31 | class DiagonalGaussianDistribution(object): |
| 32 | def __init__(self, parameters: torch.Tensor, deterministic: bool = False): |
| 33 | if parameters.ndim == 3: |
| 34 | dim = 2 # (B, L, C) |
| 35 | elif parameters.ndim == 5 or parameters.ndim == 4: |
| 36 | dim = 1 # (B, C, T, H ,W) / (B, C, H, W) |
| 37 | else: |
| 38 | raise NotImplementedError |
| 39 | self.parameters = parameters |
| 40 | self.mean, self.logvar = torch.chunk(parameters, 2, dim=dim) |
| 41 | self.logvar = torch.clamp(self.logvar, -30.0, 20.0) |
| 42 | self.deterministic = deterministic |
| 43 | self.std = torch.exp(0.5 * self.logvar) |
| 44 | self.var = torch.exp(self.logvar) |
| 45 | if self.deterministic: |
| 46 | self.var = self.std = torch.zeros_like( |
| 47 | self.mean, device=self.parameters.device, dtype=self.parameters.dtype |
| 48 | ) |
| 49 | |
| 50 | def sample(self, generator: Optional[torch.Generator] = None) -> torch.FloatTensor: |
| 51 | # make sure sample is on the same device as the parameters and has same dtype |
| 52 | sample = randn_tensor( |
| 53 | self.mean.shape, |
| 54 | generator=generator, |
| 55 | device=self.parameters.device, |
| 56 | dtype=self.parameters.dtype, |
| 57 | ) |
| 58 | x = self.mean + self.std * sample |
| 59 | return x |
| 60 | |
| 61 | def kl(self, other: "DiagonalGaussianDistribution" = None) -> torch.Tensor: |
| 62 | if self.deterministic: |
| 63 | return torch.Tensor([0.0]) |
| 64 | else: |
| 65 | reduce_dim = list(range(1, self.mean.ndim)) |
| 66 | if other is None: |
| 67 | return 0.5 * torch.sum( |
| 68 | torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar, |
| 69 | dim=reduce_dim, |
| 70 | ) |
| 71 | else: |
| 72 | return 0.5 * torch.sum( |
| 73 | torch.pow(self.mean - other.mean, 2) / other.var + |
| 74 | self.var / other.var - |
| 75 | 1.0 - |
| 76 | self.logvar + |
| 77 | other.logvar, |
| 78 | dim=reduce_dim, |
| 79 | ) |
| 80 | |
| 81 | def nll(self, sample: torch.Tensor, dims: Tuple[int, ...] = [1, 2, 3]) -> torch.Tensor: |
| 82 | if self.deterministic: |
| 83 | return torch.Tensor([0.0]) |
| 84 | logtwopi = np.log(2.0 * np.pi) |
| 85 | return 0.5 * torch.sum( |
| 86 | logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, |
| 87 | dim=dims, |
| 88 | ) |
| 89 | |
| 90 | def mode(self) -> torch.Tensor: |
| 91 | return self.mean |
| 92 | |
| 93 | |
| 94 | @dataclass |
| 95 | class DecoderOutput(BaseOutput): |
| 96 | sample: torch.FloatTensor |
| 97 | posterior: Optional[DiagonalGaussianDistribution] = None |
| 98 | |
| 99 | |
| 100 | def swish(x: Tensor) -> Tensor: |
| 101 | return x * torch.sigmoid(x) |
| 102 | |
| 103 | |
| 104 | def forward_with_checkpointing(module, *inputs, use_checkpointing=False): |
| 105 | def create_custom_forward(module): |
| 106 | def custom_forward(*inputs): |
| 107 | return module(*inputs) |
| 108 | return custom_forward |
| 109 | |
| 110 | if use_checkpointing: |
| 111 | return torch.utils.checkpoint.checkpoint(create_custom_forward(module), *inputs, use_reentrant=False) |
| 112 | else: |
| 113 | return module(*inputs) |
| 114 | |
| 115 | |
| 116 | class Conv3d(nn.Conv3d): |
| 117 | """ |
| 118 | Perform Conv3d on patches with numerical differences from nn.Conv3d within 1e-5. |
| 119 | Only symmetric padding is supported. |
| 120 | """ |
| 121 | |
| 122 | def forward(self, input): |
| 123 | B, C, T, H, W = input.shape |
| 124 | memory_count = (C * T * H * W) * 2 / 1024**3 |
| 125 | if memory_count > 2: |
| 126 | n_split = math.ceil(memory_count / 2) |
| 127 | assert n_split >= 2 |
| 128 | chunks = torch.chunk(input, chunks=n_split, dim=-3) |
| 129 | padded_chunks = [] |
| 130 | for i in range(len(chunks)): |
| 131 | if self.padding[0] > 0: |
| 132 | padded_chunk = F.pad( |
| 133 | chunks[i], |
| 134 | (0, 0, 0, 0, self.padding[0], self.padding[0]), |
| 135 | mode="constant" if self.padding_mode == "zeros" else self.padding_mode, |
| 136 | value=0, |
| 137 | ) |
| 138 | if i > 0: |
| 139 | padded_chunk[:, :, :self.padding[0]] = chunks[i - 1][:, :, -self.padding[0]:] |
| 140 | if i < len(chunks) - 1: |
| 141 | padded_chunk[:, :, -self.padding[0]:] = chunks[i + 1][:, :, :self.padding[0]] |
| 142 | else: |
| 143 | padded_chunk = chunks[i] |
| 144 | padded_chunks.append(padded_chunk) |
| 145 | padding_bak = self.padding |
| 146 | self.padding = (0, self.padding[1], self.padding[2]) |
| 147 | outputs = [] |
| 148 | for i in range(len(padded_chunks)): |
| 149 | outputs.append(super().forward(padded_chunks[i])) |
| 150 | self.padding = padding_bak |
| 151 | return torch.cat(outputs, dim=-3) |
| 152 | else: |
| 153 | return super().forward(input) |
| 154 | |
| 155 | |
| 156 | class AttnBlock(nn.Module): |
| 157 | """ Attention with torch sdpa implementation. """ |
| 158 | def __init__(self, in_channels: int): |
| 159 | super().__init__() |
| 160 | self.in_channels = in_channels |
| 161 | |
| 162 | self.norm = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) |
| 163 | |
| 164 | self.q = Conv3d(in_channels, in_channels, kernel_size=1) |
| 165 | self.k = Conv3d(in_channels, in_channels, kernel_size=1) |
| 166 | self.v = Conv3d(in_channels, in_channels, kernel_size=1) |
| 167 | self.proj_out = Conv3d(in_channels, in_channels, kernel_size=1) |
| 168 | |
| 169 | def attention(self, h_: Tensor) -> Tensor: |
| 170 | h_ = self.norm(h_) |
| 171 | q = self.q(h_) |
| 172 | k = self.k(h_) |
| 173 | v = self.v(h_) |
| 174 | |
| 175 | b, c, f, h, w = q.shape |
| 176 | q = rearrange(q, "b c f h w -> b 1 (f h w) c").contiguous() |
| 177 | k = rearrange(k, "b c f h w -> b 1 (f h w) c").contiguous() |
| 178 | v = rearrange(v, "b c f h w -> b 1 (f h w) c").contiguous() |
| 179 | h_ = nn.functional.scaled_dot_product_attention(q, k, v) |
| 180 | |
| 181 | return rearrange(h_, "b 1 (f h w) c -> b c f h w", f=f, h=h, w=w, c=c, b=b) |
| 182 | |
| 183 | def forward(self, x: Tensor) -> Tensor: |
| 184 | return x + self.proj_out(self.attention(x)) |
| 185 | |
| 186 | |
| 187 | class ResnetBlock(nn.Module): |
| 188 | def __init__(self, in_channels: int, out_channels: int): |
| 189 | super().__init__() |
| 190 | self.in_channels = in_channels |
| 191 | out_channels = in_channels if out_channels is None else out_channels |
| 192 | self.out_channels = out_channels |
| 193 | |
| 194 | self.norm1 = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) |
| 195 | self.conv1 = Conv3d(in_channels, out_channels, kernel_size=3, stride=1, padding=1) |
| 196 | self.norm2 = nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=1e-6, affine=True) |
| 197 | self.conv2 = Conv3d(out_channels, out_channels, kernel_size=3, stride=1, padding=1) |
| 198 | if self.in_channels != self.out_channels: |
| 199 | self.nin_shortcut = Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0) |
| 200 | |
| 201 | def forward(self, x): |
| 202 | h = x |
| 203 | h = self.norm1(h) |
| 204 | h = swish(h) |
| 205 | h = self.conv1(h) |
| 206 | |
| 207 | h = self.norm2(h) |
| 208 | h = swish(h) |
| 209 | h = self.conv2(h) |
| 210 | |
| 211 | if self.in_channels != self.out_channels: |
| 212 | x = self.nin_shortcut(x) |
| 213 | return x + h |
| 214 | |
| 215 | |
| 216 | class Downsample(nn.Module): |
| 217 | def __init__(self, in_channels: int, add_temporal_downsample: bool = True): |
| 218 | super().__init__() |
| 219 | self.add_temporal_downsample = add_temporal_downsample |
| 220 | stride = (2, 2, 2) if add_temporal_downsample else (1, 2, 2) # THW |
| 221 | # no asymmetric padding in torch conv, must do it ourselves |
| 222 | self.conv = Conv3d(in_channels, in_channels, kernel_size=3, stride=stride, padding=0) |
| 223 | |
| 224 | def forward(self, x: Tensor): |
| 225 | spatial_pad = (0, 1, 0, 1, 0, 0) # WHT |
| 226 | x = nn.functional.pad(x, spatial_pad, mode="constant", value=0) |
| 227 | |
| 228 | temporal_pad = (0, 0, 0, 0, 0, 1) if self.add_temporal_downsample else (0, 0, 0, 0, 1, 1) |
| 229 | x = nn.functional.pad(x, temporal_pad, mode="replicate") |
| 230 | |
| 231 | x = self.conv(x) |
| 232 | return x |
| 233 | |
| 234 | |
| 235 | class DownsampleDCAE(nn.Module): |
| 236 | def __init__(self, in_channels: int, out_channels: int, add_temporal_downsample: bool = True): |
| 237 | super().__init__() |
| 238 | factor = 2 * 2 * 2 if add_temporal_downsample else 1 * 2 * 2 |
| 239 | assert out_channels % factor == 0 |
| 240 | self.conv = Conv3d(in_channels, out_channels // factor, kernel_size=3, stride=1, padding=1) |
| 241 | |
| 242 | self.add_temporal_downsample = add_temporal_downsample |
| 243 | self.group_size = factor * in_channels // out_channels |
| 244 | |
| 245 | def forward(self, x: Tensor): |
| 246 | r1 = 2 if self.add_temporal_downsample else 1 |
| 247 | h = self.conv(x) |
| 248 | h = rearrange(h, "b c (f r1) (h r2) (w r3) -> b (r1 r2 r3 c) f h w", r1=r1, r2=2, r3=2) |
| 249 | shortcut = rearrange(x, "b c (f r1) (h r2) (w r3) -> b (r1 r2 r3 c) f h w", r1=r1, r2=2, r3=2) |
| 250 | |
| 251 | B, C, T, H, W = shortcut.shape |
| 252 | shortcut = shortcut.view(B, h.shape[1], self.group_size, T, H, W).mean(dim=2) |
| 253 | return h + shortcut |
| 254 | |
| 255 | |
| 256 | class Upsample(nn.Module): |
| 257 | def __init__(self, in_channels: int, add_temporal_upsample: bool = True): |
| 258 | super().__init__() |
| 259 | self.add_temporal_upsample = add_temporal_upsample |
| 260 | self.scale_factor = (2, 2, 2) if add_temporal_upsample else (1, 2, 2) # THW |
| 261 | self.conv = Conv3d(in_channels, in_channels, kernel_size=3, stride=1, padding=1) |
| 262 | |
| 263 | def forward(self, x: Tensor): |
| 264 | x = nn.functional.interpolate(x, scale_factor=self.scale_factor, mode="nearest") |
| 265 | x = self.conv(x) |
| 266 | return x |
| 267 | |
| 268 | |
| 269 | class UpsampleDCAE(nn.Module): |
| 270 | def __init__(self, in_channels: int, out_channels: int, add_temporal_upsample: bool = True): |
| 271 | super().__init__() |
| 272 | factor = 2 * 2 * 2 if add_temporal_upsample else 1 * 2 * 2 |
| 273 | self.conv = Conv3d(in_channels, out_channels * factor, kernel_size=3, stride=1, padding=1) |
| 274 | |
| 275 | self.add_temporal_upsample = add_temporal_upsample |
| 276 | self.repeats = factor * out_channels // in_channels |
| 277 | |
| 278 | def forward(self, x: Tensor): |
| 279 | r1 = 2 if self.add_temporal_upsample else 1 |
| 280 | h = self.conv(x) |
| 281 | h = rearrange(h, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2) |
| 282 | shortcut = x.repeat_interleave(repeats=self.repeats, dim=1) |
| 283 | shortcut = rearrange(shortcut, "b (r1 r2 r3 c) f h w -> b c (f r1) (h r2) (w r3)", r1=r1, r2=2, r3=2) |
| 284 | return h + shortcut |
| 285 | |
| 286 | |
| 287 | class Encoder(nn.Module): |
| 288 | """ |
| 289 | The encoder network of AutoencoderKLConv3D. |
| 290 | """ |
| 291 | def __init__( |
| 292 | self, |
| 293 | in_channels: int, |
| 294 | z_channels: int, |
| 295 | block_out_channels: Tuple[int, ...], |
| 296 | num_res_blocks: int, |
| 297 | ffactor_spatial: int, |
| 298 | ffactor_temporal: int, |
| 299 | downsample_match_channel: bool = True, |
| 300 | ): |
| 301 | super().__init__() |
| 302 | assert block_out_channels[-1] % (2 * z_channels) == 0 |
| 303 | |
| 304 | self.z_channels = z_channels |
| 305 | self.block_out_channels = block_out_channels |
| 306 | self.num_res_blocks = num_res_blocks |
| 307 | |
| 308 | # downsampling |
| 309 | self.conv_in = Conv3d(in_channels, block_out_channels[0], kernel_size=3, stride=1, padding=1) |
| 310 | |
| 311 | self.down = nn.ModuleList() |
| 312 | block_in = block_out_channels[0] |
| 313 | for i_level, ch in enumerate(block_out_channels): |
| 314 | block = nn.ModuleList() |
| 315 | block_out = ch |
| 316 | for _ in range(self.num_res_blocks): |
| 317 | block.append(ResnetBlock(in_channels=block_in, out_channels=block_out)) |
| 318 | block_in = block_out |
| 319 | down = nn.Module() |
| 320 | down.block = block |
| 321 | |
| 322 | add_spatial_downsample = bool(i_level < np.log2(ffactor_spatial)) |
| 323 | add_temporal_downsample = (add_spatial_downsample and |
| 324 | bool(i_level >= np.log2(ffactor_spatial // ffactor_temporal))) |
| 325 | if add_spatial_downsample or add_temporal_downsample: |
| 326 | assert i_level < len(block_out_channels) - 1 |
| 327 | block_out = block_out_channels[i_level + 1] if downsample_match_channel else block_in |
| 328 | down.downsample = DownsampleDCAE(block_in, block_out, add_temporal_downsample) |
| 329 | block_in = block_out |
| 330 | self.down.append(down) |
| 331 | |
| 332 | # middle |
| 333 | self.mid = nn.Module() |
| 334 | self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in) |
| 335 | self.mid.attn_1 = AttnBlock(block_in) |
| 336 | self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in) |
| 337 | |
| 338 | # end |
| 339 | self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True) |
| 340 | self.conv_out = Conv3d(block_in, 2 * z_channels, kernel_size=3, stride=1, padding=1) |
| 341 | |
| 342 | self.gradient_checkpointing = False |
| 343 | |
| 344 | def forward(self, x: Tensor) -> Tensor: |
| 345 | use_checkpointing = bool(self.training and self.gradient_checkpointing) |
| 346 | |
| 347 | # downsampling |
| 348 | h = self.conv_in(x) |
| 349 | for i_level in range(len(self.block_out_channels)): |
| 350 | for i_block in range(self.num_res_blocks): |
| 351 | h = forward_with_checkpointing( |
| 352 | self.down[i_level].block[i_block], h, use_checkpointing=use_checkpointing) |
| 353 | if hasattr(self.down[i_level], "downsample"): |
| 354 | h = forward_with_checkpointing(self.down[i_level].downsample, h, use_checkpointing=use_checkpointing) |
| 355 | |
| 356 | # middle |
| 357 | h = forward_with_checkpointing(self.mid.block_1, h, use_checkpointing=use_checkpointing) |
| 358 | h = forward_with_checkpointing(self.mid.attn_1, h, use_checkpointing=use_checkpointing) |
| 359 | h = forward_with_checkpointing(self.mid.block_2, h, use_checkpointing=use_checkpointing) |
| 360 | |
| 361 | # end |
| 362 | group_size = self.block_out_channels[-1] // (2 * self.z_channels) |
| 363 | shortcut = rearrange(h, "b (c r) f h w -> b c r f h w", r=group_size).mean(dim=2) |
| 364 | h = self.norm_out(h) |
| 365 | h = swish(h) |
| 366 | h = self.conv_out(h) |
| 367 | h += shortcut |
| 368 | return h |
| 369 | |
| 370 | |
| 371 | class Decoder(nn.Module): |
| 372 | """ |
| 373 | The decoder network of AutoencoderKLConv3D. |
| 374 | """ |
| 375 | def __init__( |
| 376 | self, |
| 377 | z_channels: int, |
| 378 | out_channels: int, |
| 379 | block_out_channels: Tuple[int, ...], |
| 380 | num_res_blocks: int, |
| 381 | ffactor_spatial: int, |
| 382 | ffactor_temporal: int, |
| 383 | upsample_match_channel: bool = True, |
| 384 | ): |
| 385 | super().__init__() |
| 386 | assert block_out_channels[0] % z_channels == 0 |
| 387 | |
| 388 | self.z_channels = z_channels |
| 389 | self.block_out_channels = block_out_channels |
| 390 | self.num_res_blocks = num_res_blocks |
| 391 | |
| 392 | # z to block_in |
| 393 | block_in = block_out_channels[0] |
| 394 | self.conv_in = Conv3d(z_channels, block_in, kernel_size=3, stride=1, padding=1) |
| 395 | |
| 396 | # middle |
| 397 | self.mid = nn.Module() |
| 398 | self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in) |
| 399 | self.mid.attn_1 = AttnBlock(block_in) |
| 400 | self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in) |
| 401 | |
| 402 | # upsampling |
| 403 | self.up = nn.ModuleList() |
| 404 | for i_level, ch in enumerate(block_out_channels): |
| 405 | block = nn.ModuleList() |
| 406 | block_out = ch |
| 407 | for _ in range(self.num_res_blocks + 1): |
| 408 | block.append(ResnetBlock(in_channels=block_in, out_channels=block_out)) |
| 409 | block_in = block_out |
| 410 | up = nn.Module() |
| 411 | up.block = block |
| 412 | |
| 413 | add_spatial_upsample = bool(i_level < np.log2(ffactor_spatial)) |
| 414 | add_temporal_upsample = bool(i_level < np.log2(ffactor_temporal)) |
| 415 | if add_spatial_upsample or add_temporal_upsample: |
| 416 | assert i_level < len(block_out_channels) - 1 |
| 417 | block_out = block_out_channels[i_level + 1] if upsample_match_channel else block_in |
| 418 | up.upsample = UpsampleDCAE(block_in, block_out, add_temporal_upsample) |
| 419 | block_in = block_out |
| 420 | self.up.append(up) |
| 421 | |
| 422 | # end |
| 423 | self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True) |
| 424 | self.conv_out = Conv3d(block_in, out_channels, kernel_size=3, stride=1, padding=1) |
| 425 | |
| 426 | self.gradient_checkpointing = False |
| 427 | |
| 428 | def forward(self, z: Tensor) -> Tensor: |
| 429 | use_checkpointing = bool(self.training and self.gradient_checkpointing) |
| 430 | |
| 431 | # z to block_in |
| 432 | repeats = self.block_out_channels[0] // (self.z_channels) |
| 433 | h = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1) |
| 434 | |
| 435 | # middle |
| 436 | h = forward_with_checkpointing(self.mid.block_1, h, use_checkpointing=use_checkpointing) |
| 437 | h = forward_with_checkpointing(self.mid.attn_1, h, use_checkpointing=use_checkpointing) |
| 438 | h = forward_with_checkpointing(self.mid.block_2, h, use_checkpointing=use_checkpointing) |
| 439 | |
| 440 | # upsampling |
| 441 | for i_level in range(len(self.block_out_channels)): |
| 442 | for i_block in range(self.num_res_blocks + 1): |
| 443 | h = forward_with_checkpointing(self.up[i_level].block[i_block], h, use_checkpointing=use_checkpointing) |
| 444 | if hasattr(self.up[i_level], "upsample"): |
| 445 | h = forward_with_checkpointing(self.up[i_level].upsample, h, use_checkpointing=use_checkpointing) |
| 446 | |
| 447 | # end |
| 448 | h = self.norm_out(h) |
| 449 | h = swish(h) |
| 450 | h = self.conv_out(h) |
| 451 | return h |
| 452 | |
| 453 | |
| 454 | class AutoencoderKLConv3D(ModelMixin, ConfigMixin): |
| 455 | """ |
| 456 | Autoencoder model with KL-regularized latent space based on 3D convolutions. |
| 457 | """ |
| 458 | _supports_gradient_checkpointing = True |
| 459 | |
| 460 | @register_to_config |
| 461 | def __init__( |
| 462 | self, |
| 463 | in_channels: int, |
| 464 | out_channels: int, |
| 465 | latent_channels: int, |
| 466 | block_out_channels: Tuple[int, ...], |
| 467 | layers_per_block: int, |
| 468 | ffactor_spatial: int, |
| 469 | ffactor_temporal: int, |
| 470 | sample_size: int, |
| 471 | sample_tsize: int, |
| 472 | scaling_factor: float = None, |
| 473 | shift_factor: Optional[float] = None, |
| 474 | downsample_match_channel: bool = True, |
| 475 | upsample_match_channel: bool = True, |
| 476 | only_encoder: bool = False, # only build encoder for saving memory |
| 477 | only_decoder: bool = False, # only build decoder for saving memory |
| 478 | ): |
| 479 | super().__init__() |
| 480 | self.ffactor_spatial = ffactor_spatial |
| 481 | self.ffactor_temporal = ffactor_temporal |
| 482 | self.scaling_factor = scaling_factor |
| 483 | self.shift_factor = shift_factor |
| 484 | |
| 485 | # build model |
| 486 | if not only_decoder: |
| 487 | self.encoder = Encoder( |
| 488 | in_channels=in_channels, |
| 489 | z_channels=latent_channels, |
| 490 | block_out_channels=block_out_channels, |
| 491 | num_res_blocks=layers_per_block, |
| 492 | ffactor_spatial=ffactor_spatial, |
| 493 | ffactor_temporal=ffactor_temporal, |
| 494 | downsample_match_channel=downsample_match_channel, |
| 495 | ) |
| 496 | if not only_encoder: |
| 497 | self.decoder = Decoder( |
| 498 | z_channels=latent_channels, |
| 499 | out_channels=out_channels, |
| 500 | block_out_channels=list(reversed(block_out_channels)), |
| 501 | num_res_blocks=layers_per_block, |
| 502 | ffactor_spatial=ffactor_spatial, |
| 503 | ffactor_temporal=ffactor_temporal, |
| 504 | upsample_match_channel=upsample_match_channel, |
| 505 | ) |
| 506 | |
| 507 | # slicing and tiling related |
| 508 | self.use_slicing = False |
| 509 | self.slicing_bsz = 1 |
| 510 | self.use_spatial_tiling = False |
| 511 | self.use_temporal_tiling = False |
| 512 | self.use_tiling_during_training = False |
| 513 | |
| 514 | # only relevant if vae tiling is enabled |
| 515 | self.tile_sample_min_size = sample_size |
| 516 | self.tile_latent_min_size = sample_size // ffactor_spatial |
| 517 | self.tile_sample_min_tsize = sample_tsize |
| 518 | self.tile_latent_min_tsize = sample_tsize // ffactor_temporal |
| 519 | self.tile_overlap_factor = 0.25 |
| 520 | |
| 521 | # use torch.compile for faster encode speed |
| 522 | self.use_compile = False |
| 523 | |
| 524 | def _set_gradient_checkpointing(self, module, value=False): |
| 525 | if isinstance(module, (Encoder, Decoder)): |
| 526 | module.gradient_checkpointing = value |
| 527 | |
| 528 | def enable_tiling_during_training(self, use_tiling: bool = True): |
| 529 | self.use_tiling_during_training = use_tiling |
| 530 | |
| 531 | def disable_tiling_during_training(self): |
| 532 | self.enable_tiling_during_training(False) |
| 533 | |
| 534 | def enable_temporal_tiling(self, use_tiling: bool = True): |
| 535 | self.use_temporal_tiling = use_tiling |
| 536 | |
| 537 | def disable_temporal_tiling(self): |
| 538 | self.enable_temporal_tiling(False) |
| 539 | |
| 540 | def enable_spatial_tiling(self, use_tiling: bool = True): |
| 541 | self.use_spatial_tiling = use_tiling |
| 542 | |
| 543 | def disable_spatial_tiling(self): |
| 544 | self.enable_spatial_tiling(False) |
| 545 | |
| 546 | def enable_tiling(self, use_tiling: bool = True): |
| 547 | self.enable_spatial_tiling(use_tiling) |
| 548 | |
| 549 | def disable_tiling(self): |
| 550 | self.disable_spatial_tiling() |
| 551 | |
| 552 | def enable_slicing(self): |
| 553 | self.use_slicing = True |
| 554 | |
| 555 | def disable_slicing(self): |
| 556 | self.use_slicing = False |
| 557 | |
| 558 | def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int): |
| 559 | blend_extent = min(a.shape[-1], b.shape[-1], blend_extent) |
| 560 | for x in range(blend_extent): |
| 561 | b[:, :, :, :, x] = \ |
| 562 | a[:, :, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, :, x] * (x / blend_extent) |
| 563 | return b |
| 564 | |
| 565 | def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int): |
| 566 | blend_extent = min(a.shape[-2], b.shape[-2], blend_extent) |
| 567 | for y in range(blend_extent): |
| 568 | b[:, :, :, y, :] = \ |
| 569 | a[:, :, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, :, y, :] * (y / blend_extent) |
| 570 | return b |
| 571 | |
| 572 | def blend_t(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int): |
| 573 | blend_extent = min(a.shape[-3], b.shape[-3], blend_extent) |
| 574 | for x in range(blend_extent): |
| 575 | b[:, :, x, :, :] = \ |
| 576 | a[:, :, -blend_extent + x, :, :] * (1 - x / blend_extent) + b[:, :, x, :, :] * (x / blend_extent) |
| 577 | return b |
| 578 | |
| 579 | def spatial_tiled_encode(self, x: torch.Tensor): |
| 580 | """ spatial tailing for frames """ |
| 581 | B, C, T, H, W = x.shape |
| 582 | overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor)) # 256 * (1 - 0.25) = 192 |
| 583 | blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor) # 8 * 0.25 = 2 |
| 584 | row_limit = self.tile_latent_min_size - blend_extent # 8 - 2 = 6 |
| 585 | |
| 586 | rows = [] |
| 587 | for i in range(0, H, overlap_size): |
| 588 | row = [] |
| 589 | for j in range(0, W, overlap_size): |
| 590 | tile = x[:, :, :, i: i + self.tile_sample_min_size, j: j + self.tile_sample_min_size] |
| 591 | tile = self.encoder(tile) |
| 592 | row.append(tile) |
| 593 | rows.append(row) |
| 594 | result_rows = [] |
| 595 | for i, row in enumerate(rows): |
| 596 | result_row = [] |
| 597 | for j, tile in enumerate(row): |
| 598 | if i > 0: |
| 599 | tile = self.blend_v(rows[i - 1][j], tile, blend_extent) |
| 600 | if j > 0: |
| 601 | tile = self.blend_h(row[j - 1], tile, blend_extent) |
| 602 | result_row.append(tile[:, :, :, :row_limit, :row_limit]) |
| 603 | result_rows.append(torch.cat(result_row, dim=-1)) |
| 604 | moments = torch.cat(result_rows, dim=-2) |
| 605 | return moments |
| 606 | |
| 607 | def temporal_tiled_encode(self, x: torch.Tensor): |
| 608 | """ temporal tailing for frames """ |
| 609 | B, C, T, H, W = x.shape |
| 610 | overlap_size = int(self.tile_sample_min_tsize * (1 - self.tile_overlap_factor)) # 64 * (1 - 0.25) = 48 |
| 611 | blend_extent = int(self.tile_latent_min_tsize * self.tile_overlap_factor) # 8 * 0.25 = 2 |
| 612 | t_limit = self.tile_latent_min_tsize - blend_extent # 8 - 2 = 6 |
| 613 | |
| 614 | row = [] |
| 615 | for i in range(0, T, overlap_size): |
| 616 | tile = x[:, :, i: i + self.tile_sample_min_tsize, :, :] |
| 617 | if self.use_spatial_tiling and ( |
| 618 | tile.shape[-1] > self.tile_sample_min_size or tile.shape[-2] > self.tile_sample_min_size): |
| 619 | tile = self.spatial_tiled_encode(tile) |
| 620 | else: |
| 621 | tile = self.encoder(tile) |
| 622 | row.append(tile) |
| 623 | result_row = [] |
| 624 | for i, tile in enumerate(row): |
| 625 | if i > 0: |
| 626 | tile = self.blend_t(row[i - 1], tile, blend_extent) |
| 627 | result_row.append(tile[:, :, :t_limit, :, :]) |
| 628 | moments = torch.cat(result_row, dim=-3) |
| 629 | return moments |
| 630 | |
| 631 | def spatial_tiled_decode(self, z: torch.Tensor): |
| 632 | """ spatial tailing for frames """ |
| 633 | B, C, T, H, W = z.shape |
| 634 | overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor)) # 8 * (1 - 0.25) = 6 |
| 635 | blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor) # 256 * 0.25 = 64 |
| 636 | row_limit = self.tile_sample_min_size - blend_extent # 256 - 64 = 192 |
| 637 | |
| 638 | rows = [] |
| 639 | for i in range(0, H, overlap_size): |
| 640 | row = [] |
| 641 | for j in range(0, W, overlap_size): |
| 642 | tile = z[:, :, :, i: i + self.tile_latent_min_size, j: j + self.tile_latent_min_size] |
| 643 | decoded = self.decoder(tile) |
| 644 | row.append(decoded) |
| 645 | rows.append(row) |
| 646 | |
| 647 | result_rows = [] |
| 648 | for i, row in enumerate(rows): |
| 649 | result_row = [] |
| 650 | for j, tile in enumerate(row): |
| 651 | if i > 0: |
| 652 | tile = self.blend_v(rows[i - 1][j], tile, blend_extent) |
| 653 | if j > 0: |
| 654 | tile = self.blend_h(row[j - 1], tile, blend_extent) |
| 655 | result_row.append(tile[:, :, :, :row_limit, :row_limit]) |
| 656 | result_rows.append(torch.cat(result_row, dim=-1)) |
| 657 | dec = torch.cat(result_rows, dim=-2) |
| 658 | return dec |
| 659 | |
| 660 | def temporal_tiled_decode(self, z: torch.Tensor): |
| 661 | """ temporal tailing for frames """ |
| 662 | B, C, T, H, W = z.shape |
| 663 | overlap_size = int(self.tile_latent_min_tsize * (1 - self.tile_overlap_factor)) # 8 * (1 - 0.25) = 6 |
| 664 | blend_extent = int(self.tile_sample_min_tsize * self.tile_overlap_factor) # 64 * 0.25 = 16 |
| 665 | t_limit = self.tile_sample_min_tsize - blend_extent # 64 - 16 = 48 |
| 666 | assert 0 < overlap_size < self.tile_latent_min_tsize |
| 667 | |
| 668 | row = [] |
| 669 | for i in range(0, T, overlap_size): |
| 670 | tile = z[:, :, i: i + self.tile_latent_min_tsize, :, :] |
| 671 | if self.use_spatial_tiling and ( |
| 672 | tile.shape[-1] > self.tile_latent_min_size or tile.shape[-2] > self.tile_latent_min_size): |
| 673 | decoded = self.spatial_tiled_decode(tile) |
| 674 | else: |
| 675 | decoded = self.decoder(tile) |
| 676 | row.append(decoded) |
| 677 | |
| 678 | result_row = [] |
| 679 | for i, tile in enumerate(row): |
| 680 | if i > 0: |
| 681 | tile = self.blend_t(row[i - 1], tile, blend_extent) |
| 682 | result_row.append(tile[:, :, :t_limit, :, :]) |
| 683 | dec = torch.cat(result_row, dim=-3) |
| 684 | return dec |
| 685 | |
| 686 | def encode(self, x: Tensor, return_dict: bool = True): |
| 687 | """ |
| 688 | Encodes the input by passing through the encoder network. |
| 689 | Support slicing and tiling for memory efficiency. |
| 690 | """ |
| 691 | def _encode(x): |
| 692 | if self.use_temporal_tiling and x.shape[-3] > self.tile_sample_min_tsize: |
| 693 | return self.temporal_tiled_encode(x) |
| 694 | if self.use_spatial_tiling and ( |
| 695 | x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size): |
| 696 | return self.spatial_tiled_encode(x) |
| 697 | |
| 698 | if self.use_compile: |
| 699 | @torch.compile |
| 700 | def encoder(x): |
| 701 | return self.encoder(x) |
| 702 | return encoder(x) |
| 703 | return self.encoder(x) |
| 704 | |
| 705 | if len(x.shape) != 5: # (B, C, T, H, W) |
| 706 | x = x[:, :, None] |
| 707 | assert len(x.shape) == 5 # (B, C, T, H, W) |
| 708 | if x.shape[2] == 1: |
| 709 | x = x.expand(-1, -1, self.ffactor_temporal, -1, -1) |
| 710 | else: |
| 711 | assert x.shape[2] != self.ffactor_temporal and x.shape[2] % self.ffactor_temporal == 0 |
| 712 | |
| 713 | if self.use_slicing and x.shape[0] > 1: |
| 714 | if self.slicing_bsz == 1: |
| 715 | encoded_slices = [_encode(x_slice) for x_slice in x.split(1)] |
| 716 | else: |
| 717 | sections = [self.slicing_bsz] * (x.shape[0] // self.slicing_bsz) |
| 718 | if x.shape[0] % self.slicing_bsz != 0: |
| 719 | sections.append(x.shape[0] % self.slicing_bsz) |
| 720 | encoded_slices = [_encode(x_slice) for x_slice in x.split(sections)] |
| 721 | h = torch.cat(encoded_slices) |
| 722 | else: |
| 723 | h = _encode(x) |
| 724 | posterior = DiagonalGaussianDistribution(h) |
| 725 | |
| 726 | if not return_dict: |
| 727 | return (posterior,) |
| 728 | |
| 729 | return AutoencoderKLOutput(latent_dist=posterior) |
| 730 | |
| 731 | def decode(self, z: Tensor, return_dict: bool = True, generator=None): |
| 732 | """ |
| 733 | Decodes the input by passing through the decoder network. |
| 734 | Support slicing and tiling for memory efficiency. |
| 735 | """ |
| 736 | def _decode(z): |
| 737 | if self.use_temporal_tiling and z.shape[-3] > self.tile_latent_min_tsize: |
| 738 | return self.temporal_tiled_decode(z) |
| 739 | if self.use_spatial_tiling and ( |
| 740 | z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size): |
| 741 | return self.spatial_tiled_decode(z) |
| 742 | return self.decoder(z) |
| 743 | |
| 744 | if self.use_slicing and z.shape[0] > 1: |
| 745 | decoded_slices = [_decode(z_slice) for z_slice in z.split(1)] |
| 746 | decoded = torch.cat(decoded_slices) |
| 747 | else: |
| 748 | decoded = _decode(z) |
| 749 | |
| 750 | if z.shape[-3] == 1: |
| 751 | decoded = decoded[:, :, -1:] |
| 752 | |
| 753 | if not return_dict: |
| 754 | return (decoded,) |
| 755 | |
| 756 | return DecoderOutput(sample=decoded) |
| 757 | |
| 758 | def forward( |
| 759 | self, |
| 760 | sample: torch.Tensor, |
| 761 | sample_posterior: bool = False, |
| 762 | return_posterior: bool = True, |
| 763 | return_dict: bool = True |
| 764 | ): |
| 765 | posterior = self.encode(sample).latent_dist |
| 766 | z = posterior.sample() if sample_posterior else posterior.mode() |
| 767 | dec = self.decode(z).sample |
| 768 | return DecoderOutput(sample=dec, posterior=posterior) if return_dict else (dec, posterior) |
| 769 | |
| 770 | def random_reset_tiling(self, x: torch.Tensor): |
| 771 | if x.shape[-3] == 1: |
| 772 | self.disable_spatial_tiling() |
| 773 | self.disable_temporal_tiling() |
| 774 | return |
| 775 | |
| 776 | # Use fixed shape here |
| 777 | min_sample_size = int(1 / self.tile_overlap_factor) * self.ffactor_spatial |
| 778 | min_sample_tsize = int(1 / self.tile_overlap_factor) * self.ffactor_temporal |
| 779 | sample_size = random.choice([None, 1 * min_sample_size, 2 * min_sample_size, 3 * min_sample_size]) |
| 780 | if sample_size is None: |
| 781 | self.disable_spatial_tiling() |
| 782 | else: |
| 783 | self.tile_sample_min_size = sample_size |
| 784 | self.tile_latent_min_size = sample_size // self.ffactor_spatial |
| 785 | self.enable_spatial_tiling() |
| 786 | |
| 787 | sample_tsize = random.choice([None, 1 * min_sample_tsize, 2 * min_sample_tsize, 3 * min_sample_tsize]) |
| 788 | if sample_tsize is None: |
| 789 | self.disable_temporal_tiling() |
| 790 | else: |
| 791 | self.tile_sample_min_tsize = sample_tsize |
| 792 | self.tile_latent_min_tsize = sample_tsize // self.ffactor_temporal |
| 793 | self.enable_temporal_tiling() |
| 794 | |