File size: 16,393 Bytes
f0533a5 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 |
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from typing import Optional, Tuple
import numpy as np
import torch
import torch.nn as nn
from einops import rearrange
from diffusers.utils import BaseOutput, is_torch_version
from diffusers.utils.torch_utils import randn_tensor
from diffusers.models.attention_processor import SpatialNorm
from .modeling_block import (
UNetMidBlock2D,
CausalUNetMidBlock2D,
get_down_block,
get_up_block,
get_input_layer,
get_output_layer,
)
from .modeling_resnet import (
Downsample2D,
Upsample2D,
TemporalDownsample2x,
TemporalUpsample2x,
)
from .modeling_causal_conv import CausalConv3d, CausalGroupNorm
@dataclass
class DecoderOutput(BaseOutput):
r"""
Output of decoding method.
Args:
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
The decoded output sample from the last layer of the model.
"""
sample: torch.FloatTensor
class CausalVaeEncoder(nn.Module):
r"""
The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation.
Args:
in_channels (`int`, *optional*, defaults to 3):
The number of input channels.
out_channels (`int`, *optional*, defaults to 3):
The number of output channels.
down_block_types (`Tuple[str, ...]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
The types of down blocks to use. See `~diffusers.models.unet_2d_blocks.get_down_block` for available
options.
block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
The number of output channels for each block.
layers_per_block (`int`, *optional*, defaults to 2):
The number of layers per block.
norm_num_groups (`int`, *optional*, defaults to 32):
The number of groups for normalization.
act_fn (`str`, *optional*, defaults to `"silu"`):
The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
double_z (`bool`, *optional*, defaults to `True`):
Whether to double the number of output channels for the last block.
"""
def __init__(
self,
in_channels: int = 3,
out_channels: int = 3,
down_block_types: Tuple[str, ...] = ("DownEncoderBlockCausal3D",),
spatial_down_sample: Tuple[bool, ...] = (True,),
temporal_down_sample: Tuple[bool, ...] = (False,),
block_out_channels: Tuple[int, ...] = (64,),
layers_per_block: Tuple[int, ...] = (2,),
norm_num_groups: int = 32,
act_fn: str = "silu",
double_z: bool = True,
block_dropout: Tuple[int, ...] = (0.0,),
mid_block_add_attention=True,
):
super().__init__()
self.layers_per_block = layers_per_block
self.conv_in = CausalConv3d(
in_channels,
block_out_channels[0],
kernel_size=3,
stride=1,
)
self.mid_block = None
self.down_blocks = nn.ModuleList([])
# 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]
down_block = get_down_block(
down_block_type,
num_layers=self.layers_per_block[i],
in_channels=input_channel,
out_channels=output_channel,
add_spatial_downsample=spatial_down_sample[i],
add_temporal_downsample=temporal_down_sample[i],
resnet_eps=1e-6,
downsample_padding=0,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
attention_head_dim=output_channel,
temb_channels=None,
dropout=block_dropout[i],
)
self.down_blocks.append(down_block)
# mid
self.mid_block = CausalUNetMidBlock2D(
in_channels=block_out_channels[-1],
resnet_eps=1e-6,
resnet_act_fn=act_fn,
output_scale_factor=1,
resnet_time_scale_shift="default",
attention_head_dim=block_out_channels[-1],
resnet_groups=norm_num_groups,
temb_channels=None,
add_attention=mid_block_add_attention,
dropout=block_dropout[-1],
)
# out
self.conv_norm_out = CausalGroupNorm(num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6)
self.conv_act = nn.SiLU()
conv_out_channels = 2 * out_channels if double_z else out_channels
self.conv_out = CausalConv3d(block_out_channels[-1], conv_out_channels, kernel_size=3, stride=1)
self.gradient_checkpointing = False
def forward(self, sample: torch.FloatTensor, is_init_image=True, temporal_chunk=False) -> torch.FloatTensor:
r"""The forward method of the `Encoder` class."""
sample = self.conv_in(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
# down
if is_torch_version(">=", "1.11.0"):
for down_block in self.down_blocks:
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(down_block), sample, is_init_image,
temporal_chunk, use_reentrant=False
)
# middle
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(self.mid_block), sample, is_init_image,
temporal_chunk, use_reentrant=False
)
else:
for down_block in self.down_blocks:
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(down_block), sample, is_init_image, temporal_chunk)
# middle
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample, is_init_image, temporal_chunk)
else:
# down
for down_block in self.down_blocks:
sample = down_block(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk)
# middle
sample = self.mid_block(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk)
# post-process
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk)
return sample
class CausalVaeDecoder(nn.Module):
r"""
The `Decoder` layer of a variational autoencoder that decodes its latent representation into an output sample.
Args:
in_channels (`int`, *optional*, defaults to 3):
The number of input channels.
out_channels (`int`, *optional*, defaults to 3):
The number of output channels.
up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
The number of output channels for each block.
layers_per_block (`int`, *optional*, defaults to 2):
The number of layers per block.
norm_num_groups (`int`, *optional*, defaults to 32):
The number of groups for normalization.
act_fn (`str`, *optional*, defaults to `"silu"`):
The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
norm_type (`str`, *optional*, defaults to `"group"`):
The normalization type to use. Can be either `"group"` or `"spatial"`.
"""
def __init__(
self,
in_channels: int = 3,
out_channels: int = 3,
up_block_types: Tuple[str, ...] = ("UpDecoderBlockCausal3D",),
spatial_up_sample: Tuple[bool, ...] = (True,),
temporal_up_sample: Tuple[bool, ...] = (False,),
block_out_channels: Tuple[int, ...] = (64,),
layers_per_block: Tuple[int, ...] = (2,),
norm_num_groups: int = 32,
act_fn: str = "silu",
mid_block_add_attention=True,
interpolate: bool = True,
block_dropout: Tuple[int, ...] = (0.0,),
):
super().__init__()
self.layers_per_block = layers_per_block
self.conv_in = CausalConv3d(
in_channels,
block_out_channels[-1],
kernel_size=3,
stride=1,
)
self.mid_block = None
self.up_blocks = nn.ModuleList([])
# mid
self.mid_block = CausalUNetMidBlock2D(
in_channels=block_out_channels[-1],
resnet_eps=1e-6,
resnet_act_fn=act_fn,
output_scale_factor=1,
resnet_time_scale_shift="default",
attention_head_dim=block_out_channels[-1],
resnet_groups=norm_num_groups,
temb_channels=None,
add_attention=mid_block_add_attention,
dropout=block_dropout[-1],
)
# up
reversed_block_out_channels = list(reversed(block_out_channels))
output_channel = reversed_block_out_channels[0]
for i, up_block_type in enumerate(up_block_types):
prev_output_channel = output_channel
output_channel = reversed_block_out_channels[i]
is_final_block = i == len(block_out_channels) - 1
up_block = get_up_block(
up_block_type,
num_layers=self.layers_per_block[i],
in_channels=prev_output_channel,
out_channels=output_channel,
prev_output_channel=None,
add_spatial_upsample=spatial_up_sample[i],
add_temporal_upsample=temporal_up_sample[i],
resnet_eps=1e-6,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
attention_head_dim=output_channel,
temb_channels=None,
resnet_time_scale_shift='default',
interpolate=interpolate,
dropout=block_dropout[i],
)
self.up_blocks.append(up_block)
prev_output_channel = output_channel
# out
self.conv_norm_out = CausalGroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
self.conv_act = nn.SiLU()
self.conv_out = CausalConv3d(block_out_channels[0], out_channels, kernel_size=3, stride=1)
self.gradient_checkpointing = False
def forward(
self,
sample: torch.FloatTensor,
is_init_image=True,
temporal_chunk=False,
) -> torch.FloatTensor:
r"""The forward method of the `Decoder` class."""
sample = self.conv_in(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk)
upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
# middle
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(self.mid_block),
sample,
is_init_image=is_init_image,
temporal_chunk=temporal_chunk,
use_reentrant=False,
)
sample = sample.to(upscale_dtype)
# up
for up_block in self.up_blocks:
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(up_block),
sample,
is_init_image=is_init_image,
temporal_chunk=temporal_chunk,
use_reentrant=False,
)
else:
# middle
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(self.mid_block), sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk,
)
sample = sample.to(upscale_dtype)
# up
for up_block in self.up_blocks:
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample,
is_init_image=is_init_image, temporal_chunk=temporal_chunk,)
else:
# middle
sample = self.mid_block(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk)
sample = sample.to(upscale_dtype)
# up
for up_block in self.up_blocks:
sample = up_block(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk,)
# post-process
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample, is_init_image=is_init_image, temporal_chunk=temporal_chunk)
return sample
class DiagonalGaussianDistribution(object):
def __init__(self, parameters: torch.Tensor, deterministic: bool = False):
self.parameters = parameters
self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
self.deterministic = deterministic
self.std = torch.exp(0.5 * self.logvar)
self.var = torch.exp(self.logvar)
if self.deterministic:
self.var = self.std = torch.zeros_like(
self.mean, device=self.parameters.device, dtype=self.parameters.dtype
)
def sample(self, generator: Optional[torch.Generator] = None) -> torch.FloatTensor:
# make sure sample is on the same device as the parameters and has same dtype
sample = randn_tensor(
self.mean.shape,
generator=generator,
device=self.parameters.device,
dtype=self.parameters.dtype,
)
x = self.mean + self.std * sample
return x
def kl(self, other: "DiagonalGaussianDistribution" = None) -> torch.Tensor:
if self.deterministic:
return torch.Tensor([0.0])
else:
if other is None:
return 0.5 * torch.sum(
torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar,
dim=[2, 3, 4],
)
else:
return 0.5 * torch.sum(
torch.pow(self.mean - other.mean, 2) / other.var
+ self.var / other.var
- 1.0
- self.logvar
+ other.logvar,
dim=[2, 3, 4],
)
def nll(self, sample: torch.Tensor, dims: Tuple[int, ...] = [1, 2, 3]) -> torch.Tensor:
if self.deterministic:
return torch.Tensor([0.0])
logtwopi = np.log(2.0 * np.pi)
return 0.5 * torch.sum(
logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
dim=dims,
)
def mode(self) -> torch.Tensor:
return self.mean |