File size: 15,366 Bytes
cdcfdd8 |
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 |
from typing import Optional, Tuple, Union
from einops import rearrange
import torch
import torch.nn as nn
from diffusers.models.attention_processor import Attention
from diffusers.models.resnet import ResnetBlock2D
from diffusers.models.upsampling import Upsample2D
from diffusers.models.downsampling import Downsample2D
class TemporalConvBlock(nn.Module):
"""
Temporal convolutional layer that can be used for video (sequence of images) input Code mostly copied from:
https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/models/multi_modal/video_synthesis/unet_sd.py#L1016
"""
def __init__(self, in_dim, out_dim=None, dropout=0.0, up_sample=False, down_sample=False, spa_stride=1):
super().__init__()
out_dim = out_dim or in_dim
self.in_dim = in_dim
self.out_dim = out_dim
spa_pad = int((spa_stride-1)*0.5)
temp_pad = 0
self.temp_pad = temp_pad
if down_sample:
self.conv1 = nn.Sequential(
nn.GroupNorm(32, in_dim),
nn.SiLU(),
nn.Conv3d(in_dim, out_dim, (2, spa_stride, spa_stride), stride=(2,1,1), padding=(0, spa_pad, spa_pad))
)
elif up_sample:
self.conv1 = nn.Sequential(
nn.GroupNorm(32, in_dim),
nn.SiLU(),
nn.Conv3d(in_dim, out_dim*2, (1, spa_stride, spa_stride), padding=(0, spa_pad, spa_pad))
)
else:
self.conv1 = nn.Sequential(
nn.GroupNorm(32, in_dim),
nn.SiLU(),
nn.Conv3d(in_dim, out_dim, (3, spa_stride, spa_stride), padding=(temp_pad, spa_pad, spa_pad))
)
self.conv2 = nn.Sequential(
nn.GroupNorm(32, out_dim),
nn.SiLU(),
nn.Dropout(dropout),
nn.Conv3d(out_dim, in_dim, (3, spa_stride, spa_stride), padding=(temp_pad, spa_pad, spa_pad)),
)
self.conv3 = nn.Sequential(
nn.GroupNorm(32, out_dim),
nn.SiLU(),
nn.Dropout(dropout),
nn.Conv3d(out_dim, in_dim, (3, spa_stride, spa_stride), padding=(temp_pad, spa_pad, spa_pad)),
)
self.conv4 = nn.Sequential(
nn.GroupNorm(32, out_dim),
nn.SiLU(),
nn.Conv3d(out_dim, in_dim, (3, spa_stride, spa_stride), padding=(temp_pad, spa_pad, spa_pad)),
)
# zero out the last layer params,so the conv block is identity
nn.init.zeros_(self.conv4[-1].weight)
nn.init.zeros_(self.conv4[-1].bias)
self.down_sample = down_sample
self.up_sample = up_sample
def forward(self, hidden_states):
identity = hidden_states
if self.down_sample:
identity = identity[:,:,::2]
elif self.up_sample:
hidden_states_new = torch.cat((hidden_states,hidden_states),dim=2)
hidden_states_new[:, :, 0::2] = hidden_states
hidden_states_new[:, :, 1::2] = hidden_states
identity = hidden_states_new
del hidden_states_new
if self.down_sample or self.up_sample:
hidden_states = self.conv1(hidden_states)
else:
hidden_states = torch.cat((hidden_states[:,:,0:1], hidden_states), dim=2)
hidden_states = torch.cat((hidden_states,hidden_states[:,:,-1:]), dim=2)
hidden_states = self.conv1(hidden_states)
if self.up_sample:
hidden_states = rearrange(hidden_states, 'b (d c) f h w -> b c (f d) h w', d=2)
hidden_states = torch.cat((hidden_states[:,:,0:1], hidden_states), dim=2)
hidden_states = torch.cat((hidden_states,hidden_states[:,:,-1:]), dim=2)
hidden_states = self.conv2(hidden_states)
hidden_states = torch.cat((hidden_states[:,:,0:1], hidden_states), dim=2)
hidden_states = torch.cat((hidden_states,hidden_states[:,:,-1:]), dim=2)
hidden_states = self.conv3(hidden_states)
hidden_states = torch.cat((hidden_states[:,:,0:1], hidden_states), dim=2)
hidden_states = torch.cat((hidden_states,hidden_states[:,:,-1:]), dim=2)
hidden_states = self.conv4(hidden_states)
hidden_states = identity + hidden_states
return hidden_states
class DownEncoderBlock3D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor=1.0,
add_downsample=True,
add_temp_downsample=False,
downsample_padding=1,
):
super().__init__()
resnets = []
temp_convs = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=None,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
temp_convs.append(
TemporalConvBlock(
out_channels,
out_channels,
dropout=0.1,
)
)
self.resnets = nn.ModuleList(resnets)
self.temp_convs = nn.ModuleList(temp_convs)
if add_temp_downsample:
self.temp_convs_down = TemporalConvBlock(
out_channels,
out_channels,
dropout=0.1,
down_sample=True,
spa_stride=3
)
self.add_temp_downsample = add_temp_downsample
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
def _set_partial_grad(self):
for temp_conv in self.temp_convs:
temp_conv.requires_grad_(True)
if self.downsamplers:
for down_layer in self.downsamplers:
down_layer.requires_grad_(True)
def forward(self, hidden_states):
bz = hidden_states.shape[0]
for resnet, temp_conv in zip(self.resnets, self.temp_convs):
hidden_states = rearrange(hidden_states, 'b c n h w -> (b n) c h w')
hidden_states = resnet(hidden_states, temb=None)
hidden_states = rearrange(hidden_states, '(b n) c h w -> b c n h w', b=bz)
hidden_states = temp_conv(hidden_states)
if self.add_temp_downsample:
hidden_states = self.temp_convs_down(hidden_states)
if self.downsamplers is not None:
hidden_states = rearrange(hidden_states, 'b c n h w -> (b n) c h w')
for upsampler in self.downsamplers:
hidden_states = upsampler(hidden_states)
hidden_states = rearrange(hidden_states, '(b n) c h w -> b c n h w', b=bz)
return hidden_states
class UpDecoderBlock3D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default", # default, spatial
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor=1.0,
add_upsample=True,
add_temp_upsample=False,
temb_channels=None,
):
super().__init__()
self.add_upsample = add_upsample
resnets = []
temp_convs = []
for i in range(num_layers):
input_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=input_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
temp_convs.append(
TemporalConvBlock(
out_channels,
out_channels,
dropout=0.1,
)
)
self.resnets = nn.ModuleList(resnets)
self.temp_convs = nn.ModuleList(temp_convs)
self.add_temp_upsample = add_temp_upsample
if add_temp_upsample:
self.temp_conv_up = TemporalConvBlock(
out_channels,
out_channels,
dropout=0.1,
up_sample=True,
spa_stride=3
)
if self.add_upsample:
# self.upsamplers = nn.ModuleList([PSUpsample2D(out_channels, use_conv=True, use_pixel_shuffle=True, out_channels=out_channels)])
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
def _set_partial_grad(self):
for temp_conv in self.temp_convs:
temp_conv.requires_grad_(True)
if self.add_upsample:
self.upsamplers.requires_grad_(True)
def forward(self, hidden_states):
bz = hidden_states.shape[0]
for resnet, temp_conv in zip(self.resnets, self.temp_convs):
hidden_states = rearrange(hidden_states, 'b c n h w -> (b n) c h w')
hidden_states = resnet(hidden_states, temb=None)
hidden_states = rearrange(hidden_states, '(b n) c h w -> b c n h w', b=bz)
hidden_states = temp_conv(hidden_states)
if self.add_temp_upsample:
hidden_states = self.temp_conv_up(hidden_states)
if self.upsamplers is not None:
hidden_states = rearrange(hidden_states, 'b c n h w -> (b n) c h w')
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states)
hidden_states = rearrange(hidden_states, '(b n) c h w -> b c n h w', b=bz)
return hidden_states
class UNetMidBlock3DConv(nn.Module):
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default", # default, spatial
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
add_attention: bool = True,
attention_head_dim=1,
output_scale_factor=1.0,
):
super().__init__()
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
self.add_attention = add_attention
# there is always at least one resnet
resnets = [
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
]
temp_convs = [
TemporalConvBlock(
in_channels,
in_channels,
dropout=0.1,
)
]
attentions = []
if attention_head_dim is None:
attention_head_dim = in_channels
for _ in range(num_layers):
if self.add_attention:
attentions.append(
Attention(
in_channels,
heads=in_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups if resnet_time_scale_shift == "default" else None,
spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
else:
attentions.append(None)
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
temp_convs.append(
TemporalConvBlock(
in_channels,
in_channels,
dropout=0.1,
)
)
self.resnets = nn.ModuleList(resnets)
self.temp_convs = nn.ModuleList(temp_convs)
self.attentions = nn.ModuleList(attentions)
def _set_partial_grad(self):
for temp_conv in self.temp_convs:
temp_conv.requires_grad_(True)
def forward(
self,
hidden_states,
):
bz = hidden_states.shape[0]
hidden_states = rearrange(hidden_states, 'b c n h w -> (b n) c h w')
hidden_states = self.resnets[0](hidden_states, temb=None)
hidden_states = rearrange(hidden_states, '(b n) c h w -> b c n h w', b=bz)
hidden_states = self.temp_convs[0](hidden_states)
hidden_states = rearrange(hidden_states, 'b c n h w -> (b n) c h w')
for attn, resnet, temp_conv in zip(
self.attentions, self.resnets[1:], self.temp_convs[1:]
):
hidden_states = attn(hidden_states)
hidden_states = resnet(hidden_states, temb=None)
hidden_states = rearrange(hidden_states, '(b n) c h w -> b c n h w', b=bz)
hidden_states = temp_conv(hidden_states)
return hidden_states
|