forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathvae.py
More file actions
356 lines (295 loc) 路 12.3 KB
/
Copy pathvae.py
File metadata and controls
356 lines (295 loc) 路 12.3 KB
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
# 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
import numpy as np
import torch
import torch.nn as nn
from ..utils import BaseOutput, randn_tensor
from .unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block
@dataclass
class DecoderOutput(BaseOutput):
"""
Output of decoding method.
Args:
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Decoded output sample of the model. Output of the last layer of the model.
"""
sample: torch.FloatTensor
class Encoder(nn.Module):
def __init__(
self,
in_channels=3,
out_channels=3,
down_block_types=("DownEncoderBlock2D",),
block_out_channels=(64,),
layers_per_block=2,
norm_num_groups=32,
act_fn="silu",
double_z=True,
):
super().__init__()
self.layers_per_block = layers_per_block
self.conv_in = torch.nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, stride=1, padding=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]
is_final_block = i == len(block_out_channels) - 1
down_block = get_down_block(
down_block_type,
num_layers=self.layers_per_block,
in_channels=input_channel,
out_channels=output_channel,
add_downsample=not is_final_block,
resnet_eps=1e-6,
downsample_padding=0,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
attn_num_head_channels=None,
temb_channels=None,
)
self.down_blocks.append(down_block)
# mid
self.mid_block = UNetMidBlock2D(
in_channels=block_out_channels[-1],
resnet_eps=1e-6,
resnet_act_fn=act_fn,
output_scale_factor=1,
resnet_time_scale_shift="default",
attn_num_head_channels=None,
resnet_groups=norm_num_groups,
temb_channels=None,
)
# out
self.conv_norm_out = nn.GroupNorm(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 = nn.Conv2d(block_out_channels[-1], conv_out_channels, 3, padding=1)
def forward(self, x):
sample = x
sample = self.conv_in(sample)
# down
for down_block in self.down_blocks:
sample = down_block(sample)
# middle
sample = self.mid_block(sample)
# post-process
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
return sample
class Decoder(nn.Module):
def __init__(
self,
in_channels=3,
out_channels=3,
up_block_types=("UpDecoderBlock2D",),
block_out_channels=(64,),
layers_per_block=2,
norm_num_groups=32,
act_fn="silu",
):
super().__init__()
self.layers_per_block = layers_per_block
self.conv_in = nn.Conv2d(in_channels, block_out_channels[-1], kernel_size=3, stride=1, padding=1)
self.mid_block = None
self.up_blocks = nn.ModuleList([])
# mid
self.mid_block = UNetMidBlock2D(
in_channels=block_out_channels[-1],
resnet_eps=1e-6,
resnet_act_fn=act_fn,
output_scale_factor=1,
resnet_time_scale_shift="default",
attn_num_head_channels=None,
resnet_groups=norm_num_groups,
temb_channels=None,
)
# 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 + 1,
in_channels=prev_output_channel,
out_channels=output_channel,
prev_output_channel=None,
add_upsample=not is_final_block,
resnet_eps=1e-6,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
attn_num_head_channels=None,
temb_channels=None,
)
self.up_blocks.append(up_block)
prev_output_channel = output_channel
# out
self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
self.conv_act = nn.SiLU()
self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
def forward(self, z):
sample = z
sample = self.conv_in(sample)
# middle
sample = self.mid_block(sample)
# up
for up_block in self.up_blocks:
sample = up_block(sample)
# post-process
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
return sample
class VectorQuantizer(nn.Module):
"""
Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly avoids costly matrix
multiplications and allows for post-hoc remapping of indices.
"""
# NOTE: due to a bug the beta term was applied to the wrong term. for
# backwards compatibility we use the buggy version by default, but you can
# specify legacy=False to fix it.
def __init__(
self, n_e, vq_embed_dim, beta, remap=None, unknown_index="random", sane_index_shape=False, legacy=True
):
super().__init__()
self.n_e = n_e
self.vq_embed_dim = vq_embed_dim
self.beta = beta
self.legacy = legacy
self.embedding = nn.Embedding(self.n_e, self.vq_embed_dim)
self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
self.remap = remap
if self.remap is not None:
self.register_buffer("used", torch.tensor(np.load(self.remap)))
self.re_embed = self.used.shape[0]
self.unknown_index = unknown_index # "random" or "extra" or integer
if self.unknown_index == "extra":
self.unknown_index = self.re_embed
self.re_embed = self.re_embed + 1
print(
f"Remapping {self.n_e} indices to {self.re_embed} indices. "
f"Using {self.unknown_index} for unknown indices."
)
else:
self.re_embed = n_e
self.sane_index_shape = sane_index_shape
def remap_to_used(self, inds):
ishape = inds.shape
assert len(ishape) > 1
inds = inds.reshape(ishape[0], -1)
used = self.used.to(inds)
match = (inds[:, :, None] == used[None, None, ...]).long()
new = match.argmax(-1)
unknown = match.sum(2) < 1
if self.unknown_index == "random":
new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
else:
new[unknown] = self.unknown_index
return new.reshape(ishape)
def unmap_to_all(self, inds):
ishape = inds.shape
assert len(ishape) > 1
inds = inds.reshape(ishape[0], -1)
used = self.used.to(inds)
if self.re_embed > self.used.shape[0]: # extra token
inds[inds >= self.used.shape[0]] = 0 # simply set to zero
back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
return back.reshape(ishape)
def forward(self, z):
# reshape z -> (batch, height, width, channel) and flatten
z = z.permute(0, 2, 3, 1).contiguous()
z_flattened = z.view(-1, self.vq_embed_dim)
# distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
min_encoding_indices = torch.argmin(torch.cdist(z_flattened, self.embedding.weight), dim=1)
z_q = self.embedding(min_encoding_indices).view(z.shape)
perplexity = None
min_encodings = None
# compute loss for embedding
if not self.legacy:
loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + torch.mean((z_q - z.detach()) ** 2)
else:
loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean((z_q - z.detach()) ** 2)
# preserve gradients
z_q = z + (z_q - z).detach()
# reshape back to match original input shape
z_q = z_q.permute(0, 3, 1, 2).contiguous()
if self.remap is not None:
min_encoding_indices = min_encoding_indices.reshape(z.shape[0], -1) # add batch axis
min_encoding_indices = self.remap_to_used(min_encoding_indices)
min_encoding_indices = min_encoding_indices.reshape(-1, 1) # flatten
if self.sane_index_shape:
min_encoding_indices = min_encoding_indices.reshape(z_q.shape[0], z_q.shape[2], z_q.shape[3])
return z_q, loss, (perplexity, min_encodings, min_encoding_indices)
def get_codebook_entry(self, indices, shape):
# shape specifying (batch, height, width, channel)
if self.remap is not None:
indices = indices.reshape(shape[0], -1) # add batch axis
indices = self.unmap_to_all(indices)
indices = indices.reshape(-1) # flatten again
# get quantized latent vectors
z_q = self.embedding(indices)
if shape is not None:
z_q = z_q.view(shape)
# reshape back to match original input shape
z_q = z_q.permute(0, 3, 1, 2).contiguous()
return z_q
class DiagonalGaussianDistribution(object):
def __init__(self, parameters, deterministic=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=None):
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=[1, 2, 3])
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=[1, 2, 3],
)
def nll(self, sample, dims=[1, 2, 3]):
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):
return self.mean