forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathunet_1d_blocks.py
More file actions
384 lines (292 loc) 路 14.5 KB
/
Copy pathunet_1d_blocks.py
File metadata and controls
384 lines (292 loc) 路 14.5 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
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
# Copyright 2022 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.
import math
import torch
import torch.nn.functional as F
from torch import nn
_kernels = {
"linear": [1 / 8, 3 / 8, 3 / 8, 1 / 8],
"cubic": [-0.01171875, -0.03515625, 0.11328125, 0.43359375, 0.43359375, 0.11328125, -0.03515625, -0.01171875],
"lanczos3": [
0.003689131001010537,
0.015056144446134567,
-0.03399861603975296,
-0.066637322306633,
0.13550527393817902,
0.44638532400131226,
0.44638532400131226,
0.13550527393817902,
-0.066637322306633,
-0.03399861603975296,
0.015056144446134567,
0.003689131001010537,
],
}
class Downsample1d(nn.Module):
def __init__(self, kernel="linear", pad_mode="reflect"):
super().__init__()
self.pad_mode = pad_mode
kernel_1d = torch.tensor(_kernels[kernel])
self.pad = kernel_1d.shape[0] // 2 - 1
self.register_buffer("kernel", kernel_1d)
def forward(self, hidden_states):
hidden_states = F.pad(hidden_states, (self.pad,) * 2, self.pad_mode)
weight = hidden_states.new_zeros([hidden_states.shape[1], hidden_states.shape[1], self.kernel.shape[0]])
indices = torch.arange(hidden_states.shape[1], device=hidden_states.device)
weight[indices, indices] = self.kernel.to(weight)
return F.conv1d(hidden_states, weight, stride=2)
class Upsample1d(nn.Module):
def __init__(self, kernel="linear", pad_mode="reflect"):
super().__init__()
self.pad_mode = pad_mode
kernel_1d = torch.tensor(_kernels[kernel]) * 2
self.pad = kernel_1d.shape[0] // 2 - 1
self.register_buffer("kernel", kernel_1d)
def forward(self, hidden_states):
hidden_states = F.pad(hidden_states, ((self.pad + 1) // 2,) * 2, self.pad_mode)
weight = hidden_states.new_zeros([hidden_states.shape[1], hidden_states.shape[1], self.kernel.shape[0]])
indices = torch.arange(hidden_states.shape[1], device=hidden_states.device)
weight[indices, indices] = self.kernel.to(weight)
return F.conv_transpose1d(hidden_states, weight, stride=2, padding=self.pad * 2 + 1)
class SelfAttention1d(nn.Module):
def __init__(self, in_channels, n_head=1, dropout_rate=0.0):
super().__init__()
self.channels = in_channels
self.group_norm = nn.GroupNorm(1, num_channels=in_channels)
self.num_heads = n_head
self.query = nn.Linear(self.channels, self.channels)
self.key = nn.Linear(self.channels, self.channels)
self.value = nn.Linear(self.channels, self.channels)
self.proj_attn = nn.Linear(self.channels, self.channels, 1)
self.dropout = nn.Dropout(dropout_rate, inplace=True)
def transpose_for_scores(self, projection: torch.Tensor) -> torch.Tensor:
new_projection_shape = projection.size()[:-1] + (self.num_heads, -1)
# move heads to 2nd position (B, T, H * D) -> (B, T, H, D) -> (B, H, T, D)
new_projection = projection.view(new_projection_shape).permute(0, 2, 1, 3)
return new_projection
def forward(self, hidden_states):
residual = hidden_states
batch, channel_dim, seq = hidden_states.shape
hidden_states = self.group_norm(hidden_states)
hidden_states = hidden_states.transpose(1, 2)
query_proj = self.query(hidden_states)
key_proj = self.key(hidden_states)
value_proj = self.value(hidden_states)
query_states = self.transpose_for_scores(query_proj)
key_states = self.transpose_for_scores(key_proj)
value_states = self.transpose_for_scores(value_proj)
scale = 1 / math.sqrt(math.sqrt(key_states.shape[-1]))
attention_scores = torch.matmul(query_states * scale, key_states.transpose(-1, -2) * scale)
attention_probs = torch.softmax(attention_scores, dim=-1)
# compute attention output
hidden_states = torch.matmul(attention_probs, value_states)
hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous()
new_hidden_states_shape = hidden_states.size()[:-2] + (self.channels,)
hidden_states = hidden_states.view(new_hidden_states_shape)
# compute next hidden_states
hidden_states = self.proj_attn(hidden_states)
hidden_states = hidden_states.transpose(1, 2)
hidden_states = self.dropout(hidden_states)
output = hidden_states + residual
return output
class ResConvBlock(nn.Module):
def __init__(self, in_channels, mid_channels, out_channels, is_last=False):
super().__init__()
self.is_last = is_last
self.has_conv_skip = in_channels != out_channels
if self.has_conv_skip:
self.conv_skip = nn.Conv1d(in_channels, out_channels, 1, bias=False)
self.conv_1 = nn.Conv1d(in_channels, mid_channels, 5, padding=2)
self.group_norm_1 = nn.GroupNorm(1, mid_channels)
self.gelu_1 = nn.GELU()
self.conv_2 = nn.Conv1d(mid_channels, out_channels, 5, padding=2)
if not self.is_last:
self.group_norm_2 = nn.GroupNorm(1, out_channels)
self.gelu_2 = nn.GELU()
def forward(self, hidden_states):
residual = self.conv_skip(hidden_states) if self.has_conv_skip else hidden_states
hidden_states = self.conv_1(hidden_states)
hidden_states = self.group_norm_1(hidden_states)
hidden_states = self.gelu_1(hidden_states)
hidden_states = self.conv_2(hidden_states)
if not self.is_last:
hidden_states = self.group_norm_2(hidden_states)
hidden_states = self.gelu_2(hidden_states)
output = hidden_states + residual
return output
def get_down_block(down_block_type, out_channels, in_channels):
if down_block_type == "DownBlock1D":
return DownBlock1D(out_channels=out_channels, in_channels=in_channels)
elif down_block_type == "AttnDownBlock1D":
return AttnDownBlock1D(out_channels=out_channels, in_channels=in_channels)
elif down_block_type == "DownBlock1DNoSkip":
return DownBlock1DNoSkip(out_channels=out_channels, in_channels=in_channels)
raise ValueError(f"{down_block_type} does not exist.")
def get_up_block(up_block_type, in_channels, out_channels):
if up_block_type == "UpBlock1D":
return UpBlock1D(in_channels=in_channels, out_channels=out_channels)
elif up_block_type == "AttnUpBlock1D":
return AttnUpBlock1D(in_channels=in_channels, out_channels=out_channels)
elif up_block_type == "UpBlock1DNoSkip":
return UpBlock1DNoSkip(in_channels=in_channels, out_channels=out_channels)
raise ValueError(f"{up_block_type} does not exist.")
def get_mid_block(mid_block_type, in_channels, mid_channels, out_channels):
if mid_block_type == "UNetMidBlock1D":
return UNetMidBlock1D(in_channels=in_channels, mid_channels=mid_channels, out_channels=out_channels)
raise ValueError(f"{mid_block_type} does not exist.")
class UNetMidBlock1D(nn.Module):
def __init__(self, mid_channels, in_channels, out_channels=None):
super().__init__()
out_channels = in_channels if out_channels is None else out_channels
# there is always at least one resnet
self.down = Downsample1d("cubic")
resnets = [
ResConvBlock(in_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, out_channels),
]
attentions = [
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(out_channels, out_channels // 32),
]
self.up = Upsample1d(kernel="cubic")
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
def forward(self, hidden_states):
hidden_states = self.down(hidden_states)
for attn, resnet in zip(self.attentions, self.resnets):
hidden_states = resnet(hidden_states)
hidden_states = attn(hidden_states)
hidden_states = self.up(hidden_states)
return hidden_states
class AttnDownBlock1D(nn.Module):
def __init__(self, out_channels, in_channels, mid_channels=None):
super().__init__()
mid_channels = out_channels if mid_channels is None else mid_channels
self.down = Downsample1d("cubic")
resnets = [
ResConvBlock(in_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, out_channels),
]
attentions = [
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(out_channels, out_channels // 32),
]
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
def forward(self, hidden_states, temb=None):
hidden_states = self.down(hidden_states)
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states)
hidden_states = attn(hidden_states)
return hidden_states, (hidden_states,)
class DownBlock1D(nn.Module):
def __init__(self, out_channels, in_channels, mid_channels=None):
super().__init__()
mid_channels = out_channels if mid_channels is None else mid_channels
self.down = Downsample1d("cubic")
resnets = [
ResConvBlock(in_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, out_channels),
]
self.resnets = nn.ModuleList(resnets)
def forward(self, hidden_states, temb=None):
hidden_states = self.down(hidden_states)
for resnet in self.resnets:
hidden_states = resnet(hidden_states)
return hidden_states, (hidden_states,)
class DownBlock1DNoSkip(nn.Module):
def __init__(self, out_channels, in_channels, mid_channels=None):
super().__init__()
mid_channels = out_channels if mid_channels is None else mid_channels
resnets = [
ResConvBlock(in_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, out_channels),
]
self.resnets = nn.ModuleList(resnets)
def forward(self, hidden_states, temb=None):
hidden_states = torch.cat([hidden_states, temb], dim=1)
for resnet in self.resnets:
hidden_states = resnet(hidden_states)
return hidden_states, (hidden_states,)
class AttnUpBlock1D(nn.Module):
def __init__(self, in_channels, out_channels, mid_channels=None):
super().__init__()
mid_channels = out_channels if mid_channels is None else mid_channels
resnets = [
ResConvBlock(2 * in_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, out_channels),
]
attentions = [
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(mid_channels, mid_channels // 32),
SelfAttention1d(out_channels, out_channels // 32),
]
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
self.up = Upsample1d(kernel="cubic")
def forward(self, hidden_states, res_hidden_states_tuple):
res_hidden_states = res_hidden_states_tuple[-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states)
hidden_states = attn(hidden_states)
hidden_states = self.up(hidden_states)
return hidden_states
class UpBlock1D(nn.Module):
def __init__(self, in_channels, out_channels, mid_channels=None):
super().__init__()
mid_channels = in_channels if mid_channels is None else mid_channels
resnets = [
ResConvBlock(2 * in_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, out_channels),
]
self.resnets = nn.ModuleList(resnets)
self.up = Upsample1d(kernel="cubic")
def forward(self, hidden_states, res_hidden_states_tuple):
res_hidden_states = res_hidden_states_tuple[-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
for resnet in self.resnets:
hidden_states = resnet(hidden_states)
hidden_states = self.up(hidden_states)
return hidden_states
class UpBlock1DNoSkip(nn.Module):
def __init__(self, in_channels, out_channels, mid_channels=None):
super().__init__()
mid_channels = in_channels if mid_channels is None else mid_channels
resnets = [
ResConvBlock(2 * in_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, mid_channels),
ResConvBlock(mid_channels, mid_channels, out_channels, is_last=True),
]
self.resnets = nn.ModuleList(resnets)
def forward(self, hidden_states, res_hidden_states_tuple):
res_hidden_states = res_hidden_states_tuple[-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
for resnet in self.resnets:
hidden_states = resnet(hidden_states)
return hidden_states