{{ message }}
Repository navigation
Expand file tree
/
Copy pathtest_layers_utils.py
More file actions
executable file
·326 lines (261 loc) · 13.2 KB
/
Copy pathtest_layers_utils.py
File metadata and controls
executable file
·326 lines (261 loc) · 13.2 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
# coding=utf-8
# Copyright 2022 HuggingFace Inc.
#
# 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 inspect
import tempfile
import unittest
import numpy as np
import torch
from diffusers.models.embeddings import get_timestep_embedding
from diffusers.models.resnet import Downsample1D, Downsample2D, Upsample1D, Upsample2D
from diffusers.testing_utils import floats_tensor, slow, torch_device
torch.backends.cuda.matmul.allow_tf32 = False
class EmbeddingsTests(unittest.TestCase):
def test_timestep_embeddings(self):
embedding_dim = 256
timesteps = torch.arange(16)
t1 = get_timestep_embedding(timesteps, embedding_dim)
# first vector should always be composed only of 0's and 1's
assert (t1[0, : embedding_dim // 2] - 0).abs().sum() < 1e-5
assert (t1[0, embedding_dim // 2 :] - 1).abs().sum() < 1e-5
# last element of each vector should be one
assert (t1[:, -1] - 1).abs().sum() < 1e-5
# For large embeddings (e.g. 128) the frequency of every vector is higher
# than the previous one which means that the gradients of later vectors are
# ALWAYS higher than the previous ones
grad_mean = np.abs(np.gradient(t1, axis=-1)).mean(axis=1)
prev_grad = 0.0
for grad in grad_mean:
assert grad > prev_grad
prev_grad = grad
def test_timestep_defaults(self):
embedding_dim = 16
timesteps = torch.arange(10)
t1 = get_timestep_embedding(timesteps, embedding_dim)
t2 = get_timestep_embedding(
timesteps, embedding_dim, flip_sin_to_cos=False, downscale_freq_shift=1, max_period=10_000
)
assert torch.allclose(t1.cpu(), t2.cpu(), 1e-3)
def test_timestep_flip_sin_cos(self):
embedding_dim = 16
timesteps = torch.arange(10)
t1 = get_timestep_embedding(timesteps, embedding_dim, flip_sin_to_cos=True)
t1 = torch.cat([t1[:, embedding_dim // 2 :], t1[:, : embedding_dim // 2]], dim=-1)
t2 = get_timestep_embedding(timesteps, embedding_dim, flip_sin_to_cos=False)
assert torch.allclose(t1.cpu(), t2.cpu(), 1e-3)
def test_timestep_downscale_freq_shift(self):
embedding_dim = 16
timesteps = torch.arange(10)
t1 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=0)
t2 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=1)
# get cosine half (vectors that are wrapped into cosine)
cosine_half = (t1 - t2)[:, embedding_dim // 2 :]
# cosine needs to be negative
assert (np.abs((cosine_half <= 0).numpy()) - 1).sum() < 1e-5
def test_sinoid_embeddings_hardcoded(self):
embedding_dim = 64
timesteps = torch.arange(128)
# standard unet, score_vde
t1 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=1, flip_sin_to_cos=False)
# glide, ldm
t2 = get_timestep_embedding(timesteps, embedding_dim, downscale_freq_shift=0, flip_sin_to_cos=True)
# grad-tts
t3 = get_timestep_embedding(timesteps, embedding_dim, scale=1000)
assert torch.allclose(
t1[23:26, 47:50].flatten().cpu(),
torch.tensor([0.9646, 0.9804, 0.9892, 0.9615, 0.9787, 0.9882, 0.9582, 0.9769, 0.9872]),
1e-3,
)
assert torch.allclose(
t2[23:26, 47:50].flatten().cpu(),
torch.tensor([0.3019, 0.2280, 0.1716, 0.3146, 0.2377, 0.1790, 0.3272, 0.2474, 0.1864]),
1e-3,
)
assert torch.allclose(
t3[23:26, 47:50].flatten().cpu(),
torch.tensor([-0.9801, -0.9464, -0.9349, -0.3952, 0.8887, -0.9709, 0.5299, -0.2853, -0.9927]),
1e-3,
)
class Upsample2DBlockTests(unittest.TestCase):
def test_upsample_default(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32, 32)
upsample = Upsample2D(channels=32, use_conv=False)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 32, 64, 64)
output_slice = upsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor([-0.2173, -1.2079, -1.2079, 0.2952, 1.1254, 1.1254, 0.2952, 1.1254, 1.1254])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_upsample_with_conv(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32, 32)
upsample = Upsample2D(channels=32, use_conv=True)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 32, 64, 64)
output_slice = upsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor([0.7145, 1.3773, 0.3492, 0.8448, 1.0839, -0.3341, 0.5956, 0.1250, -0.4841])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_upsample_with_conv_out_dim(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32, 32)
upsample = Upsample2D(channels=32, use_conv=True, out_channels=64)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 64, 64, 64)
output_slice = upsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor([0.2703, 0.1656, -0.2538, -0.0553, -0.2984, 0.1044, 0.1155, 0.2579, 0.7755])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_upsample_with_transpose(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32, 32)
upsample = Upsample2D(channels=32, use_conv=False, use_conv_transpose=True)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 32, 64, 64)
output_slice = upsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor([-0.3028, -0.1582, 0.0071, 0.0350, -0.4799, -0.1139, 0.1056, -0.1153, -0.1046])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
class Downsample2DBlockTests(unittest.TestCase):
def test_downsample_default(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64, 64)
downsample = Downsample2D(channels=32, use_conv=False)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 32, 32, 32)
output_slice = downsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor([-0.0513, -0.3889, 0.0640, 0.0836, -0.5460, -0.0341, -0.0169, -0.6967, 0.1179])
max_diff = (output_slice.flatten() - expected_slice).abs().sum().item()
assert max_diff <= 1e-3
# assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-1)
def test_downsample_with_conv(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64, 64)
downsample = Downsample2D(channels=32, use_conv=True)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 32, 32, 32)
output_slice = downsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor(
[0.9267, 0.5878, 0.3337, 1.2321, -0.1191, -0.3984, -0.7532, -0.0715, -0.3913],
)
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_downsample_with_conv_pad1(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64, 64)
downsample = Downsample2D(channels=32, use_conv=True, padding=1)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 32, 32, 32)
output_slice = downsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor([0.9267, 0.5878, 0.3337, 1.2321, -0.1191, -0.3984, -0.7532, -0.0715, -0.3913])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_downsample_with_conv_out_dim(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64, 64)
downsample = Downsample2D(channels=32, use_conv=True, out_channels=16)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 16, 32, 32)
output_slice = downsampled[0, -1, -3:, -3:]
expected_slice = torch.tensor([-0.6586, 0.5985, 0.0721, 0.1256, -0.1492, 0.4436, -0.2544, 0.5021, 1.1522])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
class Upsample1DBlockTests(unittest.TestCase):
def test_upsample_default(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32)
upsample = Upsample1D(channels=32, use_conv=False)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 32, 64)
output_slice = upsampled[0, -1, -8:]
expected_slice = torch.tensor([-1.6340, -1.6340, 0.5374, 0.5374, 1.0826, 1.0826, -1.7105, -1.7105])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_upsample_with_conv(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32)
upsample = Upsample1D(channels=32, use_conv=True)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 32, 64)
output_slice = upsampled[0, -1, -8:]
expected_slice = torch.tensor([-0.4546, -0.5010, -0.2996, 0.2844, 0.4040, -0.7772, -0.6862, 0.3612])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_upsample_with_conv_out_dim(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32)
upsample = Upsample1D(channels=32, use_conv=True, out_channels=64)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 64, 64)
output_slice = upsampled[0, -1, -8:]
expected_slice = torch.tensor([-0.0516, -0.0972, 0.9740, 1.1883, 0.4539, -0.5285, -0.5851, 0.1152])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_upsample_with_transpose(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 32)
upsample = Upsample1D(channels=32, use_conv=False, use_conv_transpose=True)
with torch.no_grad():
upsampled = upsample(sample)
assert upsampled.shape == (1, 32, 64)
output_slice = upsampled[0, -1, -8:]
expected_slice = torch.tensor([-0.2238, -0.5842, -0.7165, 0.6699, 0.1033, -0.4269, -0.8974, -0.3716])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
class Downsample1DBlockTests(unittest.TestCase):
def test_downsample_default(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64)
downsample = Downsample1D(channels=32, use_conv=False)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 32, 32)
output_slice = downsampled[0, -1, -8:]
expected_slice = torch.tensor([-0.8796, 1.0945, -0.3434, 0.2910, 0.3391, -0.4488, -0.9568, -0.2909])
max_diff = (output_slice.flatten() - expected_slice).abs().sum().item()
assert max_diff <= 1e-3
# assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-1)
def test_downsample_with_conv(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64)
downsample = Downsample1D(channels=32, use_conv=True)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 32, 32)
output_slice = downsampled[0, -1, -8:]
expected_slice = torch.tensor(
[0.1723, 0.0811, -0.6205, -0.3045, 0.0666, -0.2381, -0.0238, 0.2834],
)
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_downsample_with_conv_pad1(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64)
downsample = Downsample1D(channels=32, use_conv=True, padding=1)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 32, 32)
output_slice = downsampled[0, -1, -8:]
expected_slice = torch.tensor([0.1723, 0.0811, -0.6205, -0.3045, 0.0666, -0.2381, -0.0238, 0.2834])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
def test_downsample_with_conv_out_dim(self):
torch.manual_seed(0)
sample = torch.randn(1, 32, 64)
downsample = Downsample1D(channels=32, use_conv=True, out_channels=16)
with torch.no_grad():
downsampled = downsample(sample)
assert downsampled.shape == (1, 16, 32)
output_slice = downsampled[0, -1, -8:]
expected_slice = torch.tensor([1.1067, -0.5255, -0.4451, 0.0487, -0.3664, -0.7945, -0.4495, -0.3129])
assert torch.allclose(output_slice.flatten(), expected_slice, atol=1e-3)
You can’t perform that action at this time.