-
Notifications
You must be signed in to change notification settings - Fork 81
/
Copy pathcct.py
356 lines (298 loc) · 14.3 KB
/
cct.py
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
from torch.hub import load_state_dict_from_url
import torch.nn as nn
from .utils.transformers import TransformerClassifier
from .utils.tokenizer import Tokenizer
from .utils.helpers import pe_check, fc_check
try:
from timm.models.registry import register_model
except ImportError:
from .registry import register_model
model_urls = {
'cct_7_3x1_32':
'https://shi-labs.com/projects/cct/checkpoints/pretrained/cct_7_3x1_32_cifar10_300epochs.pth',
'cct_7_3x1_32_sine':
'https://shi-labs.com/projects/cct/checkpoints/pretrained/cct_7_3x1_32_sine_cifar10_5000epochs.pth',
'cct_7_3x1_32_c100':
'https://shi-labs.com/projects/cct/checkpoints/pretrained/cct_7_3x1_32_cifar100_300epochs.pth',
'cct_7_3x1_32_sine_c100':
'https://shi-labs.com/projects/cct/checkpoints/pretrained/cct_7_3x1_32_sine_cifar100_5000epochs.pth',
'cct_7_7x2_224_sine':
'https://shi-labs.com/projects/cct/checkpoints/pretrained/cct_7_7x2_224_flowers102.pth',
'cct_14_7x2_224':
'https://shi-labs.com/projects/cct/checkpoints/pretrained/cct_14_7x2_224_imagenet.pth',
'cct_14_7x2_384':
'https://shi-labs.com/projects/cct/checkpoints/finetuned/cct_14_7x2_384_imagenet.pth',
'cct_14_7x2_384_fl':
'https://shi-labs.com/projects/cct/checkpoints/finetuned/cct_14_7x2_384_flowers102.pth',
}
class CCT(nn.Module):
def __init__(self,
img_size=224,
embedding_dim=768,
n_input_channels=3,
n_conv_layers=1,
kernel_size=7,
stride=2,
padding=3,
pooling_kernel_size=3,
pooling_stride=2,
pooling_padding=1,
dropout=0.,
attention_dropout=0.1,
stochastic_depth=0.1,
num_layers=14,
num_heads=6,
mlp_ratio=4.0,
num_classes=1000,
positional_embedding='learnable',
*args, **kwargs):
super(CCT, self).__init__()
self.tokenizer = Tokenizer(n_input_channels=n_input_channels,
n_output_channels=embedding_dim,
kernel_size=kernel_size,
stride=stride,
padding=padding,
pooling_kernel_size=pooling_kernel_size,
pooling_stride=pooling_stride,
pooling_padding=pooling_padding,
max_pool=True,
activation=nn.ReLU,
n_conv_layers=n_conv_layers,
conv_bias=False)
self.classifier = TransformerClassifier(
sequence_length=self.tokenizer.sequence_length(n_channels=n_input_channels,
height=img_size,
width=img_size),
embedding_dim=embedding_dim,
seq_pool=True,
dropout=dropout,
attention_dropout=attention_dropout,
stochastic_depth=stochastic_depth,
num_layers=num_layers,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
num_classes=num_classes,
positional_embedding=positional_embedding
)
def forward(self, x):
x = self.tokenizer(x)
return self.classifier(x)
def _cct(arch, pretrained, progress,
num_layers, num_heads, mlp_ratio, embedding_dim,
kernel_size=3, stride=None, padding=None,
positional_embedding='learnable',
*args, **kwargs):
stride = stride if stride is not None else max(1, (kernel_size // 2) - 1)
padding = padding if padding is not None else max(1, (kernel_size // 2))
model = CCT(num_layers=num_layers,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
embedding_dim=embedding_dim,
kernel_size=kernel_size,
stride=stride,
padding=padding,
*args, **kwargs)
if pretrained:
if arch in model_urls:
state_dict = load_state_dict_from_url(model_urls[arch],
progress=progress)
if positional_embedding == 'learnable':
state_dict = pe_check(model, state_dict)
elif positional_embedding == 'sine':
state_dict['classifier.positional_emb'] = model.state_dict()['classifier.positional_emb']
state_dict = fc_check(model, state_dict)
model.load_state_dict(state_dict)
else:
raise RuntimeError(f'Variant {arch} does not yet have pretrained weights.')
return model
@register_model
def cct_2(arch, pretrained, progress, *args, **kwargs):
return _cct(arch, pretrained, progress, num_layers=2, num_heads=2, mlp_ratio=1, embedding_dim=128,
*args, **kwargs)
@register_model
def cct_4(arch, pretrained, progress, *args, **kwargs):
return _cct(arch, pretrained, progress, num_layers=4, num_heads=2, mlp_ratio=1, embedding_dim=128,
*args, **kwargs)
@register_model
def cct_6(arch, pretrained, progress, *args, **kwargs):
return _cct(arch, pretrained, progress, num_layers=6, num_heads=4, mlp_ratio=2, embedding_dim=256,
*args, **kwargs)
@register_model
def cct_7(arch, pretrained, progress, *args, **kwargs):
return _cct(arch, pretrained, progress, num_layers=7, num_heads=4, mlp_ratio=2, embedding_dim=256,
*args, **kwargs)
@register_model
def cct_14(arch, pretrained, progress, *args, **kwargs):
return _cct(arch, pretrained, progress, num_layers=14, num_heads=6, mlp_ratio=3, embedding_dim=384,
*args, **kwargs)
@register_model
def cct_2_3x2_32(pretrained=False, progress=False,
img_size=32, positional_embedding='learnable', num_classes=10,
*args, **kwargs):
return cct_2('cct_2_3x2_32', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_2_3x2_32_sine(pretrained=False, progress=False,
img_size=32, positional_embedding='sine', num_classes=10,
*args, **kwargs):
return cct_2('cct_2_3x2_32_sine', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_4_3x2_32(pretrained=False, progress=False,
img_size=32, positional_embedding='learnable', num_classes=10,
*args, **kwargs):
return cct_4('cct_4_3x2_32', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_4_3x2_32_sine(pretrained=False, progress=False,
img_size=32, positional_embedding='sine', num_classes=10,
*args, **kwargs):
return cct_4('cct_4_3x2_32_sine', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_6_3x1_32(pretrained=False, progress=False,
img_size=32, positional_embedding='learnable', num_classes=10,
*args, **kwargs):
return cct_6('cct_6_3x1_32', pretrained, progress,
kernel_size=3, n_conv_layers=1,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_6_3x1_32_sine(pretrained=False, progress=False,
img_size=32, positional_embedding='sine', num_classes=10,
*args, **kwargs):
return cct_6('cct_6_3x1_32_sine', pretrained, progress,
kernel_size=3, n_conv_layers=1,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_6_3x2_32(pretrained=False, progress=False,
img_size=32, positional_embedding='learnable', num_classes=10,
*args, **kwargs):
return cct_6('cct_6_3x2_32', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_6_3x2_32_sine(pretrained=False, progress=False,
img_size=32, positional_embedding='sine', num_classes=10,
*args, **kwargs):
return cct_6('cct_6_3x2_32_sine', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_3x1_32(pretrained=False, progress=False,
img_size=32, positional_embedding='learnable', num_classes=10,
*args, **kwargs):
return cct_7('cct_7_3x1_32', pretrained, progress,
kernel_size=3, n_conv_layers=1,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_3x1_32_sine(pretrained=False, progress=False,
img_size=32, positional_embedding='sine', num_classes=10,
*args, **kwargs):
return cct_7('cct_7_3x1_32_sine', pretrained, progress,
kernel_size=3, n_conv_layers=1,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_3x1_32_c100(pretrained=False, progress=False,
img_size=32, positional_embedding='learnable', num_classes=100,
*args, **kwargs):
return cct_7('cct_7_3x1_32_c100', pretrained, progress,
kernel_size=3, n_conv_layers=1,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_3x1_32_sine_c100(pretrained=False, progress=False,
img_size=32, positional_embedding='sine', num_classes=100,
*args, **kwargs):
return cct_7('cct_7_3x1_32_sine_c100', pretrained, progress,
kernel_size=3, n_conv_layers=1,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_3x2_32(pretrained=False, progress=False,
img_size=32, positional_embedding='learnable', num_classes=10,
*args, **kwargs):
return cct_7('cct_7_3x2_32', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_3x2_32_sine(pretrained=False, progress=False,
img_size=32, positional_embedding='sine', num_classes=10,
*args, **kwargs):
return cct_7('cct_7_3x2_32_sine', pretrained, progress,
kernel_size=3, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_7x2_224(pretrained=False, progress=False,
img_size=224, positional_embedding='learnable', num_classes=102,
*args, **kwargs):
return cct_7('cct_7_7x2_224', pretrained, progress,
kernel_size=7, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_7_7x2_224_sine(pretrained=False, progress=False,
img_size=224, positional_embedding='sine', num_classes=102,
*args, **kwargs):
return cct_7('cct_7_7x2_224_sine', pretrained, progress,
kernel_size=7, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_14_7x2_224(pretrained=False, progress=False,
img_size=224, positional_embedding='learnable', num_classes=1000,
*args, **kwargs):
return cct_14('cct_14_7x2_224', pretrained, progress,
kernel_size=7, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_14_7x2_384(pretrained=False, progress=False,
img_size=384, positional_embedding='learnable', num_classes=1000,
*args, **kwargs):
return cct_14('cct_14_7x2_384', pretrained, progress,
kernel_size=7, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)
@register_model
def cct_14_7x2_384_fl(pretrained=False, progress=False,
img_size=384, positional_embedding='learnable', num_classes=102,
*args, **kwargs):
return cct_14('cct_14_7x2_384_fl', pretrained, progress,
kernel_size=7, n_conv_layers=2,
img_size=img_size, positional_embedding=positional_embedding,
num_classes=num_classes,
*args, **kwargs)