-
Notifications
You must be signed in to change notification settings - Fork 22
/
isonet.py
422 lines (363 loc) · 14.9 KB
/
isonet.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
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
416
417
418
419
420
421
422
#!/usr/bin/env python3
# This file is modified from https://github.com/facebookresearch/pycls/blob/master/pycls/models/resnet.py
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from isonet.utils.config import C
# Stage depths for ImageNet models
_IN_STAGE_DS = {
18: (2, 2, 2, 2),
34: (3, 4, 6, 3),
46: (3, 4, 12, 3),
50: (3, 4, 6, 3),
101: (3, 4, 23, 3),
152: (3, 8, 36, 3),
}
def get_trans_fun(name):
"""Retrieves the transformation function by name."""
trans_funs = {
'basic_transform': BasicTransform,
'bottleneck_transform': BottleneckTransform,
}
assert name in trans_funs.keys(), \
'Transformation function \'{}\' not supported'.format(name)
return trans_funs[name]
class SReLU(nn.Module):
"""Shifted ReLU"""
def __init__(self, nc):
super(SReLU, self).__init__()
self.srelu_bias = nn.Parameter(torch.Tensor(1, nc, 1, 1))
self.srelu_relu = nn.ReLU(inplace=True)
nn.init.constant_(self.srelu_bias, -1.0)
def forward(self, x):
return self.srelu_relu(x - self.srelu_bias) + self.srelu_bias
class SharedScale(nn.Module):
"""Channel-shared scalar"""
def __init__(self):
super(SharedScale, self).__init__()
self.scale = nn.Parameter(torch.ones(1, 1, 1, 1) * C.ISON.RES_MULTIPLIER)
def forward(self, x):
return x * self.scale
class ResHead(nn.Module):
"""ResNet head."""
def __init__(self, w_in, nc):
super(ResHead, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
if C.ISON.DROPOUT:
self.dropout = nn.Dropout(p=C.ISON.DROPOUT_RATE, inplace=True)
self.fc = nn.Linear(w_in, nc, bias=True)
def forward(self, x):
x = self.avg_pool(x)
x = x.view(x.size(0), -1)
if C.ISON.DROPOUT:
x = self.dropout(x)
x = self.fc(x)
return x
class BasicTransform(nn.Module):
"""Basic transformation: 3x3, 3x3"""
def __init__(self, w_in, w_out, stride, w_b=None, num_gs=1):
assert w_b is None and num_gs == 1, \
'Basic transform does not support w_b and num_gs options'
super(BasicTransform, self).__init__()
self._construct(w_in, w_out, stride)
def _construct(self, w_in, w_out, stride):
# 3x3, BN, ReLU
self.a = nn.Conv2d(
w_in, w_out, kernel_size=3,
stride=stride, padding=1, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.a_bn = nn.BatchNorm2d(w_out)
self.a_relu = nn.ReLU(inplace=True) if not C.ISON.SReLU else SReLU(w_out)
# 3x3, BN
self.b = nn.Conv2d(
w_out, w_out, kernel_size=3,
stride=1, padding=1, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.b_bn = nn.BatchNorm2d(w_out)
self.b_bn.final_bn = True
if C.ISON.HAS_RES_MULTIPLIER:
self.shared_scalar = SharedScale()
def forward(self, x):
for layer in self.children():
x = layer(x)
return x
class BottleneckTransform(nn.Module):
"""Bottleneck transformation: 1x1, 3x3, 1x1"""
def __init__(self, w_in, w_out, stride, w_b, num_gs):
super(BottleneckTransform, self).__init__()
self._construct(w_in, w_out, stride, w_b, num_gs)
def _construct(self, w_in, w_out, stride, w_b, num_gs):
# MSRA -> stride=2 is on 1x1; TH/C2 -> stride=2 is on 3x3
(str1x1, str3x3) = (1, stride)
# 1x1, BN, ReLU
self.a = nn.Conv2d(
w_in, w_b, kernel_size=1,
stride=str1x1, padding=0, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.a_bn = nn.BatchNorm2d(w_b)
self.a_relu = nn.ReLU(inplace=True) if not C.ISON.SReLU else SReLU(w_b)
# 3x3, BN, ReLU
self.b = nn.Conv2d(
w_b, w_b, kernel_size=3,
stride=str3x3, padding=1, groups=num_gs, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.b_bn = nn.BatchNorm2d(w_b)
self.b_relu = nn.ReLU(inplace=True) if not C.ISON.SReLU else SReLU(w_b)
# 1x1, BN
self.c = nn.Conv2d(
w_b, w_out, kernel_size=1,
stride=1, padding=0, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.c_bn = nn.BatchNorm2d(w_out)
self.c_bn.final_bn = True
if C.ISON.HAS_RES_MULTIPLIER:
self.shared_scalar = SharedScale()
def forward(self, x):
for layer in self.children():
x = layer(x)
return x
class ResBlock(nn.Module):
"""Residual block: x + F(x)"""
def __init__(
self, w_in, w_out, stride, trans_fun, w_b=None, num_gs=1
):
super(ResBlock, self).__init__()
self._construct(w_in, w_out, stride, trans_fun, w_b, num_gs)
def _add_skip_proj(self, w_in, w_out, stride):
self.proj = nn.Conv2d(
w_in, w_out, kernel_size=1,
stride=stride, padding=0, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.bn = nn.BatchNorm2d(w_out)
def _construct(self, w_in, w_out, stride, trans_fun, w_b, num_gs):
# Use skip connection with projection if shape changes
self.proj_block = (w_in != w_out) or (stride != 1)
if self.proj_block and C.ISON.HAS_ST:
self._add_skip_proj(w_in, w_out, stride)
self.f = trans_fun(w_in, w_out, stride, w_b, num_gs)
self.relu = nn.ReLU(True) if not C.ISON.SReLU else SReLU(w_out)
def forward(self, x):
if self.proj_block:
if C.ISON.HAS_BN and C.ISON.HAS_ST:
x = self.bn(self.proj(x)) + self.f(x)
elif not C.ISON.HAS_BN and C.ISON.HAS_ST:
x = self.proj(x) + self.f(x)
else:
x = self.f(x)
else:
if C.ISON.HAS_ST:
x = x + self.f(x)
else:
x = self.f(x)
x = self.relu(x)
return x
class ResStage(nn.Module):
"""Stage of ResNet."""
def __init__(self, w_in, w_out, stride, d, w_b=None, num_gs=1):
super(ResStage, self).__init__()
self._construct(w_in, w_out, stride, d, w_b, num_gs)
def _construct(self, w_in, w_out, stride, d, w_b, num_gs):
# Construct the blocks
for i in range(d):
# Stride and w_in apply to the first block of the stage
b_stride = stride if i == 0 else 1
b_w_in = w_in if i == 0 else w_out
# Retrieve the transformation function
trans_fun = get_trans_fun(C.ISON.TRANS_FUN)
# Construct the block
res_block = ResBlock(
b_w_in, w_out, b_stride, trans_fun, w_b, num_gs
)
self.add_module('b{}'.format(i + 1), res_block)
def forward(self, x):
for block in self.children():
x = block(x)
return x
class ResStem(nn.Module):
"""Stem of ResNet."""
def __init__(self, w_in, w_out):
super(ResStem, self).__init__()
if 'CIFAR' in C.DATASET.NAME:
self._construct_cifar(w_in, w_out)
else:
self._construct_imagenet(w_in, w_out)
def _construct_cifar(self, w_in, w_out):
# 3x3, BN, ReLU
self.conv = nn.Conv2d(
w_in, w_out, kernel_size=3,
stride=1, padding=1, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.bn = nn.BatchNorm2d(w_out)
self.relu = nn.ReLU(True) if not C.ISON.SReLU else SReLU(w_out)
def _construct_imagenet(self, w_in, w_out):
# 7x7, BN, ReLU, maxpool
self.conv = nn.Conv2d(
w_in, w_out, kernel_size=7,
stride=2, padding=3, bias=not C.ISON.HAS_BN and not C.ISON.SReLU
)
if C.ISON.HAS_BN:
self.bn = nn.BatchNorm2d(w_out)
self.relu = nn.ReLU(True) if not C.ISON.SReLU else SReLU(w_out)
self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
def forward(self, x):
for layer in self.children():
x = layer(x)
return x
class ISONet(nn.Module):
"""ResNet model."""
def __init__(self):
super(ISONet, self).__init__()
# define network structures
if 'CIFAR' in C.DATASET.NAME:
self._construct_cifar()
elif C.ISON.TRANS_FUN == 'basic_transform':
self._construct_imagenet_basic()
elif C.ISON.TRANS_FUN == 'bottleneck_transform':
self._construct_imagenet()
else:
raise NotImplementedError
# initialization
self._network_init()
def _construct_cifar(self):
assert (C.ISON.DEPTH - 2) % 6 == 0, \
'Model depth should be of the format 6n + 2 for cifar'
# Each stage has the same number of blocks for cifar
d = int((C.ISON.DEPTH - 2) / 6)
# Stem: (N, 3, 32, 32) -> (N, 16, 32, 32)
self.stem = ResStem(w_in=3, w_out=16)
# Stage 1: (N, 16, 32, 32) -> (N, 16, 32, 32)
self.s1 = ResStage(w_in=16, w_out=16, stride=1, d=d)
# Stage 2: (N, 16, 32, 32) -> (N, 32, 16, 16)
self.s2 = ResStage(w_in=16, w_out=32, stride=2, d=d)
# Stage 3: (N, 32, 16, 16) -> (N, 64, 8, 8)
self.s3 = ResStage(w_in=32, w_out=64, stride=2, d=d)
# Head: (N, 64, 8, 8) -> (N, num_classes)
self.head = ResHead(w_in=64, nc=C.DATASET.NUM_CLASSES)
def _construct_imagenet_basic(self):
# Retrieve the number of blocks per stage
(d1, d2, d3, d4) = _IN_STAGE_DS[C.ISON.DEPTH]
# Compute the initial bottleneck width
# Stem: (N, 3, 224, 224) -> (N, 64, 56, 56)
self.stem = ResStem(w_in=3, w_out=64)
# Stage 1: (N, 64, 56, 56) -> (N, 256, 56, 56)
self.s1 = ResStage(w_in=64, w_out=64, stride=1, d=d1)
# Stage 2: (N, 256, 56, 56) -> (N, 512, 28, 28)
self.s2 = ResStage(w_in=64, w_out=128, stride=2, d=d2)
# Stage 3: (N, 512, 56, 56) -> (N, 1024, 14, 14)
self.s3 = ResStage(w_in=128, w_out=256, stride=2, d=d3)
# Stage 4: (N, 1024, 14, 14) -> (N, 2048, 7, 7)
self.s4 = ResStage(w_in=256, w_out=512, stride=2, d=d4)
# Head: (N, 2048, 7, 7) -> (N, num_classes)
self.head = ResHead(w_in=512, nc=C.DATASET.NUM_CLASSES)
def _construct_imagenet(self):
# Retrieve the number of blocks per stage
(d1, d2, d3, d4) = _IN_STAGE_DS[C.ISON.DEPTH]
# Compute the initial bottleneck width
num_gs = 1 # C.RESNET.NUM_GROUPS
w_b = 64 # C.RESNET.WIDTH_PER_GROUP * num_gs
# Stem: (N, 3, 224, 224) -> (N, 64, 56, 56)
self.stem = ResStem(w_in=3, w_out=64)
# Stage 1: (N, 64, 56, 56) -> (N, 256, 56, 56)
self.s1 = ResStage(
w_in=64, w_out=256, stride=1, d=d1,
w_b=w_b, num_gs=num_gs
)
# Stage 2: (N, 256, 56, 56) -> (N, 512, 28, 28)
self.s2 = ResStage(
w_in=256, w_out=512, stride=2, d=d2,
w_b=w_b * 2, num_gs=num_gs
)
# Stage 3: (N, 512, 56, 56) -> (N, 1024, 14, 14)
self.s3 = ResStage(
w_in=512, w_out=1024, stride=2, d=d3,
w_b=w_b * 4, num_gs=num_gs
)
# Stage 4: (N, 1024, 14, 14) -> (N, 2048, 7, 7)
self.s4 = ResStage(
w_in=1024, w_out=2048, stride=2, d=d4,
w_b=w_b * 8, num_gs=num_gs
)
# Head: (N, 2048, 7, 7) -> (N, num_classes)
self.head = ResHead(w_in=2048, nc=C.DATASET.NUM_CLASSES)
def _network_init(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
if C.ISON.DIRAC_INIT:
# the first 7x7 convolution we use pytorch default initialization
# and not enforce orthogonality since the large input/output channel difference
if m.kernel_size != (7, 7):
nn.init.dirac_(m.weight)
else:
# kaiming initialization used for ResNet results
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(mean=0.0, std=np.sqrt(2.0 / fan_out))
if hasattr(m, 'bias') and m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.BatchNorm2d):
zero_init_gamma = (
hasattr(m, 'final_bn') and m.final_bn
)
m.weight.data.fill_(0.0 if zero_init_gamma else 1.0)
m.bias.data.zero_()
def forward(self, x):
for module in self.children():
x = module(x)
return x
def forward_naive(self, x):
# used for visualization of feature map
feature_list = []
residual_list = []
shortcut_list = []
x = self.stem(x)
for s_name, b_max in zip(['s1', 's2', 's3', 's4'], [3, 4, 6, 3]):
for b_num in range(1, b_max + 1):
identity = x
x = eval(f'self.{s_name}.b{b_num}.f.a')(x)
feature_list.append(x)
x = eval(f'self.{s_name}.b{b_num}.f.a_relu')(x)
x = eval(f'self.{s_name}.b{b_num}.f.b')(x)
feature_list.append(x)
if C.ISON.HAS_ST:
x = eval(f'self.{s_name}.b{b_num}.f.shared_scalar')(x)
if eval(f'self.{s_name}.b{b_num}').proj_block:
identity = eval(f'self.{s_name}.b{b_num}.proj')(identity)
shortcut_list.append(identity)
residual_list.append(x)
x = x + identity
x = eval(f'self.{s_name}.b{b_num}.relu')(x)
x = self.head(x)
return x, feature_list, shortcut_list, residual_list
def ortho(self):
ortho_penalty = []
cnt = 0
for m in self.modules():
if isinstance(m, nn.Conv2d):
if m.kernel_size == (7, 7) or m.weight.shape[1] == 3:
continue
o = self.ortho_conv(m)
cnt += 1
ortho_penalty.append(o)
ortho_penalty = sum(ortho_penalty)
return ortho_penalty
def ortho_conv(self, m, device='cuda'):
operator = m.weight
operand = torch.cat(torch.chunk(m.weight, m.groups, dim=0), dim=1)
transposed = m.weight.shape[1] < m.weight.shape[0]
num_channels = m.weight.shape[1] if transposed else m.weight.shape[0]
if transposed:
operand = operand.transpose(1, 0)
operator = operator.transpose(1, 0)
gram = F.conv2d(operand, operator, padding=(m.kernel_size[0] - 1, m.kernel_size[1] - 1),
stride=m.stride, groups=m.groups)
identity = torch.zeros(gram.shape).to(device)
identity[:, :, identity.shape[2] // 2, identity.shape[3] // 2] = torch.eye(num_channels).repeat(1, m.groups)
out = torch.sum((gram - identity) ** 2.0) / 2.0
return out