-
Notifications
You must be signed in to change notification settings - Fork 9
/
model.py
324 lines (253 loc) · 11.9 KB
/
model.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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as f
from softdtw_cuda import SoftDTW
class FCN(nn.Module):
def __init__(self, num_classes, num_segments, input_size, hidden_sizes=[128, 256, 128], kernel_sizes=[9, 5, 3], cost_type='cosine', pooling_op='avg', gamma=1.0):
super(FCN, self).__init__()
self.num_classes = num_classes
self.num_segments = num_segments
self.input_size = input_size
self.hidden_sizes = hidden_sizes
self.kernel_sizes = kernel_sizes
self.pooling_op = pooling_op
self.softdtw = SoftDTW(use_cuda=True, gamma=gamma, cost_type=cost_type, normalize=False)
self._build_model(num_classes, num_segments, input_size, hidden_sizes, kernel_sizes)
self._init_model()
def _build_model(self, num_classes, num_segments, input_size, hidden_sizes, kernel_sizes):
self.conv1 = nn.Conv1d(in_channels=input_size,
out_channels=hidden_sizes[0],
kernel_size=kernel_sizes[0])
self.conv2 = nn.Conv1d(in_channels=hidden_sizes[0],
out_channels=hidden_sizes[1],
kernel_size=kernel_sizes[1])
self.conv3 = nn.Conv1d(in_channels=hidden_sizes[1],
out_channels=hidden_sizes[2],
kernel_size=kernel_sizes[2])
self.norm1 = nn.BatchNorm1d(num_features=hidden_sizes[0])
self.norm2 = nn.BatchNorm1d(num_features=hidden_sizes[1])
self.norm3 = nn.BatchNorm1d(num_features=hidden_sizes[2])
self.fc = nn.Linear(hidden_sizes[2]*num_segments, num_classes)
self.protos = nn.Parameter(torch.zeros(hidden_sizes[2], num_segments), requires_grad=True)
def _init_model(self):
for m in self.modules():
if isinstance(m, nn.Linear):
init.xavier_normal_(m.weight.data)
init.normal_(m.bias.data)
if isinstance(m, nn.Conv1d):
init.xavier_normal_(m.weight.data)
if m.bias is not None:
init.normal_(m.bias.data)
def init_protos(self, data_loader):
for itr, batch in enumerate(data_loader):
data = batch['data'].cuda()
h = self.get_htensor(data)
self.protos.data += self.stpool(h, 'avg').mean(dim=0)
self.protos.data /= len(data_loader)
def get_htensor(self, x):
h = x
h = f.pad(h, (int(self.kernel_sizes[0]/2), int(self.kernel_sizes[0]/2)), "constant", 0)
h = f.relu(self.norm1(self.conv1(h)))
h = f.pad(h, (int(self.kernel_sizes[1]/2), int(self.kernel_sizes[1]/2)), "constant", 0)
h = f.relu(self.norm2(self.conv2(h)))
h = f.pad(h, (int(self.kernel_sizes[2]/2), int(self.kernel_sizes[2]/2)), "constant", 0)
h = f.relu(self.norm3(self.conv3(h)))
return h
# global temporal pooling
def gtpool(self, h, op):
if op == 'avg':
return torch.mean(h, dim=2)
if op == 'sum':
return torch.sum(h, dim=2)
elif op == 'max':
return torch.max(h, dim=2)[0]
# static temporal pooling
def stpool(self, h, op):
segment_sizes = [int(h.shape[2]/self.num_segments)] * self.num_segments
segment_sizes[-1] += h.shape[2] - sum(segment_sizes)
hs = torch.split(h, segment_sizes, dim=2)
if op == 'avg':
hs = [h_.mean(dim=2, keepdim=True) for h_ in hs]
elif op == 'sum':
hs = [h_.sum(dim=2, keepdim=True) for h_ in hs]
elif op == 'max':
hs = [h_.max(dim=2)[0].unsqueeze(dim=2) for h_ in hs]
hs = torch.cat(hs, dim=2)
return hs
# dynamic temporal pooling
def dtpool(self, h, op):
A = self.softdtw.align(self.protos.repeat(h.shape[0], 1, 1), h)
if op == 'avg':
A /= A.sum(dim=2, keepdim=True)
h = torch.bmm(h, A.transpose(1, 2))
elif op == 'sum':
h = h.unsqueeze(dim=2) * A.unsqueeze(dim=1)
h = h.sum(dim=3)
elif op == 'max':
h = h.unsqueeze(dim=2) * A.unsqueeze(dim=1)
h = h.max(dim=3)[0]
return h
def compute_gradcam(self, x, labels):
def hook_func(grad):
self.h_grad = grad
h, logits = self.forward(x)
h.register_hook(hook_func)
self.zero_grad()
scores = torch.gather(logits, 1, labels.unsqueeze(dim=1))
scores.mean().backward()
gradcam = (h * self.h_grad).sum(dim=1, keepdim=True)
# min-max normalization
gradcam_min = torch.min(gradcam, dim=2, keepdim=True)[0]
gradcam_max = torch.max(gradcam, dim=2, keepdim=True)[0]
gradcam = (gradcam - gradcam_min) / (gradcam_max - gradcam_min)
A = self.softdtw.align(self.protos.unsqueeze(dim=0).repeat(h.shape[0], 1, 1), h).sum(dim=2)
return gradcam, A
def forward(self, x):
h = self.get_htensor(x)
op = self.pooling_op
out = self.fc(self.dtpool(h, op).reshape(h.shape[0], -1))
return h, out
def compute_aligncost(self, h):
cost = self.softdtw(self.protos.repeat(h.shape[0], 1, 1), h.detach())
return cost.mean() / h.shape[2]
class ResidualBlock(nn.Module):
"""
Args:
hidden_size: input dimension (Channel) of 1d convolution
output_size: output dimension (Channel) of 1d convolution
kernel_size: kernel size
"""
def __init__(self, input_size, output_size, kernel_sizes=[9, 5, 3]):
super(ResidualBlock, self).__init__()
self.kernel_sizes = kernel_sizes
self.conv1 = nn.Conv1d(in_channels=input_size,
out_channels=output_size,
kernel_size=kernel_sizes[0])
self.conv2 = nn.Conv1d(in_channels=output_size,
out_channels=output_size,
kernel_size=kernel_sizes[1])
self.conv3 = nn.Conv1d(in_channels=output_size,
out_channels=output_size,
kernel_size=kernel_sizes[2])
self.conv_skip = nn.Conv1d(in_channels=input_size,
out_channels=output_size,
kernel_size=1)
self.norm1 = nn.BatchNorm1d(num_features=output_size)
self.norm2 = nn.BatchNorm1d(num_features=output_size)
self.norm3 = nn.BatchNorm1d(num_features=output_size)
self.norm_skip = nn.BatchNorm1d(num_features=output_size)
def forward(self, x):
h = x
h = f.pad(h, (int(self.kernel_sizes[0]/2), int(self.kernel_sizes[0]/2)), "constant", 0)
h = f.relu(self.norm1(self.conv1(h)))
h = f.pad(h, (int(self.kernel_sizes[1]/2), int(self.kernel_sizes[1]/2)), "constant", 0)
h = f.relu(self.norm2(self.conv2(h)))
h = f.pad(h, (int(self.kernel_sizes[2]/2), int(self.kernel_sizes[2]/2)), "constant", 0)
h = self.norm3(self.conv3(h))
s = self.norm_skip(self.conv_skip(x))
h += s
h = f.relu(h)
return h
class ResNet(nn.Module):
def __init__(self, num_classes, num_segments, input_size, hidden_sizes=[64, 128, 128], kernel_sizes=[9, 5, 3], cost_type='cosine', pooling_op='avg', gamma=1.0):
super(ResNet, self).__init__()
self.num_classes = num_classes
self.num_segments = num_segments
self.input_size = input_size
self.hidden_sizes = hidden_sizes
self.kernel_sizes = kernel_sizes
self.pooling_op = pooling_op
self.softdtw = SoftDTW(use_cuda=True, gamma=gamma, cost_type=cost_type, normalize=False)
self._build_model(num_classes, num_segments, input_size, hidden_sizes, kernel_sizes)
self._init_model()
def _build_model(self, num_classes, num_segments, input_size, hidden_sizes, kernel_sizes):
self.resblock1 = ResidualBlock(input_size=input_size,
output_size=hidden_sizes[0],
kernel_sizes=kernel_sizes)
self.resblock2 = ResidualBlock(input_size=hidden_sizes[0],
output_size=hidden_sizes[1],
kernel_sizes=kernel_sizes)
self.resblock3 = ResidualBlock(input_size=hidden_sizes[1],
output_size=hidden_sizes[2],
kernel_sizes=kernel_sizes)
self.fc = nn.Linear(hidden_sizes[2]*num_segments, num_classes)
self.protos = nn.Parameter(torch.zeros(hidden_sizes[2], num_segments), requires_grad=True)
def _init_model(self):
for m in self.modules():
if isinstance(m, nn.Linear):
init.xavier_normal_(m.weight.data)
init.normal_(m.bias.data)
if isinstance(m, nn.Conv1d):
init.xavier_normal_(m.weight.data)
if m.bias is not None:
init.normal_(m.bias.data)
def init_protos(self, data_loader):
for itr, batch in enumerate(data_loader):
data = batch['data'].cuda()
h = self.get_htensor(data)
self.protos.data += self.stpool(h, 'avg').mean(dim=0)
self.protos.data /= len(data_loader)
def get_htensor(self, x):
h = self.resblock1(x)
h = self.resblock2(h)
h = self.resblock3(h)
return h
# global temporal pooling
def gtpool(self, h, op):
if op == 'avg':
return torch.mean(h, dim=2)
if op == 'sum':
return torch.sum(h, dim=2)
elif op == 'max':
return torch.max(h, dim=2)[0]
# static temporal pooling
def stpool(self, h, op):
segment_sizes = [int(h.shape[2]/self.num_segments)] * self.num_segments
segment_sizes[-1] += h.shape[2] - sum(segment_sizes)
hs = torch.split(h, segment_sizes, dim=2)
if op == 'avg':
hs = [h_.mean(dim=2, keepdim=True) for h_ in hs]
if op == 'sum':
hs = [h_.sum(dim=2, keepdim=True) for h_ in hs]
elif op == 'max':
hs = [h_.max(dim=2)[0].unsqueeze(dim=2) for h_ in hs]
hs = torch.cat(hs, dim=2)
return hs
# dynamic temporal pooling
def dtpool(self, h, op):
A = self.softdtw.align(self.protos.repeat(h.shape[0], 1, 1), h)
if op == 'avg':
A /= A.sum(dim=2, keepdim=True)
h = torch.bmm(h, A.transpose(1, 2))
elif op == 'sum':
h = h.unsqueeze(dim=2) * A.unsqueeze(dim=1)
h = h.sum(dim=3)
elif op == 'max':
h = h.unsqueeze(dim=2) * A.unsqueeze(dim=1)
h = h.max(dim=3)[0]
return h
def compute_gradcam(self, x, labels):
def hook_func(grad):
self.h_grad = grad
h, logits = self.forward(x)
h.register_hook(hook_func)
self.zero_grad()
scores = torch.gather(logits, 1, labels.unsqueeze(dim=1))
scores.mean().backward()
gradcam = (h * self.h_grad).sum(dim=1, keepdim=True)
# min-max normalization
gradcam_min = torch.min(gradcam, dim=2, keepdim=True)[0]
gradcam_max = torch.max(gradcam, dim=2, keepdim=True)[0]
gradcam = (gradcam - gradcam_min) / (gradcam_max - gradcam_min)
A = self.softdtw.align(self.protos.unsqueeze(dim=0).repeat(h.shape[0], 1, 1), h).sum(dim=2)
return gradcam, A
def forward(self, x):
h = self.get_htensor(x)
op = self.pooling_op
out = self.fc(self.dtpool(h, op).reshape(h.shape[0], -1))
return h, out
def compute_aligncost(self, h):
cost = self.softdtw(self.protos.repeat(h.shape[0], 1, 1), h.detach())
return cost.mean() / h.shape[2]