-
Notifications
You must be signed in to change notification settings - Fork 1
/
exp_classification_archs.py
292 lines (231 loc) · 11.2 KB
/
exp_classification_archs.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
import math
import numpy as np
import torch as th
import torch.nn.functional as F
from CLTLayer import CLTLinear, CLTConv
from VBLayer import VBLinear, VBConv
from torch import nn
from utils import Phi
from tqdm import tqdm
class BEDL(nn.Module):
def __init__(self, n_channels, n_classes, delta=0.1, prior_prec=10):
super(BEDL, self).__init__()
self.n_channels = n_channels
self.n_classes = n_classes
self.prior_prec = prior_prec
self.n_samples = 20
relu_act = True
elu_act = False
if n_channels == 1:
self.conv1 = CLTConv(1, 20, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.conv2 = CLTConv(20, 50, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense1 = CLTLinear(4*4*50, 500, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense2 = CLTLinear(500, n_classes, prior_prec=prior_prec, relu_act=False, elu_act=elu_act)
elif n_channels == 3:
self.conv1 = CLTConv(3, 192, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.conv2 = CLTConv(192, 192, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense1 = CLTLinear(5 * 5 * 192, 1000, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense2 = CLTLinear(1000, n_classes, prior_prec=prior_prec, relu_act=False, elu_act=False)
def forward(self, inp):
mu_h, var_h = self.conv1(inp, None)
mu_h, var_h = self.conv2(mu_h, var_h)
if self.n_channels == 1:
mu_h, var_h = mu_h.view(-1, 4*4*50), var_h.view(-1, 4*4*50)
else:
mu_h, var_h = mu_h.view(-1, 5*5*192), var_h.view(-1, 5*5*192)
mu_h, var_h = self.dense1(mu_h, var_h)
mu_h, var_h = self.dense2(mu_h, var_h)
return mu_h, var_h
def predict(self, data):
mu, var = self.forward(data)
prob = sum([F.softmax(mu + var.sqrt() * th.randn_like(mu), 1) for _ in range(self.n_samples)]) / self.n_samples
return prob
def loss(self, data, target, n_train):
if target.is_cuda:
target = th.eye(self.n_classes).cuda()[target]
else:
target = th.eye(self.n_classes)[target]
prob = self.predict(data)
log_probs = th.log(prob + 1e-8)
data_fit = - th.sum(target * log_probs, 1).mean()
return data_fit
class BEDLPAC(nn.Module):
def __init__(self, n_channels, n_classes, delta=0.1, prior_prec=10):
super(BEDLPAC, self).__init__()
self.n_channels = n_channels
self.n_classes = n_classes
self.delta = delta
self.prior_prec = prior_prec
self.n_samples = 5
relu_act = True
elu_act = False
if n_channels == 1:
self.conv1 = CLTConv(1, 20, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.conv2 = CLTConv(20, 50, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense1 = CLTLinear(4*4*50, 500, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense2 = CLTLinear(500, n_classes, prior_prec=prior_prec, relu_act=False, elu_act=elu_act)
elif n_channels == 3:
self.conv1 = CLTConv(3, 192, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.conv2 = CLTConv(192, 192, 5, stride=2, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense1 = CLTLinear(5 * 5 * 192, 1000, prior_prec=prior_prec, relu_act=relu_act, elu_act=elu_act)
self.dense2 = CLTLinear(1000, n_classes, prior_prec=prior_prec, relu_act=False, elu_act=False)
def forward(self, inp):
mu_h, var_h = self.conv1(inp, None)
mu_h, var_h = self.conv2(mu_h, var_h)
if self.n_channels == 1:
mu_h, var_h = mu_h.view(-1, 4*4*50), var_h.view(-1, 4*4*50)
else:
mu_h, var_h = mu_h.view(-1, 5*5*192), var_h.view(-1, 5*5*192)
mu_h, var_h = self.dense1(mu_h, var_h)
mu_h, var_h = self.dense2(mu_h, var_h)
return mu_h, var_h
def predict(self, data):
mu, var = self.forward(data)
prob = sum([F.softmax(mu + var.sqrt() * th.randn_like(mu), 1) for _ in range(self.n_samples)]) / self.n_samples
return prob
def loss(self, data, target, n_train):
mu_pred, var_pred = self.forward(data)
kl = 0
for _ in range(self.n_samples):
alpha = th.exp(mu_pred + var_pred.sqrt() * th.randn_like(mu_pred))
alpha_0 = alpha.sum(1) # I.e. the number of classes
tmp = th.lgamma(alpha_0) - th.lgamma(alpha).sum(1) - math.lgamma(alpha.shape[1])
tmp += ((alpha - 1) * (th.digamma(alpha) - th.digamma(alpha_0[:,None]))).sum(1)
kl += tmp
kl /= self.n_samples
kl = kl.mean()
if target.is_cuda:
target = th.eye(self.n_classes).cuda()[target]
else:
target = th.eye(self.n_classes)[target]
prob = sum([F.softmax(mu_pred + var_pred.sqrt() * th.randn_like(mu_pred), 1) for _ in range(self.n_samples)]) / self.n_samples
log_probs = th.log(prob + 1e-8)
data_fit = - th.sum(target * log_probs, 1).mean()
B = n_train
reg_term = th.sqrt((kl + math.log(B) - math.log(self.delta))/n_train)
return data_fit + reg_term
class VBNet(nn.Module):
def __init__(self, n_channels, n_classes, loguniform=False, prior_prec=10):
super(VBNet, self).__init__()
self.n_channels = n_channels
self.n_classes = n_classes
self.loguniform = loguniform
self.n_samples = 5
self.prior_prec=prior_prec
if n_channels == 1:
self.conv1 = VBConv(1, 20, 5, stride=2, prior_prec=prior_prec)
self.conv2 = VBConv(20, 50, 5, stride=2, prior_prec=prior_prec)
self.dense1 = VBLinear(4*4*50, 500, prior_prec=prior_prec)
self.dense2 = VBLinear(500, n_classes, prior_prec=prior_prec)
elif n_channels == 3:
self.conv1 = VBConv(3, 192, 5, stride=2, prior_prec=prior_prec)
self.conv2 = VBConv(192, 192, 5, stride=2, prior_prec=prior_prec)
self.dense1 = VBLinear(5 * 5 * 192, 1000, prior_prec=prior_prec)
self.dense2 = VBLinear(1000, n_classes, prior_prec=prior_prec)
def forward(self, inp):
out = F.relu(self.conv1(inp))
out = F.relu(self.conv2(out))
if self.n_channels == 1:
out = out.view(-1, 4*4*50)
else:
out = out.view(-1, 5*5*192)
out = F.relu(self.dense1(out))
out = self.dense2(out)
return out
def predict(self, data):
self.sample = True
if self.sample:
prob = sum([F.softmax(self.forward(data), 1) for _ in range(self.n_samples)]) / self.n_samples
return prob
def loss(self, data, target, n_train):
kl = sum(l.KL(loguniform=self.loguniform) for l in [self.conv1, self.conv2, self.dense1, self.dense2])
if target.is_cuda:
target = th.eye(self.n_classes).cuda()[target]
else:
target = th.eye(self.n_classes)[target]
log_probs = sum([F.log_softmax(self.forward(data), 1) for _ in range(self.n_samples)]) / self.n_samples
data_fit = - th.sum(target * log_probs, 1).mean()
return data_fit + kl/n_train
class DropNet(nn.Module):
def __init__(self, n_channels, n_classes):
super(DropNet, self).__init__()
self.n_channels = n_channels
self.n_classes = n_classes
if n_channels == 1:
self.conv1 = nn.Conv2d(1, 20, 5, stride=2)
self.conv2 = nn.Conv2d(20, 50, 5, stride=2)
self.dense1 = nn.Linear(4*4*50, 500)
self.dense2 = nn.Linear(500, n_classes)
elif n_channels == 3:
self.conv1 = nn.Conv2d(3, 192, 5, stride=2)
self.conv2 = nn.Conv2d(192, 192, 5, stride=2)
self.dense1 = nn.Linear(5 * 5 * 192, 1000)
self.dense2 = nn.Linear(1000, n_classes)
def forward(self, inp):
drop_rate = 0.5
out = F.dropout(F.relu(self.conv1(inp)), drop_rate, self.training)
out = F.dropout(F.relu(self.conv2(out)), drop_rate, self.training)
if self.n_channels == 1:
out = out.view(-1, 4*4*50)
else:
out = out.view(-1, 5*5*192)
out = F.dropout(F.relu(self.dense1(out)), drop_rate, self.training)
out = self.dense2(out)
return out
def predict(self, data):
self.sample = True
if self.sample:
prob = sum([F.softmax(self.forward(data), 1) for _ in range(20)]) / 20
return prob
def loss(self, data, target, n_train):
if target.is_cuda:
target = th.eye(self.n_classes).cuda()[target]
else:
target = th.eye(self.n_classes)[target]
n_samples = 3
log_probs = sum([F.log_softmax(self.forward(data), 1) for _ in range(n_samples)]) / n_samples
data_fit = - th.sum(target * log_probs, 1).mean()
return data_fit
class DVINet(nn.Module):
def __init__(self, n_channels, n_classes, prior_prec=10):
super(DVINet, self).__init__()
self.n_channels = n_channels
self.n_classes = n_classes
self.prior_prec = prior_prec
if n_channels == 1:
self.conv1 = CLTConv(1, 20, 5, stride=2, prior_prec=prior_prec)
self.conv2 = CLTConv(20, 50, 5, stride=2, prior_prec=prior_prec)
self.dense1 = CLTLinear(4*4*50, 500, prior_prec=prior_prec)
self.dense2 = CLTLinear(500, n_classes, relu_act=False, prior_prec=prior_prec)
elif n_channels == 3:
self.conv1 = CLTConv(3, 192, 5, stride=2, prior_prec=prior_prec)
self.conv2 = CLTConv(192, 192, 5, stride=2, prior_prec=prior_prec)
self.dense1 = CLTLinear(5 * 5 * 192, 1000, prior_prec=prior_prec)
self.dense2 = CLTLinear(1000, n_classes, relu_act=False, prior_prec=prior_prec)
def forward(self, inp):
mu_h, var_h = self.conv1(inp, None)
mu_h, var_h = self.conv2(mu_h, var_h)
if self.n_channels == 1:
mu_h, var_h = mu_h.view(-1, 4*4*50), var_h.view(-1, 4*4*50)
else:
mu_h, var_h = mu_h.view(-1, 5*5*192), var_h.view(-1, 5*5*192)
mu_h, var_h = self.dense1(mu_h, var_h)
mu_h, var_h = self.dense2(mu_h, var_h)
return mu_h, var_h
def predict(self, data):
mu, var = self.forward(data)
self.sample = True
if self.sample:
prob = sum([F.softmax(mu + var.sqrt() * th.randn_like(mu), 1) for _ in range(20)]) / 20
return prob
def loss(self, data, target, n_train):
mu, var = self.forward(data)
kl = sum(l.KL() for l in [self.conv1, self.conv2, self.dense1, self.dense2])
if target.is_cuda:
target = th.eye(self.n_classes).cuda()[target]
else:
target = th.eye(self.n_classes)[target]
n_samples = 20
log_probs = sum([F.log_softmax(mu + var.sqrt() * th.randn_like(mu), 1) for _ in range(n_samples)]) / n_samples
data_fit = - th.sum(target * log_probs, 1).mean()
return data_fit + kl/n_train