-
Notifications
You must be signed in to change notification settings - Fork 189
/
model.py
144 lines (111 loc) · 4.29 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
import torch
import torch.nn as nn
import torch.nn.functional as F
class Attention(nn.Module):
def __init__(self):
super(Attention, self).__init__()
self.M = 500
self.L = 128
self.ATTENTION_BRANCHES = 1
self.feature_extractor_part1 = nn.Sequential(
nn.Conv2d(1, 20, kernel_size=5),
nn.ReLU(),
nn.MaxPool2d(2, stride=2),
nn.Conv2d(20, 50, kernel_size=5),
nn.ReLU(),
nn.MaxPool2d(2, stride=2)
)
self.feature_extractor_part2 = nn.Sequential(
nn.Linear(50 * 4 * 4, self.M),
nn.ReLU(),
)
self.attention = nn.Sequential(
nn.Linear(self.M, self.L), # matrix V
nn.Tanh(),
nn.Linear(self.L, self.ATTENTION_BRANCHES) # matrix w (or vector w if self.ATTENTION_BRANCHES==1)
)
self.classifier = nn.Sequential(
nn.Linear(self.M*self.ATTENTION_BRANCHES, 1),
nn.Sigmoid()
)
def forward(self, x):
x = x.squeeze(0)
H = self.feature_extractor_part1(x)
H = H.view(-1, 50 * 4 * 4)
H = self.feature_extractor_part2(H) # KxM
A = self.attention(H) # KxATTENTION_BRANCHES
A = torch.transpose(A, 1, 0) # ATTENTION_BRANCHESxK
A = F.softmax(A, dim=1) # softmax over K
Z = torch.mm(A, H) # ATTENTION_BRANCHESxM
Y_prob = self.classifier(Z)
Y_hat = torch.ge(Y_prob, 0.5).float()
return Y_prob, Y_hat, A
# AUXILIARY METHODS
def calculate_classification_error(self, X, Y):
Y = Y.float()
_, Y_hat, _ = self.forward(X)
error = 1. - Y_hat.eq(Y).cpu().float().mean().data.item()
return error, Y_hat
def calculate_objective(self, X, Y):
Y = Y.float()
Y_prob, _, A = self.forward(X)
Y_prob = torch.clamp(Y_prob, min=1e-5, max=1. - 1e-5)
neg_log_likelihood = -1. * (Y * torch.log(Y_prob) + (1. - Y) * torch.log(1. - Y_prob)) # negative log bernoulli
return neg_log_likelihood, A
class GatedAttention(nn.Module):
def __init__(self):
super(GatedAttention, self).__init__()
self.M = 500
self.L = 128
self.ATTENTION_BRANCHES = 1
self.feature_extractor_part1 = nn.Sequential(
nn.Conv2d(1, 20, kernel_size=5),
nn.ReLU(),
nn.MaxPool2d(2, stride=2),
nn.Conv2d(20, 50, kernel_size=5),
nn.ReLU(),
nn.MaxPool2d(2, stride=2)
)
self.feature_extractor_part2 = nn.Sequential(
nn.Linear(50 * 4 * 4, self.M),
nn.ReLU(),
)
self.attention_V = nn.Sequential(
nn.Linear(self.M, self.L), # matrix V
nn.Tanh()
)
self.attention_U = nn.Sequential(
nn.Linear(self.M, self.L), # matrix U
nn.Sigmoid()
)
self.attention_w = nn.Linear(self.L, self.ATTENTION_BRANCHES) # matrix w (or vector w if self.ATTENTION_BRANCHES==1)
self.classifier = nn.Sequential(
nn.Linear(self.M*self.ATTENTION_BRANCHES, 1),
nn.Sigmoid()
)
def forward(self, x):
x = x.squeeze(0)
H = self.feature_extractor_part1(x)
H = H.view(-1, 50 * 4 * 4)
H = self.feature_extractor_part2(H) # KxM
A_V = self.attention_V(H) # KxL
A_U = self.attention_U(H) # KxL
A = self.attention_w(A_V * A_U) # element wise multiplication # KxATTENTION_BRANCHES
A = torch.transpose(A, 1, 0) # ATTENTION_BRANCHESxK
A = F.softmax(A, dim=1) # softmax over K
Z = torch.mm(A, H) # ATTENTION_BRANCHESxM
Y_prob = self.classifier(Z)
Y_hat = torch.ge(Y_prob, 0.5).float()
return Y_prob, Y_hat, A
# AUXILIARY METHODS
def calculate_classification_error(self, X, Y):
Y = Y.float()
_, Y_hat, _ = self.forward(X)
error = 1. - Y_hat.eq(Y).cpu().float().mean().item()
return error, Y_hat
def calculate_objective(self, X, Y):
Y = Y.float()
Y_prob, _, A = self.forward(X)
Y_prob = torch.clamp(Y_prob, min=1e-5, max=1. - 1e-5)
neg_log_likelihood = -1. * (Y * torch.log(Y_prob) + (1. - Y) * torch.log(1. - Y_prob)) # negative log bernoulli
return neg_log_likelihood, A