-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
214 lines (177 loc) · 8.12 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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
def create_resmodel(num_layers: int, num_classes: int,
pretrain: bool) -> torchvision.models.resnet.ResNet:
assert num_layers in (
18, 34, 50, 101, 152
), f"Invalid number of ResNet Layers. Must be one of [18, 34, 50, 101, 152] and not {num_layers}"
feature_extractor_part1_constructor = getattr(torchvision.models, f"resnet{num_layers}")
feature_extractor_part1 = feature_extractor_part1_constructor(num_classes=num_classes)
if pretrain:
pretrained = feature_extractor_part1_constructor(pretrained=True).state_dict()
if num_classes != pretrained["fc.weight"].size(0):
del pretrained["fc.weight"], pretrained["fc.bias"]
feature_extractor_part1.load_state_dict(state_dict=pretrained, strict=False)
return feature_extractor_part1
class Attention(nn.Module):
def __init__(self, interlayer_classes, num_layers, pretrain):
super(Attention, self).__init__()
self.L = 512
self.D = 128
self.K = 1
self.interlayer_classes = interlayer_classes
self.num_layers = num_layers
self.pretrain = pretrain
# self.feature_extractor_part1 = nn.Sequential(
# nn.Conv2d(3, 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_part1 = create_resmodel(num_classes=self.interlayer_classes,
num_layers=self.num_layers,
pretrain=self.pretrain)
self.feature_extractor_part2 = nn.Sequential(
nn.Linear(self.interlayer_classes, self.L),
nn.Dropout(0.5),
nn.ReLU(),
)
self.attention = nn.Sequential(
nn.Linear(self.L, self.D),
nn.Tanh(),
nn.Linear(self.D, self.K)
)
self.classifier = nn.Sequential(
nn.Linear(self.L*self.K, 2),
nn.Sigmoid()
)
def forward(self, x):
x = x.squeeze(0)
H = self.feature_extractor_part1(x)
H = H.view(-1, 1000)
H = self.feature_extractor_part2(H) # NxL
A = self.attention(H) # NxK
A = torch.transpose(A, 1, 0) # KxN
A = F.softmax(A, dim=1) # softmax over N
M = torch.mm(A, H) # KxL
train_output = self.classifier(M)
# Y_hat = torch.ge(Y_prob[:,1], 0.5).float()
# Y_hat = torch.ge(Y_prob, 0.5).float()
return train_output, A
# AUXILIARY METHODS
# def calculate_classification_error(self, X, Y):
# Y = Y.float()
# Y_prob, Y_hat, A = self.forward(X)
# error = 1. - Y_hat.eq(Y).cpu().float().mean().item()
# 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 error, Y_hat,neg_log_likelihood,A
# return error, Y_hat,neg_log_likelihood,A
def calculate_classification_error(self, X, Y):
Y = Y.float()
train_output, A = self.forward(X)
confidence_train, train_pred = torch.max(train_output, dim=1)
error = 1. - train_pred.eq(Y).cpu().float().mean().item()
# error = 1. - Y_hat.eq(Y).cpu().float().mean().item()
# 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
criterion = nn.CrossEntropyLoss()
loss = criterion(train_output, Y.unsqueeze(0).long())
# return error, Y_hat,neg_log_likelihood,A
return error, train_pred, loss, A
def calculate_classification_error_fortest(self, X, Y):
Y = Y.float()
train_output, A = self.forward(X)
confidence_train, train_pred = torch.max(train_output, dim=1)
error = 1. - train_pred.eq(Y).cpu().float().mean().item()
# error = 1. - Y_hat.eq(Y).cpu().float().mean().item()
# 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
criterion = nn.CrossEntropyLoss()
loss = criterion(train_output, Y.unsqueeze(0).long())
# return error, Y_hat,neg_log_likelihood,A
return error, train_pred, loss, A, confidence_train
def calculate_classification_error_fortest_thresholds(self, X, Y):
Y = Y.float()
train_output, A = self.forward(X)
confidence_train, train_pred = torch.max(train_output, dim=1)
train_pred_original = train_pred
if (train_pred.cpu().numpy() == 0) and (confidence_train.cpu().numpy() <= (1-0.41899818)):
train_pred = torch.tensor([1.]) - train_pred.cpu()
error = 1. - train_pred.eq(Y).cpu().float().mean().item()
# error = 1. - Y_hat.eq(Y).cpu().float().mean().item()
# 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
criterion = nn.CrossEntropyLoss()
loss = criterion(train_output, Y.unsqueeze(0).long())
# return error, Y_hat,neg_log_likelihood,A
return error, train_pred, loss, A, confidence_train, train_pred_original
# 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.L = 500
self.D = 128
self.K = 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.L),
nn.ReLU(),
)
self.attention_V = nn.Sequential(
nn.Linear(self.L, self.D),
nn.Tanh()
)
self.attention_U = nn.Sequential(
nn.Linear(self.L, self.D),
nn.Sigmoid()
)
self.attention_weights = nn.Linear(self.D, self.K)
self.classifier = nn.Sequential(
nn.Linear(self.L*self.K, 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) # NxL
A_V = self.attention_V(H) # NxD
A_U = self.attention_U(H) # NxD
A = self.attention_weights(A_V * A_U) # element wise multiplication # NxK
A = torch.transpose(A, 1, 0) # KxN
A = F.softmax(A, dim=1) # softmax over N
M = torch.mm(A, H) # KxL
Y_prob = self.classifier(M)
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