-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathLoss.py
219 lines (161 loc) · 8.65 KB
/
Loss.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
# -*- coding: utf-8 -*
import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
import gc
from typing import Optional, List, Tuple
def MI(outputs_target):
batch_size = outputs_target.size(0)
softmax_outs_t = nn.Softmax(dim=1)(outputs_target)
avg_softmax_outs_t = torch.sum(softmax_outs_t, dim=0) / float(batch_size)
log_avg_softmax_outs_t = torch.log(avg_softmax_outs_t + 1e-7)
item1 = -torch.sum(avg_softmax_outs_t * log_avg_softmax_outs_t)
item2 = -torch.sum(softmax_outs_t * torch.log(softmax_outs_t + 1e-7)) / float(batch_size)
return item1 - item2
class EstimatorMean():
def __init__(self, feature_num, class_num, alpha=0.1):
super(EstimatorMean, self).__init__()
self.class_num = class_num
self.Ave = torch.zeros(class_num, feature_num).cuda()
self.Amount = torch.zeros(class_num).cuda()
self.alpha = alpha
self.t = 1
def update_Mean(self, features, labels):
N = features.size(0)
C = self.class_num
A = features.size(1)
NxCxFeatures = features.view(N, 1, A).expand(N, C, A)
onehot = torch.zeros(N, C).cuda()
onehot.scatter_(1, labels.view(-1, 1), 1)
NxCxA_onehot = onehot.view(N, C, 1).expand(N, C, A)
features_by_sort = NxCxFeatures.mul(NxCxA_onehot)
Amount_CxA = NxCxA_onehot.sum(0)
Amount_CxA[Amount_CxA == 0] = 1
ave_CxA = features_by_sort.sum(0) / Amount_CxA
if self.t == 1:
self.Ave = ave_CxA.detach()
else:
self.Ave = (self.alpha * self.Ave + (1 - self.alpha) * ave_CxA).detach()
self.t = self.t + 1
class EstimatorCV():
def __init__(self, feature_num, class_num, alpha=0.3):
super(EstimatorCV, self).__init__()
self.class_num = class_num
self.CoVariance = torch.zeros(class_num, feature_num, feature_num).cuda()
self.Ave = torch.zeros(class_num, feature_num).cuda()
self.Amount = torch.zeros(class_num).cuda()
self.alpha = alpha
self.V1 = torch.zeros(class_num, feature_num, feature_num).cuda()
self.U1 = torch.zeros(class_num, feature_num, feature_num).cuda()
self.U2 = torch.zeros(class_num, feature_num).cuda()
self.initial_Ave = torch.zeros(class_num, feature_num).cuda()
self.initial_covariance = torch.zeros(class_num, feature_num, feature_num).cuda()
self.t = 1
def update_CV(self, features, labels):
N = features.size(0)
C = self.class_num
A = features.size(1)
NxCxFeatures = features.view(N, 1, A).expand(N, C, A)
onehot = torch.zeros(N, C).cuda()
onehot.scatter_(1, labels.view(-1, 1), 1)
NxCxA_onehot = onehot.view(N, C, 1).expand(N, C, A)
features_by_sort = NxCxFeatures.mul(NxCxA_onehot)
Amount_CxA = NxCxA_onehot.sum(0)
Amount_CxA[Amount_CxA == 0] = 1
ave_CxA = features_by_sort.sum(0) / Amount_CxA
var_temp = features_by_sort - ave_CxA.expand(N, C, A).mul(NxCxA_onehot)
var_temp = torch.bmm(var_temp.permute(1, 2, 0), var_temp.permute(1, 0, 2)).div(Amount_CxA.view(C, A, 1).expand(C, A, A))
self.V1 = self.alpha * self.V1 + var_temp
self.U1 = self.alpha * self.U1 + torch.bmm(ave_CxA.view(C, A, 1), ave_CxA.view(C, 1, A))
self.U2 = self.alpha * self.U2 + ave_CxA
if self.t == 1:
self.Ave = ave_CxA.detach()
self.CoVariance = var_temp.detach()
self.initial_covariance = var_temp.detach()
self.initial_Ave = ave_CxA.detach()
else:
self.Ave = (self.alpha * self.Ave + (1 - self.alpha) * ave_CxA).detach()
A1 = (self.alpha ** self.t) * self.initial_covariance
A2 = (1 - self.alpha) * self.V1
A3 = (self.alpha ** self.t) * torch.bmm((self.initial_Ave).view(C, A, 1), (self.initial_Ave).view(C, 1, A))
A4 = (1 - self.alpha) * self.U1
A5 = torch.bmm(self.Ave.view(C, A, 1), self.Ave.view(C, 1, A))
A6 = (self.alpha ** self.t) * torch.bmm((self.initial_Ave).view(C, A, 1), (self.Ave).view(C, 1, A)) + \
(1 - self.alpha) * torch.bmm((self.U2).view(C, A, 1), (self.Ave).view(C, 1, A))
self.CoVariance = (A1 + A2 + A3 + A4 + A5 - (A6 + A6.transpose(1, 2))).detach()
self.t = self.t + 1
class Loss_aug_pro(nn.Module):
def __init__(self, feature_num, class_num, alpha):
super(Loss_aug_pro, self).__init__()
self.target_estimator = EstimatorCV(feature_num, class_num, alpha)
self.source_estimator_mean = EstimatorMean(feature_num, class_num, alpha)
self.class_num = class_num
self.cross_entropy = nn.CrossEntropyLoss()
def aug_s(self, s_mean_matrix, t_mean_matrix, fc, f_s, y_s, labels_s, t_cv_matrix, Lambda, margin):
N = f_s.size(0)
C = self.class_num
A = f_s.size(1)
weight_m = list(fc.parameters())[0]
assert weight_m.shape[0] == C
assert weight_m.shape[1] == A
normalized_weight_m = F.normalize(weight_m, p=2, dim=1)
NxW_ij = normalized_weight_m.expand(N, C, A)
NxW_kj = torch.gather(NxW_ij, 1, labels_s.view(N, 1, 1).expand(N, C, A))
t_CV_temp = t_cv_matrix[labels_s]
sigma2 = Lambda * torch.bmm(torch.bmm(NxW_ij - NxW_kj, t_CV_temp), (NxW_ij - NxW_kj).permute(0, 2, 1))
sigma2 = sigma2.mul(torch.eye(C).cuda().expand(N, C, C)).sum(2).view(N, C)
sourceMean_NxA = s_mean_matrix[labels_s]
targetMean_NxA = t_mean_matrix[labels_s]
dataMean_NxA = (targetMean_NxA - sourceMean_NxA)
dataMean_NxAx1 = dataMean_NxA.expand(1, N, A).permute(1, 2, 0)
del t_CV_temp, sourceMean_NxA, targetMean_NxA, dataMean_NxA
gc.collect()
dataW_NxCxA = NxW_ij - NxW_kj
dataW_x_detaMean_NxCx1 = torch.bmm(dataW_NxCxA, dataMean_NxAx1)
datW_x_detaMean_NxC = dataW_x_detaMean_NxCx1.view(N, C)
aug_result = y_s + 0.5 * sigma2 + Lambda * datW_x_detaMean_NxC
one_hot = torch.zeros_like(y_s)
one_hot.scatter_(1, labels_s.view(-1, 1), 1.0)
margin_aug_result = aug_result - one_hot * margin
return margin_aug_result
def PrototypeConstraint(self, class_num, f_s, f_t, labels_s, pseudo_labels_t, s_mean_matrix, t_mean_matrix, t_cv_matrix, Lambda):
n_s = f_s.shape[0]
n_t = f_t.shape[0]
A = f_s.shape[1]
C = class_num
# class-wisely calculate the center of target samples
NxCxFeaturesT = f_t.view(n_t, 1, A).expand(n_t, C, A)
onehotT = torch.zeros(n_t, C).cuda()
onehotT.scatter_(1, pseudo_labels_t.view(-1, 1), 1)
NxCxA_onehotT = onehotT.view(n_t, C, 1).expand(n_t, C, A)
featuresT_by_sort = NxCxFeaturesT.mul(NxCxA_onehotT)
AmountT_CxA = NxCxA_onehotT.sum(0)
flag = AmountT_CxA[:, 0]
AmountT_CxA[AmountT_CxA == 0] = 1
aveT_CxA = featuresT_by_sort.sum(0) / AmountT_CxA
count = 0
aug_align_loss = 0.0
for i in range(n_s):
c = labels_s[i]
if flag[c] == 0:
continue
count = count + 1
mean_matrix = f_s[i] + Lambda * (t_mean_matrix[c] - s_mean_matrix[c])
variance_matrix = Lambda * torch.diagonal(t_cv_matrix[c], dim1=-2, dim2=-1)
aug_align_loss = aug_align_loss + torch.sum(torch.pow(mean_matrix - aveT_CxA[c], 2) + variance_matrix)
if count == 0:
return aug_align_loss
else:
aug_align_loss = aug_align_loss / float(count)
return aug_align_loss
def forward(self, fc, f_s, f_t, y_s, labels_source, pseudo_label_target, Lambda, margin):
self.target_estimator.update_CV(f_t.detach(), pseudo_label_target) # update target covariance and target mean
self.source_estimator_mean.update_Mean(f_s.detach(), labels_source) # update source mean
aug_y_s = self.aug_s(self.source_estimator_mean.Ave.detach(), self.target_estimator.Ave.detach(), fc, \
f_s, y_s, labels_source, self.target_estimator.CoVariance.detach(), Lambda, margin)
aug_loss = self.cross_entropy(aug_y_s, labels_source)
pro_loss = self.PrototypeConstraint(y_s.shape[1], f_s, f_t, labels_source, pseudo_label_target, \
self.source_estimator_mean.Ave.detach(), self.target_estimator.Ave.detach(),\
self.target_estimator.CoVariance.detach(), Lambda)
return aug_loss, pro_loss