-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathloss.py
181 lines (110 loc) · 5.32 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
import torch
import torch.nn.functional as F
########################### GausianMLR Losses #################################
sqrt_two = 2**0.5
eps = 1.0e-4
# The probability of a Gaussian variable being positive
def gaussian_variable_positive_probability(z_mean, z_std):
return 0.5 * (1 - torch.erf(-z_mean / (z_std * sqrt_two)))
# GaussianMLR Classification Loss
def GaussianMLRClassification(z_mean, z_std, labels):
N, K = labels.shape
bigger_zero_prob = gaussian_variable_positive_probability(z_mean, z_std) + eps
smaller_zero_prob = 1 - bigger_zero_prob + 2 * eps
bigger_zero_loss = torch.sum(-torch.log(bigger_zero_prob[labels > 0]))
smaller_zero_loss = torch.sum(-torch.log(smaller_zero_prob[labels == 0]))
return (bigger_zero_loss + smaller_zero_loss) / (N * K)
# GaussianMLR Ranking Loss
def GaussianMLRRanking(z_mean, z_logvar, labels):
N, K = labels.shape
pair_map = torch.tensor([(i, j) for i in range(K - 1) for j in range(i + 1, K)])
left_means = z_mean[:, pair_map[:, 0]]
left_logvars = z_logvar[:, pair_map[:, 0]]
right_means = z_mean[:, pair_map[:, 1]]
right_logvars = z_logvar[:, pair_map[:, 1]]
diff_mean = left_means - right_means
diff_std = torch.sqrt(torch.exp(left_logvars) + torch.exp(right_logvars))
bigger_prob = gaussian_variable_positive_probability(diff_mean, diff_std) + eps
smaller_prob = 1 - bigger_prob + 2 * eps
left_labels = labels[:, pair_map[:, 0]]
right_labels = labels[:, pair_map[:, 1]]
gt_bigger_map = left_labels > right_labels
gt_smaller_map = left_labels < right_labels
bigger_loss = torch.sum(-torch.log(bigger_prob[gt_bigger_map]))
smaller_loss = torch.sum(-torch.log(smaller_prob[gt_smaller_map]))
norm_coeff_pairs = torch.sum(gt_bigger_map) + torch.sum(gt_smaller_map)
return (bigger_loss + smaller_loss) / norm_coeff_pairs
def GaussianMLR(z_mean, z_logvar, labels):
z_std = torch.exp(z_logvar / 2)
classification_loss = GaussianMLRClassification(z_mean, z_std, labels)
ranking_loss = GaussianMLRRanking(z_mean, z_logvar, labels)
return {
"classification_loss": classification_loss,
"ranking_loss": ranking_loss,
}
def weak_GaussianMLR(z_mean, z_logvar, labels):
labels[labels > 0] = 1
z_std = torch.exp(z_logvar / 2)
classification_loss = GaussianMLRClassification(z_mean, z_std, labels)
ranking_loss = GaussianMLRRanking(z_mean, z_logvar, labels)
return {
"classification_loss": classification_loss,
"ranking_loss": ranking_loss,
}
###############################################################################
################################ CLR Losses ###################################
def weak_CLR(pair_logits, labels):
labels[labels > 0] = 1
N, K = labels.shape
labels = labels.float()
labels = torch.cat((labels, torch.ones(N, 1, device=labels.device) * 0.5), dim=1)
K += 1
pair_map = torch.tensor([(i, j) for i in range(K - 1) for j in range(i + 1, K)])
bigger_map = (labels[:, pair_map[:, 0]] > labels[:, pair_map[:, 1]]).float()
return F.binary_cross_entropy_with_logits(pair_logits, bigger_map)
def strong_CLR(pair_logits, labels):
N, K = labels.shape
labels = labels.float()
labels = torch.cat((labels, torch.ones(N, 1, device=labels.device) * 0.5), dim=1)
K += 1
pair_map = torch.tensor([(i, j) for i in range(K - 1) for j in range(i + 1, K)])
bigger_map = (labels[:, pair_map[:, 0]] > labels[:, pair_map[:, 1]]).float()
return F.binary_cross_entropy_with_logits(pair_logits, bigger_map)
###############################################################################
################################# LSEP ########################################
def MultiThresholdLoss(scores, threshold, labels):
binary_labels = (labels != 0).float()
diff = scores - threshold
return F.binary_cross_entropy(F.sigmoid(diff), binary_labels)
def weak_LSEP(scores, labels):
N, K = labels.shape
binary_labels = (labels != 0).float()
pair_map = torch.tensor([(i, j) for i in range(K - 1) for j in range(i + 1, K)])
left_labels = binary_labels[:, pair_map[:, 0]]
right_labels = binary_labels[:, pair_map[:, 1]]
neg_map = left_labels > right_labels
zero_map = left_labels == right_labels
left_scores = scores[:, pair_map[:, 0]]
right_scores = scores[:, pair_map[:, 1]]
diff_scores = left_scores - right_scores
diff_scores[neg_map] *= -1
diff_scores[zero_map] = -float("inf")
exp_scores = torch.exp(diff_scores)
instance_sum = torch.log(exp_scores.sum(1) + 1)
return torch.mean(instance_sum)
def strong_LSEP(scores, labels):
N, K = labels.shape
pair_map = torch.tensor([(i, j) for i in range(K - 1) for j in range(i + 1, K)])
left_labels = labels[:, pair_map[:, 0]]
right_labels = labels[:, pair_map[:, 1]]
neg_map = left_labels > right_labels
zero_map = left_labels == right_labels
left_scores = scores[:, pair_map[:, 0]]
right_scores = scores[:, pair_map[:, 1]]
diff_scores = left_scores - right_scores
diff_scores[neg_map] *= -1
diff_scores[zero_map] = -float("inf")
exp_scores = torch.exp(diff_scores)
instance_sum = torch.log(exp_scores.sum(1) + 1)
return torch.mean(instance_sum)
###############################################################################