-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlosses.py
134 lines (101 loc) · 4.77 KB
/
losses.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
from __future__ import print_function
import torch
import torch.nn as nn
# clear those instances that have no positive instances to avoid training error：
class SupConLoss_clear_new(nn.Module):
def __init__(self,opt):
super(SupConLoss_clear_new, self).__init__()
self.temperature = opt.ins_temp
self.temperature_a = opt.ins_tempa
def forward(self, features, attributes, labels):
device = (torch.device('cuda')
if features.is_cuda
else torch.device('cpu'))
batch_size = features.shape[0]
labels = labels.contiguous().view(-1, 1)
mask = torch.eq(labels, labels.T).float().to(device)
mask_n = 1 - mask
anchor_dot_contrast = torch.div(
torch.matmul(features, features.T),
self.temperature)
aa__dot_contrast=torch.div(
torch.matmul(attributes, attributes.T),
self.temperature_a)
# normalize the logits for numerical stability
logits_max, _ = torch.max(anchor_dot_contrast, dim=1, keepdim=True)
logits = anchor_dot_contrast - logits_max.detach()
aa_max, _ = torch.max(aa__dot_contrast, dim=1, keepdim=True)
aa_logits = aa__dot_contrast - aa_max.detach()
logits_mask = torch.scatter(# scatter_(dim, index, src):
torch.ones_like(mask),
1,#dim=1
torch.arange(batch_size).view(-1, 1).to(device),#index
0
)
mask = mask * logits_mask
single_samples = (mask.sum(1) == 0).float()#
# compute log_prob (logSoftmax)
exp_logits = torch.exp(logits) * logits_mask
log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True))
exp_vvlogits = torch.exp(logits) * logits_mask
log_vvprob = logits - torch.log(exp_vvlogits.sum(1, keepdim=True))
exp_aalogits = torch.exp(aa_logits) * logits_mask
log_aaprob = aa_logits - torch.log(exp_aalogits.sum(1, keepdim=True))
# compute mean of log-likelihood over positive
# invoid to devide the zero
mean_log_prob_pos = (mask * log_prob).sum(1) / (mask.sum(1)+single_samples)
# loss
# filter those single sample
loss_pos = - mean_log_prob_pos*(1-single_samples)
loss_pos = loss_pos.sum()/(loss_pos.shape[0]-single_samples.sum())
loss_neg=torch.nn.functional.mse_loss(mask_n * log_vvprob, mask_n * log_aaprob)
return loss_pos, loss_neg
class SupConLoss_clear_mask0(nn.Module):
def __init__(self,opt):
super(SupConLoss_clear_mask0, self).__init__()
self.temperature = opt.ins_temp
self.temperature_a = opt.ins_tempa
def forward(self, features, attributes, labels):
device = (torch.device('cuda')
if features.is_cuda
else torch.device('cpu'))
batch_size = features.shape[0]
labels = labels.contiguous().view(-1, 1)
mask = torch.eq(labels, labels.T).float().to(device)
# mask_t = torch.eq(mask, torch.zeros_like(mask)).float().to(device)
mask_n = 1 - mask
anchor_dot_contrast = torch.div(
torch.matmul(features, features.T),
self.temperature)
aa__dot_contrast=torch.div(
torch.matmul(attributes, attributes.T),
self.temperature_a)
# normalize the logits for numerical stability
logits_max, _ = torch.max(anchor_dot_contrast, dim=1, keepdim=True)
logits = anchor_dot_contrast - logits_max.detach()
aa_max, _ = torch.max(aa__dot_contrast, dim=1, keepdim=True)
aa_logits = aa__dot_contrast - aa_max.detach()
logits_mask = torch.scatter(# scatter_(dim, index, src):
torch.ones_like(mask),
1,
torch.arange(batch_size).view(-1, 1).to(device),#index
0
)
mask = mask * logits_mask
single_samples = (mask.sum(1) == 0).float()#
# compute log_prob (logSoftmax)
exp_logits = torch.exp(logits) * logits_mask
log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True))
exp_vvlogits = torch.exp(logits) * logits_mask
log_vvprob = logits - torch.log(exp_vvlogits.sum(1, keepdim=True))
exp_aalogits = torch.exp(aa_logits) * logits_mask
log_aaprob = aa_logits - torch.log(exp_aalogits.sum(1, keepdim=True))
# compute mean of log-likelihood over positive
# invoid to devide the zero
mean_log_prob_pos = (mask * log_prob).sum(1) / (mask.sum(1)+single_samples)
# loss
# filter those single sample
loss_pos = - mean_log_prob_pos*(1-single_samples)
loss_pos = loss_pos.sum()/(loss_pos.shape[0]-single_samples.sum())
loss_neg = torch.nn.functional.mse_loss(log_vvprob, log_aaprob)
return loss_pos, loss_neg