-
Notifications
You must be signed in to change notification settings - Fork 0
/
loss.py
209 lines (184 loc) · 8.52 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
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class Our_SupCL_loss(nn.Module):
'''
modified based on https://github.com/HobbitLong/SupContrast/blob/master/losses.py
Supervised Contrastive Learning: https://arxiv.org/pdf/2004.11362.pdf.
It also supports the unsupervised contrastive loss in SimCLR
'''
def __init__(self, args, contrast_mode='all', base_temperature=0.07):
super(Our_SupCL_loss, self).__init__()
self.temperature = args.temp
self.contrast_mode = contrast_mode
self.base_temperature = base_temperature
self.c = args.c
self.device = args.device
# self.gamma = args.gamma
gamma_schedule = torch.ones(args.n_epoch)
# Linear
# gamma_schedule = torch.linspace(0, 1, args.warm_up)
# Exponent
gamma_schedule[:args.warm_up] = torch.logspace(-args.warm_up, 0, args.warm_up, np.e)
self.gamma = gamma_schedule * args.gamma
self.eta = args.eta
self.warmup = args.warm_up
self.aug_views = args.aug_views
def forward(self, features, logits, labels, index, delta_smooth, epoch):
"""Compute loss for model. If both `labels` and `mask` are None,
it degenerates to SimCLR unsupervised loss:
https://arxiv.org/pdf/2002.05709.pdf
Args:
features: hidden vector of shape [bsz, n_views, ...].
labels: ground truth of shape [bsz].
mask: contrastive mask of shape [bsz, bsz], mask_{i,j}=1 if sample j
has the same class as sample i. Can be asymmetric.
Returns:
A loss scalar.
"""
#TODO: l_classification
delta_smooth = delta_smooth.to(self.device)
logits_norm = F.softmax(logits, dim=1)
if len(labels.size()) == 1:
db_labels = labels.repeat(self.aug_views)
target_oh = torch.zeros(db_labels.size(0), self.c + 1).to(self.device).scatter_(1, db_labels.view(-1, 1), (
1 - delta_smooth[index].repeat(self.aug_views)).view(-1, 1)) # convert label to one-hot
target_oh[:, -1] = delta_smooth[index].repeat(self.aug_views)
# calculate as cross-entropy loss
loss_cla = -torch.mean(torch.sum(torch.log_softmax(logits, dim=1) * target_oh, dim=1))
# loss_cla = -torch.mean(torch.sum(torch.log(logits_norm) * target_oh, dim=1))
if len(features.shape) < 3:
raise ValueError('`features` needs to be [bsz, n_views, ...],'
'at least 3 dimensions are required')
if len(features.shape) > 3:
features = features.view(features.shape[0], features.shape[1], -1)
batch_size = features.shape[0]
labels = labels.contiguous().view(-1, 1)
if labels.shape[0] != batch_size:
raise ValueError('Num of labels does not match num of features')
mask = torch.eq(labels, labels.T).float().to(self.device)
contrast_count = features.shape[1]
contrast_feature = torch.cat(torch.unbind(features, dim=1), dim=0)
if self.contrast_mode == 'one':
anchor_feature = features[:, 0]
anchor_count = 1
elif self.contrast_mode == 'all':
anchor_feature = contrast_feature
anchor_count = contrast_count
else:
raise ValueError('Unknown mode: {}'.format(self.contrast_mode))
# compute logits
anchor_dot_contrast = torch.div(
torch.matmul(anchor_feature, contrast_feature.T),
self.temperature)
# for numerical stability
logits_max, _ = torch.max(anchor_dot_contrast, dim=1, keepdim=True)
logits = anchor_dot_contrast - logits_max.detach()
# tile mask
mask = mask.repeat(anchor_count, contrast_count)
# mask-out self-contrast cases
logits_mask = torch.scatter(
torch.ones_like(mask),
1,
torch.arange(batch_size * anchor_count).view(-1, 1).to(self.device),
0
)
# TODO: newly added, mask-out high uncertainty cases
low_unc = logits_norm[:, -1] <= self.eta
low_uncertain_mask = low_unc.view(-1,1)&low_unc.view(1,-1)
# low_uncertain_mask = torch.zeros_like(mask)
# low_uncertain_mask[low_unc] = True
low_uncertain_mask.float().to(self.device)
mask = mask * low_uncertain_mask
mask = mask * logits_mask
# compute log_prob
# for numerical stability
tmp = 1e-6
exp_logits = torch.exp(logits) * logits_mask
log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True)+tmp)
# compute mean of log-likelihood over positive
mean_log_prob_pos = (mask * log_prob).sum(1) / (mask.sum(1)+tmp)
# loss
loss_cl = - (self.temperature / self.base_temperature) * mean_log_prob_pos
loss_cl = loss_cl.view(anchor_count, batch_size).mean()
#TODO: combine loss_cla and loss_cl
loss = loss_cla + self.gamma[epoch] * loss_cl
return loss
class CELoss(nn.Module):
'''
CE, class_num=10
'''
def __init__(self, num_class, device):
super(CELoss, self).__init__()
self.c = num_class
self.device = device
def forward(self, logits, targets):
# logits_norm = F.softmax(logits, dim=1)
if len(targets.size())==1:
target_oh = torch.zeros(targets.size(0), self.c).to(self.device).scatter_(1, targets.view(-1,1), 1) # convert label to one-hot
loss = -torch.mean(torch.sum(torch.log_softmax(logits, dim=1) * target_oh, dim=1))
return loss
class CE_OurLoss(nn.Module):
'''
CE for ours, class_num=11
'''
def __init__(self, num_class, device):
super(CE_OurLoss, self).__init__()
self.c = num_class
self.device = device
def forward(self, logits, targets):
# logits_norm = F.softmax(logits, dim=1)
if len(targets.size())==1:
target_oh = torch.zeros(targets.size(0), self.c).to(self.device).scatter_(1, targets.view(-1,1), 1) # convert label to one-hot
# TODO for debug
# print(logits.min(), logits.max(), logits_norm.min())
loss = -torch.mean(torch.sum(torch.log_softmax(logits, dim=1)[:,:-1] * target_oh, dim=1))
# loss = -torch.mean(torch.sum(torch.log(logits_norm[:,:-1]) * target_oh, dim=1))
return loss
class OurLoss(nn.Module):
'''
our loss with regularization
'''
def __init__(self, num_class, device):
super(OurLoss, self).__init__()
self.c = num_class
self.device = device
#self.epsilon = sys.float_info.epsilon
#self.init_lambda = init_lambda
#self.lamb = lamb
#self.max_eps = max_eps
# def exponential_decay_lambda(self, cur_eps, max_eps, initial_lambda):
# lambda_val = initial_lambda * (0.1 ** (cur_eps / max_eps))
# lambda_val = max(lambda_val, 0)
# return lambda_val
#
# def linear_decay_lambda(self, cur_eps, max_eps, initial_lambda):
# #keep max_eps > 0
# if max_eps < self.epsilon:
# max_eps = self.epsilon
# lambda_val = max(initial_lambda - (cur_eps / max_eps) * initial_lambda, 0)
# return lambda_val
def forward(self, logits, targets, index, delta_smooth):
delta_smooth = delta_smooth.to(self.device)
# logits_norm = F.softmax(logits, dim=1)
if len(targets.size()) == 1:
target_oh = torch.zeros(targets.size(0), self.c+1).to(self.device).scatter_(1, targets.view(-1,1), (1-delta_smooth[index]).view(-1,1)) # convert label to one-hot
target_oh[:,-1] = delta_smooth[index]
#calculate as cross-entropy loss
loss = -torch.mean(torch.sum(torch.log_softmax(logits, dim=1) * target_oh, dim=1))
# loss = -torch.mean(torch.sum(torch.log(logits_norm) * target_oh, dim=1))
# ce+regularization term
#l_ce = -torch.sum(torch.log(logits_norm[:, :-1]) * target_oh, dim=1)
# args.lam should not be too large, keep l_reg smaller(or comparable to) l_ce
#l_reg = torch.sum(self.lamb * (1 - logits_norm[:, :-1]), dim=1)
# lamb = self.linear_decay_lambda(cur_eps, self.max_eps, self.init_lambda)
# l_reg = torch.sum(lamb * (1 - logits_norm[:, :-1]), dim=1)
# l_reg = torch.sum(
# lamb * (1 - logits_norm[:, :-1]) / logits_norm[:, -1].unsqueeze(1).expand_as(logits_norm[:, :-1]),
# dim=1)
# if cur_eps < self.max_eps:
# loss = torch.mean(l_ce + l_reg)
# else:
# loss = torch.mean(l_ce)
return loss