-
Notifications
You must be signed in to change notification settings - Fork 1
/
loss.py
201 lines (162 loc) · 8.62 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import numpy as np
# Loss functions
def loss_softcoteaching(y_1, y_2, t, forget_rate, ind, noise_or_not, num_classes, loss_rate, split_rate):
temp = F.one_hot(t,num_classes)
pred_1 = torch.tensor(y_1.tolist(),requires_grad=False)
pred_1 = pred_1.cuda()
pred_2 = torch.tensor(y_2.tolist(),requires_grad=False)
pred_2 = pred_2.cuda()
t_1 = torch.argmax((1-loss_rate)*temp+loss_rate*pred_2,dim=1)
t_1 = t_1.cuda()
t_2 = torch.argmax((1-loss_rate)*temp+loss_rate*pred_1,dim=1)
t_2 = t_2.cuda()
# print(t_1,t_2)
# loss_1 = F.cross_entropy(y_1, t_1, reduce = False)
loss_1 = F.cross_entropy(y_1, t, reduce = False)
ind_1_sorted = torch.argsort(loss_1).cuda()
ind_1_cum = ind_1_sorted.cpu()
loss_1_sorted = loss_1[ind_1_sorted]
# loss_2 = F.cross_entropy(y_2, t_2, reduce = False)
loss_2 = F.cross_entropy(y_2, t, reduce = False)
ind_2_sorted = torch.argsort(loss_2).cuda()
ind_2_cum = ind_2_sorted.cpu()
loss_2_sorted = loss_2[ind_2_sorted]
remember_rate = 1 - max(forget_rate,split_rate)
throw_rate = 1 - min(forget_rate,split_rate)
# remember_rate = 1 - forget_rate
num_remember = int(remember_rate * len(loss_1_sorted))
num_throw = int(throw_rate * len(loss_1_sorted))
pure_ratio_1 = np.sum(noise_or_not[ind[ind_1_cum[:num_remember]]])/float(num_remember)
pure_ratio_2 = np.sum(noise_or_not[ind[ind_2_cum[:num_remember]]])/float(num_remember)
ind_1_update=ind_1_sorted[:num_remember]
ind_2_update=ind_2_sorted[:num_remember]
ind_1_midate=ind_1_sorted[num_remember:num_throw]
ind_2_midate=ind_2_sorted[num_remember:num_throw]
ind_1_downdate=ind_1_sorted[num_throw:]
ind_2_downdate=ind_2_sorted[num_throw:]
# exchange
if len(ind_1_midate)==0:
loss_1_update = F.cross_entropy(y_1[ind_2_update], t[ind_2_update])
loss_2_update = F.cross_entropy(y_2[ind_1_update], t[ind_1_update])
else:
loss_1_update = F.cross_entropy(y_1[ind_2_update], t[ind_2_update]) + F.cross_entropy(y_1[ind_2_midate], t_2[ind_2_midate])
loss_2_update = F.cross_entropy(y_2[ind_1_update], t[ind_1_update]) + F.cross_entropy(y_2[ind_1_midate], t_1[ind_1_midate])
return torch.sum(loss_1_update)/num_remember, torch.sum(loss_2_update)/num_remember, pure_ratio_1, pure_ratio_2
def loss_curve(y_1, y_2, t, forget_rate, ind, noise_or_not):
loss_1 = F.cross_entropy(y_1, t, reduce = False)
ind_1_sorted = np.argsort(loss_1.data).cuda()
loss_1_sorted = loss_1[ind_1_sorted]
loss_2 = F.cross_entropy(y_2, t, reduce = False)
ind_2_sorted = np.argsort(loss_2.data).cuda()
loss_2_sorted = loss_2[ind_2_sorted]
batch_size=len(loss_1_sorted)
remember_rate = 1 - forget_rate
num_remember = int(remember_rate * batch_size)
ind1=np.arange(128,dtype=np.int)
ind1=ind1[noise_or_not[ind]]
pure_ratio_1 = np.sum(noise_or_not[ind[ind_1_sorted[:num_remember]]])/float(num_remember)
pure_ratio_2 = np.sum(noise_or_not[ind[ind_2_sorted[:num_remember]]])/float(num_remember)
ind_1_update=ind_1_sorted[:num_remember]
ind_2_update=ind_2_sorted[:num_remember]
# exchange
loss_1_update = F.cross_entropy(y_1[ind_2_update], t[ind_2_update])
loss_2_update = F.cross_entropy(y_2[ind_1_update], t[ind_1_update])
return torch.sum(loss_1_update)/num_remember, torch.sum(loss_2_update)/num_remember, pure_ratio_1, pure_ratio_2, (torch.sum(loss_1[ind1])+torch.sum(loss_2[ind1]))/2, (torch.sum(loss_1)+torch.sum(loss_2)-torch.sum(loss_1[ind1])-torch.sum(loss_2[ind1]))/2
def loss_selfteaching(y_1, y_2, t, forget_rate, ind, noise_or_not, num_classes):
temp = F.one_hot(t,num_classes)
pred_1 = torch.tensor(y_1.tolist(),requires_grad=False)
pred_1 = pred_1.cuda()
pred_2 = torch.tensor(y_2.tolist(),requires_grad=False)
pred_2 = pred_2.cuda()
t_1 = torch.argmax((1-forget_rate)*temp+forget_rate*pred_2,dim=1)
t_1 = t_1.cuda()
t_2 = torch.argmax((1-forget_rate)*temp+forget_rate*pred_1,dim=1)
t_2 = t_2.cuda()
# print(t_1,t_2)
loss_1 = F.cross_entropy(y_1, t_1, reduce = False)
# ind_1_sorted = torch.argsort(loss_1).cuda()
# ind_1_cum = ind_1_sorted.cpu()
# ind_1_cum = ind_1_sorted
# loss_1_sorted = loss_1[ind_1_sorted]
loss_2 = F.cross_entropy(y_2, t_2, reduce = False)
# ind_2_sorted = torch.argsort(loss_2).cuda()
# ind_2_cum = ind_2_sorted.cpu()
# ind_2_cum = ind_2_sorted
# loss_2_sorted = loss_2[ind_2_sorted]
# remember_rate = 1 - forget_rate
num_remember = len(loss_1)
pure_ratio_1 = 1
pure_ratio_2 = 1
# ind_1_update=ind_1_sorted[:num_remember]
# ind_2_update=ind_2_sorted[:num_remember]
# exchange
loss_1_update = loss_1.cuda()
loss_2_update = loss_2.cuda()
# loss_1_update = F.cross_entropy(y_1[ind_2_update], t[ind_2_update])
# loss_2_update = F.cross_entropy(y_2[ind_1_update], t[ind_1_update])
return torch.sum(loss_1_update)/num_remember, torch.sum(loss_2_update)/num_remember, pure_ratio_1, pure_ratio_2
def loss_coteaching(y_1, y_2, t, forget_rate, ind, noise_or_not):
loss_1 = F.cross_entropy(y_1, t, reduce = False)
ind_1_sorted = torch.argsort(loss_1).cuda()
ind_1_cum = ind_1_sorted.cpu()
# ind_1_cum = ind_1_sorted
loss_1_sorted = loss_1[ind_1_sorted]
loss_2 = F.cross_entropy(y_2, t, reduce = False)
ind_2_sorted = torch.argsort(loss_2).cuda()
ind_2_cum = ind_2_sorted.cpu()
# ind_2_cum = ind_2_sorted
loss_2_sorted = loss_2[ind_2_sorted]
remember_rate = 1 - forget_rate
num_remember = int(remember_rate * len(loss_1_sorted))
pure_ratio_1 = np.sum(noise_or_not[ind[ind_1_cum[:num_remember]]])/float(num_remember)
pure_ratio_2 = np.sum(noise_or_not[ind[ind_2_cum[:num_remember]]])/float(num_remember)
ind_1_update=ind_1_sorted[:num_remember]
ind_2_update=ind_2_sorted[:num_remember]
# exchange
loss_1_update = F.cross_entropy(y_1[ind_2_update], t[ind_2_update])
loss_2_update = F.cross_entropy(y_2[ind_1_update], t[ind_1_update])
return torch.sum(loss_1_update)/num_remember, torch.sum(loss_2_update)/num_remember, pure_ratio_1, pure_ratio_2
def loss_curriculum(y_1, y_2, t, forget_rate, ind):
loss_1 = F.cross_entropy(y_1, t, reduce = False)
ind_1_sorted = np.argsort(loss_1.data)
loss_1_sorted = loss_1[ind_1_sorted]
loss_2 = F.cross_entropy(y_2, t, reduce = False)
ind_2_sorted = np.argsort(loss_2.data)
loss_2_sorted = loss_2[ind_2_sorted]
remember_rate = 1 - forget_rate
num_remember = int(remember_rate * len(loss_1_sorted))
ind_1_update=ind_1_sorted[:num_remember]
ind_2_update=ind_2_sorted[:num_remember]
# exchange
# loss_1_update = F.cross_entropy(y_1[ind_2_update], t[ind_2_update])
# loss_2_update = F.cross_entropy(y_2[ind_1_update], t[ind_1_update])
loss_1_update = F.cross_entropy(y_1[:num_remember], t[:num_remember])
loss_2_update = F.cross_entropy(y_2[:num_remember], t[:num_remember])
return torch.sum(loss_1_update)/num_remember, torch.sum(loss_2_update)/num_remember, 1, 1
def loss_3teaching(y_1, y_2, y_3, t, forget_rate, ind, noise_or_not):
loss_1 = F.cross_entropy(y_1, t, reduce = False)
ind_1_sorted = np.argsort(loss_1.data).cuda()
loss_1_sorted = loss_1[ind_1_sorted]
loss_2 = F.cross_entropy(y_2, t, reduce = False)
ind_2_sorted = np.argsort(loss_2.data).cuda()
loss_2_sorted = loss_2[ind_2_sorted]
loss_3 = F.cross_entropy(y_3, t, reduce = False)
ind_3_sorted = np.argsort(loss_3.data).cuda()
loss_3_sorted = loss_3[ind_3_sorted]
remember_rate = 1 - forget_rate
num_remember = int(remember_rate * len(loss_1_sorted))
pure_ratio_1 = np.sum(noise_or_not[ind[ind_1_sorted[:num_remember]]])/float(num_remember)
pure_ratio_2 = np.sum(noise_or_not[ind[ind_2_sorted[:num_remember]]])/float(num_remember)
pure_ratio_3 = np.sum(noise_or_not[ind[ind_3_sorted[:num_remember]]])/float(num_remember)
ind_1_update=np.intersect1d(ind_2_sorted[:num_remember],ind_3_sorted[:num_remember],assume_unique=True)
ind_2_update=np.intersect1d(ind_3_sorted[:num_remember],ind_1_sorted[:num_remember],assume_unique=True)
ind_3_update=np.intersect1d(ind_1_sorted[:num_remember],ind_2_sorted[:num_remember],assume_unique=True)
# exchange
loss_1_update = F.cross_entropy(y_1[ind_1_update], t[ind_1_update])
loss_2_update = F.cross_entropy(y_2[ind_2_update], t[ind_2_update])
loss_3_update = F.cross_entropy(y_3[ind_3_update], t[ind_3_update])
return torch.sum(loss_1_update)/num_remember, torch.sum(loss_2_update)/num_remember, torch.sum(loss_3_update)/num_remember, pure_ratio_1, pure_ratio_2, pure_ratio_3