-
Notifications
You must be signed in to change notification settings - Fork 5
/
losses.py
180 lines (144 loc) · 6.72 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
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
import numpy
import torch
import torch.nn as nn
import pdb
class crossentropy_loss(torch.nn.Module):
def __init__(self):
super(crossentropy_loss, self).__init__()
self.name = "crossentropy_loss"
self.loss = nn.CrossEntropyLoss(reduction='mean')
def forward(self, positive, negative_1):
scores = torch.cat([positive, negative_1], dim=-1)
truth = torch.zeros(positive.shape[0],
dtype=torch.long).cuda() # positive.shape[1]+negative_1.shape[1]+negative_2.shape[1]).cuda()
# truth[:, 0] = 1
truth = torch.autograd.Variable(truth, requires_grad=False)
losses = self.loss(scores, truth)
return losses
class crossentropy_loss_AllNeg(torch.nn.Module):
def __init__(self):
super(crossentropy_loss_AllNeg, self).__init__()
self.name = "crossentropy_loss_bothNegSep"
self.loss = nn.CrossEntropyLoss(reduction='mean')
def forward(self, truth, scores):
# '''
# -----------#
truth = torch.autograd.Variable(truth, requires_grad=False)
truth = truth.view(truth.shape[0])
losses = self.loss(scores, truth)
# -----------#
# '''
return losses
class softmax_loss_wtpos(torch.nn.Module):
def __init__(self):
super(softmax_loss_wtpos, self).__init__()
def forward(self, positive, negative_1, negative_2):
negative = torch.cat([positive, negative_1], dim=-1) # , negative_2], dim=-1)
max_den = negative.max(dim=1, keepdim=True)[0].detach()
den = (negative - max_den).exp().sum(dim=-1, keepdim=True)
losses = (positive - max_den) - den.log()
return - (losses.mean())
class softmax_loss_AllNeg(torch.nn.Module):
def __init__(self):
super(softmax_loss_AllNeg, self).__init__()
def forward(self, positive, scores):
# negative = torch.cat([positive, negative_1],dim=-1)#, negative_2], dim=-1)
max_den = scores.max(dim=1, keepdim=True)[0].detach()
den = (scores - max_den).exp().sum(dim=-1, keepdim=True)
print("positive", positive.shape, "scores", scores.shape, "max_den", max_den.shape)
truth = positive.view(positive.shape[0])
losses = (scores[:, truth] - max_den) - den.log()
return - (losses.mean())
class test(torch.nn.Module):
def __init__(self):
super(test, self).__init__()
self.loss = nn.CrossEntropyLoss(reduction='mean')
def forward(self, positive, negative_1, negative_2):
# '''
max_den_e1 = negative_1.max(dim=1, keepdim=True)[0].detach()
##max_den_e2 = negative_2.max(dim=1, keepdim=True)[0].detach()
# print("max_den_e1",max_den_e1.shape)
# print("(negative_1-max_den_e1)",(negative_1-max_den_e1).shape)
den_e1 = (negative_1 - max_den_e1).exp().sum(dim=-1, keepdim=True)
##den_e2 = (negative_2-max_den_e2).exp().sum(dim=-1, keepdim=True)
# print("den_e1",den_e1.shape)
##losses = ((2*positive-max_den_e1-max_den_e2) - den_e1.log() - den_e2.log())
losses = ((positive - max_den_e1) - den_e1.log())
# print("positive-max_den_e1", (positive-max_den_e1).shape)
# print("(positive-max_den_e1)-den_e1.log()",((positive-max_den_e1)-den_e1.log()).shape)
# '''
den_e1_noOverflow = (negative_1).exp().sum(dim=-1, keepdim=True)
losses_noOverflow = ((positive) - den_e1_noOverflow.log())
##
# part 1c
scores_denPos = torch.cat([positive, negative_1], dim=-1)
max_den_e1_denPos = scores_denPos.max(dim=1, keepdim=True)[0].detach()
den_e1_denPos = (scores_denPos - max_den_e1_denPos).exp().sum(dim=-1, keepdim=True)
losses_denPos = ((positive - max_den_e1_denPos) - den_e1_denPos.log())
#
##part2
##scores = torch.cat([positive, negative_1, negative_2], dim=-1)
scores = torch.cat([positive, negative_1], dim=-1)
print("scores for pre-built functions", scores.shape)
truth = torch.zeros(positive.shape[0],
dtype=torch.long).cuda() # positive.shape[1]+negative_1.shape[1]+negative_2.shape[1]).cuda()
# truth[:, 0] = 1
truth = torch.autograd.Variable(truth, requires_grad=False)
losses_ce = self.loss(scores, truth)
# part3
loss1 = nn.LogSoftmax(dim=1)
loss2 = nn.NLLLoss()
scores3 = loss1(scores)
losses_ce_2 = loss2(scores, truth)
##
print("!!!")
print("nn.LogSoftmax", scores3[:, 0], scores3[:, 0].shape)
print("manual log softMax", losses, losses.shape)
print("manual log softMax noOverflow", losses_noOverflow, losses_noOverflow.shape)
print("manual log softMax with denPos", losses_denPos, losses_denPos.shape)
print("!!!")
##
print("CE loss:", losses_ce)
print("SM loss:", -losses.mean())
print("SM noOverflow loss:", -losses_noOverflow.mean())
print("SM denPos loss:", -losses_denPos.mean())
print("CE 2 loss:", losses_ce_2)
return -losses.mean()
class logistic_loss(torch.nn.Module):
def __init__(self):
super(logistic_loss, self).__init__()
def forward(self, positive, negative_1):
scores = torch.cat([positive, negative_1], dim=-1)
truth = torch.ones(1, positive.shape[1] + negative_1.shape[1]).cuda()
truth[0, 0] = -1
truth = -truth
truth = torch.autograd.Variable(truth, requires_grad=False)
# print("Logistic loss forward:",scores*truth)
x = torch.log(1 + torch.exp(-scores * truth))
total = x.sum()
return total / ((positive.shape[1] + negative_1.shape[1]) * positive.shape[0])
class hinge_loss(torch.nn.Module):
def __init__(self):
super(hinge_loss, self).__init__()
def forward(self, positive, negative_1):
scores = torch.cat([positive, negative_1], dim=-1)
truth = torch.ones(1, positive.shape[1] + negative_1.shape[1]).cuda()
truth[0, 0] = -1
truth = -truth
truth = torch.autograd.Variable(truth, requires_grad=False)
return nn.HingeEmbeddingLoss(margin=5)(scores, truth)
class margin_pairwise_loss(torch.nn.Module):
def __init__(self, margin=10.0):
super(margin_pairwise_loss, self).__init__()
self.margin = margin
def forward(self, positive, negative):
# print("positive:{}, negative:{}".format(positive.size(), negative.size()))
# print("diff: {}".format(diff.size()))
# pdb.set_trace()
diff = positive - negative + self.margin
#diff = torch.max(diff, torch.tensor([0.0]))
diff = torch.max(diff, torch.tensor([0.0]).cuda())
# print("max_diff:{}".format(max_diff.size()))
loss = diff.sum()
# xx=input()
return loss