-
Notifications
You must be signed in to change notification settings - Fork 0
/
loss.py
executable file
·198 lines (150 loc) · 6.18 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
"""Loss function"""
import torch
import torch.nn as nn
from torch.nn import functional as F
def kl_loss(pred1, pred2, eps=1e-8):
single_loss = torch.sum(pred2 * torch.log(eps + pred2 / (pred1 + eps)), 1)
return single_loss
def ce_loss(logit, pred):
single_loss = torch.sum(-pred * F.log_softmax(logit, dim=-1), dim=-1)
return single_loss
class BirankLoss(nn.Module):
"""
Compute contrastive loss
"""
def __init__(self, margin=0, max_violation=False):
super(BirankLoss, self).__init__()
self.margin = margin
self.max_violation = max_violation
def forward(self, scores):
# compute image-sentence score matrix
diagonal = scores.diag().view(scores.size(0), 1)
d1 = diagonal.expand_as(scores)
d2 = diagonal.t().expand_as(scores)
# compare every diagonal score to scores in its column
# caption retrieval
cost_s = (self.margin + scores - d1).clamp(min=0)
# compare every diagonal score to scores in its row
# image retrieval
cost_im = (self.margin + scores - d2).clamp(min=0)
# clear diagonals
mask = torch.eye(scores.size(0)) > .5
if torch.cuda.is_available():
I = mask.cuda()
cost_s = cost_s.masked_fill_(I, 0)
cost_im = cost_im.masked_fill_(I, 0)
# keep the maximum violating negative for each query
if self.max_violation:
cost_s = cost_s.max(1)[0]
cost_im = cost_im.max(0)[0]
return cost_s.sum() + cost_im.sum()
class CMPMLoss(nn.Module):
def __init__(self, smooth=10.0, eps=1e-8):
super(CMPMLoss, self).__init__()
self.smooth = smooth
self.eps = eps
self.softmax = nn.Softmax(dim=-1)
self.relu = nn.ReLU()
def forward(self, sims, labels):
label_mask = labels.view(-1, 1) - labels.view(1, -1)
label_mask = (torch.abs(label_mask) < 0.5).float()
label_mask_norm = F.normalize(label_mask, p=1, dim=-1)
v2t_pred = self.softmax(self.smooth * sims)
t2v_pred = self.softmax(self.smooth * sims.t())
v2t_loss = kl_loss(label_mask_norm, v2t_pred, self.eps)
t2v_loss = kl_loss(label_mask_norm, t2v_pred, self.eps)
loss = v2t_loss.mean() + t2v_loss.mean()
return loss
class BoostabsLoss(nn.Module):
"""
Compute contrastive loss
"""
def __init__(self, margin=0, beta=1.0, gamma=0.5):
super(BoostabsLoss, self).__init__()
self.margin = margin * beta
self.margin_pos = self.margin * gamma
self.margin_neg = self.margin * (1.0 - gamma)
def forward(self, scores, scores_anchor):
# compute image-sentence score matrix
pos_sims = scores.diag().view(scores.size(0), 1)
pos_sims_anchor = scores_anchor.diag().view(scores_anchor.size(0), 1)
# compare every diagonal score to scores in its column
cost_pos = (self.margin_pos + pos_sims_anchor - pos_sims).clamp(min=0).sum()
cost_neg = (self.margin_neg + scores - scores_anchor).clamp(min=0)
# clear diagonals
mask = torch.eye(scores.size(0)) > .5
if torch.cuda.is_available():
I = mask.cuda()
cost_neg = cost_neg.masked_fill_(I, 0)
cost_neg = cost_neg.max(1)[0].sum() + cost_neg.max(0)[0].sum()
return cost_pos * 2 + cost_neg
class BoostrelLoss(nn.Module):
"""
Compute contrastive loss
"""
def __init__(self, margin=0, beta=1.0):
super(BoostrelLoss, self).__init__()
self.margin = margin * beta
def forward(self, scores, scores_anchor):
# compute image-sentence score matrix
diagonal = scores.diag().view(scores.size(0), 1)
d1 = diagonal.expand_as(scores)
d2 = diagonal.t().expand_as(scores)
diagonal_anchor = scores_anchor.diag().view(scores_anchor.size(0), 1)
d1_anchor = diagonal_anchor.expand_as(scores_anchor)
d2_anchor = diagonal_anchor.t().expand_as(scores_anchor)
# compare every diagonal score to scores in its column
cost_s = (self.margin + (d1_anchor-scores_anchor) - (d1-scores)).clamp(min=0)
cost_im = (self.margin + (d2_anchor-scores_anchor) - (d2-scores)).clamp(min=0)
# clear diagonals
mask = torch.eye(scores.size(0)) > .5
if torch.cuda.is_available():
I = mask.cuda()
cost_s = cost_s.masked_fill_(I, 0)
cost_im = cost_im.masked_fill_(I, 0)
cost_s = cost_s.max(1)[0]
cost_im = cost_im.max(0)[0]
return cost_s.sum() + cost_im.sum()
class BoostDINOLoss(nn.Module):
def __init__(self, smooth=10.0):
super(BoostDINOLoss, self).__init__()
self.smooth = smooth
self.softmax = nn.Softmax(dim=-1)
self.relu = nn.ReLU()
def forward(self, scores, scores_t):
v2t_pred = self.softmax(self.smooth * scores_t)
t2v_pred = self.softmax(self.smooth * scores_t.t())
v2t_loss = ce_loss(self.smooth * scores, v2t_pred)
t2v_loss = ce_loss(self.smooth * scores.t(), t2v_pred)
loss = v2t_loss.mean() + t2v_loss.mean()
return loss
class BoostDINOppLoss(nn.Module):
def __init__(self, smooth=10.0, eps=1e-8):
super(BoostDINOppLoss, self).__init__()
self.smooth = smooth
self.eps = eps
self.softmax = nn.Softmax(dim=-1)
self.relu = nn.ReLU()
def forward(self, scores, scores_t):
v2t_pred = self.softmax(self.smooth * scores)
t2v_pred = self.softmax(self.smooth * scores.t())
v2t_pred_t = self.softmax(self.smooth * scores_t)
t2v_pred_t = self.softmax(self.smooth * scores_t.t())
v2t_loss = kl_loss(v2t_pred_t, v2t_pred, self.eps)
t2v_loss = kl_loss(t2v_pred_t, t2v_pred, self.eps)
loss = v2t_loss.mean() + t2v_loss.mean()
return loss
_loss = {
'birank': BirankLoss,
'cmpm': CMPMLoss,
'boostabs': BoostabsLoss,
'boostrel': BoostrelLoss,
'dino': BoostDINOLoss,
'dinopp': BoostDINOppLoss,
}
def init_loss(name, **kwargs):
"""Initializes an dataset."""
avai_losses = list(_loss.keys())
if name not in avai_losses:
raise ValueError('Invalid loss function. Received "{}"'.format(name))
return _loss[name](**kwargs)