sxtforreal
commited on
Create loss.py
Browse filesThis file holds 4 loss functions for the 4 models respectively.
loss.py
ADDED
@@ -0,0 +1,228 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import config
|
5 |
+
|
6 |
+
|
7 |
+
class ContrastiveLoss_simcse(nn.Module):
|
8 |
+
"""SimCSE loss"""
|
9 |
+
|
10 |
+
def __init__(self):
|
11 |
+
super(ContrastiveLoss_simcse, self).__init__()
|
12 |
+
self.temperature = config.temperature
|
13 |
+
|
14 |
+
def forward(self, feature_vectors, labels):
|
15 |
+
normalized_features = F.normalize(
|
16 |
+
feature_vectors, p=2, dim=0
|
17 |
+
) # normalize along columns
|
18 |
+
|
19 |
+
# Identify indices for each label
|
20 |
+
anchor_indices = (labels == 0).nonzero().squeeze(dim=1)
|
21 |
+
positive_indices = (labels == 1).nonzero().squeeze(dim=1)
|
22 |
+
negative_indices = (labels == 2).nonzero().squeeze(dim=1)
|
23 |
+
|
24 |
+
# Extract tensors based on labels
|
25 |
+
anchor = normalized_features[anchor_indices]
|
26 |
+
positives = normalized_features[positive_indices]
|
27 |
+
negatives = normalized_features[negative_indices]
|
28 |
+
pos_and_neg = torch.cat([positives, negatives])
|
29 |
+
|
30 |
+
denominator = torch.sum(
|
31 |
+
torch.exp(
|
32 |
+
torch.div(
|
33 |
+
torch.matmul(anchor, torch.transpose(pos_and_neg, 0, 1)),
|
34 |
+
self.temperature,
|
35 |
+
)
|
36 |
+
)
|
37 |
+
)
|
38 |
+
|
39 |
+
numerator = torch.exp(
|
40 |
+
torch.div(
|
41 |
+
torch.matmul(anchor, torch.transpose(positives, 0, 1)),
|
42 |
+
self.temperature,
|
43 |
+
)
|
44 |
+
)
|
45 |
+
|
46 |
+
loss = -torch.log(
|
47 |
+
torch.div(
|
48 |
+
numerator,
|
49 |
+
denominator,
|
50 |
+
)
|
51 |
+
)
|
52 |
+
|
53 |
+
return loss
|
54 |
+
|
55 |
+
|
56 |
+
class ContrastiveLoss_simcse_w(nn.Module):
|
57 |
+
"""SimCSE loss with weighting."""
|
58 |
+
|
59 |
+
def __init__(self):
|
60 |
+
super(ContrastiveLoss_simcse_w, self).__init__()
|
61 |
+
self.temperature = config.temperature
|
62 |
+
|
63 |
+
def forward(self, feature_vectors, labels, scores):
|
64 |
+
normalized_features = F.normalize(
|
65 |
+
feature_vectors, p=2, dim=0
|
66 |
+
) # normalize along columns
|
67 |
+
|
68 |
+
# Identify indices for each label
|
69 |
+
anchor_indices = (labels == 0).nonzero().squeeze(dim=1)
|
70 |
+
positive_indices = (labels == 1).nonzero().squeeze(dim=1)
|
71 |
+
negative_indices = (labels == 2).nonzero().squeeze(dim=1)
|
72 |
+
|
73 |
+
pos_scores = scores[positive_indices].float()
|
74 |
+
normalized_neg_scores = F.normalize(
|
75 |
+
scores[negative_indices].float(), p=2, dim=0
|
76 |
+
) # l2-norm
|
77 |
+
normalized_neg_scores += 1
|
78 |
+
scores = torch.cat([pos_scores, normalized_neg_scores])
|
79 |
+
|
80 |
+
# Extract tensors based on labels
|
81 |
+
anchor = normalized_features[anchor_indices]
|
82 |
+
positives = normalized_features[positive_indices]
|
83 |
+
negatives = normalized_features[negative_indices]
|
84 |
+
pos_and_neg = torch.cat([positives, negatives])
|
85 |
+
|
86 |
+
denominator = torch.sum(
|
87 |
+
torch.exp(
|
88 |
+
scores
|
89 |
+
* torch.div(
|
90 |
+
torch.matmul(anchor, torch.transpose(pos_and_neg, 0, 1)),
|
91 |
+
self.temperature,
|
92 |
+
)
|
93 |
+
)
|
94 |
+
)
|
95 |
+
|
96 |
+
numerator = torch.exp(
|
97 |
+
torch.div(
|
98 |
+
torch.matmul(anchor, torch.transpose(positives, 0, 1)),
|
99 |
+
self.temperature,
|
100 |
+
)
|
101 |
+
)
|
102 |
+
|
103 |
+
loss = -torch.log(
|
104 |
+
torch.div(
|
105 |
+
numerator,
|
106 |
+
denominator,
|
107 |
+
)
|
108 |
+
)
|
109 |
+
|
110 |
+
return loss
|
111 |
+
|
112 |
+
|
113 |
+
class ContrastiveLoss_samp(nn.Module):
|
114 |
+
"""Supervised contrastive loss without weighting."""
|
115 |
+
|
116 |
+
def __init__(self):
|
117 |
+
super(ContrastiveLoss_samp, self).__init__()
|
118 |
+
self.temperature = config.temperature
|
119 |
+
|
120 |
+
def forward(self, feature_vectors, labels):
|
121 |
+
# Normalize feature vectors
|
122 |
+
normalized_features = F.normalize(
|
123 |
+
feature_vectors, p=2, dim=0
|
124 |
+
) # normalize along columns
|
125 |
+
|
126 |
+
# Identify indices for each label
|
127 |
+
anchor_indices = (labels == 0).nonzero().squeeze(dim=1)
|
128 |
+
positive_indices = (labels == 1).nonzero().squeeze(dim=1)
|
129 |
+
negative_indices = (labels == 2).nonzero().squeeze(dim=1)
|
130 |
+
|
131 |
+
# Extract tensors based on labels
|
132 |
+
anchor = normalized_features[anchor_indices]
|
133 |
+
positives = normalized_features[positive_indices]
|
134 |
+
negatives = normalized_features[negative_indices]
|
135 |
+
pos_and_neg = torch.cat([positives, negatives])
|
136 |
+
|
137 |
+
pos_cardinal = positives.shape[0]
|
138 |
+
|
139 |
+
denominator = torch.sum(
|
140 |
+
torch.exp(
|
141 |
+
torch.div(
|
142 |
+
torch.matmul(anchor, torch.transpose(pos_and_neg, 0, 1)),
|
143 |
+
self.temperature,
|
144 |
+
)
|
145 |
+
)
|
146 |
+
)
|
147 |
+
|
148 |
+
sum_log_ent = torch.sum(
|
149 |
+
torch.log(
|
150 |
+
torch.div(
|
151 |
+
torch.exp(
|
152 |
+
torch.div(
|
153 |
+
torch.matmul(anchor, torch.transpose(positives, 0, 1)),
|
154 |
+
self.temperature,
|
155 |
+
)
|
156 |
+
),
|
157 |
+
denominator,
|
158 |
+
)
|
159 |
+
)
|
160 |
+
)
|
161 |
+
|
162 |
+
scale = -1 / pos_cardinal
|
163 |
+
|
164 |
+
return scale * sum_log_ent
|
165 |
+
|
166 |
+
|
167 |
+
class ContrastiveLoss_samp_w(nn.Module):
|
168 |
+
"""Supervised contrastive loss with weighting."""
|
169 |
+
|
170 |
+
def __init__(self):
|
171 |
+
super(ContrastiveLoss_samp_w, self).__init__()
|
172 |
+
self.temperature = config.temperature
|
173 |
+
|
174 |
+
def forward(self, feature_vectors, labels, scores):
|
175 |
+
# Normalize feature vectors
|
176 |
+
normalized_features = F.normalize(
|
177 |
+
feature_vectors, p=2, dim=0
|
178 |
+
) # normalize along columns
|
179 |
+
|
180 |
+
# Identify indices for each label
|
181 |
+
anchor_indices = (labels == 0).nonzero().squeeze(dim=1)
|
182 |
+
positive_indices = (labels == 1).nonzero().squeeze(dim=1)
|
183 |
+
negative_indices = (labels == 2).nonzero().squeeze(dim=1)
|
184 |
+
|
185 |
+
# Normalize score vector
|
186 |
+
num_skip = len(positive_indices) + 1
|
187 |
+
pos_scores = scores[: (num_skip - 1)].float() # exclude anchor
|
188 |
+
normalized_neg_scores = F.normalize(
|
189 |
+
scores[num_skip:].float(), p=2, dim=0
|
190 |
+
) # l2-norm
|
191 |
+
normalized_neg_scores += 1
|
192 |
+
scores = torch.cat([pos_scores, normalized_neg_scores])
|
193 |
+
|
194 |
+
# Extract tensors based on labels
|
195 |
+
anchor = normalized_features[anchor_indices]
|
196 |
+
positives = normalized_features[positive_indices]
|
197 |
+
negatives = normalized_features[negative_indices]
|
198 |
+
pos_and_neg = torch.cat([positives, negatives])
|
199 |
+
|
200 |
+
pos_cardinal = positives.shape[0]
|
201 |
+
|
202 |
+
denominator = torch.sum(
|
203 |
+
torch.exp(
|
204 |
+
scores
|
205 |
+
* torch.div(
|
206 |
+
torch.matmul(anchor, torch.transpose(pos_and_neg, 0, 1)),
|
207 |
+
self.temperature,
|
208 |
+
)
|
209 |
+
)
|
210 |
+
)
|
211 |
+
|
212 |
+
sum_log_ent = torch.sum(
|
213 |
+
torch.log(
|
214 |
+
torch.div(
|
215 |
+
torch.exp(
|
216 |
+
torch.div(
|
217 |
+
torch.matmul(anchor, torch.transpose(positives, 0, 1)),
|
218 |
+
self.temperature,
|
219 |
+
)
|
220 |
+
),
|
221 |
+
denominator,
|
222 |
+
)
|
223 |
+
)
|
224 |
+
)
|
225 |
+
|
226 |
+
scale = -1 / pos_cardinal
|
227 |
+
|
228 |
+
return scale * sum_log_ent
|