-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
44 lines (35 loc) · 1.53 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
import torch
import torch.nn as nn
def angular_softmax_loss():
"""Loss function for SphereFace"""
pass
def large_margin_cos_loss():
"""Loss function for CosFace"""
pass
def additive_angular_margin_loss():
"""Loss function for ArcFace"""
def triplet_loss():
"""Loss function for FaceNet"""
class NT_Xent(nn.Module):
"""Loss function used for SimCLR (Mix of PyTorch Lightning and Spijkervet implementations)"""
def __init__(self, batch_size, temperature=0.5):
super(NT_Xent, self).__init__()
self.batch_size = batch_size
self.temperature = temperature
self.N = batch_size * 2
self.mask = torch.eye(self.N).bool()
self.criterion = nn.CrossEntropyLoss(reduction="sum")
self.cos_sim = nn.CosineSimilarity(dim=2)
def __call__(self, x_i, x_j):
z = torch.cat((x_i, x_j), dim=0)
sim = self.cos_sim(z.unsqueeze(1), z.unsqueeze(0)) / self.temperature
#extract positive pairs: (x_i, x_j)
sim_i_j = torch.diag(sim, self.batch_size)
sim_j_i = torch.diag(sim, -self.batch_size)
positive_samples = torch.cat((sim_i_j, sim_j_i), dim=0).reshape(self.N, 1)
# remove left digonal (x_i, x_i) to get negative pairs
negative_samples = sim[self.mask].reshape(self.N, -1)
labels = torch.zeros(self.N).to(positive_samples.device).long()
logits = torch.cat((positive_samples, negative_samples), dim=1)
loss = self.criterion(logits, labels) / self.N
return loss