-
Notifications
You must be signed in to change notification settings - Fork 15
/
loss.py
33 lines (24 loc) · 978 Bytes
/
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
import torch as torch
import torch.nn.functional as F
class Contrastive_Loss(torch.nn.Module):
def __init__(self):
super(Contrastive_Loss, self).__init__()
self.ce_loss = torch.nn.CrossEntropyLoss()
def forward(self, x, target):
return self.ce_loss(x, target)
class LogSoftmax(torch.nn.Module):
def __init__(self, dim):
super(LogSoftmax, self).__init__()
self.dim = dim
def forward(self, x, a):
nll = -F.log_softmax(x, self.dim, _stacklevel=5)
return (nll * a / a.sum(1, keepdim=True).clamp(min=1)).sum(dim=1).mean()
class NCELoss(torch.nn.Module):
def __init__(self, batch_size=4096):
super(NCELoss, self).__init__()
self.ce_loss = torch.nn.CrossEntropyLoss()
def forward(self, x):
batch_size = len(x)
target = torch.arange(batch_size).cuda()
x = torch.cat((x, x.t()), dim=1)
return self.ce_loss(x, target)