forked from hutchresearch/fine-grained-cad
-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
46 lines (38 loc) · 1.73 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
import numpy as np
import torch
import torch.nn.functional as F
from torch.autograd import Variable
def focal_loss(logits, labels, gamma=2, alpha=None):
logpt = F.log_softmax(logits, dim=1)
logpt = logpt.gather(1, labels)
logpt = logpt.view(-1)
pt = Variable(logpt.data.exp())
if alpha is not None:
if isinstance(alpha, list):
alpha = torch.Tensor(alpha)
if alpha.type() != logits.data.type():
alpha = alpha.type_as(logits.data)
at = alpha.gather(0, labels.data.view(-1))
logpt = logpt * Variable(at)
loss = -1 * (1-pt)**gamma * logpt
return torch.mean(loss)
def cb_loss(logits, labels, samples_per_cls, beta, gamma):
"""Compute the Class Balanced Loss between `logits` and the ground truth `labels`.
Class Balanced Loss: ((1-beta)/(1-beta^n))*Loss(labels, logits)
where Loss is one of the standard losses used for Neural Networks. Whenever
gamma = 0, beta > 0, this is equivalent to focal loss. Whenever gamma = 0,
beta = 0, this is equivalent to multi-class cross-entropy loss.
Args:
labels: A int tensor of size [batch].
logits: A float tensor of size [batch, no_of_classes].
samples_per_cls: A python list of size [no_of_classes].
beta: float. 0 -> no class balancing, 1 -> Inverse class freq
gamma: float. 0 -> Disable downweighting, inf -> Downweight 'easy' samples.
Returns:
cb_loss: A float tensor representing class balanced loss
"""
no_of_classes = len(samples_per_cls)
effective_num = 1.0 - np.power(beta, samples_per_cls)
weights = (1.0 - beta) / effective_num
weights = weights / np.sum(weights) * no_of_classes
return focal_loss(logits, labels, gamma, torch.tensor(weights))