#
# focal_loss.py, doom-net
#
# Created by Andrey Kolishchak on 03/11/18.
#
import torch
import torch.nn as nn
import torch.nn.functional as F
from device import device
class FocalLoss(nn.Module):
def __init__(self, alfa=1, gamma=2):
super().__init__()
self.alfa = alfa
self.gamma = gamma
self.epsilon = 0 if int(gamma) == gamma and gamma != 0 else 1e-5
pass
def forward(self, input, target):
target = target.view(target.shape[0], 1, *target.shape[1:])
target_one_hot = torch.zeros(*input.shape, device=device)
target_one_hot = target_one_hot.scatter_(1, target, 1.0)
logp = input * target_one_hot
p = logp.exp()+self.epsilon
output = -(1-p).pow(self.gamma) * logp
output = output.sum(dim=1).mean()
return output
def test():
loss_nll = nn.NLLLoss2d()
loss_focal = FocalLoss(gamma=0)
target = torch.Tensor(2, 1, 5).random_(3).long()
data = torch.rand(2, 3, 1, 5)
input1 = torch.Tensor(data, requires_grad=True)
loss1 = loss_nll(F.log_softmax(input1), target)
loss1.backward()
print(loss1)
print(input1.grad)
input2 = torch.Tensor(data, requires_grad=True)
loss2 = loss_focal(F.log_softmax(input2), target)
loss2.backward()
print(loss2)
print(input2.grad)
#test()
