REVA-QCAV/dice_loss.py

42 lines
1.2 KiB
Python
Raw Normal View History

import torch
from torch.autograd import Function, Variable
class DiceCoeff(Function):
"""Dice coeff for individual examples"""
def forward(self, input, target):
self.save_for_backward(input, target)
eps = 0.0001
self.inter = torch.dot(input.view(-1), target.view(-1))
self.union = torch.sum(input) + torch.sum(target) + eps
t = (2 * self.inter.float() + eps) / self.union.float()
return t
# This function has only a single output, so it gets only one gradient
def backward(self, grad_output):
input, target = self.saved_variables
grad_input = grad_target = None
if self.needs_input_grad[0]:
grad_input = grad_output * 2 * (target * self.union - self.inter) \
/ (self.union * self.union)
if self.needs_input_grad[1]:
grad_target = None
return grad_input, grad_target
def dice_coeff(input, target):
"""Dice coeff for batches"""
if input.is_cuda:
s = torch.FloatTensor(1).cuda().zero_()
else:
s = torch.FloatTensor(1).zero_()
for i, c in enumerate(zip(input, target)):
s = s + DiceCoeff().forward(c[0], c[1])
return s / (i + 1)