I am trying to understand how Binary Cross Entropy is counted in PyTorch. I've tried the same code from the PyTorch documentation here, but I get a different result comparing to mathematical implementation of this function.
Code (I made little changes comparing to documentation):
import torch import torch.nn as nn m = nn.Sigmoid() loss = nn.BCELoss() input = torch.randn(1, requires_grad=True) print(input) print(m(input)) target = torch.empty(1).random_(2) print(target) output = loss(m(input), target) print(output) output.backward() print(output)
The mathematical formula for BCELoss is:
output = -[target*log(our_value) + (1-target)*log(1-our_value)]
From the mathematical formula above I should get 'output'=0.3215 for 'our_value'=0.4770 and 'target'=1. But PyTorch shows that the 'output'=0.7403.
I've also found a C code here of this function but the formula is almost the same (only very small epsila should be added and it gives no difference in output). So, it's still unclear for me.
How does it come to? Any hints are welcome.