BCELoss.py

import numpy as np
import torch

crit = torch.nn.BCELoss()

preds   = torch.tensor([[1, 0.2, 0.3, 1, 1], [1, 0.1, 0.7, 1, 1]]).float() # n classes
targets = torch.tensor([[1, 1, 1, 0, 0], [1, 0, 1, 1, 1]]).float() # n classes 0 or 1

print(preds, targets)

print(crit(preds, targets))

#---
EPS = 1e-12

res = targets * torch.log(preds + EPS) + (1 - targets) * torch.log(1 - preds + EPS)
res *= -1
res[res >= -np.log(EPS)] = 100


print(res.sum() / np.product(targets.shape[:2]))