losses.py 文件源码

def binaryXloss(logits, label):
    mask = (label.view(-1) != VOID_LABEL)
    nonvoid = mask.long().sum()
    if nonvoid == 0:
        # only void pixels, the gradients should be 0
        return logits.sum() * 0.
    # if nonvoid == mask.numel():
    #     # no void pixel, use builtin
    #     return F.cross_entropy(logits, Variable(label))
    target = label.contiguous().view(-1)[mask]
    logits = logits.contiguous().view(-1)[mask]
    # loss = F.binary_cross_entropy(logits, Variable(target.float()))
    loss = StableBCELoss()(logits, Variable(target.float()))
    return loss