losses.py 文件源码

def softmax_kl_loss(input_logits, target_logits):
    """Takes softmax on both sides and returns KL divergence

    Note:
    - Returns the sum over all examples. Divide by the batch size afterwards
      if you want the mean.
    - Sends gradients to inputs but not the targets.
    """
    assert input_logits.size() == target_logits.size()
    input_log_softmax = F.log_softmax(input_logits, dim=1)
    target_softmax = F.softmax(target_logits, dim=1)
    return F.kl_div(input_log_softmax, target_softmax, size_average=False)