多标签分类的 Focal loss 的实现

Question

尝试为多标签分类编写焦点损失

class FocalLoss(nn.Module):
    def __init__(self, gamma=2, alpha=0.25):
        self._gamma = gamma
        self._alpha = alpha

    def forward(self, y_true, y_pred):
        cross_entropy_loss = torch.nn.BCELoss(y_true, y_pred)
        p_t = ((y_true * y_pred) +
               ((1 - y_true) * (1 - y_pred)))
        modulating_factor = 1.0
        if self._gamma:
            modulating_factor = torch.pow(1.0 - p_t, self._gamma)
        alpha_weight_factor = 1.0
        if self._alpha is not None:
            alpha_weight_factor = (y_true * self._alpha +
                                   (1 - y_true) * (1 - self._alpha))
        focal_cross_entropy_loss = (modulating_factor * alpha_weight_factor *
                                    cross_entropy_loss)
        return focal_cross_entropy_loss.mean()

但是当我运行这个时，我得到

  File "train.py", line 82, in <module>
    loss = loss_fn(output, target)
  File "/home/bubbles/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 538, in __call__
    for hook in self._forward_pre_hooks.values():
  File "/home/bubbles/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 591, in __getattr__
    type(self).__name__, name))
AttributeError: 'FocalLoss' object has no attribute '_forward_pre_hooks'

任何建议都会非常有帮助，提前致谢。

Answer 1

您不应该继承，torch.nn.Module因为它是为具有可学习参数的模块（例如神经网络）而设计的。

只需创建正常的函子或函数就可以了。

顺便提一句。如果你继承它，你应该调用super().__init__()你的__init__().

编辑

实际上继承nn.Module可能是一个好主意，它允许您将损失用作神经网络的一部分，并且在 PyTorch 实现/PyTorch Lightning 中很常见。