Pytorch框架学习 -3 pytorch损失函数

Pytorch框架学习 -3 pytorch损失函数

前置

nn.Module 简单分析

无参数损失函数

定义

class _Loss(Module):
    reduction: str

pytorch设计是真的强，我原本以为损失函数是单独定义的类，没想到他居然也是nn.Module的子类。

初始化方法

    def __init__(self, size_average=None, reduce=None, reduction: str = 'mean') -> None:
        super(_Loss, self).__init__()
        if size_average is not None or reduce is not None:
            self.reduction = _Reduction.legacy_get_string(size_average, reduce)
        else:
            self.reduction = reduction

1. 实现对基类的继承

2. 如果size_average或者reduce非空调用

   def legacy_get_string(size_average, reduce, emit_warning=True):
       # type: (Optional[bool], Optional[bool], bool) -> str
       warning = "size_average and reduce args will be deprecated, please use reduction='{}' instead."
   
       if size_average is None:
           size_average = True
       if reduce is None:
           reduce = True
   
       if size_average and reduce:
           ret = 'mean'
       elif reduce:
           ret = 'sum'
       else:
           ret = 'none'
       if emit_warning:
           warnings.warn(warning.format(ret))
       return ret
   

   来推断reduction的应该是的情况

3. 否则直接使用传入的reduce

这估计是最友好的pytorch基类了吧！

一个实现的例子L1范数损失

class L1Loss(_Loss):
   
    __constants__ = ['reduction']

    def __init__(self, size_average=None,