Pytorch MSELoss

CLASS torch.nn.MSELoss(size_average=None, reduce=None, reduction='mean')

Creates a criterion that measures the mean squared error (squared L2 norm) between each element in the input x and target yy.

创建一个准则测量输入x和目标y每个元素间的平方二范数。

x和y可以是任意形状的张量，每个张量总的元素个数为n。

Note: size_average and reduce are in the process of being deprecated, and in the meantime, specifying either of those two args will override reduction. Default: 'mean'

size_average和reduce正在被弃用，因此我们只看reduction参数。

reduction='none'

The unreduced (i.e. with reduction set to 'none') loss can be described as:

返回每个元素的平方二范数，shape与输入相同。

reduction='sum'和reduction='mean'

对所有对应元素的平方二范数求和或求平均（sum / n）。

a = torch.tensor([[[1., 2.]], [[1, 2]]])  # [2, 1, 2]
b = torch.tensor([[[2., 4.]], [[2, 4.]]]) # [2, 1, 2]

C =  torch.nn.MSELoss(reduction='none')

loss = C(a, b)

print(a)
print(b)
print(loss)

>>>

tensor([[[1., 2.]],
        [[1., 2.]]])

tensor([[[2., 4.]],
        [[2., 4.]]])

tensor([[[1., 4.]],
        [[1., 4.]]])  # 分别计算对应元素的平方二范数

C =  torch.nn.MSELoss(reduction='sum')

loss = C(a, b)
print(loss)

>>> tensor(10.) # 所有元素平方二范数之和

C =  torch.nn.MSELoss(reduction='mean')

loss = C(a, b)
print(loss)
>>> tensor(2.5000) # 所有元素平方二范数的均值