backword

tensor的backward()用于求导，如下对y=x^2求导，dy=2x=6.0

import torch


x = torch.tensor(3., requires_grad=True)
y = x*x
y.backward()
print(x.grad)  # dy/dx = 2x=6.0

如果x的值非标量，即有多个值，会出现错误：RuntimeError: grad can be implicitly created only for scalar outputs

import torch


x = torch.tensor([3., 2., 4.], requires_grad=True)
y = x*x
y.backward()
print(x.grad)  # RuntimeError: grad can be implicitly created only for scalar outputs

一种解决方法是，对y进行求平均，相当于在原先求导之后再进行平均

import torch


x = torch.tensor([3., 2., 4.], requires_grad=True)
y = x*x
y = y.mean()
y.backward()
print(x.grad)  # dy/dx = 2x=  [6.0/2, 4./3, 8./3]

方法二是在backward指定参数，计算导数相当于在增加一个权重值， 参数retain_graph表示是否对前一次求导时进行累加。如果retain_graph的值不指定为True，同时y不再次进行赋值，则无法进行再次y.backward()

import torch


x = torch.tensor([3., 2., 4.], requires_grad=True)
y = x*x

weight = torch.ones(3)
y.backward(weight, retain_graph=True)
print(x.grad)  # dy/dx = 2x*weight= [6.0*1, 4*1, 8*1]

 

 

来源：oschina

链接：https://my.oschina.net/u/4228078/blog/4319267