pytorch中如何只让指定变量向后传播梯度?

pytorch中如何只让指定变量向后传播梯度？

（或者说如何让指定变量不参与后向传播？）

有以下公式，假如要让L对xvar求导：

（1）中，L对xvar的求导将同时计算out1部分和out2部分；

（2）中，L对xvar的求导只计算out2部分，因为out1的requires_grad=False；

（3）中，L对xvar的求导只计算out1部分，因为out2的requires_grad=False；

验证如下：

#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Wed May 23 10:02:04 2018

@author: hy
"""

import torch
from torch.autograd import Variable
print("Pytorch version: {}".format(torch.__version__))
x=torch.Tensor([1])
xvar=Variable(x,requires_grad=True)
y1=torch.Tensor([2])
y2=torch.Tensor([7])
y1var=Variable(y1)
y2var=Variable(y2)
#(1)
print("For (1)")
print("xvar requres_grad: {}".format(xvar.requires_grad))
print("y1var requres_grad: {}".format(y1var.requires_grad))
print("y2var requres_grad: {}".format(y2var.requires_grad))
out1 = xvar*y1var
print("out1 requres_grad: {}".format(out1.requires_grad))
out2 = xvar*y2var
print("out2 requres_grad: {}".format(out2.requires_grad))
L=torch.pow(out1-out2,2)
L.backward()
print("xvar.grad: {}".format(xvar.grad))
xvar.grad.data.zero_()
#(2)
print("For (2)")
print("xvar requres_grad: {}".format(xvar.requires_grad))
print("y1var requres_grad: {}".format(y1var.requires_grad))
print("y2var requres_grad: {}".format(y2var.requires_grad))
out1 = xvar*y1var
print("out1 requres_grad: {}".format(out1.requires_grad))
out2 = xvar*y2var
print("out2 requres_grad: {}".format(out2.requires_grad))
out1 = out1.detach()
print("after out1.detach(), out1 requres_grad: {}".format(out1.requires_grad))
L=torch.pow(out1-out2,2)
L.backward()
print("xvar.grad: {}".format(xvar.grad))
xvar.grad.data.zero_()
#(3)
print("For (3)")
print("xvar requres_grad: {}".format(xvar.requires_grad))
print("y1var requres_grad: {}".format(y1var.requires_grad))
print("y2var requres_grad: {}".format(y2var.requires_grad))
out1 = xvar*y1var
print("out1 requres_grad: {}".format(out1.requires_grad))
out2 = xvar*y2var
print("out2 requres_grad: {}".format(out2.requires_grad))
#out1 = out1.detach()
out2 = out2.detach()
print("after out2.detach(), out2 requres_grad: {}".format(out1.requires_grad))
L=torch.pow(out1-out2,2)
L.backward()
print("xvar.grad: {}".format(xvar.grad))
xvar.grad.data.zero_()

pytorch中，将变量的requires_grad设为False，即可让变量不参与梯度的后向传播；

但是不能直接将out1.requires_grad=False;

其实，Variable类型提供了detach()方法，所返回变量的requires_grad为False。

 

注意：如果out1和out2的requires_grad都为False的话，那么xvar.grad就出错了，因为梯度没有传到xvar

补充: volatile=True表示这个变量不计算梯度, 参考:Volatile is recommended for purely inference mode, when you’re sure you won’t be even calling .backward(). It’s more efficient than any other autograd setting - it will use the absolute minimal amount of memory to evaluate the model. volatile also determines that requires_grad is False.