How to get the gradient of output w.r.t. model parameters?

autograd
1

I want to get the gradient of output w.r.t. model parameters. For example, y = xW where x is a vector of size 1x5, and W is a vector of size 5x1, and W is model parameter. I don’t have a loss function, just want to get the gradient of y w.r.t W. Obviously, in this case, dy/dW should be x. However, using pytorch backward, the value I got is wrong. What’s the reason? Here is my code.

import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as torchF
from torch.autograd import Variable

class Net(nn.Module):
    def __init__(self, dim_action=1): # assumes a [256,256,3] input
        super(Net, self).__init__()
        self.fc = nn.Linear(5,1, bias=False)
    
    def forward(self,x):
        x = self.fc(x)
        return x

net = Net()
net.double()
x = Variable(torch.from_numpy(np.array([1.0,2.0,3.0,4.0,5.0]).reshape((1,5))))
out = net(x)
outvalue = np.array(out.data)
net.zero_grad()
y = torch.randn(1,1)
y.data = outvalue
y = y.double()
out.backward(y)

for f in net.parameters():
    print('data is')
    print(f.data)
    print('grad is')
    print(f.grad)

and the output is,
data is

-0.3839 -0.1894 -0.3376 -0.3934 -0.2313
[torch.DoubleTensor of size 1x5]

grad is
Variable containing:
 0.2647  0.5293  0.7940  1.0587  1.3233
[torch.DoubleTensor of size 1x5]
1 Like
2

y is now a function of Variable W. Hence a simple y.backward should be able to compute the gradients of y with respect to W for you. You might want to perform net.zero_grad() before performing y.backward() so as to clear the previous gradients if any.

In your case: the code can be written in fewer lines:
x = Variable(torch.from_numpy(np.array(range(1, 6)).reshape((1, 5))).type(torch.FloatTensor))
net.zero_grad()
y = net(x)
y.backward()

Now please check the gradients.

1 Like
3

Thanks. Actually y here should be torch.ones(1,1) instead of output value. After replace y with the value, the result is correct.