How to calculate gradients wrt one of inputs

autograd
1

Suppose that F(x) is a NN with 3 inputs and 1 output and I try to compute the gradient of F wrt the first input. But I got the error message: “One of the differentiated Tensors appears to not have been used in the graph. Set allow_unused=True if this is the desired behavior.” What is the problem?

My code is as follows:

import torch
import torch.nn as nn
from torch.autograd import Variable,grad

NN = nn.Sequential(
torch.nn.Linear(3, 64),
torch.nn.ReLU(),
torch.nn.Linear(64, 64),
torch.nn.ReLU(),
torch.nn.Linear(64, 64),
torch.nn.ReLU(),
torch.nn.Linear(64, 1)
)

X = Variable(torch.rand(5,3), requires_grad=True)
F = NN(X)
G = grad(torch.sum(F),X[:,0],create_graph=True)[0]
print(G)

1 Like
2

Hi,

The problem here is that your function takes a single Tensor as input. The autograd works with Tensors as a unit and so it does not recognize X[:,0] as being a subset of X, but as a different Tensor. So for him, this new Tensor was not used during the backward.
You will have to either give each entry separately as a different tensor (and use torch.cat in your forward to recreate a single Tensor) to be able to ask gradients for a single one.
Otherwise, you could ask the gradients for X and then slice the gradients to get the part you want.

2 Likes