CODEWITHSUNDEEP
machine-learningpytorchgradient-descent
Penguin
Penguin

Reputation: 2411

How to flow gradient through parameter updates?

I have a parameter

import torch
import torch.nn as nn
x = nn.Parameter(torch.tensor([1.0]))

I need to update the value of this parameter while maintaining the gradient flow (for a meta learning problem).

PyTorch does not allow in-place operations on leaf-variable like

a = torch.tensor([2.0], requires_grad=True)
x.copy_(a)
x.backward()
>>> RuntimeError: a leaf Variable that requires grad is being used in an in-place operation.

In addition, I also tried the following, but while working, the gradient didn't flow:

x = nn.Parameter(torch.tensor([1.0]))
a = torch.tensor([2.0], requires_grad=True)
x = nn.Parameter(a)
x.backward()
print(a.grad)
>>> None

How can I update the value of a parameter while maintaining gradient flow?

Upvotes: 0

Views: 497

Answers (1)

Ivan
Ivan

Reputation: 40638

What you are looking for is torch.Tensor.clone:

This function is differentiable, so gradients will flow back from the result of this operation to input. To create a tensor without an autograd relationship to input see detach().

>>> a = torch.tensor([2.0], requires_grad=True)
>>> x = a.clone()
>>> x.backward()

>>> a.grad
tensor([1.])

Upvotes: 1

Related Questions