Chainer.grad() how to implement this function in Pytorch

autograd
#1

I am writing to migrate code from chainer framework to pytorch. I came across the below code:

chainer.grad([loss_func(F.clipped_relu(X2,z=1.0),Yp)], [X2], set_grad=True, retain_grad=True)
I tired looking at the pytorch documentation but was not able to find an equivalent function to implement the above code.

https://docs.chainer.org/en/stable/reference/generated/chainer.grad.html
The above link explains what this function does in chainer framework. I would appreciate if someone could guide me on how to go about migrating the above chainer code to pytorch.

#2

I think torch.autograd.grad would be the corresponding function.

1 Like
#3

Hello, thank you for this answer.

I am able to calculate the gradient using torch.autograd.grad() function but this does not populate .grad field.

Could you tell me how I can populate this feild using the gradient calculate from torch.autograd.grad()?

Any help will be highly appreciated

#4

If you want to populate the .grad attribute of specific parameters, you could use the inputs argument in the backward call as e.g. given here.

1 Like
#5

Thank you for this. So this will update the .grad feild of those parameters using the gradient that was calculated using autograd,grad() function?

Thank you once again

#6

Yes, that should be the case.

1 Like