Hi! I was trying to do backward on the first derivative (Jacobian). I observed that the usage of memory continues to grow if
I have read this post https://discuss.pytorch.org/t/how-to-free-the-graph-after-create-graph-true/58476/4, where it is said the graph will be deteted if the referece is deleted.
But I also found this post: https://github.com/pytorch/pytorch/issues/4661, stating that the leakage issue is still open.
I am confused. Could you please help me out? And I can’t use
torch.autograd.grad, since my outputs
y are vectors, not scalar outputs.
The conclusion from the issue you linked is that this is expected behavior mostly (or something we should forbid people from doing).
torch.autograd.grad works for vectors as well. What is the issue you encounter when trying to use it?
Ah! I see! So I should stick to
I just figured out how to use
torch.autograd.grad for vectors couple minutes ago.
Yes you should stick to
torch.autograd.grad and all will be good.
Hi, I meet the same problem, but I want to backward on the first derivative w.r.t. network parameters. Since both
torch.autograd.functional.jacobian only takes vector inputs while network parameters are tuple of tensors, is there a feasible way to do this? Thanks in advance for any possible help!
This post: Get gradient and Jacobian wrt the parameters helps get jacobian but I’m trying to backward further on Jacobian. It would be really nice if PyTorch supports gradient w.r.t PyTree objects like jax.
Both these functions take either a single Tensor or a tuple of Tensors as input. So it should work just fine.