Hessian vector product with create_graph flag

Hi, I am currently implemented Trust Region Policy Optimization using the newest version of pytorch (0.2), which support higher order gradients.
I spend several hours to debug my code and find it’s mainly due to one flag I didn’t pass.

My code is kind of like this:

surr_loss = surr_loss_fn(policy)

kl_div = kl_div_func(policy)
grad = torch.cat([param.grad.view(-1) for param in policy.parameters()])
gradient_vector_product = torch.sum(grad * Variable(vector))

I find this is not working, and the code will complains

RuntimeError: element 0 of variables tuple is volatile

on the last backward().

If I insert this before the last two lines of code

grad.volatile = False

then it will complain

RuntimeError: there are no graph nodes that require computing gradients

Finally, I find the way to have the code not to complain is to following at the second line of code:


I am curious if this is a desired behavior of getting higher order gradient? Or in other words, did I do this right?
And if it is, what is the reason for setting create_graph flag in the first backward call?

1 Like

yes, you do need to give create_graph where appropriate to be able to compute higher order gradients.

So I guess since I compute the second order gradient of kl_div w.r.t. policy parameter after I compute gradient of surr_loss, even I do not need the second order gradient of surr_loss, I still need set the create_graph to True when compute its gradient?