I’m on a kind of meta-learning project and I want to update a parameter (theta) of a module with a gradient (d_loss(theta, x)/d_theta) that is differentiable wrt x.
I could get the differentiable gradient using autograd.grad, but I see no way to update my parameter by hand without detaching the gradient.
module = nn.Linear(4,2) loss = (module(input)*x).sum() grad = autograd.grad(loss, (module.weight), retain_graph=True) # this does not work: module.weight = module.weight - alpha*grad
Is there any simple way to do that?
Another solution to my problem would be the equivalent of an nn.Module but with tensors instead of parameters (I just don’t want to re-implement a conv2d mechanism by hand with tensors)