Assume an image classification task using a neural net where loss function is defined as L .Consider an image x ,while performing gradient descent on the loss, the gradient of L with respect to x is dL/dx . Now about my doubt. After this I want to calculate the gradient of norm of dL/dx wrt to weights of neural net. how do I do that?
You can do the following:
x = torch.rand() output = net(x) L = crit(output, label) dLdx = autograd.grad(L, x, create_graph=True) dLdx.backward() # Now you have the grads as usual in the .grad fields
Thanks,this works perfectly