I want to implement a neural network on pytorch where gradients are not computed over all the weights.