Update only a middle layer of a neural network

autograd
3

Thanks for your answer.
But, in order to calculate layer2 gradients, I think we need to calculate layer3 gradients. If I set layer3 grads to False, how it would be possible?