How to take the features of net1 as the input of net2

autograd
10

If you need to have more flexibility, you could have a look at forward hooks in this post.
Since you need the gradients, make sure not to detach the tensor.

1 Like