How to extract features of an image from a trained model

Thank you, @apaszke, I think I start to understand (sorry, I’m slow…).
Could you give me a hint about these “better ways” you are referring at? I’m just super curious now :grin:

Answering you question “Why do you even need to know what Module instantiated which Function?” is pretty simple: so that I know to which Moudle I should register a forward hook in order to inspect the Tensor's flow in the graph.
Even better, would it make sense to have a register_forward/backward_hook() for the Variable class instead of the current register[_backward]_hook()?