Back propagation problem

1

I am new to Pytorch. When I try to create my own module with Pytorch. I meet some problem.First, I want to write a backward() function in python,but I can not use nn.SpatialAdaptiveMaxPooling().backwrad().Then how can i get the gradinput？Thank you for your help.

2

You don’t need to write your own backward function. You just need to write your forward function and your backward function will be generated automatically via autograd. See http://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html for an example.

3

Thank you for your replay, I am sorry that I did not express meaning correctly. Actually，I want to write a new layer,not a module, then how to write the backward function?

4

See http://pytorch.org/docs/notes/extending.html