How to do backpropagation for feature?

1

Hi, guys. I recently read this paper. <>http://personal.ie.cuhk.edu.hk/~pluo/pdf/luoWLWiccv17.pdf
It does the bp for features rather than parameters, can any one give a suitable solution? Thanks!

2

You would need input = autograd.Variable(tensor, requires_grad = True). Then backward will save gradient at input.grad

1 Like