Implementation of weighted softmax by extending mx.autograd.Function fails

[ThomasDelteil]

See my other answer to your post: I would recommend using a custom op:

Implementing backward for a custom layer in gluon

It looks like that gluon custom layers don’t need backward function. (ref ) But I want to write a custom backward. On trying to use the CustomOp interface from mx.nd, I get an error which seems to indicate it’ll not work with gluon (Implementation of weighted softmax by extending mx.autograd.Function fails) How can I implement backward for a custom layer in gluon?