eldercrow / additions_mxnet Goto Github PK

Hi, Thanks very much for your sharing your codes.
I read your twn python ops and do some test on it.

I think there's a big problem is that when we update weight with twn-ops, although conv ops using the updated equalization weights, the backward did not update it. Loss only do effect on the original weights.

This will make the weights updated not effective.

With this I want to assign the in_data[0] from out_data[0] when forward. I failed because in_data[0] is read only.....

I'm not familiar with MXNet yet. Can you help me solve the problem?

Recently, I am using mxnet to do something! When I found your coding, that's greate!
Best Thanks !

Thanks!