Is there any plan to refactor the legacy Ops into nnvm?

[mbrookhart]

Right now, primitive Tensor operations are directly registered with nnvm, but the legacy layer operations go through a more convoluted forward/backward wrapper to interact with nnvm and thus the graph executor. After #6928 Is there any thought to refactor the legacy ops into the nnvm tensor primitive ops for easier graph optimization?

[piiswrong]

It's in the backlog. Currently we don't have people assigned on this

[mbrookhart]

Thanks.