NervanaSystems / ngraph-mxnet

MXNet - nGraph integration

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Enable Gluon hybridize cached_op with ngraph inference and training

ashokei opened this issue · comments

Enable Gluon hybridize cached_op with ngraph inference and training.

Goal is to

  1. Gluon hybridize inference and training working , without performance loss
  2. Unit-tests working, including RNN bidirectional and other gluon tests that were failing previously

For this Issue, should we also make MXNET_NGRAPH_GLUON=1 be the default behavior?

So...DCGAN gets a huge performance boost on nGraph by hybridizing (according to the Poland team). Are we sure we don't already support this through the subgraph API?

It definitely looks like hybridizing is sending the graph to nGraph.

@mbrookhart did you check with your hybridize script, just want to make sure if the poland team script is not converting gluon to symbol somewhere.

Based on source code last time i looked at, it doesnt seem gluon hybridize goes through subgraph, but it may have changed.