samskalicky commented on a change in pull request #15886: Graph Partition API
URL: https://github.com/apache/incubator-mxnet/pull/15886#discussion_r317869524
##########
File path: python/mxnet/symbol/symbol.py
##########
@@ -1437,6 +1437,54 @@ def _gen_atomic_symbol(self):
return Symbol(handle)
+ def optimize_for(self, backend, args=None, **kwargs):
Review comment:
Sorry, updated the labels.
You can call partitioning on your symbol without passing args. It will just
not have shape/type info during partitioning.
I guess we can add optional arguments for: shape/type/stype dicts so that
users can do something similar to simple_bind.
Our initial use-case is for inference, so we're assuming that the model is
pre-trained. But we dont want to pigeon-hole ourselves away from general
support (ie. training). Since we're accepting kwargs we could add the ability
for users to pass in shape/type/stype dicts in a follow-up PR without a
breaking API change. Would this be acceptable, or can you explain the need for
it be available immediately?
Agree on the partial inference point, we'll make changes to support that.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
