zhiics commented on a change in pull request #5770:
URL: https://github.com/apache/incubator-tvm/pull/5770#discussion_r441850480
##########
File path: python/tvm/contrib/graph_runtime.py
##########
@@ -162,7 +162,11 @@ def set_input(self, key=None, value=None, **params):
keys = list(params.keys())
keys.sort(key=lambda x: -np.prod(params[x].shape))
for k in keys:
- self._get_input(k).copyfrom(params[k])
+ # TODO(zhiics) Skip the weights for submodule in a better way.
+ # We could get all inputs required by graphruntime first,
+ # we should use MetadataModule for initialization.
Review comment:
Yeah, we can do some checking at the C++ side as well. We will need to
change graph runtime later to let it have a "__init" function so that we can
initialize it with params. And we will need to remove some of the current ways
to set the params up. For example, the `set_input` API should not take
`**params` and the chunk of hacky code will be removed. That needs some more
thinking and may worth an RFC.
----------------------------------------------------------------
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:
us...@infra.apache.org
