acphile commented on a change in pull request #18619:
URL: https://github.com/apache/incubator-mxnet/pull/18619#discussion_r453431154
##########
File path: python/mxnet/gluon/block.py
##########
@@ -1573,29 +1534,31 @@ def __init__(self, outputs, inputs, params=None):
arg_types, aux_types = _infer_param_types(syms, out, arg_params,
aux_params)
- def _set_params_attr(name, **kwargs):
- if params.get(name) is None:
- param = Parameter(**kwargs)
- param._name = name
- else:
- param = params[name]
- param._check_and_setattr(**kwargs)
- if self._structured_named:
- lis = structure[name]
- assert len(lis) > 0, "Can not find structured name for
Parameter %s in 'params'. " \
- "Please check 'params' is complete!" % name
- for structured_name in lis:
- self._reg_params[structured_name] = param
- else:
- self._reg_params[name] = param
+ if params is None:
+ params = {}
+ unused_params = set(params.keys()) - set(arg_params) - set(aux_params)
+ if len(unused_params) > 0:
+ raise ValueError('{} params are unused by the
model.'.format(unused_params))
Review comment:
I‘m worried about this case: when a model's forward returns multiple
outputs and we want to build SymbolBlock using one of the outputs, it may be
necessary for users to precisely pass the intended parameters which the output
uses. For example, a joint-learning model may have different output layers for
different tasks and maybe we just want to use one specific output to build a
SymbolBlock.
----------------------------------------------------------------
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]
