Infer input_names in mxnet.gluon.SymbolBlock.imports API

[liusy182]

The API mxnet.gluon.SymbolBlock.export exports input names by following the convention below according to documentation:

When there are only one input, it will have name data. When there Are more than one inputs, 
they will be named as data0, data1, etc.

The counterpart mxnet.gluon.SymbolBlock.imports requires the input_names to be specified as the second parameter.

This is problematic for clients who use this set of APIs. E.g. a client who have previously exported the model may not know the format of input_names when it tries to do import later.

I see there are 2 solutions for this problem:

1. let the mxnet.gluon.SymbolBlock.imports being able to infer the input_names. (preferred)
2. have mxnet.gluon.SymbolBlock.export return the input_names which can later be used by the import.

Does gluon support any of the above functionality in anyways? If not, would it be a good idea to support this?

[szha]

@liusy182 thanks for bringing this up. I agree this is an issue to be solved. Between the two proposed approaches, I think having imports infer input_names seems like a good idea. To keep backward compatibility, I think we can set imports's argument input_names to be optional and defaults to None.

[leezu]

Would you like to contribute a PR @liusy182?

[liusy182]

Sure. I would love to. What are the things to watch out on this area as someone new to the code base?

[szha]

You can check out the contribution guides here for tips: https://mxnet.apache.org/versions/master/community

[liusy182]

I have created #19569 to keep track.

I assume the inference should happen right before here. What should be the C API to use to get the inferred parameters?