[BUGFIX] fix unknown parameter shapes when np_shape is turned on.

[zheng-da]

Description

When np_shape is turned on, the unknown parameter shapes should be -1 instead of 0. So when a parameter is created but isn't initialized, the dimension of unknown size is -1 when np_shape is turned on, and is 0, otherwise.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

[reminisce]

I think the following two places checking whether the shape is known should be modified too.
/~https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/gluon/parameter.py#L272
/~https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/gluon/parameter.py#L383

[reminisce]

I'm worried about this line
/~https://github.com/apache/incubator-mxnet/blob/5fc4fc53df74f276aafa51208142e657e9cfe42d/python/mxnet/gluon/parameter.py#L168

Users may notice that we started to use -1 to represent unknown dim size in np shape semantics and some of them in practice may use shapes like (-1, 10) to create parameters for deferred initialization, instead of (0, 10). In those cases, we need to support that unknown dim size can be either 0 or -1, so this line should become

all(j in (-1, 0, i) for i, j in zip(new_shape, self._shape))

when is_np_shape = True.

@szha What do you think?

[abhinavs95]

@mxnet-label-bot add [Gluon]

[szha]

@reminisce I had similar concern though I thought the API doc should do the job. another option is to expose an internal interface for infer_shape to use, instead of piggybacking on the .shape attribute

[roywei]

@reminisce @szha @zheng-da what's the next step here? There is nothing to prevent users from using -1 as unknown dim. so maybe state clearly in comment or API doc and note experimental?

[reminisce]

I think it would be safer to note np_shape as experimental and still under development.

[zheng-da]

where should we document -1 is unknown dim and is experimental?

[roywei]

I believe for NP shape APIs, it's already documented as experimental, so just add some comments in this PR about -1 and refer to the np documentation here /~https://github.com/apache/incubator-mxnet/blob/master/python/mxnet/util.py#L84