Add a contrib operator for Constant

[vandanavk]

Description

MXNet doesn't have a constant op for sym API (similar to gluon.constant). The closest is to map it to BlockGrad as discussed in #6087. It would be useful to have a constant operator in general, and it is also required to convert Pytorch models to MXNet through ONNX.

Adding support for this operator here.

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

Changes

• Add a constant operator in contrib, add relevant tests
• Add ONNX conversion for contrib.constant, and relevant tests

Comments

[vandanavk]

@mxnet-label-bot add [Operator, pr-awaiting-review, ONNX]

[ZhennanQin]

For performance optimization purpose, I think it's better to have constant against with variable. Make constant as a operator is an option, but the question is, how to set the value for it? Especially when the value is not a broad casted value.

[vandanavk]

For performance optimization purpose, I think it's better to have constant against with variable. Make constant as a operator is an option, but the question is, how to set the value for it? Especially when the value is not a broad casted value.

One of the main reasons for adding this as an operator was for ONNX conversion. I tried blockgrad with constant initializer - but it dint convert well.
Could you give an example for the second part of your question - the one about when the value is not broadcasted.

[ZhennanQin]

Here's the most confusing part. I assume v1 is the shape of constant output, but where to specify the output value? On the other words, why the output is [-1, 2]?

[vandanavk]

v1 is the value we want to store. since this operator is a constant, the output is equal to the input ([-1,2])

[ZhennanQin]

I mean, If we did some constant folding optimization in the graph, and generate a new tensor which is not filled with same values. For example,

scaled_weight = weight * scale, 
out = convolution(weight = scaled_weight)

For this code, that scale multiply will execute many times during each iteration of inference. Because weight and scale are all constant, so we can fold them before inference, and use a constant node to replace the folded result in the graph. The value holding by constant node is vary, which can't be generated with simple pattern.

[anirudhacharya]

can you reference this issue in the PR description - #6087

[vandanavk]

I mean, If we did some constant folding optimization in the graph, and generate a new tensor which is not filled with same values. For example,

scaled_weight = weight * scale, 
out = convolution(weight = scaled_weight)

For this code, that scale multiply will execute many times during each iteration of inference. Because weight and scale are all constant, so we can fold them before inference, and use a constant node to replace the folded result in the graph. The value holding by constant node is vary, which can't be generated with simple pattern.

@ZhennanQin which operator is currently used in this scenario?