[Java] Implement non-sparse tensors

[asfimport]

We'd like to be able to round-trip NumPy ndarrays through Java, and create tensors in Java that can be eventually mapped to ndarrays in Python. Having even a basic Tensor implementation, with extension types, as a contrib module would help greatly.

Some prior discussions

• https://lists.apache.org/thread.html/9b142c1709aa37dc35f1ce8db4e1ced94fcc4cdd96cc72b5772b373b%40%3Cdev.arrow.apache.org%3E

Reporter: David Li / @lidavidm

PRs and other links:

• GitHub Pull Request #8278

_{Note: This issue was originally created as ARROW-10101. Please see the migration documentation for further details.}

[asfimport]

Joris Van den Bossche / @jorisvandenbossche: @lidavidm From the description above, it's not fully clear to me if you are talking about the (standalone) Tensor message type of the IPC protocol, or about storing a tensor as a value in a RecordBatch field.

Your description seems to talk about the first, but the mailing list thread talks about the second I think. There are also some open issues about defining a standard ExtensionType for storing arrays in RecordBatch fields (ARROW-1614, ARROW-8714)

[asfimport]

David Li / @lidavidm: Hey @jorisvandenbossche this is about the standalone Tensor type - I'd like both eventually, but having the Tensor type itself implemented is a prerequisite to that, at least for our use cases (Python <-> Java). Thanks for the pointers!

 

[asfimport]

Micah Kornfield / @emkornfield: @lidavidm  i took a very cursory look at the code and it seems straight-forward.  But one question, I had is if there is an existing OSS tensor model that makes sense for us to re-use or is the Arrow off-heap/object model enough of a snowflake to make that impractical?

[asfimport]

David Li / @lidavidm: Thanks [~emkornfield@gmail.com]. I'm not aware of an existing model. Honestly, my intent here is not really to provide an API to manipulate them in Java, but to just make it possible to round-trip them and convert to/from other APIs, hence why the methods on this Tensor are pretty sparse.

A brief search turns up these:

• nd4j - looks abandoned - https://github.com/deeplearning4j/nd4j - but it seems to be based on an off-heap buffer model like Arrow with shape/stride info
• djl.ai from AWSLabs - https://github.com/awslabs/djl/tree/master/api/src/main/java/ai/djl/ndarray - seems to be ByteBuffer based with shape/stride info

• Vectorz - double[] based - https://github.com/mikera/vectorz/blob/develop/src/main/java/mikera/arrayz/impl/BaseNDArray.java

  Maybe we should consider if our Tensor can be easily (zero-copy) wrapped by djl.ai's since they seem to have a similar structure, though it seems they also have their own memory management model.

[asfimport]

Micah Kornfield / @emkornfield: Thanks for investigating, I'm not an expert in this space, but I can try to take a look at the PR if no one else has provided feedback.

[asfimport]

David Li / @lidavidm: Thanks Micah - I'd appreciate any feedback on the API and approach before I go implement the rest of the tensor classes. I think like Python, it mostly suffices to have something that makes it easy to convert to/from the API actually being used.