[R] Support for Tensor class

[asfimport]

It would be nice to have support for the Tensor class in the R package. 

I'm interested in using it for a reliable way to transfer torch tensors between R sessions.

I can submit a PR in a few days for discussion.

Thanks very much!

 

 

Reporter: Daniel Falbel

_{Note: This issue was originally created as ARROW-11636. Please see the migration documentation for further details.}

[asfimport]

Neal Richardson / @nealrichardson:
Thanks! Happy to help if you have any questions, feel free to ping me.

[multimeric]

I would be happy to work on this, but I guess there's a lot of design to be done first? I'm envisaging Tensor will be an R6 class holding a Tensor pointer. Then we write S3 methods for Tensor so it can behave kind of like a native R array. Is that roughly right? Only thing I'm not sure about is where it will fit into the R6 inheritance hierarchy. Is it an ArrowObject? An ArrayDatum?

[multimeric]

@nealrichardson?

[thisisnic]

CC @paleolimbot too as I know you experimented with something relating to this.

[paleolimbot]

I think the prototype you're referring to is here: https://gist.github.com/paleolimbot/c42f068c2b8b98255dbfbe379d905607

This prototype is for the tensor extension type, which is more like a tensor as a column in a table and is, I believe the preferred way to deal with tensors since the extension type works with the C data interface and ArrowTensor does not (as I understand it, ArrowTensor is a C++ class and an item in the IPC spec but has little support elsewhere).

The extension type will have C++ support so it should be straightforward to add a wrapper (i.e., the prototype I did probably won't be necessary). It should work like any other parameterized type (i.e., you could use something like the wrapper around ListType as an example for how to implement it). I don't remember if that PR has merged yet.

We could also wrap the ArrowTensor C++ class, which is I believe what you're referring to. I'm a little concerned that adding support for that particular class is adding support for something that will be deprecated soon and hasn't seen much use elsewhere.

[multimeric]

Okay I see your point about there being two different types involved here. The reason I want ArrowTensor in R is because I want to be able to store a single tensor on disk (in the Feather/IPC format) without the context of a table. Is the extension type capable of doing this? Does anyone know for sure if the ArrowTensor class is going away? If so, why?

I worry that, to use the tensor extension type for my use case, I would have to add some redundant metadata such as putting it in a 1-column, 1-row table which I don't particularly want to do. But if it works with Feather and it will insulate this implementation from deprecation, then I'm okay with using it.

[paleolimbot]

I would have to add some redundant metadata such as putting it in a 1-column, 1-row table which I don't particularly want to do

It is true, that's what the tradeoff is. In memory and in an IPC/feather file it's still just a bunch of memory-mapped bytes, and with the extension type version you would also have the ability to zero-copy ship it to Python/Rust/Java/Whatever (whereas that's not supported and may never be supported).

You could start a PR with the prototype I linked above...the final version is still in review but at the R level it will still be an R6 class with the same methods as the C++ class ( https://github.com/apache/arrow/pull/8510/files#diff-2961e877000b1d44dd785688a0bda16061d7536facc15d762d412ea18875609d ). The final version might look more like the FixedSizeListType class: https://github.com/apache/arrow/blob/main/r/R/type.R#L683-L699 ; however, the "convert from R matrix/array" and "convert to R matrix/array" can (and maybe should) be implemented in R. We will also need tests, and with an R prototype we can write those tests so that when the C++ version is merged we're ready to go.

[multimeric]

Okay, I'm happy to be guided by your experience. If you think the tensor extension type is going to be more stable and more portable than ArrowTensor and offer similar value then I'll help with that. It just seems odd that Arrow should only provide these high-level data-frame types ("table containing a column which is a list of tensors") via IPC and not the lower level ones ("a tensor").

Only thing I'd like to know is the time frame for merging the extension type in C++, because I might hold off on the R until the interface is stable. It seems like a bit of wasted effort to work on it right now when things may still change.

[paleolimbot]

As far as I know @rok is hot on the case!

[rok]

Indeed! #8510 just merged and Python bindings are incoming #34883. There's also an incoming utility (FromTensor, ToTensor, strides) PR that might be interesting here #34797

[rok]

As for the Tensor type - it's not clear to me how much use it gets in the wild and depending on your use case you're probably better off with the tensor extension array type.

[github-actions]

This issue has been marked as stale because it has had no activity in the past 365 days. Please remove the stale label or comment below, or this issue will be closed in 14 days. If this improvement is still desired but has no current owner, please add the 'Status: needs champion' label.