Support GatherND operator in ONNX

[hariharans29]

This change is to support GatherND operator in ONNX.

[gramalingam]

TF's gather-nd seems to have an extra batch_dims attribute for batched gather-nd … is that required or useful?

[houseroad]

Also could you provide some brief summary about the PR in the PR's description?

[hariharans29]

Also could you provide some brief summary about the PR in the PR's description?

Well sure - but the description is same as the title - to support GatherND op in ONNX. Do you mean you want a justification to include this by any chance ?

[hariharans29]

TF's gather-nd seems to have an extra batch_dims attribute for batched gather-nd … is that required or useful?

Not sure about this - let me take a look. Thanks!

[hariharans29]

TF's gather-nd seems to have an extra batch_dims attribute for batched gather-nd … is that required or useful?

Not sure about this - let me take a look. Thanks!

@gramalingam - I took a look at the batch_dims concept in the tf Gather_nd op and like you said, it is to support batched inputs ('params' and 'indices' in tf op) and does seem like a good extension to the op's abstraction. However, I see no immediate use-case for us to justify including that. If need be, we can version this op later to support batch_dims. Or I can make the spec adjustment now and modify the contrib implementation of this op in ORT. Any thoughts?

CC: @wschin @linkerzhang @spandantiwari

[gramalingam]

I personally think it is okay to leave it for future extension if we don't see a use-case for it.

[hariharans29]

I personally think it is okay to leave it for future extension if we don't see a use-case for it.

I think this is a good idea too. We can always version the op later.

[CLAassistant]

All committers have signed the CLA.

[spandantiwari]

@hariharans29 - Overall, this looks OK. I don't have any specific blocking issues on this. Could you please take a look at the recently approved guideline document for adding new ops and update (if needed) based on that.
https://github.com/onnx/onnx/blob/master/docs/AddNewOp.md
Some suggestions in line with the guidelines:

1. Add more details in the op definition; could we include some equations?
2. Add more test examples.
3. Add a reference implementation in the test script.

[spandantiwari]

LGTM.

[wschin]

LGTM. Just one suspicious typo.

[gramalingam]

Change to "The output tensor is obtained by mapping each index-tuple in the indices tensor to the corresponding slice of the input data"

[hariharans29]

Let me also add - This operator is the inverse of ScatterND ,since ScatterND has a similar line in it

[gramalingam]

By the way: there was a discussion about supporting negative indices in Gather (since pytorch supports that). Do we want to extend this to allow negative indices?

[hariharans29]

By the way: there was a discussion about supporting negative indices in Gather (since pytorch supports that). Do we want to extend this to allow negative indices?

This is a good question - in this spec version, I explicitly did not allow it and pretty much followed the spec clarification in Gather. Any comments @spandantiwari @BowenBao ? Btw - Does GatherElements, ScatterElements, and ScatterND support negative indices ?

[hariharans29]

Spoke offline to @spandantiwari and @BowenBao - We will support negative indexing in this op after a consensus on it is reached and we will add negative indexing support in a more streamlined manner across similar ops.

For now, it makes sense to stay consistent with Gather which doesn't support negative indexing.