Add Split-20 - change uneven split behavior to be more torch-like

[p-wysocki]

Description

Introduce Split-20, which performs uneven tensor split in torch's manner: instead of reducing only the size of the last dimension of the output, it reduces the last n dimensions' sized by 1 each.

This was done using a mode attribute, which has 2 possible values: torch (default) and legacy. It's up to discussion if it should stay like that or maybe it should be renamed or changed to a boolean.

Motivation and Context

Issue: #4742

TODO

• Update reference evaluator
• Confirm the correctness of setting a 0-size dimension in ONNX (test_split_uneven_split_3d_4)
• Add more test cases and assert backend tests are correct

[xadupre]

Should we keep the legacy as default?

[p-wysocki]

As far as I know torch is the desired one, and I set legacy to be default in version converter when coming from opset 19 to 20. I can change it.

[gramalingam]

Good question. I think the current choice is ok. Especially if it is only the torch exporter that uses this, and presumably the torch converter needs the new option, not the old one.

[gramalingam]

Actually, this doesn't always work. Eg., if we split 7 values into 5, then this will produce 4, 4, 4, 4, -1 which doesn't make sense. We should do 4, 3, 0, 0, 0, I think. We may need to change the description accordingly. In the other mode, we will do 2, 2, 1, 1, 1.

[gramalingam]

Instead of changing the axis, it would make sense to change the new attribute introduced minimize_diff or whatever it is called. At least, I am looking for some test-case to show the difference that attribute makes, whether that is this test or some other test.

[gramalingam]

Hi, is this ready for review? Would it be possible to complete this PR for the upcoming release?

[justinchuby]

@p-wysocki

[p-wysocki]

Added new tests, generated docs and test files are on the way - I'm currently having some issues with protobuf in my environment, which stops me from generating them.

[gramalingam]

It will also have to be moved to a later opset (21, I guess?)

[zhenhuaw-me]

Spec like mode:torch requires background of how Torch implements it.

We can make it more specific by:

1. Keep num_outputs, and change the semantic to match numpy.array_split and torch.tensor_split

If indices_or_sections is an integer n or a zero dimensional long tensor with value n, input is split into n sections along dimension dim. If input is divisible by n along dimension dim, each section will be of equal size, input.size(dim) / n. If input is not divisible by n, the sizes of the first int(input.size(dim) % n) sections will have size int(input.size(dim) / n) + 1, and the rest will have size int(input.size(dim) / n).

2. Introduce split_size, which is the same as torch.split:

If split_size_or_sections is an integer type, then tensor will be split into equally sized chunks (if possible). Last chunk will be smaller if the tensor size along the given dimension dim is not divisible by split_size.