Add onnx export of models with a multiple choice classification head

[echarlaix]

This PR adds the export support of models with a multiple-choice classification head, resolving #16695

This includes the following additions:

• The "multiple-choice" feature was added to the corresponding model topologies
• The dummy inputs are generated to match the expected inputs shape which includes an extra dimension corresponding to the number of candidates answers
• The inputs method of the models corresponding OnnxConfig were modified to support the additional dynamic axis corresponding to the number of candidates

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Looks good on my side, thanks for working on this! Just make sure to run make fix-copies and make style on your branch to fix the quality issues.

[echarlaix]

Thanks for the review as well as the reminder @sgugger !

[lewtun]

Thanks for adding this @echarlaix - it looks great to me!

I think there's a few more models that support multiple choice classification and have an ONNX export:

• BigBird
• Data2VecText
• Electra
• FlauBERT

Would you like to include them in this PR too?

We should also do a final run of the slow tests before merging to make sure everything passes:

RUN_SLOW=1 pytest tests/onnx/test_onnx_v2.py

[michaelbenayoun]

Perfect, thanks a lot for taking care of this @echarlaix !

[michaelbenayoun]

I would just add a small comment to make clear what's happening here.
What do you think?

Suggested change

	tokenized_input = preprocessor(dummy_input, text_pair=dummy_input)
	# Shape is [batch_size x num_choices, seq_length]
	tokenized_input = preprocessor(dummy_input, text_pair=dummy_input)

[echarlaix]

Thanks for the reviews @lewtun and @michaelbenayoun.
I have added some comments to make things clearer as well as the BigBird, Data2VecText, Electra and FlauBERT models support. Also when running the command line RUN_SLOW=1 pytest tests/onnx/test_onnx_v2.py, all the tests are passing.

[lewtun]

Thanks for iterating on this @echarlaix - it looks great!