[TFLite Export] Adds TFLite support for LayoutLMv3

[salmanmaq]

What does this PR do?

Adds support for exporting LayoutLMv3 to TFLite. Some tests for all quantization approaches were failing for some tasks, so I limited the supported quantization schemes for now.

It was primarily based on #813 and the ONNX export code already part of the library.

It can be tested as:

pytest tests/exporters/tflite/test_*.py -k "layoutlmv3" -s --exitfirst

You can export a trained model as:

optimum-cli export tflite --model <input_dir> --task <> --sequence_length <> --width <> --height <> <output_dir>
optimum-cli export tflite --model microsoft/layoutlmv3-base --task token-classification --sequence_length 512 --width 224 --height 224 ~/tflite_conversion

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes? (Not needed as far as I can tell)
• Did you write any new necessary tests? (Not needed as far as I can tell)

[fxmarty]

Hi @salmanmaq thank you for your contribution, LGTM!

[michaelbenayoun]

LGTM!

Will wait that the CI goes green to merge. Except for int8 quantization tests that do not seem related to your changes at all.

[salmanmaq]

Hey @fxmarty @michaelbenayoun Would you like to merge this before it goes stale? Happy to update the PR in case it has

[michaelbenayoun]

Hi @salmanmaq ,
Happy to do it once the tests related to LayoutLMv3 pass. I think some tests in export with quantization do not pass.

[salmanmaq]

Hi @salmanmaq , Happy to do it once the tests related to LayoutLMv3 pass. I think some tests in export with quantization do not pass.

Thanks @michaelbenayoun. Is it something with the PR, or at your end? You had previously mentioned

Except for int8 quantization tests that do not seem related to your changes at all.

Would be checking it out though. If you can give some pointers, I am happy to have a look and attempt a fix.

[ahmet-sabbagh]

@salmanmaq, since there is a suggestion the issue might not be related to your PR, I noticed that your fork is a bit behind, maybe you could pull the latest commits try again?

[salmanmaq]

Hi @fxmarty @michaelbenayoun . Bringing this to your attention again for a potential merge. I've merged and pushed the latest changes. I've tested LayoutLMv3 conversion (without quantization) and with the supported quantization types (FP16, INT8 dynamic). I am able to successfully convert and use those models. I think the test errors are not related to my code. So I suppose you can consider merging it. If the test errors are due to my code, I am happy to take some guidance and work on those (but I don't think that's the case).

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.