Quantization with TFLite (#854)
* Cleaning DatasetProcessor
* Cache
* Cache
* Starting working
* WIP
* [WIP] Big refactor of TaskProcessing classes
* [WIP] tests
* [WIP] tests almost done
* [WIP] tests almost done
* [WIP] tests done
* Quantization working
* Remove dependency on evaluate
* Renaming file
* Adding torchvision as a dependency
* [WIP] quantization tests
* Fix get_task_from_model
* Styling
* Fix bad argument name
* Fix batching
* Add quantization approach
* Fix stable diffusion test
* Add CLI tests
* Load smallest split if not provided
* Fix fallback_to_float argument
* Skipping benchmark tests
* Decouple github actions to make tests faster
* Mark test in tests/exporters/tflite/test_tflite_export.py as well
* Add docstrings
* Add argument description
* Styling
* Skipping tests for unsupported tasks
* [WIP] make export quantization arguments a dataclass
* Apply suggestions
* Fix token-classification for failing models
* Fix question-answering tests
* Filter which models support int8x16
* Fix image-classification image-key
* Styling
* Remove int8x16 for roformer
* Fix question answering issue with null columns
* Disable int8 quantization for Deberta and XLM-Roberta on question-answering for now
* Disable int8 quantization for Deberta
* Rename QuantizationConfig to TFLiteQuantizationConfig
* Add warnings when guessing the data keys
* fix import in test
* Remove pytest.ini
* Fuse get_task_from_model with infer_task_from_model
* Add log if smallest split is used
