Simpletransformers

Added

- Added support for `BERTweet` with `ClassificationModel`. [manueltonneau](https://github.com/manueltonneau)

Added

- Added support for multilabel classification with the CamemBERT model. [adrienrenaud](https://github.com/adrienrenaud)

Changed

- Output arrays in classification evaluate/predict now avoids `np.append()`. This should be more time and memory efficient.

Added

- Added `layoutlm` model for NER

Fixed

- Potential fix for inconsistent `eval_loss` calculation

Added

- Added `convert_to_onnx` function to the following models:
- ClassificationModel
- NERModel
- Converted ONNX models can be loaded (requires specifying `onnx: True` in model_args) and used for prediction.
- Added `fp16` support for evaluation and prediction (requires Pytorch >= 1.6) for the following models:
- ClassificationModel
- ConvAI
- MultiModalClassificationModel
- NERModel
- QuestionAnsweringModel
- Seq2Seq
- T5Model
- Added multigpu prediction/eval in
- ClassificationModel
- ConvAI
- MultiModalClassificationModel
- NERModel
- QuestionAnsweringModel
- Seq2Seq
- T5Model

Fixed

- Thread count can now be specified for MultiLabelClassificationModel.

Fixed

- Fixed compatibility issue with transformers 3.2. (BertPreTrainedModel was being imported from an incompatible path)