Taycir Yahmed

https://github.com/iterative/cml

Ideas: * Check scikit-learn speed guide: https://scikit-learn.org/stable/developers/performance.html * Check pybind: https://github.com/pybind/pybind11

Consider huggingface Trainer and/or PyTorch lightning. Add ONNX support.

Change the doc strings to Google style.

check the tutorial [here](https://realpython.com/python-type-checking/). The idea is to add type hints in all methods and functions, avoiding duck typing.

LMs to add : * [openai gpt](https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/modeling_openai.py) * [gpt2](https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/modeling_gpt2.py) Also, try to use `pytorch hub` instead. * [GPT-2](https://pytorch.org/hub/huggingface_pytorch-pretrained-bert_gpt2/) * [GPT](https://pytorch.org/hub/huggingface_pytorch-pretrained-bert_gpt/) * [BERT](https://pytorch.org/hub/huggingface_pytorch-pretrained-bert_bert/)

The idea is to create example notebooks for use cases using the library. You can start by benchmark training on Swag or CoS-E datasets.

The idea is to write a hyper-parameters selection wrapper like `sklearn`'s `GridSearch` and `RandomizedGridSearch` that can ensure training the `mcQA` model with various parameters and selecting one of the combinations...

The idea is to write command line entry points to : `train`, `eval` and `train_eval` multiple choice question answering models.