Adaptnlp

FastAPI endpoints can now be stood up and deployed with any huggingface sequence classification or question answering model specified as an env var arg.


Dependencies
Transformers pinned for stable updates

New Features
- NER Fine-Tuner

Bugs Squashed

- QA queries that did not contain `n_best_answers` worth of errors would throw a key error. Changed to return only what is there
- Some issues with the tutorials have been upstreamed, and inference tutorials now run under the CI

Bug Squashed

- Embeddings were conjoined rather than seperated out by word
- Question Answering Results would only return the first instance, rather than top `n` instances
- `AdaptiveTuner` can accept a `label_names` parameter for where the labels in a batch are present

New Features

- A new [Data API](https://novetta.github.io/adaptnlp/tuner.core#TaskDatasets) that integrates with HuggingFace's Dataset class

- A new [Tuner API](https://novetta.github.io/adaptnlp/tuner.core#AdaptiveTuner) for training and fine-tuning Transformer models

- Full integration of the latest [fastai library](https://docs.fast.ai) for full access to state-of-the-art practices when training and fine-tuning a model. As improvements are made to the library AdaptNLP will update to accomodate them

- A new [Result API](https://novetta.github.io/adaptnlp/result) that most inference modules return. This is a filterable result ensuring that you only get the most relevent information when returning a prediction from the `Easy*` modules

Breaking Changes

- The `train` and `eval` capabilities in the `Easy*` modules no longer exist, and all training related functionalities have migrated to the [Tuner](https://novetta.github.io/adaptnlp/tuner.core) API
- `LanguageModelFineTuner` no longer exists, and the same tuning functionality is in `LanguageModelTuner`

Bugs Squashed

- `max_len` Attribute Error ([127](https://github.com/Novetta/adaptnlp/issues/127)
- Integrate a complete Data API (milestone) ([129](https://github.com/Novetta/adaptnlp/issues/129)
- Use the latest fastcore ([132](https://github.com/Novetta/adaptnlp/issues/132))
- Fix unused kwarg arguments in text generation ([134](https://github.com/Novetta/adaptnlp/issues/134))
- Fix name 'df' is not defined ([135](https://github.com/Novetta/adaptnlp/issues/135))

Breaking Changes

- New versions of AdaptNLP will require a special forked and pinned version of Flair, that does not have a torch maximum requirement. Thus allowing for torch 1.7 >