Bug Fixes

• Fixes bug introduced in https://github.com/Novetta/adaptnlp/issues/149 with https://github.com/Novetta/adaptnlp/pull/151

New Features:

• NER Fine-Tuner added

Bug Squashed

• Embeddings were conjoined rather than separated 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 that integrates with HuggingFace's Dataset class

• A new Tuner API for training and fine-tuning Transformer models

• Full integration of the latest fastai library 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 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 API
• LanguageModelFineTuner no longer exists, and the same tuning functionality is in LanguageModelTuner

Bugs Squashed

• max_len Attribute Error (127
• Integrate a complete Data API (milestone) (129
• Use the latest fastcore (132)
• Fix unused kwarg arguments in text generation (134)
• Fix name 'df' is not defined (135)

Breaking Changes:

• New versions of AdaptNLP will require a minimum torch version of 1.7, and flair of 0.9 (currently we install via git until 0.9/0.81 is released)

New Features

• Complete conversion to the nbdev library format and actions

• Complete revamp of the documentation

• Inference API entirely relies on fastai_minima and is now built on fastai's Callback System

• Integration with fastcore to simplify logic

• HuggingFace and Flair ModelHubs, an easier API to interact, search, and download HF and Flair models. Uses huggingface_hub as a backend. Has logged every single Flair model, including those not in the HuggingFace API

Bugs Squashed

• Fix accessing bart-large-cnn (110)
• Fix SAVE_STATE_WARNING (114)

Official AdaptNLP Docker Images updated

• Using NVIDIA NGC Container Registry Cuda base images #101
• All images should be deployable via. Kubeflow Jupyter Servers
• Cleaner python virtualvenv setup #101
• Official readme can be found at https://github.com/Novetta/adaptnlp/blob/master/docker/README.md

Minor Bug Fixes

• Fix token tagging REST application type check #92
• Semantic fixes in readme #94
• Standalone microservice REST application images #93
• Python 3.7+ is now an official requirement #97

Updated to nlp 0.4 -> datasets 1.0+ and multi-label training for sequence classification fixes.

EasySequenceClassifier.train() Updates

• Integrates datasets.Dataset now
• Swapped order of formatting and label column renaming due to labels not showing up from torch data batches #87

Tutorials and Documentation

• Documentation and sequence classification tutorials have been updated to address nlp->datasets name change
• Broken links also updated

ODSC Europe Workshop 2020: Notebooks and Colab

• ODSC Europe 2020 workshop materials now available in repository "/tutorials/Workshop"
• Easy to run notebooks and colab links aligned with the tutorials are available

Updated to transformers 3+, nlp 0.4+, flair 0.6+, pandas 1+

New Features!

New and "easier" training framework with easy modules: EasySequenceClassifier.train() and EasySequenceClassifier.evaluate()

• Integrates nlp.Dataset and transformers.Trainer for a streamlined training workflow
• Tutorials, notebooks, and colab links available
• Sequence Classification task has been implemented, other NLP tasks are in the works
• SequenceClassifierTrainer is still available, but will be transitioned into the EasySequenceClassifier and deprecated

New and "easier" LMFineTuner

• Integrates transformers.Trainer for a streamlined training workflow
• Older LMFineTuner is still available as LMFineTunerManual, but will be deprecated in later releases
• Tutorials, notebooks, and colab links available

EasyTextGenerator

• New module for text generation. GPT models are currently supported, other models may work but still experimental
• Tutorials, notebooks, and colab links available

Tutorials and Documentation

• Documentation has been edited and updated to include additional features like the change in training frameworks and fine-tuning
• The sequence classification tutorial is a good indicator of the direction we are going with the training and fine-tuning framework

Notebooks and Colab

• Easy to run notebooks and colab links aligned with the tutorials are available

Bug fixes

• Minor bug and implementation error fixes from flair upgrades

Split dev requirements #29 #66 Pinned torch #70

Updated to Transformers 2.8.0 which now includes the ELECTRA language model

EasySummarizer and EasyTranslator Bug Fix #63

• Address mini batch output format issue for language model heads for the summarization and translation task

Tutorials and Workshop #64

• Add the ODSC Timeline Generator notebooks along with colab links
• Small touch ups in tutorial notebooks

Documentation

• Address missing model_name_or_path param in some easy modules