Multiprocessing Support for `QuestionAnsweringModel`

Added

• Added multiprocessing support for Question Answering tasks for substantial performance boost where CPU-bound tasks (E.g. prediction especially with long contexts)
• Added multiprocessing_chunksize (default 500) to global_args for finer control over chunking. Usually, the optimal value will be (roughly) number of examples / process count.

Fixed

• Fixed bug in NERModel.predict() method when split_on_space=False. @alexysdussier

Option to disable model saving

Added

• Added no_save option to model args. Setting this to True will prevent models from being saved to disk.
• Added minimal training script for Seq2Seq models in the examples directory.

Docs update and Bug fixes

Fixed

• Fixed potential bugs in loading weights when fine-tuning an ELECTRA language model. Fine-Tuning an ELECTRA language model now requires both model_name and model_type to be set to electra.
• Bugfix for generic Seq2SeqModel
• Bugfix when training language models from scratch

Changed

• Updated Seq2SeqModel to use MarianTokenizer with MarianMT models. @flozi00

Sequence-to-Sequence task support added

Added

• Sequence-to-Sequence task support added. This includes the following models:
  • BART
  • Marian
  • Generic Encoder-Decoder
• The args dict of a task-specific Simple Transformers model is now saved along with the model. When loading the model, these values will be read and used.
  Any new args passed into the model initialization will override the loaded values.

Improvements and Bug Fixes

Added

• Support for AutoModel in NER, QA, and LanguageModeling. @flozi00

Fixed

• Now predict function from NER_Model returns a value model_outputs that contains:
  A Python list of lists with dicts containing each word mapped to its list with raw model output. @flaviussn
• Fixed T5 lm_labels not being masked properly
• Fixed issue with custom evaluation metrics not being handled correctly in MultiLabelClassificationModel. @galtay

Changed

• Torchvision import is now optional. It only needs to be installed if MultiModal models are used.
• Pillow import is now optional. It only needs to be installed if MultiModal models are used.

T5 Model Added

Added

• Added support for T5 Model.
• Added do_sample arg to language generation.
• NERModel.predict() now accepts a split_on_space optional argument. If set to False, to_predict must be a a list of lists, with the inner list being a list of strings consisting of the split sequences. The outer list is the list of sequences to predict on.

Changed

• eval_df argument in NERModel.train_model() renamed to eval_data to better reflect the input format. Added Deprecation Warning.

ELECTRA model support for Classification and Question Answering tasks

Added

• Added Electra model support for sequence classification (binary, multiclass, multilabel)
• Added Electra model support for question answering
• Added Roberta model support for question answering

Changed

• Reduced logger messages during question answering evaluation

Language Generation

Language Generation is now supported!

Supported model types:

• GPT-2
• CTRL
• OpenAI-GPT
• XLNet
• Transformer-XL
• XLM

Custom Metrics for Question Answering

Added

• Added support for custom metrics with QuestionAnsweringModel.

Fixed

• Fixed issue with passing proxies to ConvAI models. @Pradhy729

Easier configuration for models and support for getting hidden layer outputs

Added

• Added option to get hidden layer outputs and embedding outputs with ClassificationModel.predict() method.
  • Setting config: {"output_hidden_states": True} will automatically return all embedding outputs and hidden layer outputs.

Changed

• global_args now has a config dictionary which can be used to override default values in the confg class.
  • This can be used with ClassificationModel, MultiLabelClassificationModel, NERModel, QuestionAnsweringModel, and LanguageModelingModel