0.4.0a1
=======

* The paraphrase generation code was extended and can now use BART instead of GPT2. It also now
  has the ability to run as a translation task as well (using the Marian models) [#26, #27, #29, #31].
* Added the ability to override the context and the question used as input to the model [#23].
* MultiGPU training was tested and fixed [#25].
* Completed support for beam search, including the ability to return multiple results for a given input [#30].
* Misc bug fixes [#32].

0.3.0
=====

* New option: sentence batching. Multiple sentences with related properties can be batched
  together in microbatches within a larger minibatch [#14, #11].
* Added option to append context and question in a single model input [#18, #20, #22].
* Updated Transformers dependency to 2.9, and fixed compatibility with newer versions [#18, #24].

0.2.0
=====

* No changes since 0.2.0b2.

Please see the development releases below for the full list of features in this release.

0.2.0b2
=======

* Misc bug fixes related to inference time [#12, #13].

0.2.0b1
=======

* Added multilingual Almond tasks [#10].

0.2.0a2
=======

* Misc bug fixes [#8, #9]

0.2.0a1
=======

New features:
* Add new tasks for Almond: almond_dialogue_nlu, almond_dialogue_nlg, almond_dialogue_policy
* Added a new encoder, "Coattention", which encodes the context and question separately, then
  coattends and applies a BiLSTM layer.
* For Coattention and Identity encoder, it is now possible to specify the context and question
  embeddings separately.
* Embeddings in context, question and answer can now be untied, by suffixing the name with '@'
  followed by an unique identifier (e.g. bert-base-uncased@0 and bert-base-uncased@1).
* Added an option to pretrain the context encoder, using MLM objective.
* Added beam search.
* New embedding option: XLM-R (XLM trained with Roberta).
* New task: paraphrasing with GPT2. This is not fully integrated with the other tasks yet,
  but it will in the future.
* New command "genienlp export" can be used to save a trained model for inference.

Incompatible changes:
* The --save flag is now required when calling train

0.1.1
=====

* Fix publishing on pypi

0.1.0
=====

* First release