Melusine Versions Save

What's Changed

New features

• An Ethics Guide is now available to evaluate AI projects, with guidelines and questionnaire. The questionnaire is based on criteria derived in particular from the work of the European Commission and grouped by categories.
• Melusine also offers an easy and nice dashboard app with StreamLit. The App contains exploratory dashboard on the email dataset and a more specific study on discrimination between the dataset and a neural model classification.

What's Changed

New features:

• Attentive Neural Networks are now available. 🎉 We propose you an original Transformer architecture as well as pre-trained BERT models (Camembert and Flaubert)
• Tutorial 13 will explain you how to get started with these models and attempt to compare them.
• Validation data can now be used to train models (See fit function from NeuralModel for usage)
• The activation function can now be modified to adapt to your needs (See NeuralModel init for usage)

What's Changed

New features:

• Added a class ExchangeConnector to interact with an Exchange Mailbox
• Added new tutorial tutorial14_exchange_connector to demonstrate the usage of the ExchangeConnector class

Updates:

• Gensim upgrade (4.0.0)
• Propagate modifications stemming from the Gensim upgrade (code and tutorials)
• Package deployment : switch from Travis CI to Github actions

What's Changed

Updates:

• Compatibility with python 3.7 and 3.8
• Optional dependencies (viz, transformers, all)
• Specify custom configurations with environment variable MELUSINE_CONFIG_DIR
• Use any number of JSON and YAML files for configurations (instead of just one config file)

Bug fix:

• Fixed bug when training transformers model without meta features

What's Changed

• Feature/tf probability by @benoitLebreton-perso in https://github.com/MAIF/melusine/pull/110 Add uncertainty quantification with tensorflow-probability
• [#134] Add "other" value to the training of label encoder by @DataFactory-Verlingue in https://github.com/MAIF/melusine/pull/135 Dummifier deal with unseen categories
• [#136] Deal with draft in file tree by @DataFactory-Verlingue in https://github.com/MAIF/melusine/pull/137 ExchangeConnector deal with Draft emails in the mail box
• Update scikit learn version by @hugo-quantmetry in https://github.com/MAIF/melusine/pull/145 Fix for API breaking changes in scikit-learn version 1.0

Full Changelog: https://github.com/MAIF/melusine/compare/v2.3.5...v2.3.6

What's Changed

Bug fix:

• PR139: Fix config search with a recursive paramater.

What's Changed

New features:

• PR 128: A Lemmatizer class has been added by the Société Générale team! Lemmatizer object is compatible with sklearn pipelines and is built around an sklearn Transformer. Details can be found in tutorial 04 and 08
• PR 132: A Stemmerclass has been added. Details can also be found in tutorial 04
• PR 132: A DeterministicEmojiFlaggerclass has been added to flag emojis. Details can be found in tutorial 08.

Updates:

• Python 3.6 is no longer supported for tensorflow compatibility issues. Melusine is now running with Tensorflow 2.8
• PR 121: Add the return of the histogram after the training (train.py)
• PR 120: Tokenizer can now be specified in a NeuralModel init. Embedding and Phraser classes have been simplified. See tutorial 04
• PR 120: Config has been split into different functionalities files that can be found in /melusine/config/parameters for more readability. See tutorial 10
• PR 120: A text_flaggerand a token_flagger class have been created to give you a glimpse of the library redesign but are not called yet.

Bug fix:

• PR 124: fixing purge of dict_attr keys while saving bert models (train.py)
• Issue 126: fixing initialisation of bert_tokenizer for cross validation (train.py)