Description

Warning : This Merge Request should not be merged now. It is created to start a discussion about the code philosophy in Melusine.

The NeuralModel class is outdated and need refactoring. Exemples of pain points in the current version are:

Model saving does not work well (requires a raw pickle dump of the NeuralModel class)
The model object used in inference contains training parameters (ex: epochs, loss, etc)
Most NeuralModel methods are forked (if/elif) to apply different processing for CNN and Transformer models
The NeuralModel class only be customized in terms of model architecture but not in terms of data preparation
The model training config is hard to find (lost in long conf file)

Suggestions

Object Oriented Programming : Create a BaseMelusineModel class that defines a signature for all Melusine models
Create subclasses for the CNN and Transformers models
Make it easy for a Melusine user to customize the default models (model architecture and preprocessing)
Create a Trainer class to separate training from inference parameters
Improve the model saving to make it robust and make configs readable by humans
Make the tokenizer a part of the model object

Fixes # (issue) Several Issues are concerned:

Please delete options that are not relevant.

The refactored code should come with an extensive set of new tests