MAJOR change in design outlined here
MLJis an attempt to create a framework capable of easily tuning machine learning models. Thanks to a solid abstraction layer, it allows user to easily add new models to its framework, without losing any of the features.
Landmarks:
- Implement first basic structure
- Implement tuning for continuous parameters
- Implement tuning for discrete parameters
- Basic custom sampling method (K-fold)
- Basic CV with custom score
- Wrap at least a handful of models for regression & classification
- Add multivariable regression methods
- Add automatic labelling for classifiers
- Find a way to make it clear what arguments a model expects
- Allow any sampling methods from
MLBase.jl - Add compatibility with multiple targets
Known Issues:
- Fix stacking storage
- Get packages change Float to AbstractFloat so that forward diff can work
Notes: Forward diff does not work