Contributions to MLJ are most welcome. Queries can be made through issues or the Julia slack channel, #MLJ.
We follow
this git
work-flow and, in particular, ask that all pull requests be made to
thedev branch of the appropriate repo, and not to master. This
includes changes to documentation. All pull requests onto master
come from dev and generally precede a tagged release.
Contributors are kindly requested to adhere to the Blue style guide, with line widths capped at 80 characters.
MLJ has a basement level model interface, which must be implemented
for each new learning algorithm. Formally, each model is a mutable struct storing hyperparameters and the implementer defines
model-dispatched fit and predict methods; for details, see
here. The general user
interacts using machines which bind models with data and have an
internal state reflecting the outcomes of applying fit! and
predict methods on them. The model interface is pure "functional";
the machine interface more "object-oriented".
A generalization of machine, called a nodal machine, is a key element of learning networks which combine several models together, and form the basis for specifying new composite model types. See here for more on these.
MLJ code is now spread over multiple repositories.