Very nice article. I think to elaborate, Flux has a decent and functional onehot batch implementation. The biggest advantage to using Lathe for preprocessing and modeling is ecosystem consistency. Lathe has bindings similar to Sklearn with objects that hold methods (that are literally the same.)

D = RandomForestClassifier(trainX, trainy)

D.predict(testX)

Same can be said for the scalers, all of which can be used in a pipeline object.

So in essence, Lathe is a black box model resource for Julia that has as little black box as possible with hyper parameters, just like Sklearn. I would say that Flux is the Tensor flow of the Julia world, and Lathe is the Sklearn of the Julia world.

Lathe also has a very early and non-breaking multi function module for doing neural networks, but the implementation is meant to be node by node and quite complex, I do not think it is usable in its current state.

All in all though, with all of the packages mentioned I think Julia does have a significant future for itself in this regard.