Binary fast tree doesn't return tree predictor in pigsty

[Ivanidzo4ka]

machinelearning/src/Microsoft.ML.FastTree/FastTreeStatic.cs

Line 93 in a285f8d

Action<IPredictorWithFeatureWeights<float>> onFit = null)

It returns IPredictorWithFeatureWeights which I found weird since I expect tree predictor from fast tree.
Regression and Multiclass returns FastTreeRegressionPredictor and FastTreeRankingPredictor which is right.

[Ivanidzo4ka]

@sfilipi any chance you know why?

[sfilipi]

Because that's what the trainer produces...

Similar for LightGBM

I think it is this way because of inheritance. Does it need to get re-worked?

[Ivanidzo4ka]

I don't see any reason to train Tree predictor and get weights object, but maybe I'm just silly. Let's hear our druid.
@TomFinley can you share with us the wisdom of centennial oaks?

[shauheen]

Poking the druid, @TomFinley

[wschin]

This was rediscovered in #2319. I will fix it. Note that fixing #2378 also fixes this issue.

[TomFinley]

I had missed that I was being poked. I have related a distinct but nonetheless related issue of #2378. @wschin brought this issue to my attention, I believe but am not certain that he will write there.

Anyway, I think the idea, which I think @yaeldekel also may have had some thoughts on, was a generic class on the type of model parameters, combined with an (optional) calibration object. Indeed this issue is mightily interfering with my ability to resolve #2251.