Model summary to show tree details for FastTree

[ganik]

Currently model summary doesnt show trees and split points for FastTree models.
FastTree model needs to implement ICanGetSummaryAsIDataView that would dump this info.

[yaeldekel]

Hi @ganik ,
FastTree models implement ICanGetSummaryAsIRow, so it doesn't need to implement ICanGetSummaryAsIDataView. Either way, both these interfaces are internal and are not intended to be used programatically to get information about the model.
For getting information about the model, FastTree models have a public TrainedTreeEnsemble property, that returns an object that contains information about the tree ensemble:

        public double Bias { get; }
        public IReadOnlyList<T> Trees { get; }
        public IReadOnlyList<double> TreeWeights { get; }

where T derives from RegressionTreeBase and contains information about a single tree.

[codemzs]

@ganik Can you please comment? if this needs to be fixed can you please confirm this can go in before 1.1 release next week?

[artidoro]

This issue pertains to Nimbus ML. In order to pass the model parameters to Nimbus ML the model parameters class has to implement ICanGetSummaryAsIRow or ICanGetSummaryAsIDataView.

FastTree implements ICanGetSummaryAsIRow as Yael points out. However, it does not expose all the tree structure, it only exposes the gains.

We should implement ICanGetSummaryAsIDataView to pass all the information in TrainedTreeEnsemble to Nimbus (including the tree structure).

[artidoro]

My suggestion is to construct an IDataView where each tree in the ensemble is represented in its own row. Since Wei-Sheng did a lot of work to reduce down the list of public fields defining the trees, they should all be include in the IDataView, with each column corresponding to a different field.