Accessibility Problem of FastTree through Dynamic APIs

[wschin]

The following examples shows a way to retrieved the underlying model learned by a trainer, but I failed compiling it. For static APIs, we are able to assign FastTreeRegressionModelParameters to pred but in the corresponding dynamic API, the type of p becomes

Microsoft.ML.Data.BinaryPredictionTransformer<Microsoft.ML.Internal.Internallearn.IPredictorWithFeatureWeights<float>>

, which doesn't publicly expose `FastTreeRegressionModelParameters as one of its field.

            var data = new TextLoader(Env,
                    new TextLoader.Arguments()
                    {
                        Separator = ";",
                        HasHeader = true,
                        Column = new[]
                        {
                            new TextLoader.Column("Features", DataKind.R4, 0, 10),
                            new TextLoader.Column("Label", DataKind.R4, 11)
                        }
                    }).Read(GetDataPath(TestDatasets.generatedRegressionDataset.trainFilename));


            var trainer = ML.BinaryClassification.Trainers.FastTree(
                new FastTreeBinaryClassificationTrainer.Options
                {
                    NumThreads = 1,
                    NumTrees = 10,
                    NumLeaves = 5,
                });

            FastTreeRegressionModelParameters pred = null;
            trainer.WithOnFitDelegate((p) => { pred = p; });

Any tricks I can use to make this code example compiled without using internal classes? Or we need to modify our dynamic APIs?

cc @TomFinley, @Ivanidzo4ka, @sfilipi, @zeahmed.

We also should add an example using public interfaces introduced in #2243 after fixing this issue.

[wschin]

In addition to regression, I can also demonstrate a related bug using FastTree for binary classification.

            var ml = new MLContext(seed: 1, conc: 1);
            var data = ml.Data.ReadFromTextFile<SentimentData>(GetDataPath(TestDatasets.Sentiment.trainFilename), hasHeader: true);

            var trainer = ml.BinaryClassification.Trainers.FastTree(numLeaves: 5, numTrees: 3);

            BinaryPredictionTransformer<IPredictorWithFeatureWeights<float>> pred = null;
            var decoratedTrainer = trainer.WithOnFitDelegate((p) =>
                {
                    pred = p; // This line will never be executed.
                }
            );

            var pipeline = ml.Transforms.Text.FeaturizeText("SentimentText", "Features")
                .AppendCacheCheckpoint(ml)
                .Append(decoratedTrainer); // Replace "decoratedTrainer" with "trainer" will not trigger the assignment of pred, so pred will be bull.

            // Train.
            var model = pipeline.Fit(data);

Basically, the onFit is not implemented for at least FastTree and SDCA.
[Update] This is a false alarm. Please see the change made in the last edit.