Model summary for trees

[ganik]

Currently model summary for trees in ML.NET returns vectorized output.

• Need to double check it doesnt break NimbusML
• If it does, needs a fix either in ML.NET or NimbusML

[pieths]

The following tree based predictors throw an error when retrieving the summary after a fit().

• FastTreesBinaryClassifier
• FastTreesRegressor
• FastTreesTweedieRegressor
• FastForestBinaryClassifier
• FastForestRegressor

Here is the error output. Note, the second error only shows up after the CategoricalSplitFeatures column has been dropped.

Error: *** System.NotSupportedException: 'Column has variable length vector: CategoricalSplitFeatures.
Not supported in python. Drop column before sending to Python

Error: *** System.NotSupportedException: 'Column has variable length vector: CategoricalCategoricalSplitFeatureRange.
Not supported in python. Drop column before sending to Python

It looks like the IDataView summary output for all the tree based predictors gets created in src\Microsoft.ML.FastTree\Utils\RegressionTreeBaseUtils.cs in ml.net.

[pieths]

Three of the LightGbm predictors are effected by this as well.

• LightGbmBinaryClassifier
• LightGbmRanker
• LightGbmRegressor

[pieths]

The summaries now work as expected because variable column support was added in commit df1cf21.