Tree based trainers implement ICanGetSummaryAsIDataView #3892
Conversation
| /// <see cref="TrainedTreeEnsemble"/>. | ||
| /// </summary> | ||
| public abstract class TreeEnsembleModelParametersBasedOnRegressionTree : TreeEnsembleModelParameters | ||
| public abstract class TreeEnsembleModelParametersBasedOnRegressionTree : TreeEnsembleModelParameters, ICanGetSummaryAsIDataView |
There was a problem hiding this comment.
Why do we need ICanGetSummaryAs...? IDataView RegressionTreeEnsembleAsIDataView(..) looks sufficient for generating a summary. #ByDesign
There was a problem hiding this comment.
That's the interface that Summarize entrypoint uses to get the IDataView from a trainer.
See:
And:
artidoro
force-pushed
the
fasttreesummary2
branch
from
81a5d39 to
9b0e097
Compare
| /// Used for the Summarize entrypoint. | ||
| /// </summary> | ||
| IDataView ICanGetSummaryAsIDataView.GetSummaryDataView(RoleMappedSchema schema) | ||
| => RegressionTreeBaseUtils.RegressionTreeEnsembleAsIDataView(Host, TrainedTreeEnsemble.Bias, TrainedTreeEnsemble.TreeWeights, TrainedTreeEnsemble.Trees); |
| [LightGBMFact] | ||
| public void LightGbmBinaryClassificationTestSummary() | ||
| { | ||
| var (pipeline, dataView) = GetBinaryClassificationPipeline(); |
There was a problem hiding this comment.
GetBinaryClassificationPipeline [](start = 39, length = 31)
Make sure that these work with the categorical split after one hot encoding.
codemzs
left a comment
codemzs
left a comment
There was a problem hiding this comment.
We reviewed this in person and mostly looks good. Signing off assuming you will be fix the output of the summary to not display negative values for node ids and instead use 2s complement so that it is easier to understand the tree structure.
artidoro
force-pushed
the
fasttreesummary2
branch
from
d6bf2c4 to
78f4f86
Compare
ghost
locked as resolved and limited conversation to collaborators
4 participants
Fixes #3755.
NimbusML did not have access to the details on the tree structure.
This PR implements the
ICanGetSummaryAsIDataViewinterface which is used in theSummarizeentrypoint to pass a summary of the model parameters to NimbusML in the form of anIDataView.I create a utility method that does the conversion from
RegressionTreeBasetoIDataViewwith special treatment forQuantileRegressionTreeswhich have additional information.In the
IDataView, each node has its own row and the columns correspond to the fields describing each node. To determine which tree the node belongs to there is aTreeIDcolumn.