Skip to content

Combine multiple tree ensemble models into a single tree ensemble #364

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Jun 19, 2018
Merged

Combine multiple tree ensemble models into a single tree ensemble #364

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b333e1a
Add a way to create a single tree ensemble model from multiple tree e…
yaeldMS Jun 15, 2018
6738a23
Address PR comments, and fix bugs in serializing/deserializing Regres…
yaeldMS Jun 18, 2018
3bd4b97
Address PR comments.
yaeldMS Jun 18, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions src/Microsoft.ML.Core/Prediction/ITrainer.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,6 +74,8 @@ public interface IModelCombiner<TModel, TPredictor>
TPredictor CombineModels(IEnumerable<TModel> models);
}

public delegate void SignatureModelCombiner(PredictionKind kind);

/// <summary>
/// Weakly typed interface for a trainer "session" that produces a predictor.
/// </summary>
Expand Down
1 change: 1 addition & 0 deletions src/Microsoft.ML.FastTree/Microsoft.ML.FastTree.csproj
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,7 @@
<Compile Include="TreeEnsemble\Ensemble.cs" />
<Compile Include="TreeEnsemble\QuantileRegressionTree.cs" />
<Compile Include="TreeEnsemble\RegressionTree.cs" />
<Compile Include="TreeEnsemble\TreeEnsembleCombiner.cs" />
<Compile Include="Training\Applications\GradientWrappers.cs" />
<Compile Include="Training\Applications\ObjectiveFunction.cs" />
<Compile Include="Training\BaggingProvider.cs" />
Expand Down
49 changes: 33 additions & 16 deletions src/Microsoft.ML.FastTree/TreeEnsemble/RegressionTree.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,22 +105,21 @@ public RegressionTree(byte[] buffer, ref int position)
LteChild = buffer.ToIntArray(ref position);
GtChild = buffer.ToIntArray(ref position);
SplitFeatures = buffer.ToIntArray(ref position);
int[] categoricalNodeIndices = buffer.ToIntArray(ref position);
CategoricalSplit = GetCategoricalSplitFromIndices(categoricalNodeIndices);
if (categoricalNodeIndices?.Length > 0)
byte[] categoricalSplitAsBytes = buffer.ToByteArray(ref position);
CategoricalSplit = categoricalSplitAsBytes.Select(b => b > 0).ToArray();
if (CategoricalSplit.Any(b => b))
Copy link
Member

@codemzs codemzs Jun 15, 2018 •
edited by yaeldekel
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for fixing this. We should add a test for this function. I believe this function is not used when saving the tree model to disk and reading it back in TrainTest or CV, hence it was not caught during testing. #Resolved

{
CategoricalSplitFeatures = new int[NumNodes][];
CategoricalSplitFeatureRanges = new int[NumNodes][];
foreach (var index in categoricalNodeIndices)
for (int index = 0; index < NumNodes; index++)
{
Contracts.Assert(CategoricalSplit[index]);

CategoricalSplitFeatures[index] = buffer.ToIntArray(ref position);
CategoricalSplitFeatureRanges[index] = buffer.ToIntArray(ref position, 2);
CategoricalSplitFeatureRanges[index] = buffer.ToIntArray(ref position);
}
}

Thresholds = buffer.ToUIntArray(ref position);
RawThresholds = buffer.ToFloatArray(ref position);
_splitGain = buffer.ToDoubleArray(ref position);
_gainPValue = buffer.ToDoubleArray(ref position);
_previousLeafValue = buffer.ToDoubleArray(ref position);
Expand All @@ -144,6 +143,23 @@ private bool[] GetCategoricalSplitFromIndices(int[] indices)
return categoricalSplit;
}

private bool[] GetCategoricalSplitFromBytes(byte[] indices)
{
bool[] categoricalSplit = new bool[NumNodes];
if (indices == null)
return categoricalSplit;

Contracts.Assert(indices.Length <= NumNodes);

foreach (int index in indices)
{
Contracts.Assert(index >= 0 && index < NumNodes);
categoricalSplit[index] = true;
}

return categoricalSplit;
}

/// <summary>
/// Create a Regression Tree object from raw tree contents.
/// </summary>
Expand Down Expand Up @@ -192,7 +208,7 @@ internal RegressionTree(int[] splitFeatures, Double[] splitGain, Double[] gainPV
LeafValues = leafValues;
CategoricalSplitFeatures = categoricalSplitFeatures;
CategoricalSplitFeatureRanges = new int[CategoricalSplitFeatures.Length][];
for(int i= 0; i < CategoricalSplitFeatures.Length; ++i)
for (int i = 0; i < CategoricalSplitFeatures.Length; ++i)
{
if (CategoricalSplitFeatures[i] != null && CategoricalSplitFeatures[i].Length > 0)
{
Expand Down Expand Up @@ -500,6 +516,7 @@ public virtual int SizeInBytes()
NumNodes * sizeof(int) +
CategoricalSplit.Length * sizeof(bool) +
Thresholds.SizeInBytes() +
RawThresholds.SizeInBytes() +
_splitGain.SizeInBytes() +
_gainPValue.SizeInBytes() +
_previousLeafValue.SizeInBytes() +
Expand All @@ -514,22 +531,22 @@ public virtual void ToByteArray(byte[] buffer, ref int position)
LteChild.ToByteArray(buffer, ref position);
GtChild.ToByteArray(buffer, ref position);
SplitFeatures.ToByteArray(buffer, ref position);
CategoricalSplit.Length.ToByteArray(buffer, ref position);
foreach (var split in CategoricalSplit)
Convert.ToByte(split).ToByteArray(buffer, ref position);

if (CategoricalSplitFeatures != null)
{
foreach (var splits in CategoricalSplitFeatures)
splits.ToByteArray(buffer, ref position);
}

if (CategoricalSplitFeatureRanges != null)
{
foreach (var ranges in CategoricalSplitFeatureRanges)
ranges.ToByteArray(buffer, ref position);
Contracts.AssertValue(CategoricalSplitFeatureRanges);
for (int i = 0; i < CategoricalSplitFeatures.Length; i++)
{
CategoricalSplitFeatures[i].ToByteArray(buffer, ref position);
CategoricalSplitFeatureRanges[i].ToByteArray(buffer, ref position);
}
}

Thresholds.ToByteArray(buffer, ref position);
RawThresholds.ToByteArray(buffer, ref position);
_splitGain.ToByteArray(buffer, ref position);
_gainPValue.ToByteArray(buffer, ref position);
_previousLeafValue.ToByteArray(buffer, ref position);
Expand Down
114 changes: 114 additions & 0 deletions src/Microsoft.ML.FastTree/TreeEnsemble/TreeEnsembleCombiner.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,114 @@
// Licensed to the .NET Foundation under one or more agreements.
// The .NET Foundation licenses this file to you under the MIT license.
// See the LICENSE file in the project root for more information.

using System.Collections.Generic;
using Microsoft.ML.Runtime;
using Microsoft.ML.Runtime.FastTree.Internal;
using Microsoft.ML.Runtime.Internal.Calibration;

[assembly: LoadableClass(typeof(TreeEnsembleCombiner), null, typeof(SignatureModelCombiner), "Fast Tree Model Combiner", "FastTreeCombiner")]

namespace Microsoft.ML.Runtime.FastTree.Internal
{
public sealed class TreeEnsembleCombiner : IModelCombiner<IPredictorProducing<float>, IPredictorProducing<float>>
{
private readonly IHost _host;
private readonly PredictionKind _kind;

public TreeEnsembleCombiner(IHostEnvironment env, PredictionKind kind)
{
_host = env.Register("TreeEnsembleCombiner");
switch (kind)
{
case PredictionKind.BinaryClassification:
case PredictionKind.Regression:
case PredictionKind.Ranking:
_kind = kind;
break;
default:
throw _host.ExceptUserArg(nameof(kind), "Tree ensembles can be either binary classifiers, regressors or rankers");
Copy link
Contributor

@TomFinley TomFinley Jun 18, 2018 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

binary classifiers, regressors or rankers [](start = 90, length = 41)

Would it be better for this to be an interpolated $ style string with {nameof(PredictionKind.BinaryClassificiation)}, and so on? Not sure. #Closed

}
}

public IPredictorProducing<float> CombineModels(IEnumerable<IPredictorProducing<float>> models)
{
_host.CheckValue(models, nameof(models));

var ensemble = new Ensemble();
int modelCount = 0;
int featureCount = -1;
bool binaryClassifier = false;
foreach (var model in models)
{
modelCount++;

var predictor = model;
_host.CheckValue(predictor, nameof(models), "One of the models is null");

var calibrated = predictor as CalibratedPredictorBase;
double paramA = 1;
if (calibrated != null)
{
_host.Check(calibrated.Calibrator is PlattCalibrator,
"Combining FastTree models can only be done when the models are calibrated with Platt calibrator");
predictor = calibrated.SubPredictor;
paramA = -(calibrated.Calibrator as PlattCalibrator).ParamA;
}
var tree = predictor as FastTreePredictionWrapper;
if (tree == null)
throw _host.Except("Model is not a tree ensemble");
foreach (var t in tree.TrainedEnsemble.Trees)
{
var bytes = new byte[t.SizeInBytes()];
int position = -1;
t.ToByteArray(bytes, ref position);
position = -1;
var tNew = new RegressionTree(bytes, ref position);
if (paramA != 1)
{
for (int i = 0; i < tNew.NumLeaves; i++)
tNew.SetOutput(i, tNew.LeafValues[i] * paramA);
}
ensemble.AddTree(tNew);
}

if (modelCount == 1)
{
binaryClassifier = calibrated != null;
featureCount = tree.InputType.ValueCount;
}
else
{
_host.Check((calibrated != null) == binaryClassifier, "Ensemble contains both calibrated and uncalibrated models");
_host.Check(featureCount == tree.InputType.ValueCount, "Found models with different number of features");
}
}

var scale = 1 / (double)modelCount;

foreach (var t in ensemble.Trees)
{
for (int i = 0; i < t.NumLeaves; i++)
t.SetOutput(i, t.LeafValues[i] * scale);
}

switch (_kind)
{
case PredictionKind.BinaryClassification:
if (!binaryClassifier)
return new FastTreeBinaryPredictor(_host, ensemble, featureCount, null);

var cali = new PlattCalibrator(_host, -1, 0);
return new FeatureWeightsCalibratedPredictor(_host, new FastTreeBinaryPredictor(_host, ensemble, featureCount, null), cali);
case PredictionKind.Regression:
return new FastTreeRegressionPredictor(_host, ensemble, featureCount, null);
case PredictionKind.Ranking:
return new FastTreeRankingPredictor(_host, ensemble, featureCount, null);
default:
_host.Assert(false);
throw _host.ExceptNotSupp("PredictionKind can only be binary classification, regression or ranking");
}
Copy link
Contributor

@TomFinley TomFinley Jun 18, 2018 •
edited by yaeldekel
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Since it is impossible to reach this state anyway due to the check in the constructor, perhaps this ought to just be a throw without the message. #Resolved

}
}
}
32 changes: 20 additions & 12 deletions src/Microsoft.ML.FastTree/Utils/ToByteArrayExtensions.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
using System;
using System.Linq;
using System.Text;
using Microsoft.ML.Runtime.Internal.Utilities;

namespace Microsoft.ML.Runtime.FastTree.Internal
{
Expand Down Expand Up @@ -290,7 +291,7 @@ public static string ToString(this byte[] buffer, ref int position)

public static int SizeInBytes(this byte[] a)
{
return sizeof(int) + a.Length * sizeof(byte);
return sizeof(int) + Utils.Size(a) * sizeof(byte);
}

public static void ToByteArray(this byte[] a, byte[] buffer, ref int position)
Expand All @@ -314,7 +315,7 @@ public static byte[] ToByteArray(this byte[] buffer, ref int position)

public static int SizeInBytes(this short[] a)
{
return sizeof(int) + a.Length * sizeof(short);
return sizeof(int) + Utils.Size(a) * sizeof(short);
}

public unsafe static void ToByteArray(this short[] a, byte[] buffer, ref int position)
Expand Down Expand Up @@ -353,7 +354,7 @@ public unsafe static short[] ToShortArray(this byte[] buffer, ref int position)

public static int SizeInBytes(this ushort[] a)
{
return sizeof(int) + a.Length * sizeof(ushort);
return sizeof(int) + Utils.Size(a) * sizeof(ushort);
}

public unsafe static void ToByteArray(this ushort[] a, byte[] buffer, ref int position)
Expand Down Expand Up @@ -392,12 +393,12 @@ public unsafe static ushort[] ToUShortArray(this byte[] buffer, ref int position

public static int SizeInBytes(this int[] array)
{
return sizeof(int) + array.Length * sizeof(int);
return sizeof(int) + Utils.Size(array) * sizeof(int);
}

public unsafe static void ToByteArray(this int[] a, byte[] buffer, ref int position)
{
int length = a.Length;
int length = Utils.Size(a);
length.ToByteArray(buffer, ref position);

fixed (byte* tmpBuffer = buffer)
Expand All @@ -415,6 +416,9 @@ public unsafe static int[] ToIntArray(this byte[] buffer, ref int position)

public unsafe static int[] ToIntArray(this byte[] buffer, ref int position, int length)
{
if (length == 0)
return null;

int[] a = new int[length];

fixed (byte* tmpBuffer = buffer)
Expand All @@ -433,7 +437,7 @@ public unsafe static int[] ToIntArray(this byte[] buffer, ref int position, int

public static int SizeInBytes(this uint[] array)
{
return sizeof(int) + array.Length * sizeof(uint);
return sizeof(int) + Utils.Size(array) * sizeof(uint);
}

public unsafe static void ToByteArray(this uint[] a, byte[] buffer, ref int position)
Expand Down Expand Up @@ -472,7 +476,7 @@ public unsafe static uint[] ToUIntArray(this byte[] buffer, ref int position)

public static int SizeInBytes(this long[] array)
{
return sizeof(int) + array.Length * sizeof(long);
return sizeof(int) + Utils.Size(array) * sizeof(long);
}

public unsafe static void ToByteArray(this long[] a, byte[] buffer, ref int position)
Expand Down Expand Up @@ -511,7 +515,7 @@ public unsafe static long[] ToLongArray(this byte[] buffer, ref int position)

public static int SizeInBytes(this ulong[] array)
{
return sizeof(int) + array.Length * sizeof(ulong);
return sizeof(int) + Utils.Size(array) * sizeof(ulong);
}

public unsafe static void ToByteArray(this ulong[] a, byte[] buffer, ref int position)
Expand Down Expand Up @@ -550,7 +554,7 @@ public unsafe static ulong[] ToULongArray(this byte[] buffer, ref int position)

public static int SizeInBytes(this MD5Hash[] array)
{
return sizeof(int) + array.Length * MD5Hash.SizeInBytes();
return sizeof(int) + Utils.Size(array) * MD5Hash.SizeInBytes();
}

public static void ToByteArray(this MD5Hash[] a, byte[] buffer, ref int position)
Expand All @@ -577,7 +581,7 @@ public unsafe static MD5Hash[] ToUInt128Array(this byte[] buffer, ref int positi

public static int SizeInBytes(this float[] array)
{
return sizeof(int) + array.Length * sizeof(float);
return sizeof(int) + Utils.Size(array) * sizeof(float);
}

public unsafe static void ToByteArray(this float[] a, byte[] buffer, ref int position)
Expand Down Expand Up @@ -616,7 +620,7 @@ public unsafe static float[] ToFloatArray(this byte[] buffer, ref int position)

public static int SizeInBytes(this double[] array)
{
return sizeof(int) + array.Length * sizeof(double);
return sizeof(int) + Utils.Size(array) * sizeof(double);
}

public unsafe static void ToByteArray(this double[] a, byte[] buffer, ref int position)
Expand Down Expand Up @@ -655,6 +659,8 @@ public unsafe static double[] ToDoubleArray(this byte[] buffer, ref int position

public static int SizeInBytes(this double[][] array)
{
if (Utils.Size(array) == 0)
return sizeof(int);
return sizeof(int) + array.Sum(x => x.SizeInBytes());
}

Expand Down Expand Up @@ -683,10 +689,12 @@ public static double[][] ToDoubleJaggedArray(this byte[] buffer, ref int positio
public static long SizeInBytes(this string[] array)
{
long length = sizeof(int);
for (int i = 0; i < array.Length; ++i)
for (int i = 0; i < Utils.Size(array); ++i)
{
length += array[i].SizeInBytes();
}
if (length > 0)
Copy link
Contributor

@TomFinley TomFinley Jun 18, 2018 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

if (length > 0) [](start = 12, length = 15)

Is this intentional? long length = sizeof(int) means that it will be initialized to 8, so it should always be positive. Was this check meant to be on Utils.Size(array) instead? #Closed

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actually I'm not sure what my intention was. If the length of array is 0, then the size returned should be sizeof(int) since the length of the array is always serialized before the array itself.

In reply to: 196203608 [](ancestors = 196203608)

length += sizeof(int);
return length;
}

Expand Down
Loading