|
// 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;
using System.Linq;
using System.Text;
using Microsoft.ML;
using Microsoft.ML.Data;
using Microsoft.ML.EntryPoints;
using Microsoft.ML.Internal.Utilities;
using Microsoft.ML.Model;
using Microsoft.ML.Model.OnnxConverter;
using Microsoft.ML.Runtime;
using Microsoft.ML.Trainers.FastTree;
[assembly: LoadableClass(FastTreeTweedieTrainer.Summary, typeof(FastTreeTweedieTrainer), typeof(FastTreeTweedieTrainer.Options),
new[] { typeof(SignatureRegressorTrainer), typeof(SignatureTrainer), typeof(SignatureTreeEnsembleTrainer), typeof(SignatureFeatureScorerTrainer) },
FastTreeTweedieTrainer.UserNameValue,
FastTreeTweedieTrainer.LoadNameValue,
FastTreeTweedieTrainer.ShortName)]
[assembly: LoadableClass(typeof(FastTreeTweedieModelParameters), null, typeof(SignatureLoadModel),
"FastTree Tweedie Regression Executor",
FastTreeTweedieModelParameters.LoaderSignature)]
namespace Microsoft.ML.Trainers.FastTree
{
/// <summary>
/// The <see cref="IEstimator{TTransformer}"/> for training a decision tree regression model using Tweedie loss function.
/// This trainer is a generalization of Poisson, compound Poisson, and gamma regression.
/// </summary>
/// <remarks>
/// <format type="text/markdown"><)
/// or [FastTreeTweedie(Options)](xref:Microsoft.ML.TreeExtensions.FastTreeTweedie(Microsoft.ML.RegressionCatalog.RegressionTrainers,Microsoft.ML.Trainers.FastTree.FastTreeTweedieTrainer.Options)).
///
/// [!include[io](~/../docs/samples/docs/api-reference/io-columns-regression.md)]
///
/// ### Trainer Characteristics
/// | | |
/// | -- | -- |
/// | Machine learning task | Regression |
/// | Is normalization required? | No |
/// | Is caching required? | No |
/// | Required NuGet in addition to Microsoft.ML | Microsoft.ML.FastTree |
/// | Exportable to ONNX | Yes |
///
/// ### Training Algorithm Details
/// The Tweedie boosting model follows the mathematics established in
/// [Insurance Premium Prediction via Gradient Tree-Boosted Tweedie Compound Poisson Models</a> from Yang, Quan, and Zou](https://arxiv.org/pdf/1508.06378.pdf).
/// For an introduction to Gradient Boosting, and more information, see:
/// [Wikipedia: Gradient boosting(Gradient tree boosting)](https://en.wikipedia.org/wiki/Gradient_boosting#Gradient_tree_boosting) or
/// [Greedy function approximation: A gradient boosting machine](https://projecteuclid.org/DPubS?service=UI&version=1.0&verb=Display&handle=euclid.aos/1013203451).
///
/// Check the See Also section for links to usage examples.
/// ]]>
/// </format>
/// </remarks>
/// <seealso cref="TreeExtensions.FastTreeTweedie(RegressionCatalog.RegressionTrainers, string, string, string, int, int, int, double)"/>
/// <seealso cref="TreeExtensions.FastTreeTweedie(RegressionCatalog.RegressionTrainers, FastTreeTweedieTrainer.Options)"/>
/// <seealso cref="Options"/>
public sealed partial class FastTreeTweedieTrainer
: BoostingFastTreeTrainerBase<FastTreeTweedieTrainer.Options, RegressionPredictionTransformer<FastTreeTweedieModelParameters>, FastTreeTweedieModelParameters>
{
internal const string LoadNameValue = "FastTreeTweedieRegression";
internal const string UserNameValue = "FastTree (Boosted Trees) Tweedie Regression";
internal const string Summary = "Trains gradient boosted decision trees to fit target values using a Tweedie loss function. This learner is a generalization of Poisson, compound Poisson, and gamma regression.";
internal const string ShortName = "fttweedie";
private TestHistory _firstTestSetHistory;
private Test _trainRegressionTest;
private Test _testRegressionTest;
private protected override PredictionKind PredictionKind => PredictionKind.Regression;
private SchemaShape.Column[] _outputColumns;
/// <summary>
/// Initializes a new instance of <see cref="FastTreeTweedieTrainer"/>
/// </summary>
/// <param name="env">The private instance of <see cref="IHostEnvironment"/>.</param>
/// <param name="labelColumnName">The name of the label column.</param>
/// <param name="featureColumnName">The name of the feature column.</param>
/// <param name="exampleWeightColumnName">The name for the column containing the example weight.</param>
/// <param name="learningRate">The learning rate.</param>
/// <param name="minimumExampleCountPerLeaf">The minimal number of documents allowed in a leaf of a regression tree, out of the subsampled data.</param>
/// <param name="numberOfLeaves">The max number of leaves in each regression tree.</param>
/// <param name="numberOfTrees">Total number of decision trees to create in the ensemble.</param>
internal FastTreeTweedieTrainer(IHostEnvironment env,
string labelColumnName = DefaultColumnNames.Label,
string featureColumnName = DefaultColumnNames.Features,
string exampleWeightColumnName = null,
int numberOfLeaves = Defaults.NumberOfLeaves,
int numberOfTrees = Defaults.NumberOfTrees,
int minimumExampleCountPerLeaf = Defaults.MinimumExampleCountPerLeaf,
double learningRate = Defaults.LearningRate)
: base(env, TrainerUtils.MakeR4ScalarColumn(labelColumnName), featureColumnName, exampleWeightColumnName, null, numberOfLeaves, numberOfTrees, minimumExampleCountPerLeaf, learningRate)
{
Host.CheckNonEmpty(labelColumnName, nameof(labelColumnName));
Host.CheckNonEmpty(featureColumnName, nameof(featureColumnName));
Initialize();
}
/// <summary>
/// Initializes a new instance of <see cref="FastTreeTweedieTrainer"/> by using the <see cref="Options"/> class.
/// </summary>
/// <param name="env">The instance of <see cref="IHostEnvironment"/>.</param>
/// <param name="options">Algorithm advanced settings.</param>
internal FastTreeTweedieTrainer(IHostEnvironment env, Options options)
: base(env, options, TrainerUtils.MakeR4ScalarColumn(options.LabelColumnName))
{
Initialize();
}
private protected override FastTreeTweedieModelParameters TrainModelCore(TrainContext context)
{
Host.CheckValue(context, nameof(context));
var trainData = context.TrainingSet;
ValidData = context.ValidationSet;
TestData = context.TestSet;
using (var ch = Host.Start("Training"))
{
ch.CheckValue(trainData, nameof(trainData));
trainData.CheckRegressionLabel();
trainData.CheckFeatureFloatVector();
trainData.CheckOptFloatWeight();
FeatureCount = trainData.Schema.Feature.Value.Type.GetValueCount();
ConvertData(trainData);
TrainCore(ch);
}
return new FastTreeTweedieModelParameters(Host, TrainedEnsemble, FeatureCount, InnerOptions);
}
private protected override void CheckOptions(IChannel ch)
{
Contracts.AssertValue(ch);
base.CheckOptions(ch);
// REVIEW: In order to properly support early stopping, the early stopping metric should be a subcomponent, not just
// a simple integer, because the metric that we might want is parameterized by this floating point "index" parameter. For now
// we just leave the existing regression checks, though with a warning.
if (FastTreeTrainerOptions.EarlyStoppingMetrics > 0)
ch.Warning("For Tweedie regression, early stopping does not yet use the Tweedie distribution.");
bool doEarlyStop = FastTreeTrainerOptions.EarlyStoppingRuleFactory != null ||
FastTreeTrainerOptions.EnablePruning;
// Please do not remove it! See comment above.
if (doEarlyStop)
ch.CheckUserArg(FastTreeTrainerOptions.EarlyStoppingMetrics == 1 || FastTreeTrainerOptions.EarlyStoppingMetrics == 2,
nameof(FastTreeTrainerOptions.EarlyStoppingMetrics), "should be 1 (L1-norm) or 2 (L2-norm).");
}
private protected override ObjectiveFunctionBase ConstructObjFunc(IChannel ch)
{
return new ObjectiveImpl(TrainSet, FastTreeTrainerOptions);
}
private protected override OptimizationAlgorithm ConstructOptimizationAlgorithm(IChannel ch)
{
OptimizationAlgorithm optimizationAlgorithm = base.ConstructOptimizationAlgorithm(ch);
if (FastTreeTrainerOptions.UseLineSearch)
{
var lossCalculator = new RegressionTest(optimizationAlgorithm.TrainingScores);
// REVIEW: We should make loss indices an enum in BinaryClassificationTest.
// REVIEW: Nope, subcomponent.
optimizationAlgorithm.AdjustTreeOutputsOverride = new LineSearch(lossCalculator, 1 /*L2 error*/, FastTreeTrainerOptions.MaximumNumberOfLineSearchSteps, FastTreeTrainerOptions.MinimumStepSize);
}
return optimizationAlgorithm;
}
/// <summary>
/// Gets the regression labels that were stored in the dataset skeleton, or
/// constructs them from the ratings if absent. This returns null if the
/// dataset itself is null.
/// </summary>
/// <param name="set">The dataset</param>
/// <returns>The list of regression targets, or null if <paramref name="set"/> was null</returns>
internal static float[] GetDatasetRegressionLabels(Dataset set)
{
if (set == null)
return null;
double[] dlabels = set.Targets;
Contracts.AssertValue(dlabels);
Contracts.Assert(dlabels.Length == set.NumDocs);
// REVIEW: Seems wasteful??
return Utils.BuildArray(dlabels.Length, i => (float)dlabels[i]);
}
private protected override void PrepareLabels(IChannel ch)
{
// Intentionally empty.
}
private protected override Test ConstructTestForTrainingData()
{
return new RegressionTest(ConstructScoreTracker(TrainSet));
}
private void Initialize()
{
Host.CheckUserArg(1 <= FastTreeTrainerOptions.Index && FastTreeTrainerOptions.Index <= 2, nameof(FastTreeTrainerOptions.Index), "Must be in the range [1, 2]");
_outputColumns = new[]
{
new SchemaShape.Column(DefaultColumnNames.Score, SchemaShape.Column.VectorKind.Scalar, NumberDataViewType.Single, false)
};
}
private void AddFullRegressionTests()
{
// Always compute training L1/L2 errors.
Tests.Add(new RegressionTest(ConstructScoreTracker(TrainSet)));
RegressionTest validTest = null;
if (ValidSet != null)
{
validTest = new RegressionTest(ConstructScoreTracker(ValidSet));
Tests.Add(validTest);
}
// If external label is missing use Rating column for L1/L2 error.
// The values may not make much sense if regression value is not an actual label value.
if (TestSets != null)
{
for (int t = 0; t < TestSets.Length; ++t)
Tests.Add(new RegressionTest(ConstructScoreTracker(TestSets[t])));
}
}
private protected override void InitializeTests()
{
// Initialize regression tests.
if (FastTreeTrainerOptions.TestFrequency != int.MaxValue)
AddFullRegressionTests();
if (FastTreeTrainerOptions.PrintTestGraph)
{
// If FirstTestHistory is null (which means the tests were not initialized due to /tf==infinity),
// we need initialize first set for graph printing.
// Adding to a tests would result in printing the results after final iteration.
if (_firstTestSetHistory == null)
{
var firstTestSetTest = new RegressionTest(ConstructScoreTracker(TestSets[0]));
_firstTestSetHistory = new TestHistory(firstTestSetTest, 0);
}
}
if (FastTreeTrainerOptions.PrintTrainValidGraph && _trainRegressionTest == null)
{
Test trainRegressionTest = new RegressionTest(ConstructScoreTracker(TrainSet));
_trainRegressionTest = trainRegressionTest;
}
if (FastTreeTrainerOptions.PrintTrainValidGraph && _testRegressionTest == null && TestSets != null && TestSets.Length > 0)
_testRegressionTest = new RegressionTest(ConstructScoreTracker(TestSets[0]));
// Add early stopping if appropriate.
TrainTest = new RegressionTest(ConstructScoreTracker(TrainSet), FastTreeTrainerOptions.EarlyStoppingMetrics);
if (ValidSet != null)
ValidTest = new RegressionTest(ConstructScoreTracker(ValidSet), FastTreeTrainerOptions.EarlyStoppingMetrics);
if (FastTreeTrainerOptions.EnablePruning && ValidTest != null)
{
if (FastTreeTrainerOptions.UseTolerantPruning) // Use simple early stopping condition.
PruningTest = new TestWindowWithTolerance(ValidTest, 0, FastTreeTrainerOptions.PruningWindowSize, FastTreeTrainerOptions.PruningThreshold);
else
PruningTest = new TestHistory(ValidTest, 0);
}
}
private protected override void PrintIterationMessage(IChannel ch, IProgressChannel pch)
{
base.PrintIterationMessage(ch, pch);
}
private protected override string GetTestGraphHeader()
{
StringBuilder headerBuilder = new StringBuilder("Eval:\tFileName\tNDCG@1\tNDCG@2\tNDCG@3\tNDCG@4\tNDCG@5\tNDCG@6\tNDCG@7\tNDCG@8\tNDCG@9\tNDCG@10");
if (FastTreeTrainerOptions.PrintTrainValidGraph)
{
headerBuilder.Append("\tNDCG@20\tNDCG@40");
headerBuilder.Append("\nNote: Printing train L2 error as NDCG@20 and test L2 error as NDCG@40..\n");
}
return headerBuilder.ToString();
}
private protected override void ComputeTests()
{
if (_firstTestSetHistory != null)
_firstTestSetHistory.ComputeTests();
if (_trainRegressionTest != null)
_trainRegressionTest.ComputeTests();
if (_testRegressionTest != null)
_testRegressionTest.ComputeTests();
if (PruningTest != null)
PruningTest.ComputeTests();
}
private protected override string GetTestGraphLine()
{
StringBuilder lineBuilder = new StringBuilder();
lineBuilder.AppendFormat("Eval:\tnet.{0:D8}.ini", Ensemble.NumTrees - 1);
foreach (var r in _firstTestSetHistory.ComputeTests())
{
lineBuilder.AppendFormat("\t{0:0.0000}"{0:0.0000}", r.FinalValue);
}
double trainRegression = 0.0;
double validRegression = 0.0;
// We only print non-zero train&valid graph if earlyStoppingTruncation!=0.
// In case /es is not set, we print 0 for train and valid graph NDCG.
// Let's keeping this behavior for backward compatibility with previous FR version.
// Ideally /graphtv should enforce non-zero /es in the commandline validation.
if (_trainRegressionTest != null)
trainRegression = _trainRegressionTest.ComputeTests().Last().FinalValue;
if (_testRegressionTest != null)
validRegression = _testRegressionTest.ComputeTests().Last().FinalValue;
lineBuilder.AppendFormat("\t{0:0.0000}\t{1:0.0000}", trainRegression, validRegression);
return lineBuilder.ToString();
}
private protected override void Train(IChannel ch)
{
base.Train(ch);
// Print final last iteration.
// Note that trainNDCG printed in graph will be from copy of a value from previous iteration
// and will differ slightly from the proper final value computed by FullTest.
// We cannot compute the final NDCG here due to the fact we use FastNDCGTestForTrainSet
// computing NDCG based on label sort saved during gradient computation (and we don't have
// gradients for n+1 iteration).
// Keeping it in sync with original FR code
PrintTestGraph(ch);
}
private protected override RegressionPredictionTransformer<FastTreeTweedieModelParameters> MakeTransformer(FastTreeTweedieModelParameters model, DataViewSchema trainSchema)
=> new RegressionPredictionTransformer<FastTreeTweedieModelParameters>(Host, model, trainSchema, FeatureColumn.Name);
/// <summary>
/// Trains a <see cref="FastTreeTweedieTrainer"/> using both training and validation data, returns
/// a <see cref="RegressionPredictionTransformer{FastTreeTweedieModelParameters}"/>.
/// </summary>
public RegressionPredictionTransformer<FastTreeTweedieModelParameters> Fit(IDataView trainData, IDataView validationData)
=> TrainTransformer(trainData, validationData);
private protected override SchemaShape.Column[] GetOutputColumnsCore(SchemaShape inputSchema)
{
return new[]
{
new SchemaShape.Column(DefaultColumnNames.Score, SchemaShape.Column.VectorKind.Scalar, NumberDataViewType.Single, false, new SchemaShape(AnnotationUtils.GetTrainerOutputAnnotation()))
};
}
private sealed class ObjectiveImpl : ObjectiveFunctionBase, IStepSearch
{
private readonly float[] _labels;
private readonly Double _index1; // 1 minus the index parameter.
private readonly Double _index2; // 2 minus the index parameter.
private readonly Double _maxClamp;
public ObjectiveImpl(Dataset trainData, Options options)
: base(
trainData,
options.LearningRate,
options.Shrinkage,
options.MaximumTreeOutput,
options.GetDerivativesSampleRate,
options.BestStepRankingRegressionTrees,
options.Seed)
{
if (options.DropoutRate > 0 && LearningRate > 0) // Don't do shrinkage if dropouts are used.
Shrinkage = 1.0 / LearningRate;
_labels = GetDatasetRegressionLabels(trainData);
// Should we fail instead of clamping negative values to 0?
for (int i = 0; i < _labels.Length; ++i)
{
if (_labels[i] < 0)
_labels[i] = 0;
}
_index1 = 1 - options.Index;
_index2 = 2 - options.Index;
_maxClamp = Math.Abs(options.MaximumTreeOutput);
}
public void AdjustTreeOutputs(IChannel ch, InternalRegressionTree tree, DocumentPartitioning partitioning, ScoreTracker trainingScores)
{
double shrinkage = LearningRate * Shrinkage;
var scores = trainingScores.Scores;
var weights = trainingScores.Dataset.SampleWeights;
// Following equation 18, and line 2c of algorithm 1 in the source paper.
for (int l = 0; l < tree.NumLeaves; ++l)
{
Double num = 0;
Double denom = 0;
if (_index1 == 0)
{
// The index == 1 Poisson case.
foreach (int i in partitioning.DocumentsInLeaf(l))
{
var s = scores[i];
var w = weights == null ? 1 : weights[i];
num += w * _labels[i];
denom += w * Math.Exp(s);
}
}
else
{
// The index in (1,2] case.
foreach (int i in partitioning.DocumentsInLeaf(l))
{
var s = scores[i];
var w = weights == null ? 1 : weights[i];
num += w * _labels[i] * Math.Exp(_index1 * s);
denom += w * Math.Exp(_index2 * s);
}
}
var step = shrinkage * (Math.Log(num) - Math.Log(denom));
if (num == 0 && denom == 0)
step = 0;
// If we do not clamp, it is entirely possible for num to be 0 (with 0 labels), which
// means that we will have negative infinities in the leaf nodes. This has a number of
// bad negative effects we'd prefer to avoid. Nonetheless, we do give up a substantial
// amount of "gain" for those examples.
if (step < -_maxClamp)
step = -_maxClamp;
else if (step > _maxClamp)
step = _maxClamp;
tree.SetOutput(l, step);
}
}
protected override void GetGradientInOneQuery(int query, int threadIndex)
{
int begin = Dataset.Boundaries[query];
int end = Dataset.Boundaries[query + 1];
// Gradient.
unchecked
{
if (_index1 == 0)
{
// In the case where index=1, the Tweedie distribution is Poisson. We treat this important
// class of distribution as a special case even though the mathematics wind up being
// the same.
for (int i = begin; i < end; ++i)
{
// From "Generalized Boosted Models: A guide to the gbm package" by Greg Ridgeway, 2007, section 4.7.
Gradient[i] = Math.Exp(Scores[i]) - _labels[i];
}
}
else
{
for (int i = begin; i < end; ++i)
{
// Following equation 14, and line 2a of algorithm 1 in the source paper. The w_i
// are not incorporated here.
Gradient[i] = Math.Exp(_index2 * Scores[i]) - _labels[i] * Math.Exp(_index1 * Scores[i]);
}
}
}
}
}
}
/// <summary>
/// Model parameters for <see cref="FastTreeTweedieTrainer"/>.
/// </summary>
public sealed class FastTreeTweedieModelParameters : TreeEnsembleModelParametersBasedOnRegressionTree, ISingleCanSaveOnnx
{
internal const string LoaderSignature = "FastTreeTweedieExec";
internal const string RegistrationName = "FastTreeTweediePredictor";
private static VersionInfo GetVersionInfo()
{
return new VersionInfo(
modelSignature: "FTREE TW",
// verWrittenCur: 0x00010001, // Initial
//verWrittenCur: 0x00010002, // Add _defaultValueForMissing
verWrittenCur: 0x00010003, // Categorical splits.
verReadableCur: 0x00010002,
verWeCanReadBack: 0x00010001,
loaderSignature: LoaderSignature,
loaderAssemblyName: typeof(FastTreeTweedieModelParameters).Assembly.FullName);
}
private protected override uint VerNumFeaturesSerialized => 0x00010001;
private protected override uint VerDefaultValueSerialized => 0x00010002;
private protected override uint VerCategoricalSplitSerialized => 0x00010003;
internal FastTreeTweedieModelParameters(IHostEnvironment env, InternalTreeEnsemble trainedEnsemble, int featureCount, string innerArgs)
: base(env, RegistrationName, trainedEnsemble, featureCount, innerArgs)
{
}
private FastTreeTweedieModelParameters(IHostEnvironment env, ModelLoadContext ctx)
: base(env, RegistrationName, ctx, GetVersionInfo())
{
}
private protected override void SaveCore(ModelSaveContext ctx)
{
base.SaveCore(ctx);
ctx.SetVersionInfo(GetVersionInfo());
}
internal static FastTreeTweedieModelParameters Create(IHostEnvironment env, ModelLoadContext ctx)
{
Contracts.CheckValue(env, nameof(env));
env.CheckValue(ctx, nameof(ctx));
ctx.CheckAtModel(GetVersionInfo());
return new FastTreeTweedieModelParameters(env, ctx);
}
bool ISingleCanSaveOnnx.SaveAsOnnx(OnnxContext ctx, string[] outputNames, string featureColumn)
{
const int minimumOpSetVersion = 9;
ctx.CheckOpSetVersion(minimumOpSetVersion, LoaderSignature);
// Mapping score to prediction
var fastTreeOutput = ctx.AddIntermediateVariable(null, "FastTreeOutput", true);
base.SaveAsOnnx(ctx, new[] { fastTreeOutput }, featureColumn);
var opType = "Exp";
ctx.CreateNode(opType, new[] { fastTreeOutput }, outputNames, ctx.GetNodeName(opType), "");
return true;
}
private protected override void Map(in VBuffer<float> src, ref float dst)
{
// The value learnt and predicted by the trees is the log of the expected value,
// as seen in equation 9 of the paper. So for the actual prediction, we take its
// exponent.
base.Map(in src, ref dst);
// REVIEW: Some packages like R's GBM apparently clamp the input to the exponent
// in the range [-19, 19]. We have historically taken a dim view of this sort of thing
// ourselves, but if our views prove problematic we can reconsider. (An upper clamp of 19
// seems far too restrictive, leading to a practical upper limit of about 178 million.)
dst = MathUtils.ExpSlow(dst);
}
private protected override PredictionKind PredictionKind => PredictionKind.Regression;
}
internal static partial class FastTree
{
[TlcModule.EntryPoint(Name = "Trainers.FastTreeTweedieRegressor",
Desc = FastTreeTweedieTrainer.Summary,
UserName = FastTreeTweedieTrainer.UserNameValue,
ShortName = FastTreeTweedieTrainer.ShortName)]
public static CommonOutputs.RegressionOutput TrainTweedieRegression(IHostEnvironment env, FastTreeTweedieTrainer.Options input)
{
Contracts.CheckValue(env, nameof(env));
var host = env.Register("TrainTweeedie");
host.CheckValue(input, nameof(input));
EntryPointUtils.CheckInputArgs(host, input);
return TrainerEntryPointsUtils.Train<FastTreeTweedieTrainer.Options, CommonOutputs.RegressionOutput>(host, input,
() => new FastTreeTweedieTrainer(host, input),
() => TrainerEntryPointsUtils.FindColumn(host, input.TrainingData.Schema, input.LabelColumnName),
() => TrainerEntryPointsUtils.FindColumn(host, input.TrainingData.Schema, input.ExampleWeightColumnName),
() => TrainerEntryPointsUtils.FindColumn(host, input.TrainingData.Schema, input.RowGroupColumnName));
}
}
}
|