public void TestLDATransform()
{
var builder = new ArrayDataViewBuilder(Env);
var data = new[]
{
new[] { (Float)1.0, (Float)0.0, (Float)0.0 },
new[] { (Float)0.0, (Float)1.0, (Float)0.0 },
new[] { (Float)0.0, (Float)0.0, (Float)1.0 },
};
builder.AddColumn("F1V", NumberType.Float, data);
var srcView = builder.GetDataView();
LdaTransform.Column col = new LdaTransform.Column();
col.Source = "F1V";
col.NumTopic = 20;
col.NumTopic = 3;
col.NumSummaryTermPerTopic = 3;
col.AlphaSum = 3;
col.NumThreads = 1;
col.ResetRandomGenerator = true;
LdaTransform.Arguments args = new LdaTransform.Arguments();
args.Column = new LdaTransform.Column[] { col };
LdaTransform ldaTransform = new LdaTransform(Env, args, srcView);
using (var cursor = ldaTransform.GetRowCursor(c => true))
{
var resultGetter = cursor.GetGetter <VBuffer <Float> >(1);
VBuffer <Float> resultFirstRow = new VBuffer <Float>();
VBuffer <Float> resultSecondRow = new VBuffer <Float>();
VBuffer <Float> resultThirdRow = new VBuffer <Float>();
Assert.True(cursor.MoveNext());
resultGetter(ref resultFirstRow);
Assert.True(cursor.MoveNext());
resultGetter(ref resultSecondRow);
Assert.True(cursor.MoveNext());
resultGetter(ref resultThirdRow);
Assert.False(cursor.MoveNext());
Assert.True(resultFirstRow.Length == 3);
Assert.True(resultFirstRow.GetItemOrDefault(0) == 0);
Assert.True(resultFirstRow.GetItemOrDefault(2) == 0);
Assert.True(resultFirstRow.GetItemOrDefault(1) == 1.0);
Assert.True(resultSecondRow.Length == 3);
Assert.True(resultSecondRow.GetItemOrDefault(0) == 0);
Assert.True(resultSecondRow.GetItemOrDefault(2) == 0);
Assert.True(resultSecondRow.GetItemOrDefault(1) == 1.0);
Assert.True(resultThirdRow.Length == 3);
Assert.True(resultThirdRow.GetItemOrDefault(0) == 0);
Assert.True(resultThirdRow.GetItemOrDefault(1) == 0);
Assert.True(resultThirdRow.GetItemOrDefault(2) == 1.0);
}
}
public void RangeFilterTest()
{
var builder = new ArrayDataViewBuilder(ML);
builder.AddColumn("Strings", new[] { "foo", "bar", "baz" });
builder.AddColumn("Floats", NumberType.R4, new float[] { 1, 2, 3 });
var data = builder.GetDataView();
var data1 = ML.Data.FilterByColumn(data, "Floats", upperBound: 2.8);
var cnt = data1.GetColumn <float>(ML, "Floats").Count();
Assert.Equal(2L, cnt);
data = ML.Transforms.Conversion.Hash("Key", "Strings", hashBits: 20).Fit(data).Transform(data);
var data2 = ML.Data.FilterByKeyColumnFraction(data, "Key", upperBound: 0.5);
cnt = data2.GetColumn <float>(ML, "Floats").Count();
Assert.Equal(1L, cnt);
}
private FastForestRegressionModelParameters FitModel(IEnumerable <IRunResult> previousRuns)
{
Single[] targets = new Single[previousRuns.Count()];
Single[][] features = new Single[previousRuns.Count()][];
int i = 0;
foreach (RunResult r in previousRuns)
{
features[i] = SweeperProbabilityUtils.ParameterSetAsFloatArray(_host, _sweepParameters, r.ParameterSet, true);
targets[i] = (float)r.MetricValue;
i++;
}
ArrayDataViewBuilder dvBuilder = new ArrayDataViewBuilder(_host);
dvBuilder.AddColumn(DefaultColumnNames.Label, NumberDataViewType.Single, targets);
dvBuilder.AddColumn(DefaultColumnNames.Features, NumberDataViewType.Single, features);
IDataView view = dvBuilder.GetDataView();
_host.Assert(view.GetRowCount() == targets.Length, "This data view will have as many rows as there have been evaluations");
using (IChannel ch = _host.Start("Single training"))
{
// Set relevant random forest arguments.
// Train random forest.
var trainer = new FastForestRegressionTrainer(_host,
new FastForestRegressionTrainer.Options
{
FeatureFraction = _args.SplitRatio,
NumberOfTrees = _args.NumOfTrees,
MinimumExampleCountPerLeaf = _args.NMinForSplit,
LabelColumnName = DefaultColumnNames.Label,
FeatureColumnName = DefaultColumnNames.Features,
});
var predictor = trainer.Fit(view);
// Return random forest predictor.
return(predictor.Model);
}
}
private FastForestRegressionPredictor FitModel(IEnumerable <IRunResult> previousRuns)
{
Single[] targets = new Single[previousRuns.Count()];
Single[][] features = new Single[previousRuns.Count()][];
int i = 0;
foreach (RunResult r in previousRuns)
{
features[i] = SweeperProbabilityUtils.ParameterSetAsFloatArray(_host, _sweepParameters, r.ParameterSet, true);
targets[i] = (Float)r.MetricValue;
i++;
}
ArrayDataViewBuilder dvBuilder = new ArrayDataViewBuilder(_host);
dvBuilder.AddColumn("Label", NumberType.Float, targets);
dvBuilder.AddColumn("Features", NumberType.Float, features);
IDataView view = dvBuilder.GetDataView();
_host.Assert(view.GetRowCount() == targets.Length, "This data view will have as many rows as there have been evaluations");
RoleMappedData data = TrainUtils.CreateExamples(view, "Label", "Features");
using (IChannel ch = _host.Start("Single training"))
{
// Set relevant random forest arguments.
FastForestRegression.Arguments args = new FastForestRegression.Arguments();
args.FeatureFraction = _args.SplitRatio;
args.NumTrees = _args.NumOfTrees;
args.MinDocumentsInLeafs = _args.NMinForSplit;
// Train random forest.
FastForestRegression trainer = new FastForestRegression(_host, args);
trainer.Train(data);
FastForestRegressionPredictor predictor = trainer.CreatePredictor();
// Return random forest predictor.
ch.Done();
return(predictor);
}
}
public void TextColumnDimensionsTest()
{
var context = new MLContext();
var dataBuilder = new ArrayDataViewBuilder(context);
dataBuilder.AddColumn("categorical", new string[] { "0", "1", "0", "1", "0", "1", "2", "2", "0", "1" });
dataBuilder.AddColumn("text", new string[] { "0", "1", "0", "1", "0", "1", "2", "2", "0", "1" });
var data = dataBuilder.GetDataView();
var dimensions = DatasetDimensionsApi.CalcColumnDimensions(context, data, new[] {
new PurposeInference.Column(0, ColumnPurpose.CategoricalFeature),
new PurposeInference.Column(0, ColumnPurpose.TextFeature),
});
Assert.NotNull(dimensions);
Assert.Equal(2, dimensions.Length);
Assert.Equal(3, dimensions[0].Cardinality);
Assert.Null(dimensions[1].Cardinality);
Assert.Null(dimensions[0].HasMissing);
Assert.Null(dimensions[1].HasMissing);
}
public void FloatColumnDimensionsTest()
{
var context = new MLContext();
var dataBuilder = new ArrayDataViewBuilder(context);
dataBuilder.AddColumn("NoNan", NumberDataViewType.Single, new float[] { 0, 1, 0, 1, 0 });
dataBuilder.AddColumn("Nan", NumberDataViewType.Single, new float[] { 0, 1, 0, 1, float.NaN });
var data = dataBuilder.GetDataView();
var dimensions = DatasetDimensionsApi.CalcColumnDimensions(context, data, new[] {
new PurposeInference.Column(0, ColumnPurpose.NumericFeature),
new PurposeInference.Column(1, ColumnPurpose.NumericFeature),
});
Assert.NotNull(dimensions);
Assert.Equal(2, dimensions.Length);
Assert.Null(dimensions[0].Cardinality);
Assert.Null(dimensions[1].Cardinality);
Assert.False(dimensions[0].HasMissing);
Assert.True(dimensions[1].HasMissing);
}
public IDataView GetSummaryDataView(RoleMappedSchema schema)
{
var bldr = new ArrayDataViewBuilder(Host);
var cols = new VBuffer <Float> [_rank + 1];
var names = new string[_rank + 1];
for (var i = 0; i < _rank; ++i)
{
names[i] = "EigenVector" + i;
cols[i] = _eigenVectors[i];
}
names[_rank] = "MeanVector";
cols[_rank] = _mean;
bldr.AddColumn("VectorName", names);
bldr.AddColumn("VectorData", NumberType.R4, cols);
return(bldr.GetDataView());
}
public void ValidateEmptyValidationDataThrows()
{
// Training data
var dataViewBuilder = new ArrayDataViewBuilder(new MLContext());
dataViewBuilder.AddColumn("Number", NumberDataViewType.Single, 0f);
dataViewBuilder.AddColumn(DefaultColumnNames.Label, NumberDataViewType.Single, 0f);
var trainingData = dataViewBuilder.GetDataView();
// Validation data
var schemaBuilder = new DataViewSchema.Builder();
schemaBuilder.AddColumn("Number", NumberDataViewType.Single);
schemaBuilder.AddColumn(DefaultColumnNames.Label, NumberDataViewType.Single);
var schema = schemaBuilder.ToSchema();
var validationData = DataViewTestFixture.BuildDummyDataView(schema, createDummyRow: false);
var ex = Assert.Throws <ArgumentException>(() => UserInputValidationUtil.ValidateExperimentExecuteArgs(trainingData, new ColumnInformation(),
validationData, TaskKind.Regression));
Assert.StartsWith("Validation data has 0 rows", ex.Message);
}
public void Setup()
{
var ctx = new MLContext(1);
var builder = new ArrayDataViewBuilder(ctx);
int[] values = new int[Length];
for (int i = 0; i < values.Length; ++i)
{
values[i] = i;
}
builder.AddColumn("A", NumberDataViewType.Int32, values);
var dv = builder.GetDataView();
var cacheDv = ctx.Data.Cache(dv);
var col = cacheDv.Schema.GetColumnOrNull("A").Value;
// First do one pass through.
using (var cursor = cacheDv.GetRowCursor(col))
{
var getter = cursor.GetGetter <int>(col);
int val = 0;
int count = 0;
while (cursor.MoveNext())
{
getter(ref val);
if (val != cursor.Position)
{
throw new Exception($"Unexpected value {val} at {cursor.Position}");
}
count++;
}
if (count != Length)
{
throw new Exception($"Expected {Length} values in cache but only saw {count}");
}
}
_cacheDataView = cacheDv;
// Only needed for seeker, but may as well set it.
_positions = new long[Length];
var rand = new Random(0);
for (int i = 0; i < _positions.Length; ++i)
{
_positions[i] = rand.Next(Length);
}
_col = _cacheDataView.Schema["A"];
_seeker = ((IRowSeekable)_cacheDataView).GetSeeker(colIndex => colIndex == _col.Index);
_seekerGetter = _seeker.GetGetter <int>(_col);
}
/// <summary>
/// Save schema associations of role/column-name in <paramref name="rep"/>.
/// </summary>
internal static void SaveRoleMappings(IHostEnvironment env, IChannel ch, RoleMappedSchema schema, RepositoryWriter rep)
{
// REVIEW: Should we also save this stuff, for instance, in some portion of the
// score command or transform?
Contracts.AssertValue(env);
env.AssertValue(ch);
ch.AssertValue(schema);
ArrayDataViewBuilder builder = new ArrayDataViewBuilder(env);
List <string> rolesList = new List <string>();
List <string> columnNamesList = new List <string>();
// OrderBy is stable, so there is no danger in it "reordering" columns
// when a role is filled by multiple columns.
foreach (var role in schema.GetColumnRoleNames().OrderBy(r => r.Key.Value))
{
rolesList.Add(role.Key.Value);
columnNamesList.Add(role.Value);
}
builder.AddColumn("Role", rolesList.ToArray());
builder.AddColumn("Column", columnNamesList.ToArray());
using (var entry = rep.CreateEntry(DirTrainingInfo, RoleMappingFile))
{
// REVIEW: It seems very important that we have the role mappings
// be easily human interpretable and even manipulable, but relying on the
// text saver/loader means that special characters like '\n' won't be reinterpretable.
// On the other hand, no one is such a big lunatic that they will actually
// ever go ahead and do something so stupid as that.
var saver = new TextSaver(env, new TextSaver.Arguments()
{
Dense = true, Silent = true
});
var view = builder.GetDataView();
saver.SaveData(entry.Stream, view, Utils.GetIdentityPermutation(view.Schema.ColumnCount));
}
}
public void ValidateExperimentExecuteArgsTrainValidColCountMismatch()
{
var context = new MLContext();
var trainDataBuilder = new ArrayDataViewBuilder(context);
trainDataBuilder.AddColumn("0", NumberDataViewType.Single, new float[] { 1 });
trainDataBuilder.AddColumn("1", new string[] { "1" });
var trainData = trainDataBuilder.GetDataView();
var validDataBuilder = new ArrayDataViewBuilder(context);
validDataBuilder.AddColumn("0", NumberDataViewType.Single, new float[] { 1 });
var validData = validDataBuilder.GetDataView();
var ex = Assert.Throws <ArgumentException>(() => UserInputValidationUtil.ValidateExperimentExecuteArgs(trainData,
new ColumnInformation()
{
LabelColumnName = "0"
}, validData, TaskKind.Regression));
Assert.StartsWith("Training data and validation data schemas do not match. Train data has '2' columns,and validation data has '1' columns.", ex.Message);
}
private FastForestRegressionModelParameters FitModel(IEnumerable <IRunResult> previousRuns)
{
Single[] targets = new Single[previousRuns.Count()];
Single[][] features = new Single[previousRuns.Count()][];
int i = 0;
foreach (RunResult r in previousRuns)
{
features[i] = SweeperProbabilityUtils.ParameterSetAsFloatArray(_sweepParameters, r.ParameterSet, true);
targets[i] = (Float)r.MetricValue;
i++;
}
ArrayDataViewBuilder dvBuilder = new ArrayDataViewBuilder(_context);
dvBuilder.AddColumn(DefaultColumnNames.Label, NumberType.Float, targets);
dvBuilder.AddColumn(DefaultColumnNames.Features, NumberType.Float, features);
IDataView data = dvBuilder.GetDataView();
AutoMlUtils.Assert(data.GetRowCount() == targets.Length, "This data view will have as many rows as there have been evaluations");
// Set relevant random forest arguments.
// Train random forest.
var trainer = new FastForestRegression(_context, DefaultColumnNames.Label, DefaultColumnNames.Features, advancedSettings: s =>
{
s.FeatureFraction = _args.SplitRatio;
s.NumTrees = _args.NumOfTrees;
s.MinDocumentsInLeafs = _args.NMinForSplit;
});
var predictor = trainer.Train(data).Model;
// Return random forest predictor.
return(predictor);
}
public void LdaWorkoutEstimatorCore()
{
var ml = new MLContext();
var builder = new ArrayDataViewBuilder(Env);
var data = new[]
{
new[] { (float)1.0, (float)0.0, (float)0.0 },
new[] { (float)0.0, (float)1.0, (float)0.0 },
new[] { (float)0.0, (float)0.0, (float)1.0 },
};
builder.AddColumn("F1V", NumberDataViewType.Single, data);
var srcView = builder.GetDataView();
var est = ml.Transforms.Text.LatentDirichletAllocation("F1V");
TestEstimatorCore(est, srcView);
}
public void TransposerSaverLoaderTest()
{
const int rowCount = 1000;
Random rgen = new Random(1);
ArrayDataViewBuilder builder = new ArrayDataViewBuilder(Env);
// A is to check the splitting of a sparse-ish column.
var dataA = GenerateHelper(rowCount, 0.1, rgen, () => (int)rgen.Next(), 50, 5, 10, 15);
dataA[rowCount / 2] = new VBuffer <int>(50, 0, null, null); // Coverage for the null vbuffer case.
builder.AddColumn("A", NumberDataViewType.Int32, dataA);
// B is to check the splitting of a dense-ish column.
builder.AddColumn("B", NumberDataViewType.Double, GenerateHelper(rowCount, 0.8, rgen, rgen.NextDouble, 50, 0, 25, 49));
// C is to just have some column we do nothing with.
builder.AddColumn("C", NumberDataViewType.Int16, GenerateHelper(rowCount, 0.1, rgen, () => (short)1, 30, 3, 10, 24));
// D is to check some column we don't have to split because it's sufficiently small.
builder.AddColumn("D", NumberDataViewType.Double, GenerateHelper(rowCount, 0.1, rgen, rgen.NextDouble, 3, 1));
// E is to check a sparse scalar column.
builder.AddColumn("E", NumberDataViewType.UInt32, GenerateHelper(rowCount, 0.1, rgen, () => (uint)rgen.Next(int.MinValue, int.MaxValue)));
// F is to check a dense-ish scalar column.
builder.AddColumn("F", NumberDataViewType.Int32, GenerateHelper(rowCount, 0.8, rgen, () => (int)rgen.Next()));
IDataView view = builder.GetDataView();
IMultiStreamSource src;
using (MemoryStream mem = new MemoryStream())
{
TransposeSaver saver = new TransposeSaver(Env, new TransposeSaver.Arguments());
saver.SaveData(mem, view, Utils.GetIdentityPermutation(view.Schema.Count));
src = new BytesStreamSource(mem.ToArray());
}
TransposeLoader loader = new TransposeLoader(Env, new TransposeLoader.Arguments(), src);
// First check whether this as an IDataView yields the same values.
CheckSameValues(view, loader);
TransposeCheckHelper <int>(view, 0, loader); // A
TransposeCheckHelper <Double>(view, 1, loader); // B
TransposeCheckHelper <short>(view, 2, loader); // C
TransposeCheckHelper <Double>(view, 3, loader); // D
TransposeCheckHelper <uint>(view, 4, loader); // E
TransposeCheckHelper <int>(view, 5, loader); // F
Done();
}
public static Output ExtractSweepResult(IHostEnvironment env, ResultInput input)
{
var autoMlState = input.State as AutoInference.AutoMlMlState;
if (autoMlState == null)
{
throw env.Except("The state must be a valid AutoMlState.");
}
// Create results output dataview
var rows = autoMlState.GetAllEvaluatedPipelines().Select(p => p.ToResultRow()).ToList();
IDataView outputView;
var col1 = new KeyValuePair <string, ColumnType>("Graph", TextType.Instance);
var col2 = new KeyValuePair <string, ColumnType>("MetricValue", PrimitiveType.FromKind(DataKind.R8));
var col3 = new KeyValuePair <string, ColumnType>("PipelineId", TextType.Instance);
var col4 = new KeyValuePair <string, ColumnType>("TrainingMetricValue", PrimitiveType.FromKind(DataKind.R8));
var col5 = new KeyValuePair <string, ColumnType>("FirstInput", TextType.Instance);
var col6 = new KeyValuePair <string, ColumnType>("PredictorModel", TextType.Instance);
if (rows.Count == 0)
{
var host = env.Register("ExtractSweepResult");
outputView = new EmptyDataView(host, SimpleSchemaUtils.Create(host, col1, col2, col3, col4, col5, col6));
}
else
{
var builder = new ArrayDataViewBuilder(env);
builder.AddColumn(col1.Key, (PrimitiveType)col1.Value, rows.Select(r => r.GraphJson.AsMemory()).ToArray());
builder.AddColumn(col2.Key, (PrimitiveType)col2.Value, rows.Select(r => r.MetricValue).ToArray());
builder.AddColumn(col3.Key, (PrimitiveType)col3.Value, rows.Select(r => r.PipelineId.AsMemory()).ToArray());
builder.AddColumn(col4.Key, (PrimitiveType)col4.Value, rows.Select(r => r.TrainingMetricValue).ToArray());
builder.AddColumn(col5.Key, (PrimitiveType)col5.Value, rows.Select(r => r.FirstInput.AsMemory()).ToArray());
builder.AddColumn(col6.Key, (PrimitiveType)col6.Value, rows.Select(r => r.PredictorModel.AsMemory()).ToArray());
outputView = builder.GetDataView();
}
return(new Output {
Results = outputView, State = autoMlState
});
}
/// <summary>
/// Features: x1, x2vBuff(sparce vector), x3.
/// y = 10x1 + 10x2vBuff + 30x3 + e.
/// Within xBuff feature 2nd slot will be sparse most of the time.
/// 2nd slot of xBuff has the least importance: Evaluation metrics do not change a lot when this slot is permuted.
/// x3 has the biggest importance.
/// </summary>
private IDataView GetSparseDataset(TaskType task = TaskType.Regression, int numberOfInstances = 1000)
{
// Setup synthetic dataset.
var rand = new Random(10);
float[] yArray = new float[numberOfInstances],
x1Array = new float[numberOfInstances],
x3Array = new float[numberOfInstances];
VBuffer <float>[] vbArray = new VBuffer <float> [numberOfInstances];
for (var i = 0; i < numberOfInstances; i++)
{
var x1 = rand.Next(1000);
x1Array[i] = x1;
var x3Important = rand.Next(10000);
x3Array[i] = x3Important;
VBuffer <float> vb;
if (i % 10 != 0)
{
vb = new VBuffer <float>(4, 3, new float[] { rand.Next(1000), rand.Next(1000), rand.Next(1000) }, new int[] { 0, 2, 3 });
}
else
{
vb = new VBuffer <float>(4, 4, new float[] { rand.Next(1000), rand.Next(1000), rand.Next(1000), rand.Next(1000) }, new int[] { 0, 1, 2, 3 });
}
vbArray[i] = vb;
float vbSum = 0;
foreach (var vbValue in vb.DenseValues())
{
vbSum += vbValue * 10;
}
var noise = rand.Next(50);
yArray[i] = 10 * x1 + vbSum + 20 * x3Important + noise;
}
// If binary classification, modify the labels
if (task == TaskType.BinaryClassification ||
task == TaskType.MulticlassClassification)
{
GetBinaryClassificationLabels(yArray);
}
else if (task == TaskType.Ranking)
{
GetRankingLabels(yArray);
}
// Create data view.
var bldr = new ArrayDataViewBuilder(Env);
bldr.AddColumn("X1", NumberType.Float, x1Array);
bldr.AddColumn("X2VBuffer", NumberType.Float, vbArray);
bldr.AddColumn("X3Important", NumberType.Float, x3Array);
bldr.AddColumn("Label", NumberType.Float, yArray);
if (task == TaskType.Ranking)
{
bldr.AddColumn("GroupId", NumberType.U4, CreateGroupIds(yArray.Length));
}
var srcDV = bldr.GetDataView();
var pipeline = ML.Transforms.Concatenate("Features", "X1", "X2VBuffer", "X3Important")
.Append(ML.Transforms.Normalize("Features"));
// Create a keytype for Ranking
if (task == TaskType.Ranking)
{
return(pipeline.Append(ML.Transforms.Conversion.MapValueToKey("GroupId"))
.Fit(srcDV).Transform(srcDV));
}
return(pipeline.Fit(srcDV).Transform(srcDV));
}
public void TestFeatureImportance()
{
// Setup synthetic dataset.
const int numberOfInstances = 1000;
var rand = new Random(10);
float[] yArray = new float[numberOfInstances],
x1Array = new float[numberOfInstances],
x2Array = new float[numberOfInstances],
x3Array = new float[numberOfInstances],
x4RandArray = new float[numberOfInstances];
for (var i = 0; i < numberOfInstances; i++)
{
var x1 = rand.Next(1000);
x1Array[i] = x1;
var x2Important = rand.Next(10000);
x2Array[i] = x2Important;
var x3 = rand.Next(5000);
x3Array[i] = x3;
var x4Rand = rand.Next(1000);
x4RandArray[i] = x4Rand;
var noise = rand.Next(50);
yArray[i] = (float)(10 * x1 + 20 * x2Important + 5.5 * x3 + noise);
}
// Create data view.
var bldr = new ArrayDataViewBuilder(Env);
bldr.AddColumn("X1", NumberType.Float, x1Array);
bldr.AddColumn("X2Important", NumberType.Float, x2Array);
bldr.AddColumn("X3", NumberType.Float, x3Array);
bldr.AddColumn("X4Rand", NumberType.Float, x4RandArray);
bldr.AddColumn("Label", NumberType.Float, yArray);
var srcDV = bldr.GetDataView();
var pipeline = ML.Transforms.Concatenate("Features", "X1", "X2Important", "X3", "X4Rand")
.AppendCacheCheckpoint(ML)
.Append(ML.Transforms.Normalize("Features"));
var data = pipeline.Fit(srcDV).Transform(srcDV);
var model = ML.Regression.Trainers.OrdinaryLeastSquares().Fit(data);
var args = new FeatureContributionCalculationTransform.Arguments()
{
Bottom = 10,
Top = 10
};
var output = FeatureContributionCalculationTransform.Create(Env, args, data, model.Model, model.FeatureColumn);
// Get prediction scores and contributions
var enumerator = output.AsEnumerable <ScoreAndContribution>(Env, true).GetEnumerator();
ScoreAndContribution row = null;
var expectedValues = new List <float[]>();
expectedValues.Add(new float[4] {
0.06319684F, 1, 0.1386623F, 4.46209469E-06F
});
expectedValues.Add(new float[4] {
0.03841561F, 1, 0.1633037F, 2.68303256E-06F
});
expectedValues.Add(new float[4] {
0.12006103F, 1, 0.254072F, 1.18671605E-05F
});
expectedValues.Add(new float[4] {
0.20861618F, 0.99999994F, 0.407312155F, 6.963478E-05F
});
expectedValues.Add(new float[4] {
0.024050576F, 0.99999994F, 0.31106182F, 8.456762E-06F
});
int index = 0;
while (enumerator.MoveNext() && index < expectedValues.Count)
{
row = enumerator.Current;
// We set predicion to 6 because the limit of floating-point numbers is 7.
Assert.Equal(expectedValues[index][0], row.FeatureContributions[0], 6);
Assert.Equal(expectedValues[index][1], row.FeatureContributions[1], 6);
Assert.Equal(expectedValues[index][2], row.FeatureContributions[2], 6);
Assert.Equal(expectedValues[index++][3], row.FeatureContributions[3], 6);
}
Done();
}
public static CombinedOutput CombineMetrics(IHostEnvironment env, CombineMetricsInput input)
{
var eval = GetEvaluator(env, input.Kind);
var perInst = EvaluateUtils.ConcatenatePerInstanceDataViews(env, eval, true, true, input.PerInstanceMetrics.Select(
idv => RoleMappedData.CreateOpt(idv, new[]
{
RoleMappedSchema.CreatePair(RoleMappedSchema.ColumnRole.Label, input.LabelColumn),
RoleMappedSchema.CreatePair(RoleMappedSchema.ColumnRole.Weight, input.WeightColumn.Value),
RoleMappedSchema.CreatePair(RoleMappedSchema.ColumnRole.Group, input.GroupColumn.Value)
})).ToArray(),
out var variableSizeVectorColumnNames);
var warnings = input.Warnings != null ? new List <IDataView>(input.Warnings) : new List <IDataView>();
if (variableSizeVectorColumnNames.Length > 0)
{
var dvBldr = new ArrayDataViewBuilder(env);
var warn = $"Detected columns of variable length: {string.Join(", ", variableSizeVectorColumnNames)}." +
$" Consider setting collateMetrics- for meaningful per-Folds results.";
dvBldr.AddColumn(MetricKinds.ColumnNames.WarningText, TextType.Instance, new DvText(warn));
warnings.Add(dvBldr.GetDataView());
}
env.Assert(Utils.Size(perInst) == 1);
var overall = eval.GetOverallResults(input.OverallMetrics);
overall = EvaluateUtils.CombineFoldMetricsDataViews(env, overall, input.OverallMetrics.Length);
IDataView conf = null;
if (Utils.Size(input.ConfusionMatrix) > 0)
{
EvaluateUtils.ReconcileSlotNames <double>(env, input.ConfusionMatrix, MetricKinds.ColumnNames.Count, NumberType.R8);
for (int i = 0; i < input.ConfusionMatrix.Length; i++)
{
var idv = input.ConfusionMatrix[i];
// Find the old Count column and drop it.
for (int col = 0; col < idv.Schema.ColumnCount; col++)
{
if (idv.Schema.IsHidden(col) &&
idv.Schema.GetColumnName(col).Equals(MetricKinds.ColumnNames.Count))
{
input.ConfusionMatrix[i] = new ChooseColumnsByIndexTransform(env,
new ChooseColumnsByIndexTransform.Arguments()
{
Drop = true, Index = new[] { col }
}, idv);
break;
}
}
}
conf = EvaluateUtils.ConcatenateOverallMetrics(env, input.ConfusionMatrix);
}
var warningsIdv = warnings.Count > 0 ? AppendRowsDataView.Create(env, warnings[0].Schema, warnings.ToArray()) : null;
return(new CombinedOutput()
{
PerInstanceMetrics = perInst[0],
OverallMetrics = overall,
ConfusionMatrix = conf,
Warnings = warningsIdv
});
}
/// <summary>
/// Features: x1, x2, x3, xRand; y = 10*x1 + 20x2 + 5.5x3 + e, xRand- random and Label y is to dependant on xRand.
/// xRand has the least importance: Evaluation metrics do not change a lot when xRand is permuted.
/// x2 has the biggest importance.
/// </summary>
private IDataView GetDenseDataset(TaskType task = TaskType.Regression)
{
// Setup synthetic dataset.
const int numberOfInstances = 1000;
var rand = new Random(10);
float[] yArray = new float[numberOfInstances],
x1Array = new float[numberOfInstances],
x2Array = new float[numberOfInstances],
x3Array = new float[numberOfInstances],
x4RandArray = new float[numberOfInstances];
for (var i = 0; i < numberOfInstances; i++)
{
var x1 = rand.Next(1000);
x1Array[i] = x1;
var x2Important = rand.Next(10000);
x2Array[i] = x2Important;
var x3 = rand.Next(5000);
x3Array[i] = x3;
var x4Rand = rand.Next(1000);
x4RandArray[i] = x4Rand;
var noise = rand.Next(50);
yArray[i] = (float)(10 * x1 + 20 * x2Important + 5.5 * x3 + noise);
}
// If binary classification, modify the labels
if (task == TaskType.BinaryClassification ||
task == TaskType.MulticlassClassification)
{
GetBinaryClassificationLabels(yArray);
}
else if (task == TaskType.Ranking)
{
GetRankingLabels(yArray);
}
// Create data view.
var bldr = new ArrayDataViewBuilder(Env);
bldr.AddColumn("X1", NumberDataViewType.Single, x1Array);
bldr.AddColumn("X2Important", NumberDataViewType.Single, x2Array);
bldr.AddColumn("X3", NumberDataViewType.Single, x3Array);
bldr.AddColumn("X4Rand", NumberDataViewType.Single, x4RandArray);
bldr.AddColumn("Label", NumberDataViewType.Single, yArray);
if (task == TaskType.Ranking)
{
bldr.AddColumn("GroupId", NumberDataViewType.UInt32, CreateGroupIds(yArray.Length));
}
var srcDV = bldr.GetDataView();
var pipeline = ML.Transforms.Concatenate("Features", "X1", "X2Important", "X3", "X4Rand")
.Append(ML.Transforms.Normalize("Features"));
// Create a keytype for Ranking
if (task == TaskType.Ranking)
{
return(pipeline.Append(ML.Transforms.Conversion.MapValueToKey("GroupId"))
.Fit(srcDV).Transform(srcDV));
}
return(pipeline.Fit(srcDV).Transform(srcDV));
}
/// <summary>
/// Utility method used to represent a tree ensemble as an <see cref="IDataView"/>.
/// Every row in the <see cref="IDataView"/> corresponds to a node in the tree ensemble. The columns are the fields for each node.
/// The column TreeID specifies which tree the node belongs to. The <see cref="QuantileRegressionTree"/> gets
/// special treatment since it has some additional fields (<see cref="QuantileRegressionTree.GetLeafSamplesAt(int)"/>
/// and <see cref="QuantileRegressionTree.GetLeafSampleWeightsAt(int)"/>).
/// </summary>
public static IDataView RegressionTreeEnsembleAsIDataView(IHost host, double bias, IReadOnlyList <double> treeWeights, IReadOnlyList <RegressionTreeBase> trees)
{
var builder = new ArrayDataViewBuilder(host);
var numberOfRows = trees.Select(tree => tree.NumberOfNodes).Sum() + trees.Select(tree => tree.NumberOfLeaves).Sum();
var treeWeightsList = new List <double>();
var treeId = new List <int>();
var isLeaf = new List <ReadOnlyMemory <char> >();
var leftChild = new List <int>();
var rightChild = new List <int>();
var numericalSplitFeatureIndexes = new List <int>();
var numericalSplitThresholds = new List <float>();
var categoricalSplitFlags = new List <bool>();
var leafValues = new List <double>();
var splitGains = new List <double>();
var categoricalSplitFeatures = new List <VBuffer <int> >();
var categoricalCategoricalSplitFeatureRange = new List <VBuffer <int> >();
for (int i = 0; i < trees.Count; i++)
{
// TreeWeights column. The TreeWeight value will be repeated for all the notes in the same tree in the IDataView.
treeWeightsList.AddRange(Enumerable.Repeat(treeWeights[i], trees[i].NumberOfNodes + trees[i].NumberOfLeaves));
// Tree id indicates which tree the node belongs to.
treeId.AddRange(Enumerable.Repeat(i, trees[i].NumberOfNodes + trees[i].NumberOfLeaves));
// IsLeaf column indicates if node is a leaf node.
isLeaf.AddRange(Enumerable.Repeat(new ReadOnlyMemory <char>("Tree node".ToCharArray()), trees[i].NumberOfNodes));
isLeaf.AddRange(Enumerable.Repeat(new ReadOnlyMemory <char>("Leaf node".ToCharArray()), trees[i].NumberOfLeaves));
// LeftChild column.
leftChild.AddRange(trees[i].LeftChild.AsEnumerable());
leftChild.AddRange(Enumerable.Repeat(0, trees[i].NumberOfLeaves));
// RightChild column.
rightChild.AddRange(trees[i].RightChild.AsEnumerable());
rightChild.AddRange(Enumerable.Repeat(0, trees[i].NumberOfLeaves));
// NumericalSplitFeatureIndexes column.
numericalSplitFeatureIndexes.AddRange(trees[i].NumericalSplitFeatureIndexes.AsEnumerable());
numericalSplitFeatureIndexes.AddRange(Enumerable.Repeat(0, trees[i].NumberOfLeaves));
// NumericalSplitThresholds column.
numericalSplitThresholds.AddRange(trees[i].NumericalSplitThresholds.AsEnumerable());
numericalSplitThresholds.AddRange(Enumerable.Repeat(0f, trees[i].NumberOfLeaves));
// CategoricalSplitFlags column.
categoricalSplitFlags.AddRange(trees[i].CategoricalSplitFlags.AsEnumerable());
categoricalSplitFlags.AddRange(Enumerable.Repeat(false, trees[i].NumberOfLeaves));
// LeafValues column.
leafValues.AddRange(Enumerable.Repeat(0d, trees[i].NumberOfNodes));
leafValues.AddRange(trees[i].LeafValues.AsEnumerable());
// SplitGains column.
splitGains.AddRange(trees[i].SplitGains.AsEnumerable());
splitGains.AddRange(Enumerable.Repeat(0d, trees[i].NumberOfLeaves));
for (int j = 0; j < trees[i].NumberOfNodes; j++)
{
// CategoricalSplitFeatures column.
var categoricalSplitFeaturesArray = trees[i].GetCategoricalSplitFeaturesAt(j).ToArray();
categoricalSplitFeatures.Add(new VBuffer <int>(categoricalSplitFeaturesArray.Length, categoricalSplitFeaturesArray));
var len = trees[i].GetCategoricalSplitFeaturesAt(j).ToArray().Length;
// CategoricalCategoricalSplitFeatureRange column.
var categoricalCategoricalSplitFeatureRangeArray = trees[i].GetCategoricalCategoricalSplitFeatureRangeAt(j).ToArray();
categoricalCategoricalSplitFeatureRange.Add(new VBuffer <int>(categoricalCategoricalSplitFeatureRangeArray.Length, categoricalCategoricalSplitFeatureRangeArray));
len = trees[i].GetCategoricalCategoricalSplitFeatureRangeAt(j).ToArray().Length;
}
categoricalSplitFeatures.AddRange(Enumerable.Repeat(new VBuffer <int>(), trees[i].NumberOfLeaves));
categoricalCategoricalSplitFeatureRange.AddRange(Enumerable.Repeat(new VBuffer <int>(), trees[i].NumberOfLeaves));
}
// Bias column. This will be a repeated value for all rows in the resulting IDataView.
builder.AddColumn("Bias", NumberDataViewType.Double, Enumerable.Repeat(bias, numberOfRows).ToArray());
builder.AddColumn("TreeWeights", NumberDataViewType.Double, treeWeightsList.ToArray());
builder.AddColumn("TreeID", NumberDataViewType.Int32, treeId.ToArray());
builder.AddColumn("IsLeaf", TextDataViewType.Instance, isLeaf.ToArray());
builder.AddColumn(nameof(RegressionTreeBase.LeftChild), NumberDataViewType.Int32, leftChild.ToArray());
builder.AddColumn(nameof(RegressionTreeBase.RightChild), NumberDataViewType.Int32, rightChild.ToArray());
builder.AddColumn(nameof(RegressionTreeBase.NumericalSplitFeatureIndexes), NumberDataViewType.Int32, numericalSplitFeatureIndexes.ToArray());
builder.AddColumn(nameof(RegressionTreeBase.NumericalSplitThresholds), NumberDataViewType.Single, numericalSplitThresholds.ToArray());
builder.AddColumn(nameof(RegressionTreeBase.CategoricalSplitFlags), BooleanDataViewType.Instance, categoricalSplitFlags.ToArray());
builder.AddColumn(nameof(RegressionTreeBase.LeafValues), NumberDataViewType.Double, leafValues.ToArray());
builder.AddColumn(nameof(RegressionTreeBase.SplitGains), NumberDataViewType.Double, splitGains.ToArray());
builder.AddColumn("CategoricalSplitFeatures", NumberDataViewType.Int32, categoricalSplitFeatures.ToArray());
builder.AddColumn("CategoricalCategoricalSplitFeatureRange", NumberDataViewType.Int32, categoricalCategoricalSplitFeatureRange.ToArray());
// If the input tree array is a quantile regression tree we need to add two more columns.
var quantileTrees = trees as IReadOnlyList <QuantileRegressionTree>;
if (quantileTrees != null)
{
// LeafSamples column.
var leafSamples = new List <VBuffer <double> >();
// LeafSampleWeights column.
var leafSampleWeights = new List <VBuffer <double> >();
for (int i = 0; i < quantileTrees.Count; i++)
{
leafSamples.AddRange(Enumerable.Repeat(new VBuffer <double>(), quantileTrees[i].NumberOfNodes));
leafSampleWeights.AddRange(Enumerable.Repeat(new VBuffer <double>(), quantileTrees[i].NumberOfNodes));
for (int j = 0; j < quantileTrees[i].NumberOfLeaves; j++)
{
var leafSamplesArray = quantileTrees[i].GetLeafSamplesAt(j).ToArray();
leafSamples.Add(new VBuffer <double>(leafSamplesArray.Length, leafSamplesArray));
var len = quantileTrees[i].GetLeafSamplesAt(j).ToArray().Length;
var leafSampleWeightsArray = quantileTrees[i].GetLeafSampleWeightsAt(j).ToArray();
leafSampleWeights.Add(new VBuffer <double>(leafSampleWeightsArray.Length, leafSampleWeightsArray));
len = quantileTrees[i].GetLeafSampleWeightsAt(j).ToArray().Length;
}
}
builder.AddColumn("LeafSamples", NumberDataViewType.Double, leafSamples.ToArray());
builder.AddColumn("LeafSampleWeights", NumberDataViewType.Double, leafSampleWeights.ToArray());
}
var data = builder.GetDataView();
return(data);
}
/// <summary>
/// Features: x1, x2, x3, xRand; y = 10*x1 + 20x2 + 5.5x3 + e, xRand- random, Label y is dependant on xRand.
/// Test verifies that feature contribution scores are outputted along with a score for predicted data.
/// </summary>
private void TestFeatureContribution(
ITrainerEstimator <ISingleFeaturePredictionTransformer <IPredictor>, IPredictor> trainer,
List <float[]> expectedValues,
int precision = 6)
{
// Setup synthetic dataset.
const int numInstances = 1000;
const int numFeatures = 4;
var rand = new Random(10);
float[] yArray = new float[numInstances];
float[][] xArray = new float[numFeatures][];
int[] xRangeArray = new[] { 1000, 10000, 5000, 1000 };
float[] xWeightArray = new[] {
10,
20, // Most important feature with high weight. Should have the highest contribution.
5.5f,
0, // Least important feature. Should have the least contribution.
};
for (var instanceIndex = 0; instanceIndex < numInstances; instanceIndex++)
{
for (int featureIndex = 0; featureIndex < numFeatures; featureIndex++)
{
if (xArray[featureIndex] == null)
{
xArray[featureIndex] = new float[numInstances];
}
xArray[featureIndex][instanceIndex] = rand.Next(xRangeArray[featureIndex]);
yArray[instanceIndex] += xArray[featureIndex][instanceIndex] * xWeightArray[featureIndex];
}
var noise = rand.Next(50);
yArray[instanceIndex] += noise;
}
// Create data view.
var bldr = new ArrayDataViewBuilder(Env);
bldr.AddColumn("X1", NumberType.Float, xArray[0]);
bldr.AddColumn("X2Important", NumberType.Float, xArray[1]);
bldr.AddColumn("X3", NumberType.Float, xArray[2]);
bldr.AddColumn("X4Rand", NumberType.Float, xArray[3]);
bldr.AddColumn("Label", NumberType.Float, yArray);
var srcDV = bldr.GetDataView();
var pipeline = ML.Transforms.Concatenate("Features", "X1", "X2Important", "X3", "X4Rand")
.AppendCacheCheckpoint(ML)
.Append(ML.Transforms.Normalize("Features"));
var data = pipeline.Fit(srcDV).Transform(srcDV);
var model = trainer.Fit(data);
var args = new FeatureContributionCalculationTransform.Arguments()
{
Bottom = 10,
Top = 10
};
var output = FeatureContributionCalculationTransform.Create(Env, args, data, model.Model, model.FeatureColumn);
var transformedOutput = output.AsEnumerable <ScoreAndContribution>(Env, true);
int rowIndex = 0;
foreach (var row in transformedOutput.Take(expectedValues.Count))
{
var expectedValue = expectedValues[rowIndex++];
for (int i = 0; i < numFeatures; i++)
{
Assert.Equal(expectedValue[i], row.FeatureContributions[i], precision);
}
}
Done();
}
/// <summary>
/// Helper function that builds the IDataView given a list of keys and non-vector values
/// </summary>
internal static IDataView CreateDataView <TKey, TValue>(IHostEnvironment env,
IEnumerable <TKey> keys,
IEnumerable <TValue> values,
string keyColumnName,
string valueColumnName,
bool treatValuesAsKeyTypes)
{
var keyType = GetPrimitiveType(typeof(TKey), out bool isKeyVectorType);
var valueType = GetPrimitiveType(typeof(TValue), out bool isValueVectorType);
var dataViewBuilder = new ArrayDataViewBuilder(env);
dataViewBuilder.AddColumn(keyColumnName, keyType, keys.ToArray());
if (treatValuesAsKeyTypes)
{
// When treating the values as KeyTypes, generate the unique
// set of values. This is used for generating the metadata of
// the column.
HashSet <TValue> valueSet = new HashSet <TValue>();
HashSet <TKey> keySet = new HashSet <TKey>();
for (int i = 0; i < values.Count(); ++i)
{
var v = values.ElementAt(i);
if (valueSet.Contains(v))
{
continue;
}
valueSet.Add(v);
var k = keys.ElementAt(i);
keySet.Add(k);
}
var metaKeys = keySet.ToArray();
// Key Values are treated in one of two ways:
// If the values are of type uint or ulong, these values are used directly as the keys types and no new keys are created.
// If the values are not of uint or ulong, then key values are generated as uints starting from 1, since 0 is missing key.
if (valueType.RawKind == DataKind.U4)
{
uint[] indices = values.Select((x) => Convert.ToUInt32(x)).ToArray();
dataViewBuilder.AddColumn(valueColumnName, GetKeyValueGetter(metaKeys), 0, metaKeys.Length, indices);
}
else if (valueType.RawKind == DataKind.U8)
{
ulong[] indices = values.Select((x) => Convert.ToUInt64(x)).ToArray();
dataViewBuilder.AddColumn(valueColumnName, GetKeyValueGetter(metaKeys), 0, metaKeys.Length, indices);
}
else
{
// When generating the indices, treat each value as being unique, i.e. two values that are the same will
// be assigned the same index. The dictionary is used to maintain uniqueness, indices will contain
// the full list of indices (equal to the same length of values).
Dictionary <TValue, uint> keyTypeValueMapping = new Dictionary <TValue, uint>();
uint[] indices = new uint[values.Count()];
// Start the index at 1
uint index = 1;
for (int i = 0; i < values.Count(); ++i)
{
TValue value = values.ElementAt(i);
if (!keyTypeValueMapping.ContainsKey(value))
{
keyTypeValueMapping.Add(value, index);
index++;
}
var keyValue = keyTypeValueMapping[value];
indices[i] = keyValue;
}
dataViewBuilder.AddColumn(valueColumnName, GetKeyValueGetter(metaKeys), 0, metaKeys.Count(), indices);
}
}
else
{
dataViewBuilder.AddColumn(valueColumnName, valueType, values.ToArray());
}
return(dataViewBuilder.GetDataView());
}
public void Train(List <FeatureSubsetModel <IPredictorProducing <TOutput> > > models, RoleMappedData data, IHostEnvironment env)
{
Contracts.CheckValue(env, nameof(env));
var host = env.Register(Stacking.LoadName);
host.CheckValue(models, nameof(models));
host.CheckValue(data, nameof(data));
using (var ch = host.Start("Training stacked model"))
{
ch.Check(Meta == null, "Train called multiple times");
ch.Check(BasePredictorType != null);
var maps = new ValueMapper <VBuffer <Single>, TOutput> [models.Count];
for (int i = 0; i < maps.Length; i++)
{
Contracts.Assert(models[i].Predictor is IValueMapper);
var m = (IValueMapper)models[i].Predictor;
maps[i] = m.GetMapper <VBuffer <Single>, TOutput>();
}
// REVIEW: Should implement this better....
var labels = new Single[100];
var features = new VBuffer <Single> [100];
int count = 0;
// REVIEW: Should this include bad values or filter them?
using (var cursor = new FloatLabelCursor(data, CursOpt.AllFeatures | CursOpt.AllLabels))
{
TOutput[] predictions = new TOutput[maps.Length];
var vBuffers = new VBuffer <Single> [maps.Length];
while (cursor.MoveNext())
{
Parallel.For(0, maps.Length, i =>
{
var model = models[i];
if (model.SelectedFeatures != null)
{
EnsembleUtils.SelectFeatures(ref cursor.Features, model.SelectedFeatures, model.Cardinality, ref vBuffers[i]);
maps[i](ref vBuffers[i], ref predictions[i]);
}
else
{
maps[i](ref cursor.Features, ref predictions[i]);
}
});
Utils.EnsureSize(ref labels, count + 1);
Utils.EnsureSize(ref features, count + 1);
labels[count] = cursor.Label;
FillFeatureBuffer(predictions, ref features[count]);
count++;
}
}
ch.Info("The number of instances used for stacking trainer is {0}", count);
var bldr = new ArrayDataViewBuilder(host);
Array.Resize(ref labels, count);
Array.Resize(ref features, count);
bldr.AddColumn(DefaultColumnNames.Label, NumberType.Float, labels);
bldr.AddColumn(DefaultColumnNames.Features, NumberType.Float, features);
var view = bldr.GetDataView();
var rmd = new RoleMappedData(view, DefaultColumnNames.Label, DefaultColumnNames.Features);
var trainer = BasePredictorType.CreateComponent(host);
if (trainer.Info.NeedNormalization)
{
ch.Warning("The trainer specified for stacking wants normalization, but we do not currently allow this.");
}
Meta = trainer.Train(rmd);
CheckMeta();
}
}
public void TestDenseSGD()
{
// Setup synthetic dataset.
const int numberOfInstances = 1000;
var rand = new Random(10);
float[] yArray = new float[numberOfInstances],
x1Array = new float[numberOfInstances],
x2Array = new float[numberOfInstances],
x3Array = new float[numberOfInstances],
x4RandArray = new float[numberOfInstances];
for (var i = 0; i < numberOfInstances; i++)
{
var x1 = rand.Next(1000);
x1Array[i] = x1;
var x2Important = rand.Next(10000);
x2Array[i] = x2Important;
var x3 = rand.Next(5000);
x3Array[i] = x3;
var x4Rand = rand.Next(1000);
x4RandArray[i] = x4Rand;
var noise = rand.Next(50);
yArray[i] = (float)(10 * x1 + 20 * x2Important + 5.5 * x3 + noise);
}
// Create data view.
var bldr = new ArrayDataViewBuilder(Env);
bldr.AddColumn("X1", NumberType.Float, x1Array);
bldr.AddColumn("X2Important", NumberType.Float, x2Array);
bldr.AddColumn("X3", NumberType.Float, x3Array);
bldr.AddColumn("X4Rand", NumberType.Float, x4RandArray);
bldr.AddColumn("Label", NumberType.Float, yArray);
var srcDV = bldr.GetDataView();
var pipeline = ML.Transforms.Concatenate("Features", "X1", "X2Important", "X3", "X4Rand")
.Append(ML.Transforms.Normalize("Features"));
var data = pipeline.Fit(srcDV).Transform(srcDV);
var model = ML.Regression.Trainers.OnlineGradientDescent().Fit(data);
var pfi = ML.Regression.PermutationFeatureImportance(model, data);
// Pfi Indices:
// X1: 0
// X2Important: 1
// X3: 2
// X4Rand: 3
// For the following metrics lower is better, so maximum delta means more important feature, and vice versa
Assert.True(MinDeltaIndex(pfi, m => m.L1) == 3);
Assert.True(MaxDeltaIndex(pfi, m => m.L1) == 1);
Assert.True(MinDeltaIndex(pfi, m => m.L2) == 3);
Assert.True(MaxDeltaIndex(pfi, m => m.L2) == 1);
Assert.True(MinDeltaIndex(pfi, m => m.Rms) == 3);
Assert.True(MaxDeltaIndex(pfi, m => m.Rms) == 1);
// For the following metrics higher is better, so minimum delta means more important feature, and vice versa
Assert.True(MaxDeltaIndex(pfi, m => m.RSquared) == 3);
Assert.True(MinDeltaIndex(pfi, m => m.RSquared) == 1);
Done();
}
public void TestFeatureImportance()
{
// Setup synthetic dataset.
const int numberOfInstances = 1000;
var rand = new Random(10);
float[] yArray = new float[numberOfInstances],
x1Array = new float[numberOfInstances],
x2Array = new float[numberOfInstances],
x3Array = new float[numberOfInstances],
x4RandArray = new float[numberOfInstances];
for (var i = 0; i < numberOfInstances; i++)
{
var x1 = rand.Next(1000);
x1Array[i] = x1;
var x2Important = rand.Next(10000);
x2Array[i] = x2Important;
var x3 = rand.Next(5000);
x3Array[i] = x3;
var x4Rand = rand.Next(1000);
x4RandArray[i] = x4Rand;
var noise = rand.Next(50);
yArray[i] = (float)(10 * x1 + 20 * x2Important + 5.5 * x3 + noise);
}
// Create data view.
var bldr = new ArrayDataViewBuilder(Env);
bldr.AddColumn("X1", NumberType.Float, x1Array);
bldr.AddColumn("X2Important", NumberType.Float, x2Array);
bldr.AddColumn("X3", NumberType.Float, x3Array);
bldr.AddColumn("X4Rand", NumberType.Float, x4RandArray);
bldr.AddColumn("Label", NumberType.Float, yArray);
var srcDV = bldr.GetDataView();
var pipeline = ML.Transforms.Concatenate("Features", "X1", "X2Important", "X3", "X4Rand")
.Append(ML.Transforms.Normalize("Features"));
var data = pipeline.Fit(srcDV).Transform(srcDV);
var model = ML.Regression.Trainers.OnlineGradientDescent().Fit(data);
var args = new FeatureContributionCalculationTransform.Arguments()
{
Bottom = 10,
Top = 10
};
var output = FeatureContributionCalculationTransform.Create(Env, args, data, model.Model, model.FeatureColumn);
// Get prediction scores and contributions
var enumerator = output.AsEnumerable <ScoreAndContribution>(Env, true).GetEnumerator();
ScoreAndContribution row = null;
var expectedValues = new List <float[]>();
expectedValues.Add(new float[4] {
0.15640761F, 1, 0.155862764F, 0.07276783F
});
expectedValues.Add(new float[4] {
0.09507586F, 1, 0.1835608F, 0.0437548943F
});
expectedValues.Add(new float[4] {
0.297142357F, 1, 0.2855884F, 0.193529665F
});
expectedValues.Add(new float[4] {
0.45465675F, 0.8805887F, 0.4031663F, 1
});
expectedValues.Add(new float[4] {
0.0595234372F, 0.99999994F, 0.349647522F, 0.137912869F
});
int index = 0;
while (enumerator.MoveNext() && index < expectedValues.Count)
{
row = enumerator.Current;
Assert.True(row.FeatureContributions[0] == expectedValues[index][0]);
Assert.True(row.FeatureContributions[1] == expectedValues[index][1]);
Assert.True(row.FeatureContributions[2] == expectedValues[index][2]);
Assert.True(row.FeatureContributions[3] == expectedValues[index++][3]);
}
Done();
}
public void TestSparseSGD()
{
// Setup synthetic dataset.
const int numberOfInstances = 10000;
var rand = new Random(10);
float[] yArray = new float[numberOfInstances],
x1Array = new float[numberOfInstances],
x3Array = new float[numberOfInstances];
VBuffer <float>[] vbArray = new VBuffer <float> [numberOfInstances];
for (var i = 0; i < numberOfInstances; i++)
{
var x1 = rand.Next(1000);
x1Array[i] = x1;
var x3Important = rand.Next(10000);
x3Array[i] = x3Important;
VBuffer <float> vb;
if (i % 10 != 0)
{
vb = new VBuffer <float>(4, 3, new float[] { rand.Next(1000), rand.Next(1000), rand.Next(1000) }, new int[] { 0, 2, 3 });
}
else
{
vb = new VBuffer <float>(4, 4, new float[] { rand.Next(1000), rand.Next(1000), rand.Next(1000), rand.Next(1000) }, new int[] { 0, 1, 2, 3 });
}
vbArray[i] = vb;
float vbSum = 0;
foreach (var vbValue in vb.DenseValues())
{
vbSum += vbValue * 10;
}
var noise = rand.Next(50);
yArray[i] = 10 * x1 + vbSum + 20 * x3Important + noise;
}
// Create data view.
var bldr = new ArrayDataViewBuilder(Env);
bldr.AddColumn("X1", NumberType.Float, x1Array);
bldr.AddColumn("X2VBuffer", NumberType.Float, vbArray);
bldr.AddColumn("X3Important", NumberType.Float, x3Array);
bldr.AddColumn("Label", NumberType.Float, yArray);
var srcDV = bldr.GetDataView();
var pipeline = ML.Transforms.Concatenate("Features", "X1", "X2VBuffer", "X3Important")
.Append(ML.Transforms.Normalize("Features"));
var data = pipeline.Fit(srcDV).Transform(srcDV);
var model = ML.Regression.Trainers.OnlineGradientDescent().Fit(data);
var results = ML.Regression.PermutationFeatureImportance(model, data);
// Pfi Indices:
// X1: 0
// X2VBuffer-Slot-0: 1
// X2VBuffer-Slot-1: 2
// X2VBuffer-Slot-2: 3
// X2VBuffer-Slot-3: 4
// X3Important: 5
// Permuted X2VBuffer-Slot-1 lot (f2) should have min impact on SGD metrics, X3Important -- max impact.
// For the following metrics lower is better, so maximum delta means more important feature, and vice versa
Assert.True(MinDeltaIndex(results, m => m.L1) == 2);
Assert.True(MaxDeltaIndex(results, m => m.L1) == 5);
Assert.True(MinDeltaIndex(results, m => m.L2) == 2);
Assert.True(MaxDeltaIndex(results, m => m.L2) == 5);
Assert.True(MinDeltaIndex(results, m => m.Rms) == 2);
Assert.True(MaxDeltaIndex(results, m => m.Rms) == 5);
// For the following metrics higher is better, so minimum delta means more important feature, and vice versa
Assert.True(MaxDeltaIndex(results, m => m.RSquared) == 2);
Assert.True(MinDeltaIndex(results, m => m.RSquared) == 5);
}
public void TransposerTest()
{
const int rowCount = 1000;
Random rgen = new Random(0);
ArrayDataViewBuilder builder = new ArrayDataViewBuilder(Env);
// A is to check the splitting of a sparse-ish column.
var dataA = GenerateHelper(rowCount, 0.1, rgen, () => (int)rgen.Next(), 50, 5, 10, 15);
dataA[rowCount / 2] = new VBuffer <int>(50, 0, null, null); // Coverage for the null vbuffer case.
builder.AddColumn("A", NumberDataViewType.Int32, dataA);
// B is to check the splitting of a dense-ish column.
builder.AddColumn("B", NumberDataViewType.Double, GenerateHelper(rowCount, 0.8, rgen, rgen.NextDouble, 50, 0, 25, 49));
// C is to just have some column we do nothing with.
builder.AddColumn("C", NumberDataViewType.Int16, GenerateHelper(rowCount, 0.1, rgen, () => (short)1, 30, 3, 10, 24));
// D is to check some column we don't have to split because it's sufficiently small.
builder.AddColumn("D", NumberDataViewType.Double, GenerateHelper(rowCount, 0.1, rgen, rgen.NextDouble, 3, 1));
// E is to check a sparse scalar column.
builder.AddColumn("E", NumberDataViewType.UInt32, GenerateHelper(rowCount, 0.1, rgen, () => (uint)rgen.Next(int.MinValue, int.MaxValue)));
// F is to check a dense-ish scalar column.
builder.AddColumn("F", NumberDataViewType.Int32, GenerateHelper(rowCount, 0.8, rgen, () => rgen.Next()));
IDataView view = builder.GetDataView();
// Do not force save. This will have a mix of passthrough and saved columns. Note that duplicate
// specification of "D" to test that specifying a column twice has no ill effects.
string[] names = { "B", "A", "E", "D", "F", "D" };
using (Transposer trans = Transposer.Create(Env, view, false, names))
{
// Before checking the contents, check the names.
for (int i = 0; i < names.Length; ++i)
{
int index;
Assert.True(trans.Schema.TryGetColumnIndex(names[i], out index), $"Transpose schema couldn't find column '{names[i]}'");
int trueIndex;
bool result = view.Schema.TryGetColumnIndex(names[i], out trueIndex);
Contracts.Assert(result);
Assert.True(trueIndex == index, $"Transpose schema had column '{names[i]}' at unexpected index");
}
// Check the contents
Assert.Null(((ITransposeDataView)trans).GetSlotType(2)); // C check to see that it's not transposable.
TransposeCheckHelper <int>(view, 0, trans); // A check.
TransposeCheckHelper <Double>(view, 1, trans); // B check.
TransposeCheckHelper <Double>(view, 3, trans); // D check.
TransposeCheckHelper <uint>(view, 4, trans); // E check.
TransposeCheckHelper <int>(view, 5, trans); // F check.
}
// Force save. Recheck columns that would have previously been passthrough columns.
// The primary benefit of this check is that we check the binary saving / loading
// functionality of scalars which are otherwise always must necessarily be
// passthrough. Also exercise the select by index functionality while we're at it.
using (Transposer trans = Transposer.Create(Env, view, true, 3, 5, 4))
{
// Check to see that A, B, and C were not transposed somehow.
var itdv = (ITransposeDataView)trans;
Assert.Null(itdv.GetSlotType(0));
Assert.Null(itdv.GetSlotType(1));
Assert.Null(itdv.GetSlotType(2));
TransposeCheckHelper <Double>(view, 3, trans); // D check.
TransposeCheckHelper <uint>(view, 4, trans); // E check.
TransposeCheckHelper <int>(view, 5, trans); // F check.
}
}