internal unsafe string GetModelString()
{
int bufLen = 2 << 15;
byte[] buffer = new byte[bufLen];
int size = 0;
fixed(byte *ptr = buffer)
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterSaveModelToString(Handle, 0, BestIteration, bufLen, ref size, ptr));
// If buffer size is not enough, reallocate buffer and get again.
if (size > bufLen)
{
bufLen = size;
buffer = new byte[bufLen];
fixed(byte *ptr = buffer)
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterSaveModelToString(Handle, 0, BestIteration, bufLen, ref size, ptr));
}
byte[] content = new byte[size];
Array.Copy(buffer, content, size);
fixed(byte *ptr = content)
return(LightGbmInterfaceUtils.GetString((IntPtr)ptr));
}
private void InitParallelTraining()
{
Options = LightGbmTrainerOptions.ToDictionary(Host);
ParallelTraining = LightGbmTrainerOptions.ParallelTrainer != null?LightGbmTrainerOptions.ParallelTrainer.CreateComponent(Host) : new SingleTrainer();
if (ParallelTraining.ParallelType() != "serial" && ParallelTraining.NumMachines() > 1)
{
Options["tree_learner"] = ParallelTraining.ParallelType();
var otherParams = ParallelTraining.AdditionalParams();
if (otherParams != null)
{
foreach (var pair in otherParams)
{
Options[pair.Key] = pair.Value;
}
}
Contracts.CheckValue(ParallelTraining.GetReduceScatterFunction(), nameof(ParallelTraining.GetReduceScatterFunction));
Contracts.CheckValue(ParallelTraining.GetAllgatherFunction(), nameof(ParallelTraining.GetAllgatherFunction));
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.NetworkInitWithFunctions(
ParallelTraining.NumMachines(),
ParallelTraining.Rank(),
ParallelTraining.GetReduceScatterFunction(),
ParallelTraining.GetAllgatherFunction()
));
}
}
private protected override LightGbmRankingModelParameters CreatePredictor()
{
Host.Check(TrainedEnsemble != null, "The predictor cannot be created before training is complete");
var innerArgs = LightGbmInterfaceUtils.JoinParameters(Options);
return(new LightGbmRankingModelParameters(Host, TrainedEnsemble, FeatureCount, innerArgs));
}
private void DisposeParallelTraining()
{
if (ParallelTraining.NumMachines() > 1)
{
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.NetworkFree());
}
}
public int GetNumCols()
{
int res = 0;
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetGetNumFeature(_handle, ref res));
return(res);
}
private Dataset LoadTrainingData(IChannel ch, RoleMappedData trainData, out CategoricalMetaData catMetaData)
{
// Verifications.
Host.AssertValue(ch);
ch.CheckValue(trainData, nameof(trainData));
CheckDataValid(ch, trainData);
// Load metadata first.
var factory = CreateCursorFactory(trainData);
GetMetainfo(ch, factory, out int numRow, out float[] labels, out float[] weights, out int[] groups);
catMetaData = GetCategoricalMetaData(ch, trainData, numRow);
GetDefaultParameters(ch, numRow, catMetaData.CategoricalBoudaries != null, catMetaData.TotalCats);
Dataset dtrain;
string param = LightGbmInterfaceUtils.JoinParameters(Options);
// To reduce peak memory usage, only enable one sampling task at any given time.
lock (LightGbmShared.SampleLock)
{
CreateDatasetFromSamplingData(ch, factory, numRow,
param, labels, weights, groups, catMetaData, out dtrain);
}
// Push rows into dataset.
LoadDataset(ch, factory, dtrain, numRow, LightGbmTrainerOptions.BatchSize, catMetaData);
// Some checks.
CheckAndUpdateParametersBeforeTraining(ch, trainData, labels, groups);
return(dtrain);
}
public bool Update()
{
int isFinished = 0;
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterUpdateOneIter(Handle, ref isFinished));
return(isFinished == 1);
}
private protected override OvaModelParameters CreatePredictor()
{
Host.Check(TrainedEnsemble != null, "The predictor cannot be created before training is complete.");
Host.Assert(_numClass > 1, "Must know the number of classes before creating a predictor.");
Host.Assert(TrainedEnsemble.NumTrees % _numClass == 0, "Number of trees should be a multiple of number of classes.");
var innerArgs = LightGbmInterfaceUtils.JoinParameters(Options);
IPredictorProducing <float>[] predictors = new IPredictorProducing <float> [_tlcNumClass];
for (int i = 0; i < _tlcNumClass; ++i)
{
var pred = CreateBinaryPredictor(i, innerArgs);
var cali = new PlattCalibrator(Host, -0.5, 0);
predictors[i] = new FeatureWeightsCalibratedPredictor(Host, pred, cali);
}
string obj = (string)GetGbmParameters()["objective"];
if (obj == "multiclass")
{
return(OvaModelParameters.Create(Host, OvaModelParameters.OutputFormula.Softmax, predictors));
}
else
{
return(OvaModelParameters.Create(Host, predictors));
}
}
public void Dispose()
{
if (_handle != IntPtr.Zero)
{
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetFree(_handle));
}
_handle = IntPtr.Zero;
}
public void Dispose()
{
if (Handle != IntPtr.Zero)
{
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterFree(Handle));
}
Handle = IntPtr.Zero;
}
private protected override CalibratedModelParametersBase <LightGbmBinaryModelParameters, PlattCalibrator> CreatePredictor()
{
Host.Check(TrainedEnsemble != null, "The predictor cannot be created before training is complete");
var innerArgs = LightGbmInterfaceUtils.JoinParameters(Options);
var pred = new LightGbmBinaryModelParameters(Host, TrainedEnsemble, FeatureCount, innerArgs);
var cali = new PlattCalibrator(Host, -0.5, 0);
return(new FeatureWeightsCalibratedModelParameters <LightGbmBinaryModelParameters, PlattCalibrator>(Host, pred, cali));
}
public unsafe void SetLabel(float[] labels)
{
Contracts.AssertValue(labels);
Contracts.Assert(labels.Length == GetNumRows());
fixed(float *ptr = labels)
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetSetField(_handle, "label", (IntPtr)ptr, labels.Length,
WrappedLightGbmInterface.CApiDType.Float32));
}
public unsafe void SetGroup(int[] groups)
{
if (groups != null)
{
fixed(int *ptr = groups)
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetSetField(_handle, "group", (IntPtr)ptr, groups.Length,
WrappedLightGbmInterface.CApiDType.Int32));
}
}
/// <summary>
/// Create a <see cref="Dataset"/> for storing training and prediciton data under LightGBM framework. The main goal of this function
/// is not marshaling ML.NET data set into LightGBM format but just creates a (unmanaged) container where examples can be pushed into by calling
/// <see cref="PushRows(float[], int, int, int)"/>. It also pre-allocates memory so the actual size (number of examples and number of features)
/// of the data set is required. A sub-sampled version of the original data set is passed in to compute some statictics needed by the training
/// procedure. Note that we use "original" to indicate a property from the unsampled data set.
/// </summary>
/// <param name="sampleValuePerColumn">A 2-D array which encodes the sub-sampled data matrix. sampleValuePerColumn[i] stores
/// all the non-zero values of the i-th feature. sampleValuePerColumn[i][j] is the j-th non-zero value of i-th feature encountered when scanning
/// the values row-by-row (i.e., example-by-example) in the matrix and column-by-column (i.e., feature-by-feature) within one row. It is similar
/// to CSC format for storing sparse matrix.</param>
/// <param name="sampleIndicesPerColumn">A 2-D array which encodes sub-sampled example indexes of non-zero features stored in sampleValuePerColumn.
/// The sampleIndicesPerColumn[i][j]-th example has a non-zero i-th feature whose value is sampleValuePerColumn[i][j].</param>
/// <param name="numCol">Total number of features in the original data.</param>
/// <param name="sampleNonZeroCntPerColumn">sampleNonZeroCntPerColumn[i] is the size of sampleValuePerColumn[i].</param>
/// <param name="numSampleRow">The number of sampled examples in the sub-sampled data matrix.</param>
/// <param name="numTotalRow">The number of original examples added using <see cref="PushRows(float[], int, int, int)"/>.</param>
/// <param name="param">LightGBM parameter used in https://github.com/Microsoft/LightGBM/blob/c920e6345bcb41fc1ec6ac338f5437034b9f0d38/src/c_api.cpp#L421. </param>
/// <param name="labels">Labels of the original data. labels[i] is the label of the i-th original example.</param>
/// <param name="weights">Example weights of the original data. weights[i] is the weight of the i-th original example.</param>
/// <param name="groups">Group identifiers of the original data. groups[i] is the group ID of the i-th original example.</param>
public unsafe Dataset(double[][] sampleValuePerColumn,
int[][] sampleIndicesPerColumn,
int numCol,
int[] sampleNonZeroCntPerColumn,
int numSampleRow,
int numTotalRow,
string param, float[] labels, float[] weights = null, int[] groups = null)
{
_handle = IntPtr.Zero;
// Use GCHandle to pin the memory, avoid the memory relocation.
GCHandle[] gcValues = new GCHandle[numCol];
GCHandle[] gcIndices = new GCHandle[numCol];
try
{
double *[] ptrArrayValues = new double *[numCol];
int *[] ptrArrayIndices = new int *[numCol];
for (int i = 0; i < numCol; i++)
{
gcValues[i] = GCHandle.Alloc(sampleValuePerColumn[i], GCHandleType.Pinned);
ptrArrayValues[i] = (double *)gcValues[i].AddrOfPinnedObject().ToPointer();
gcIndices[i] = GCHandle.Alloc(sampleIndicesPerColumn[i], GCHandleType.Pinned);
ptrArrayIndices[i] = (int *)gcIndices[i].AddrOfPinnedObject().ToPointer();
}
;
fixed(double **ptrValues = ptrArrayValues)
fixed(int **ptrIndices = ptrArrayIndices)
{
// Create container. Examples will pushed in later.
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetCreateFromSampledColumn(
(IntPtr)ptrValues, (IntPtr)ptrIndices, numCol, sampleNonZeroCntPerColumn, numSampleRow, numTotalRow,
param, ref _handle));
}
}
finally
{
for (int i = 0; i < numCol; i++)
{
if (gcValues[i].IsAllocated)
{
gcValues[i].Free();
}
if (gcIndices[i].IsAllocated)
{
gcIndices[i].Free();
}
}
;
}
// Before adding examples (i.e., feature vectors of the original data set), the original labels, weights, and groups are added.
SetLabel(labels);
SetWeights(weights);
SetGroup(groups);
Contracts.Assert(GetNumCols() == numCol);
Contracts.Assert(GetNumRows() == numTotalRow);
}
/// <summary>
/// Append examples to LightGBM dataset.
/// </summary>
/// <param name="data">Dense (# of rows)-by-(# of columns) matrix flattened in a row-major format. One row per example.
/// The value at the i-th row and j-th column is stored in data[j + i * (# of columns)].</param>
/// <param name="numRow"># of rows of the data matrix.</param>
/// <param name="numCol"># of columns of the data matrix.</param>
/// <param name="startRowIdx">The actual row index of the first row pushed in. If it's 36, the first row in data would be the 37th row in <see cref="Dataset"/>.</param>
public void PushRows(float[] data, int numRow, int numCol, int startRowIdx)
{
Contracts.Assert(startRowIdx == _lastPushedRowID);
Contracts.Assert(numCol == GetNumCols());
Contracts.Assert(numRow > 0);
Contracts.Assert(startRowIdx <= GetNumRows() - numRow);
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetPushRows(_handle, data, numRow, numCol, startRowIdx));
_lastPushedRowID = startRowIdx + numRow;
}
public void PushRows(int[] indPtr, int[] indices, float[] data, int nIndptr,
long numElem, int numCol, int startRowIdx)
{
Contracts.Assert(startRowIdx == _lastPushedRowID);
Contracts.Assert(numCol == GetNumCols());
Contracts.Assert(startRowIdx < GetNumRows());
LightGbmInterfaceUtils.Check(
WrappedLightGbmInterface.DatasetPushRowsByCsr(
_handle, indPtr, indices, data, nIndptr, numElem, numCol, startRowIdx));
_lastPushedRowID = startRowIdx + nIndptr - 1;
}
// Not used now. Can use for the continued train.
public unsafe void SetInitScore(double[] initScores)
{
if (initScores != null)
{
Contracts.Assert(initScores.Length % GetNumRows() == 0);
fixed(double *ptr = initScores)
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetSetField(_handle, "init_score", (IntPtr)ptr, initScores.Length,
WrappedLightGbmInterface.CApiDType.Float64));
}
}
private unsafe double Eval(int dataIdx)
{
if (!_hasMetric)
{
return(double.NaN);
}
int outLen = 0;
double[] res = new double[1];
fixed(double *ptr = res)
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterGetEval(Handle, dataIdx, ref outLen, ptr));
return(res[0]);
}
public Dataset(Dataset reference, int numTotalRow, float[] labels, float[] weights = null, int[] groups = null)
{
IntPtr refHandle = IntPtr.Zero;
if (reference != null)
{
refHandle = reference.Handle;
}
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetCreateByReference(refHandle, numTotalRow, ref _handle));
SetLabel(labels);
SetWeights(weights);
SetGroup(groups);
}
public unsafe void SetWeights(float[] weights)
{
if (weights != null)
{
Contracts.Assert(weights.Length == GetNumRows());
// Skip SetWeights if all weights are same.
bool allSame = true;
for (int i = 1; i < weights.Length; ++i)
{
if (weights[i] != weights[0])
{
allSame = false;
break;
}
}
if (!allSame)
{
fixed(float *ptr = weights)
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.DatasetSetField(_handle, "weight", (IntPtr)ptr, weights.Length,
WrappedLightGbmInterface.CApiDType.Float32));
}
}
}
public Booster(Dictionary <string, object> parameters, Dataset trainset, Dataset validset = null)
{
var param = LightGbmInterfaceUtils.JoinParameters(parameters);
var handle = IntPtr.Zero;
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterCreate(trainset.Handle, param, ref handle));
Handle = handle;
if (validset != null)
{
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterAddValidData(Handle, validset.Handle));
_hasValid = true;
}
int numEval = 0;
BestIteration = -1;
LightGbmInterfaceUtils.Check(WrappedLightGbmInterface.BoosterGetEvalCounts(Handle, ref numEval));
// At most one metric in ML.NET.
Contracts.Assert(numEval <= 1);
if (numEval == 1)
{
_hasMetric = true;
}
}
