public Test(string datasetName, Dataset set, double[] initScores)
: this(new ScoreTracker(datasetName, set, initScores))
{
}
protected abstract ScoreTracker ConstructScoreTracker(string name, Dataset set, double[] initScores);
public void SetTrainingData(Dataset trainData, double[] initTrainScores)
{
TrainingScores = ConstructScoreTracker("train", trainData, initTrainScores);
TrackedScores[0] = TrainingScores;
}
// REVIEW: When the FastTree appliation is decoupled with tree learner and boosting logic, this class should be removed.
public RandomForestOptimizer(TreeEnsemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
: base(ensemble, trainData, initTrainScores, gradientWrapper)
{
_gradientWrapper = gradientWrapper;
}
protected override ScoreTracker ConstructScoreTracker(string name, Dataset set, double[] initScores)
{
//REVIEW: This is not necessary. We can remove this by creating dummy scorer.
return(new ScoreTracker(name, set, initScores));
}
public ConjugateGradientDescent(TreeEnsemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
: base(ensemble, trainData, initTrainScores, gradientWrapper)
{
_currentDk = new double[trainData.NumDocs];
}
public RankingBaggingProvider(Dataset completeTrainingSet, int maxLeaves, int randomSeed, double trainFraction) :
base(completeTrainingSet, maxLeaves, randomSeed, trainFraction)
{
}
// thread worker
private void WinLossRangeWorkerFromScores(Dataset dataset, short[] labels, double[] scores, double[] result, int query, int threadIndex)
{
int begin = dataset.Boundaries[query];
int count = dataset.Boundaries[query + 1] - begin;
int[] permutation = _permutationBuffers[threadIndex];
DcgPermutationComparer comparer = _comparers[threadIndex];
// set values for the comparer
comparer.Scores = scores;
comparer.Labels = labels;
comparer.ScoresOffset = begin;
comparer.LabelsOffset = begin;
// calculate the permutation
Array.Copy(_oneTwoThree, permutation, count);
Array.Sort(permutation, 0, count, comparer);
int surplus = 0;
int maxsurplus = 0;
int maxsurpluspos = 0;
for (int t = 0; t < count; ++t)
{
if (labels[begin + permutation[t]] > 0 ||
labels[begin + permutation[t]] < 0)
{
surplus += labels[begin + permutation[t]];
}
else
{
surplus--;
}
if (surplus > maxsurplus)
{
maxsurplus = surplus;
maxsurpluspos = t;
}
if (t == 100)
{
Utils.InterlockedAdd(ref result[0], surplus);
}
if (t == 200)
{
Utils.InterlockedAdd(ref result[1], surplus);
}
if (t == 300)
{
Utils.InterlockedAdd(ref result[2], surplus);
}
if (t == 400)
{
Utils.InterlockedAdd(ref result[3], surplus);
}
if (t == 500)
{
Utils.InterlockedAdd(ref result[4], surplus);
}
if (t == 1000)
{
Utils.InterlockedAdd(ref result[5], surplus);
}
}
Utils.InterlockedAdd(ref result[6], maxsurplus);
Utils.InterlockedAdd(ref result[7], maxsurpluspos);
Utils.InterlockedAdd(ref result[8], count);
}
public GradientDescent(TreeEnsemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
: base(ensemble, trainData, initTrainScores)
{
_gradientWrapper = gradientWrapper;
_treeScores = new List <double[]>();
}
public Dataset GetSubDataset(int[] docIndices, bool destroyThisDataset, FileObjectStore <IntArrayFormatter> newBinsCache)
{
#endif
int[] queryIndices = docIndices.Select(d => DocToQuery[d]).ToArray();
ulong[] uniqueQueryIds = queryIndices.Distinct().Select(q => QueryIds[q]).ToArray();
// calculate boundaries
int[] boundaries = new int[uniqueQueryIds.Length + 1];
boundaries[0] = 0;
int queryIndex = 1;
for (int q = 1; q < queryIndices.Length; ++q)
{
if (queryIndices[q] != queryIndices[q - 1])
{
boundaries[queryIndex++] = q;
}
}
boundaries[uniqueQueryIds.Length] = queryIndices.Length;
// construct skeleton
DatasetSkeleton datasetSkeleton = new DatasetSkeleton(docIndices.Select(d => Ratings[d]).ToArray(),
boundaries,
uniqueQueryIds,
docIndices.Select(d => DocIds[d]).ToArray());
// create features
FeatureFlockBase[] features = new FeatureFlockBase[NumFlocks];
int[][] assignment = new int[][] { docIndices };
Parallel.For(0, NumFlocks, new ParallelOptions {
MaxDegreeOfParallelism = BlockingThreadPool.NumThreads
},
(int flockIndex) =>
{
#if !NO_STORE
GetSubDataset_ThreadWorker(features, flockIndex, assignment, destroyThisDataset, newBinsCache);
#else
GetSubDatasetThreadWorker(features, flockIndex, assignment, destroyThisDataset);
#endif
});
uint[] filteredDupeIds = null;
// Filter the dupe ids, if any
if (DupeIds != null)
{
uint[] dupeIds = DupeIds;
filteredDupeIds = docIndices.Select(i => dupeIds[i]).ToArray();
}
// auxiliary data
Dictionary <string, DatasetSkeletonQueryDocData> auxData = _datasetSkeleton.AuxiliaryData;
Dictionary <string, DatasetSkeletonQueryDocData> newAuxData = new Dictionary <string, DatasetSkeletonQueryDocData>();
foreach (KeyValuePair <string, DatasetSkeletonQueryDocData> pair in auxData)
{
newAuxData[pair.Key] = pair.Value.GetSubset(pair.Value.IsQueryLevel ? queryIndices.Distinct().ToArray() : docIndices);
}
datasetSkeleton.AuxiliaryData = newAuxData;
// create new Dataset
Dataset dataset = new Dataset(datasetSkeleton, features);
dataset.DupeIds = filteredDupeIds;
return(dataset);
}
/// <summary>
/// Constructs partitioning object based on the documents and RegressionTree splits
/// NOTE: It has been optimized for speed and multiprocs with 10x gain on naive LINQ implementation
/// </summary>
internal DocumentPartitioning(InternalRegressionTree tree, Dataset dataset)
: this(dataset.NumDocs, tree.NumLeaves)
{
using (Timer.Time(TimerEvent.DocumentPartitioningConstruction))
{
// figure out which leaf each document belongs to
// NOTE: break it up into NumThreads chunks. This minimizes the number of re-computations necessary in
// the row-wise indexer.
int innerLoopSize = 1 + dataset.NumDocs / BlockingThreadPool.NumThreads; // +1 is to make sure we don't have a few left over at the end
// figure out the exact number of chunks, needed in pathological cases when NumDocs < NumThreads
int numChunks = dataset.NumDocs / innerLoopSize;
if (dataset.NumDocs % innerLoopSize != 0)
{
++numChunks;
}
var perChunkDocumentLists = new List <int> [numChunks][];
// REVIEW: This partitioning doesn't look optimal.
// Probably make sence to investigate better ways of splitting data?
var actions = new Action[(int)Math.Ceiling(1.0 * dataset.NumDocs / innerLoopSize)];
var actionIndex = 0;
for (int docStart = 0; docStart < dataset.NumDocs; docStart += innerLoopSize)
{
var fromDoc = docStart;
var toDoc = Math.Min(docStart + innerLoopSize, dataset.NumDocs);
var chunkIndex = docStart / innerLoopSize;
actions[actionIndex++] = () =>
{
Contracts.Assert(perChunkDocumentLists[chunkIndex] == null);
var featureBins = dataset.GetFeatureBinRowwiseIndexer();
List <int>[] perLeafDocumentLists = Enumerable.Range(0, tree.NumLeaves)
.Select(x => new List <int>(innerLoopSize / tree.NumLeaves))
.ToArray();
for (int d = fromDoc; d < toDoc; d++)
{
int leaf = tree.GetLeaf(featureBins[d]);
perLeafDocumentLists[leaf].Add(d);
}
perChunkDocumentLists[chunkIndex] = perLeafDocumentLists;
};
}
Parallel.Invoke(new ParallelOptions {
MaxDegreeOfParallelism = BlockingThreadPool.NumThreads
}, actions);
// establish leaf starts and document counts
_leafCount = Enumerable.Range(0, tree.NumLeaves)
.Select(leaf => Enumerable.Range(0, perChunkDocumentLists.Length)
.Select(thread => perChunkDocumentLists[thread][leaf].Count)
.Sum())
.ToArray();
var cumulativeLength = _leafCount.CumulativeSum <int>().Take(tree.NumLeaves - 1);
_leafBegin = Enumerable.Range(0, 1).Concat(cumulativeLength).ToArray();
// move all documents that belong to the same leaf together
Contracts.Assert(_documents.Length == _leafBegin[tree.NumLeaves - 1] + _leafCount[tree.NumLeaves - 1]);
actions = new Action[tree.NumLeaves];
actionIndex = 0;
for (int leaf = 0; leaf < tree.NumLeaves; leaf++)
{
var l = leaf;
actions[actionIndex++] = () =>
{
int documentPos = _leafBegin[l];
for (int chunkIndex = 0; chunkIndex < perChunkDocumentLists.Length; chunkIndex++)
{
foreach (int d in perChunkDocumentLists[chunkIndex][l])
{
_documents[documentPos++] = d;
}
perChunkDocumentLists[chunkIndex][l] = null;
}
};
}
Parallel.Invoke(new ParallelOptions {
MaxDegreeOfParallelism = BlockingThreadPool.NumThreads
}, actions);
}
}
public AcceleratedGradientDescent(TreeEnsemble ensemble, Dataset trainData, double[] initTrainScores, IGradientAdjuster gradientWrapper)
: base(ensemble, trainData, initTrainScores, gradientWrapper)
{
UseFastTrainingScoresUpdate = false;
}
public ScoreTracker(string datasetName, Dataset set, double[] initScores)
{
Initialize(datasetName, set, initScores);
}
} //An Xk is an alias to scores
public AgdScoreTracker(string datsetName, Dataset set, double[] initScores)
: base(datsetName, set, initScores)
{
_k = 0;
YK = (double[])XK.Clone();
}
public LogLossObjectiveFunction(Dataset trainSet, LogLossCommandLineArgs cmd)
: base(trainSet, trainSet.Ratings, cmd)
{
_mode = cmd.loglossmode;
_coef = cmd.loglosscoef;
}
protected override ScoreTracker ConstructScoreTracker(string name, Dataset set, double[] initScores)
{
return(new ScoreTracker(name, set, initScores));
}
protected TreeLearner(Dataset trainData, int numLeaves)
{
TrainData = trainData;
NumLeaves = numLeaves;
Partitioning = new DocumentPartitioning(TrainData.NumDocs, numLeaves);
}
public WinLossSurplusObjectiveFunction(Dataset trainSet, short[] labels, WinLossSurplusCommandLineArgs cmd)
: base(trainSet, labels, cmd)
{
}