// 2m0rr0w2 , local pca partion, check all dims [START]
private bool getPcaPartitionAllDim(ref int dimIndex, ref int partitionId, ref double partitionValue
, GeoWave parentNode, double error, ref List <int> sortedIds)
{
//construct PCA to parent node and save original (before transformation) data matrix
double[][] originalNodeData = DimReduction.constructNodePca(_trainingDt, parentNode);
double[][] transformedData = parentNode.localPca.Transform(originalNodeData);
parentNode.pcaDim = parentNode.localPca.Components.Count;
//break if dimention equivalent to num of points
if (parentNode.pcaDim == transformedData.Count())
{
return(false);
}
//save dim of transformed data
int pcaDim = parentNode.localPca.Components.Count;
double[] errorEachDim = new double[pcaDim];
int[] partitionIdEachDim = new int[pcaDim];
// _rc.dim replaced by index of last component
Helpers.applyFor(0, pcaDim, dim =>
{
errorEachDim[dim] = getPcaPartitionSingleDim(dim, transformedData, parentNode, partitionIdEachDim);
});
dimIndex = Enumerable.Range(0, pcaDim)
.Aggregate((a, b) => (errorEachDim[a] < errorEachDim[b]) ? a : b);
partitionId = partitionIdEachDim[dimIndex];
//save id's order in transformed data at best dimention
int bestDim = dimIndex;
sortedIds = new List <int>(parentNode.pointsIdArray); // will be sorted at best split dimention
List <int> idsClone = new List <int>(sortedIds); // id's in original position
sortedIds.Sort((c1, c2) =>
transformedData[idsClone.IndexOf(c1)][bestDim].CompareTo(transformedData[idsClone.IndexOf(c2)][bestDim]));
//save partition value
//start test
int originalSplitLocation = idsClone.IndexOf(partitionId);
if (originalSplitLocation == -1)
{
return(false);
}
//end test
partitionValue = transformedData[originalSplitLocation][bestDim];
return(errorEachDim[dimIndex] < error);
}
private void btnScript_Click(object sender, EventArgs e, StreamWriter statusWriter)
{
set2Config();
Refresh();
u_config.printConfig(@"C:\Wavelets decomposition\config.txt", null);
AmazonS3Client client = Helpers.configAmazonS3ClientS3Client();
UseS3 = UseS3CB.Checked;
rumPrallel = rumPrallelCB.Checked;
runBoosting = runBoostingCB.Checked;
runProoning = runProoningCB.Checked;
runBoostingProoning = runBoostingProoningCB.Checked;
runRFProoning = runRFProoningCB.Checked;
runRf = runRfCB.Checked;
runBoostingLearningRate = runBoostingLearningRateCB.Checked;
bucketName = bucketTB.Text;
string results_path = @ResultsTB.Text;
string db_path = @DBTB.Text + "\\";//@"C:\Users\Administrator\Dropbox\ADA\ada_valid\"; //"D:\\Phd\\Shai\\code\\tests\\helix tests\\noise_5\\noise_5\\"; // "C:\\reasearch\\tests\\lena\\";
//get dir
MainFolderName = results_path;
Helpers.createMainDirectoryOrResultPath(results_path, bucketName, client);
//READ DATA
DB db = new DB();
db.training_dt = db.getDataTable(db_path + "trainingData.txt");
db.testing_dt = db.getDataTable(db_path + "testingData.txt");
db.validation_dt = db.getDataTable(db_path + "ValidData.txt");
db.training_label = db.getDataTable(db_path + "trainingLabel.txt");
db.testing_label = db.getDataTable(db_path + "testingLabel.txt");
db.validation_label = db.getDataTable(db_path + "ValidLabel.txt");
upper_label = db.training_label.Max();
lower_label = db.training_label.Min();
double trainingPercent = double.Parse(trainingPercentTB.Text); // 0.02;
long rowToRemoveFrom = Convert.ToInt64(db.training_dt.Count() * trainingPercent);
db.training_dt = db.training_dt.Where((el, i) => i < rowToRemoveFrom).ToArray();
db.training_label = db.training_label.Where((el, i) => i < rowToRemoveFrom).ToArray();
db.testing_dt = db.testing_dt.Where((el, i) => i < rowToRemoveFrom).ToArray();
db.testing_label = db.testing_label.Where((el, i) => i < rowToRemoveFrom).ToArray();
db.validation_dt = db.training_dt.Where((el, i) => i < rowToRemoveFrom).ToArray();
db.validation_label = db.validation_label.Where((el, i) => i < rowToRemoveFrom).ToArray();
//REDUCE DIM, GLOBAL PCA
if (usePCA.Checked)
{
DimReduction dimreduction = new DimReduction(db.training_dt);
db.PCAtraining_dt = dimreduction.getGlobalPca(db.training_dt);
db.PCAtesting_dt = dimreduction.getGlobalPca(db.testing_dt);
db.PCAvalidation_dt = dimreduction.getGlobalPca(db.validation_dt);
}
else
{
//de-activate pca for dbg
db.PCAtraining_dt = db.training_dt;
db.PCAtesting_dt = db.testing_dt;
db.PCAvalidation_dt = db.validation_dt;
}
db.PCAtraining_GridIndex_dt = new long[db.PCAtraining_dt.Count()][];
for (int i = 0; i < db.PCAtraining_dt.Count(); i++)
{
db.PCAtraining_GridIndex_dt[i] = new long[db.PCAtraining_dt[i].Count()];
}
//BOUNDING BOX AND MAIN GRID
boundingBox = db.getboundingBox(db.PCAtraining_dt);
MainGrid = db.getMainGrid(db.PCAtraining_dt, boundingBox, ref db.PCAtraining_GridIndex_dt);
//READ CONFIG
methodConfig mc = new methodConfig(true);
int Nloops = int.Parse(NloopsTB.Text) - 1;
int Kfolds = 0;
if (int.TryParse(croosValidTB.Text, out Kfolds))
{
Nloops = Kfolds - 1;
}
for (int k = 0; k < Nloops; k++)
{
mc.boostlamda_0.Add(3.8); // - create variant in number of pixels
}
//mc.boostlamda_0.Add(1500);// - create variant in number of pixels
//mc.boostlamda_0.Add(2500);// - create variant in number of pixels
//mc.boostlamda_0.Add(3000);// - create variant in number of pixels
mc.generateRecordConfigArr();
for (int k = 0; k < mc.recArr.Count(); k++)
{
mc.recArr[k].dim = NfeaturesTB.Text == @"all" ? db.PCAtraining_dt[0].Count() : int.Parse(evaluateString(NfeaturesTB.Text, k));
mc.recArr[k].approxThresh = double.Parse(evaluateString(approxThreshTB.Text, k)); // 0.1;
mc.recArr[k].partitionErrType = int.Parse(evaluateString(partitionTypeTB.Text, k)); //2;
mc.recArr[k].minWaveSize = int.Parse(evaluateString(minNodeSizeTB.Text, k)); //1;//CHANGE AFTER DBG
mc.recArr[k].hopping_size = int.Parse(evaluateString(waveletsSkipEstimationTB.Text, k)); //25;// 10 + 5 * (k + 1);// +5 * (k % 10);// 1;//25;
mc.recArr[k].test_error_size = double.Parse(evaluateString(waveletsPercentEstimationTB.Text, k)); // +0.05 * (k % 10);// 1;// 0.1;//percent of waves to check
mc.recArr[k].NskipsinKfunc = double.Parse(evaluateString(boostingKfuncPercentTB.Text, k)); // 0.0025;
mc.recArr[k].rfBaggingPercent = double.Parse(evaluateString(bagginPercentTB.Text, k)); // 0.6;
mc.recArr[k].rfNum = int.Parse(evaluateString(NrfTB.Text, k)); // k + 1;//10 + k*10;// 100 / (k + 46) * 2;// int.Parse(Math.Pow(10, k + 1).ToString());
mc.recArr[k].boostNum = int.Parse(evaluateString(NboostTB.Text, k)); // 10;
mc.recArr[k].boostProoning_0 = int.Parse(evaluateString(NfirstPruninginBoostingTB.Text, k)); //13
mc.recArr[k].boostlamda_0 = double.Parse(evaluateString(boostingLamda0TB.Text, k)); // 0.01 - (k + 1) * 0.001; //0.05;// 0.0801 + k * 0.001;// Math.Pow(0.1, k);// 0.22 + k*0.005;
mc.recArr[k].NwaveletsBoosting = int.Parse(evaluateString(NfirstwaveletsBoostingTB.Text, k)); // 4;// k + 1;
//mc.recArr[k].learningRate = 0;// 0.01;
mc.recArr[k].boostNumLearningRate = int.Parse(evaluateString(NboostingLearningRateTB.Text, k)); // 55;// 18;
mc.recArr[k].percent_training_db = trainingPercent;
mc.recArr[k].BoundLevel = int.Parse(evaluateString(boundLevelTB.Text, k)); //1024;
mc.recArr[k].NDimsinRF = NfeaturesrfTB.Text == @"all" ? db.PCAtraining_dt[0].Count() : int.Parse(evaluateString(NfeaturesrfTB.Text, k));
mc.recArr[k].split_type = int.Parse(evaluateString(splitTypeTB.Text, k)); //0
mc.recArr[k].NormLPType = int.Parse(evaluateString(errTypeEstimationTB.Text, k));
mc.recArr[k].RFpruningTestRange[1] = int.Parse(evaluateString(RFpruningEstimationRange1TB.Text, k)); // 12;// k + 9;
mc.recArr[k].boundDepthTree = int.Parse(evaluateString(boundDepthTB.Text, k)); //1024;
mc.recArr[k].CrossValidFold = k;
// 2m0rr0w2 save labels dim in confif
mc.recArr[k].labelDim = db.training_label[0].Count();
//mc.recArr[k].boostNum = t ;// tmp to delete !!!!!!!
//mc.recArr[k].RFwaveletsTestRange[0] = 25;
//mc.recArr[k].RFwaveletsTestRange[1] = 50;
}
Helpers.createOutputDirectories(mc.recArr, client, u_config, bucketName, results_path);
//SET ID ARRAY LIST
List <int> trainingID = Enumerable.Range(0, db.PCAtraining_dt.Count()).ToList();
List <int> testingID = Enumerable.Range(0, db.PCAtesting_dt.Count()).ToList();
//cross validation
List <List <int> > trainingFoldId = new List <List <int> >();
List <List <int> > testingFoldId = new List <List <int> >();
Random ran = new Random(2);
List <int> training_rand = trainingID.OrderBy(x => ran.Next()).ToList().GetRange(0, trainingID.Count);
//THE LARGEST GROUP IS TRAINING
if (int.TryParse(croosValidTB.Text, out Kfolds))
{
createCrossValid(Kfolds, training_rand, trainingFoldId, testingFoldId);
}
//bounding intervals
int[][] BB = new int[2][];
BB[0] = new int[boundingBox[0].Count()];
BB[1] = new int[boundingBox[0].Count()];
for (int i = 0; i < boundingBox[0].Count(); i++)
{
BB[1][i] = MainGrid[i].Count() - 1; //set last index in each dim
}
for (int i = 0; i < mc.recArr.Count; i++)
{
Analizer Analizer = new Analizer(MainFolderName + "\\" + mc.recArr[i].getShortName(), MainGrid, db, mc.recArr[i]);
if (!croosValidCB.Checked)
{
Analizer.analize(trainingID, testingID, BB);
}
else
{
Analizer.analize(trainingFoldId[i], testingFoldId[i], BB); //cross validation
}
statusWriter.WriteLine("fold " + i + " ready!!!!");
}
//btnScript.BackColor = Color.Green;
}
/* private void recursiveBSP_WaveletsByConsts(List<GeoWave> geoWaveArr, int geoWaveId, int seed=0)
* {
* //CALC APPROX_SOLUTION FOR GEO WAVE
* double error = geoWaveArr[geoWaveId].calc_MeanValueReturnError(_trainingLabel, geoWaveArr[geoWaveId].pointsIdArray);
* if (error < _rc.approxThresh ||
* geoWaveArr[geoWaveId].pointsIdArray.Count() <= _rc.minWaveSize ||
* _rc.boundDepthTree <= geoWaveArr[geoWaveId].level)
* return;
*
* int dimIndex = -1;
* int Maingridindex = -1;
*
* bool IsPartitionOK = false;
* switch (_rc.split_type)
* {
* case 0:
* IsPartitionOK = getBestPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, _dime2Take);
* break;
* case 1:
* IsPartitionOK = getRandPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, seed);
* break;
* case 2:
* {
* Random ran1 = new Random(seed);
* Random ran2 = new Random(geoWaveId);
* int one = ran1.Next(0, int.MaxValue / 10);
* int two = ran2.Next(0, int.MaxValue / 10);
* bool[] Dim2TakeNode = getDim2Take(_rc, one + two);
* IsPartitionOK = getBestPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, Dim2TakeNode);
* }
* break;
* case 3:
* IsPartitionOK = getGiniPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, _dime2Take);
* break;
* case 4:
* {
* Random ran1 = new Random(seed);
* Random ran2 = new Random(geoWaveId);
* int one = ran1.Next(0, int.MaxValue / 10);
* int two = ran2.Next(0, int.MaxValue / 10);
* bool[] Dim2TakeNode = getDim2Take(_rc, one + two);
* IsPartitionOK = getGiniPartitionResult(ref dimIndex, ref Maingridindex, geoWaveArr, geoWaveId, error, Dim2TakeNode);
* }
* break;
*
*
* }
*
*
*
*
* if (!IsPartitionOK)
* return;
*
*
* GeoWave child0 = new GeoWave(geoWaveArr[geoWaveId].boubdingBox, _trainingLabel[0].Count(), geoWaveArr[geoWaveId].rc);
* GeoWave child1 = new GeoWave(geoWaveArr[geoWaveId].boubdingBox, _trainingLabel[0].Count(), geoWaveArr[geoWaveId].rc);
*
* //set partition
* child0.boubdingBox[1][dimIndex] = Maingridindex;
* child1.boubdingBox[0][dimIndex] = Maingridindex;
*
* //DOCUMENT ON CHILDREN
* child0.dimIndex = dimIndex;
* child0.Maingridindex = Maingridindex;
* child1.dimIndex = dimIndex;
* child1.Maingridindex = Maingridindex;
*
* child0.MaingridValue = Form1.MainGrid[dimIndex][Maingridindex];
* child1.MaingridValue = Form1.MainGrid[dimIndex][Maingridindex];
*
* //calc norm
* //calc mean value
*
* if (Form1.isBoxSingular(child0.boubdingBox, _rc.dim) || Form1.isBoxSingular(child1.boubdingBox, _rc.dim))
* return;
*
* //SHOULD I VERIFY THAT THE CHILD IS NOT ITS PARENT ? (IN CASES WHERE CAN'T MODEFY THE PARTITION)
*
* setChildrensPointsAndMeanValue(ref child0, ref child1, dimIndex, geoWaveArr[geoWaveId].pointsIdArray);
* //SET TWO CHILDS
* child0.parentID = child1.parentID = geoWaveId;
* child0.child0 = child1.child0 = -1;
* child0.child1 = child1.child1 = -1;
* child0.level = child1.level = geoWaveArr[geoWaveId].level + 1;
*
* child0.computeNormOfConsts(geoWaveArr[geoWaveId]);
* child1.computeNormOfConsts(geoWaveArr[geoWaveId]);
* geoWaveArr.Add(child0);
* geoWaveArr.Add(child1);
* geoWaveArr[geoWaveId].child0 = geoWaveArr.Count - 2;
* geoWaveArr[geoWaveId].child1 = geoWaveArr.Count - 1;
*
*
*
*
* //RECURSION STEP !!!
* recursiveBSP_WaveletsByConsts(geoWaveArr, geoWaveArr[geoWaveId].child0, seed);
* recursiveBSP_WaveletsByConsts(geoWaveArr, geoWaveArr[geoWaveId].child1, seed);
* }
*/
private SplitProps getTransformedPartitionAllDim(GeoWave parentNode, double error, SplitType splitType)
{
double[][] originalNodeData = parentNode.pointsIdArray.Select(id => _trainingDt[id]).ToArray();
double[][] transformedData;
//clean columns of categorical variables 2m0rr0w2
// originalNodeData = Helpers.copyAndRemoveCategoricalColumns(originalNodeData, _rc);
//result struct
SplitProps resultProps = new SplitProps();
switch (splitType)
{
case SplitType.LocalPca:
DimReduction.constructNodePcaByOriginalData(originalNodeData, parentNode);
transformedData = parentNode.localPca.Transform(originalNodeData);
break;
case SplitType.DiffMaps5Percent:
if (originalNodeData.Count() <= _rc.dim)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
transformedData = DiffusionMaps.getTransformedMatrix(originalNodeData, 0.05);
break;
case SplitType.DiffMaps1Percent:
if (originalNodeData.Count() <= _rc.dim)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
transformedData = DiffusionMaps.getTransformedMatrix(originalNodeData, 0.01);
break;
case SplitType.DiffMapsHalfPercent:
if (originalNodeData.Count() <= _rc.dim)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
transformedData = DiffusionMaps.getTransformedMatrix(originalNodeData, 0.005);
break;
case SplitType.MainAxes:
transformedData = originalNodeData;
break;
case SplitType.Categorical:
transformedData = originalNodeData;
break;
default:
transformedData = null;
break;
}
if (transformedData == null)
{
//throw new Exception("******TRANSFORMATION ERROR!!!");
resultProps.isPartitionOk = false;
Debug.WriteLine("*********Failed transformation");
Debug.WriteLine("*********Failed node size: " + parentNode.pointsIdArray.Count);
return(resultProps);
}
parentNode.transformedDim = transformedData.First().Length;
//save dim of transformed data
int transformedDim = parentNode.transformedDim;
double[] errorEachDim = new double[transformedDim];
int[] partitionIdEachDim = new int[transformedDim];
// _rc.dim replaced by transformedDim dimention
Helpers.applyFor(0, transformedDim, dim =>
{
errorEachDim[dim] = getTransformedDataPartitionSingleDim(dim, transformedData, parentNode, partitionIdEachDim);
});
int bestDim = Enumerable.Range(0, transformedDim)
.Aggregate((a, b) => (errorEachDim[a] < errorEachDim[b]) ? a : b);
resultProps.splitId = partitionIdEachDim[bestDim];
//save id's order in transformed data at best dimention
resultProps.sortedIds = new List <int>(parentNode.pointsIdArray); // will be sorted at best split dimention
List <int> idsClone = new List <int>(resultProps.sortedIds); // id's in original position
resultProps.sortedIds.Sort((c1, c2) =>
transformedData[idsClone.IndexOf(c1)][bestDim].CompareTo(transformedData[idsClone.IndexOf(c2)][bestDim]));
//save partition value
int originalSplitLocation = idsClone.IndexOf(resultProps.splitId);
if (originalSplitLocation == -1)
{
resultProps.isPartitionOk = false;
return(resultProps);
}
resultProps.isPartitionOk = (errorEachDim[bestDim] < error);
resultProps.splitValue = transformedData[originalSplitLocation][bestDim];
resultProps.error = errorEachDim[bestDim];
resultProps.type = splitType;
resultProps.dimIndex = bestDim;
//shift dimention if it was not categorical split 2m0rr0w2
/* foreach (int categoricalInd in _rc.indOfCategorical)
* {
* resultProps.dimIndex = (resultProps.dimIndex == categoricalInd)
* ? resultProps.dimIndex++
* : resultProps.dimIndex;
* }*/
return(resultProps);
}
