|
public static For ( int start, int end, Action |
||
| start | int | |
| end | int | |
| loopAction | Action |
|
| return | void | |
/// <summary>
/// Responsible for population migration
/// </summary>
/// <param name="iteration"></param>
private void MigratePopulation(int iteration)
{
switch (iteration)
{
case 0:
// Migrate values between populations
Parallel.For(0, 5,
migrationIteration =>
_populationArray[migrationIteration].Migrate(_populationArray[5 + migrationIteration],
15, new EliteSelection()));
break;
case 1:
// Migrate values between populations
for (int i = 0; i < 10; i += 2)
{
_populationArray[i].Migrate(_populationArray[i + 1], 15, new EliteSelection());
}
break;
//case 2:
// // Migrate values between populations
// Parallel.For(0, 5,
// migrationIteration =>
// _populationArray[migrationIteration].Migrate(_populationArray[9 - migrationIteration],
// 15, new EliteSelection()));
// break;
}
}
private void split(DecisionNode root, double[][] input, int[] output)
{
// 2. If all examples are for the same class, return the single-node
// tree with the output label corresponding to this common class.
double entropy = Statistics.Tools.Entropy(output, outputClasses);
if (entropy == 0)
{
if (output.Length > 0)
root.Output = output[0];
return;
}
// 3. If number of predicting attributes is empty, then return the single-node
// tree with the output label corresponding to the most common value of
// the target attributes in the examples.
int predictors = attributes.Count(x => x == false);
if (predictors <= attributes.Length - maxHeight)
{
root.Output = Statistics.Tools.Mode(output);
return;
}
// 4. Otherwise, try to select the attribute which
// best explains the data sample subset.
double[] scores = new double[predictors];
double[] entropies = new double[predictors];
double[] thresholds = new double[predictors];
int[][][] partitions = new int[predictors][][];
// Retrieve candidate attribute indices
int[] candidates = new int[predictors];
for (int i = 0, k = 0; i < attributes.Length; i++)
if (!attributes[i]) candidates[k++] = i;
// For each attribute in the data set
#if SERIAL
for (int i = 0; i < scores.Length; i++)
#else
Parallel.For(0, scores.Length, i =>
#endif
{
scores[i] = computeGainRatio(input, output, candidates[i],
entropy, out partitions[i], out thresholds[i]);
}
#if !SERIAL
);
#endif
// Select the attribute with maximum gain ratio
int maxGainIndex; scores.Max(out maxGainIndex);
var maxGainPartition = partitions[maxGainIndex];
var maxGainEntropy = entropies[maxGainIndex];
var maxGainAttribute = candidates[maxGainIndex];
var maxGainRange = inputRanges[maxGainAttribute];
var maxGainThreshold = thresholds[maxGainIndex];
// Mark this attribute as already used
attributes[maxGainAttribute] = true;
double[][] inputSubset;
int[] outputSubset;
// Now, create next nodes and pass those partitions as their responsibilities.
if (tree.Attributes[maxGainAttribute].Nature == DecisionVariableKind.Discrete)
{
// This is a discrete nature attribute. We will branch at each
// possible value for the discrete variable and call recursion.
DecisionNode[] children = new DecisionNode[maxGainPartition.Length];
// Create a branch for each possible value
for (int i = 0; i < children.Length; i++)
{
children[i] = new DecisionNode(tree)
{
Parent = root,
Value = i + maxGainRange.Min,
Comparison = ComparisonKind.Equal,
};
inputSubset = input.Submatrix(maxGainPartition[i]);
outputSubset = output.Submatrix(maxGainPartition[i]);
split(children[i], inputSubset, outputSubset); // recursion
}
root.Branches.AttributeIndex = maxGainAttribute;
root.Branches.AddRange(children);
}
else if (maxGainPartition.Length > 1)
{
// This is a continuous nature attribute, and we achieved two partitions
// using the partitioning scheme. We will branch on two possible settings:
// either the value is higher than a currently detected optimal threshold
// or it is lesser.
DecisionNode[] children =
{
new DecisionNode(tree)
{
Parent = root, Value = maxGainThreshold,
Comparison = ComparisonKind.LessThanOrEqual
},
new DecisionNode(tree)
{
Parent = root, Value = maxGainThreshold,
Comparison = ComparisonKind.GreaterThan
}
};
// Create a branch for lower values
inputSubset = input.Submatrix(maxGainPartition[0]);
outputSubset = output.Submatrix(maxGainPartition[0]);
split(children[0], inputSubset, outputSubset);
// Create a branch for higher values
inputSubset = input.Submatrix(maxGainPartition[1]);
outputSubset = output.Submatrix(maxGainPartition[1]);
split(children[1], inputSubset, outputSubset);
root.Branches.AttributeIndex = maxGainAttribute;
root.Branches.AddRange(children);
}
else
{
// This is a continuous nature attribute, but all variables are equal
// to a constant. If there is only a constant value as the predictor
// and there are multiple output labels associated with this constant
// value, there isn't much we can do. This node will be a leaf.
// We will set the class label for this node as the
// majority of the currently selected output classes.
outputSubset = output.Submatrix(maxGainPartition[0]);
root.Output = Statistics.Tools.Mode(outputSubset);
}
attributes[maxGainAttribute] = false;
}
private void split(DecisionNode root, double[][] input, int[] output)
{
// 2. If all examples are for the same class, return the single-node
// tree with the output label corresponding to this common class.
double entropy = Statistics.Tools.Entropy(output, outputClasses);
if (entropy == 0)
{
if (output.Length > 0)
{
root.Output = output[0];
}
return;
}
// 3. If number of predicting attributes is empty, then return the single-node
// tree with the output label corresponding to the most common value of
// the target attributes in the examples.
int predictors = attributes.Count(x => x == false);
if (predictors == 0)
{
root.Output = Statistics.Tools.Mode(output);
return;
}
// 4. Otherwise, try to select the attribute which
// best explains the data sample subset.
double[] scores = new double[predictors];
double[] entropies = new double[predictors];
double[] thresholds = new double[predictors];
int[][][] partitions = new int[predictors][][];
// Retrieve candidate attribute indices
int[] candidates = new int[predictors];
for (int i = 0, k = 0; i < attributes.Length; i++)
{
if (!attributes[i])
{
candidates[k++] = i;
}
}
// For each attribute in the data set
Parallel.For(0, scores.Length, i =>
{
scores[i] = computeGainRatio(input, output, candidates[i],
entropy, out partitions[i], out thresholds[i]);
});
// Select the attribute with maximum gain ratio
int maxGainIndex; scores.Max(out maxGainIndex);
var maxGainPartition = partitions[maxGainIndex];
var maxGainEntropy = entropies[maxGainIndex];
var maxGainAttribute = candidates[maxGainIndex];
var maxGainRange = inputRanges[maxGainAttribute];
var maxGainThreshold = thresholds[maxGainIndex];
// Mark this attribute as already used
attributes[maxGainAttribute] = true;
double[][] inputSubset;
int[] outputSubset;
// Now, create next nodes and pass those partitions as their responsabilities.
if (tree.Attributes[maxGainAttribute].Nature == DecisionAttributeKind.Discrete)
{
DecisionNode[] children = new DecisionNode[maxGainPartition.Length];
for (int i = 0; i < children.Length; i++)
{
children[i] = new DecisionNode(tree)
{
Parent = root,
Value = i + maxGainRange.Min,
Comparison = ComparisonKind.Equal,
};
inputSubset = input.Submatrix(maxGainPartition[i]);
outputSubset = output.Submatrix(maxGainPartition[i]);
split(children[i], inputSubset, outputSubset); // recursion
}
root.Branches = new DecisionBranchNodeCollection(maxGainAttribute, children);
}
else
{
DecisionNode[] children =
{
new DecisionNode(tree)
{
Parent = root, Value = maxGainThreshold,
Comparison = ComparisonKind.LessThanOrEqual
},
new DecisionNode(tree)
{
Parent = root, Value = maxGainThreshold,
Comparison = ComparisonKind.GreaterThan
}
};
inputSubset = input.Submatrix(maxGainPartition[0]);
outputSubset = output.Submatrix(maxGainPartition[0]);
split(children[0], inputSubset, outputSubset);
inputSubset = input.Submatrix(maxGainPartition[1]);
outputSubset = output.Submatrix(maxGainPartition[1]);
split(children[1], inputSubset, outputSubset);
root.Branches = new DecisionBranchNodeCollection(maxGainAttribute, children);
}
attributes[maxGainAttribute] = false;
}
private void split(DecisionNode root, double[][] input, int[] output, int height)
{
// 2. If all examples are for the same class, return the single-node
// tree with the output label corresponding to this common class.
double entropy = Measures.Entropy(output, outputClasses);
if (entropy == 0)
{
if (output.Length > 0)
{
root.Output = output[0];
}
return;
}
// 3. If number of predicting attributes is empty, then return the single-node
// tree with the output label corresponding to the most common value of
// the target attributes in the examples.
// how many variables have been used less than the limit
int[] candidates = Matrix.Find(attributeUsageCount, x => x < join);
if (candidates.Length == 0 || (maxHeight > 0 && height == maxHeight))
{
root.Output = Measures.Mode(output);
return;
}
// 4. Otherwise, try to select the attribute which
// best explains the data sample subset. If the tree
// is part of a random forest, only consider a percentage
// of the candidate attributes at each split point
if (MaxVariables > 0)
{
candidates = Vector.Sample(candidates, MaxVariables);
}
var scores = new double[candidates.Length];
var thresholds = new double[candidates.Length];
var partitions = new int[candidates.Length][][];
// For each attribute in the data set
Parallel.For(0, scores.Length, ParallelOptions, i =>
{
scores[i] = computeGainRatio(input, output, candidates[i],
entropy, out partitions[i], out thresholds[i]);
});
// Select the attribute with maximum gain ratio
int maxGainIndex; scores.Max(out maxGainIndex);
var maxGainPartition = partitions[maxGainIndex];
var maxGainAttribute = candidates[maxGainIndex];
var maxGainRange = inputRanges[maxGainAttribute];
var maxGainThreshold = thresholds[maxGainIndex];
// Mark this attribute as already used
attributeUsageCount[maxGainAttribute]++;
double[][] inputSubset;
int[] outputSubset;
// Now, create next nodes and pass those partitions as their responsibilities.
if (tree.Attributes[maxGainAttribute].Nature == DecisionVariableKind.Discrete)
{
// This is a discrete nature attribute. We will branch at each
// possible value for the discrete variable and call recursion.
DecisionNode[] children = new DecisionNode[maxGainPartition.Length];
// Create a branch for each possible value
for (int i = 0; i < children.Length; i++)
{
children[i] = new DecisionNode(tree)
{
Parent = root,
Value = i + maxGainRange.Min,
Comparison = ComparisonKind.Equal,
};
inputSubset = input.Get(maxGainPartition[i]);
outputSubset = output.Get(maxGainPartition[i]);
split(children[i], inputSubset, outputSubset, height + 1); // recursion
}
root.Branches.AttributeIndex = maxGainAttribute;
root.Branches.AddRange(children);
}
else if (maxGainPartition.Length > 1)
{
// This is a continuous nature attribute, and we achieved two partitions
// using the partitioning scheme. We will branch on two possible settings:
// either the value is greater than a currently detected optimal threshold
// or it is less.
DecisionNode[] children =
{
new DecisionNode(tree)
{
Parent = root, Value = maxGainThreshold,
Comparison = ComparisonKind.LessThanOrEqual
},
new DecisionNode(tree)
{
Parent = root, Value = maxGainThreshold,
Comparison = ComparisonKind.GreaterThan
}
};
// Create a branch for lower values
inputSubset = input.Get(maxGainPartition[0]);
outputSubset = output.Get(maxGainPartition[0]);
split(children[0], inputSubset, outputSubset, height + 1);
// Create a branch for higher values
inputSubset = input.Get(maxGainPartition[1]);
outputSubset = output.Get(maxGainPartition[1]);
split(children[1], inputSubset, outputSubset, height + 1);
root.Branches.AttributeIndex = maxGainAttribute;
root.Branches.AddRange(children);
}
else
{
// This is a continuous nature attribute, but all variables are equal
// to a constant. If there is only a constant value as the predictor
// and there are multiple output labels associated with this constant
// value, there isn't much we can do. This node will be a leaf.
// We will set the class label for this node as the
// majority of the currently selected output classes.
outputSubset = output.Get(maxGainPartition[0]);
root.Output = Measures.Mode(outputSubset);
}
attributeUsageCount[maxGainAttribute]--;
}
void InitializePCpu()
{
int N = cc.c.N;
const float DistanceScale = 100.0f;
const float eps = 2.22e-16f;
int bandSize = Math.Min(N, MaxGroupNumberHyp * GroupSizeHyp);
PBuf = gpu.CreateBufferRW(bandSize * N, 4, 1);
P2Buf = gpu.CreateBufferDynamic(bandSize * N, 4, 7); // dynamic buffer for fast uploading. Linked to Pcpu[] on HLSL.
int blockSize = 128; // Calculate so many rows per dispatch.
cpuP = new float[N][];
for (int i = 0; i < N; i++)
{
cpuP[i] = new float[N];
}
using (var distanceBuf = gpu.CreateBufferRW(blockSize * N, 4, 0))
using (var stagingBuf = gpu.CreateStagingBuffer(distanceBuf))
using (var sd = gpu.LoadShader("TsneDx.PartialDistance2.cso")) {
gpu.SetShader(sd);
for (int iBlock = 0; iBlock < N; iBlock += blockSize)
{
cc.c.blockIdx = iBlock;
cc.Upload();
gpu.Run(blockSize);
int iBlock2 = Math.Min(iBlock + blockSize, N);
int blockLen = (iBlock2 * (iBlock2 - 1) - iBlock * (iBlock - 1)) / 2;
float[] ret = gpu.ReadRange <float>(stagingBuf, distanceBuf, blockLen);
int idx = 0;
for (int row = iBlock; row < iBlock2; row++)
{
Array.Copy(ret, idx, cpuP[row], 0, row);
idx += row;
}
}
}
double distanceFactor = double.MinValue;
MT.For(1, N, i => {
float maxV = cpuP[i].Max();
lock (this)
distanceFactor = Math.Max(distanceFactor, maxV);
});
if (distanceFactor == 0)
{
throw new System.Exception("Distance metric degenerated: all components are zero.");
}
// Scale the distance to managable range [0, 100.0] to avoid degredation
// with exp function.
distanceFactor = DistanceScale / distanceFactor;
MT.For(1, N, i => {
for (int j = 0; j < i; j++)
{
cpuP[i][j] = (float)(cpuP[i][j] * distanceFactor);
}
});
MT.For(0, N, i => {
for (int j = 0; j < i; j++)
{
cpuP[j][i] = cpuP[i][j];
}
cpuP[i][i] = 0;
});
int bSize = MaxGroupNumberHyp * GroupSizeHyp;
using (var sd = gpu.LoadShader("TsneDx.Dist2Affinity.cso"))
using (var stagingBuf = gpu.CreateStagingBuffer(PBuf)) {
gpu.SetShader(sd);
for (int iBlock = 0; iBlock < N; iBlock += bSize)
{
cc.c.blockIdx = iBlock;
cc.Upload();
int iBlock2 = Math.Min(N, iBlock + bSize);
using (var ws = gpu.NewWriteStream(PBuf))
for (int row = iBlock; row < iBlock2; row++)
{
ws.WriteRange(cpuP[row]);
}
gpu.Run(MaxGroupNumberHyp);
using (var rs = gpu.NewReadStream(stagingBuf, PBuf))
for (int row = iBlock; row < iBlock2; row++)
{
rs.ReadRange(cpuP[row], 0, N);
}
}
}
double sum = 0;
MT.For(0, N, i => {
double sum2 = 0.0;
for (int j = i + 1; j < N; j++)
{
cpuP[i][j] += cpuP[j][i];
sum2 += cpuP[i][j];
}
lock (this)
sum += sum2;
});
if (sum == 0)
{
throw new System.Exception("Perplexity too small!");
}
sum *= 2;
MT.For(0, N, i => {
for (int j = i + 1; j < N; j++)
{
cpuP[i][j] = (float)Math.Max(cpuP[i][j] / sum, eps);
cpuP[j][i] = cpuP[i][j];
}
cpuP[i][i] = 1.0f;
});
}