public static void CalculateNeabours <T>
(CudaContext context,
DeviceDataSet <T> teaching,
DeviceDataSet <T> test,
CudaDeviceVariable <int> calculatedNeabours,
int threadsPerBlock
) where T : struct
{
var kernel = context.LoadKernel("kernels/VectorReduction.ptx", "calculateNearestNeabours");
kernel.GridDimensions = test.length / threadsPerBlock + 1;
kernel.BlockDimensions = threadsPerBlock;
kernel.SetConstantVariable("testVectorsCount", test.length);
kernel.SetConstantVariable("teachingVectorsCount", teaching.length);
kernel.SetConstantVariable("attributeCount", teaching.attributeCount);
using (var deviceDistanceMemory =
new CudaDeviceVariable <float>(teaching.length * test.length))
{
kernel.Run(
teaching.vectors.DevicePointer,
test.vectors.DevicePointer,
deviceDistanceMemory.DevicePointer,
calculatedNeabours.DevicePointer
);
Thrust.sort_by_key_multiple(deviceDistanceMemory, calculatedNeabours, teaching.length, test.length);
}
}
public VectorReductionAccuracy(CudaContext context, DeviceDataSet <int> teaching, DeviceDataSet <int> test, int popSize)
{
this.teaching = teaching;
this.test = test;
this.popSize = popSize;
this.context = context;
calculatedNeabours = new CudaDeviceVariable <int>(teaching.length * test.length);
deviceAccuracy = new CudaDeviceVariable <float>(popSize);
Profiler.Start("calculate neabours");
Neabours.CalculateNeabours(context, teaching, test, calculatedNeabours, ThreadsPerBlock);
Profiler.Stop("calculate neabours");
accuracyKernel = context.LoadKernel("kernels/VectorReduction.ptx", "calculateAccuracy");
dim3 gridDimension = new dim3()
{
x = (uint)(test.length / ThreadsPerBlock + 1),
y = (uint)popSize,
z = 1
};
accuracyKernel.GridDimensions = gridDimension;
accuracyKernel.BlockDimensions = ThreadsPerBlock;
accuracyKernel.SetConstantVariable("testVectorsCount", test.length);
accuracyKernel.SetConstantVariable("teachingVectorsCount", teaching.length);
accuracyKernel.SetConstantVariable("attributeCount", teaching.attributeCount);
accuracyKernel.SetConstantVariable("genLength", teaching.length);
K = 3;
CountToPass = 2;
}
static void ReduceVectors(CudaDataSet <int> teaching, CudaDataSet <int> test)
{
using (CudaContext context = new CudaContext())
{
int popSize = 100;
DeviceDataSet <int> deviceTeaching = new DeviceDataSet <int>(teaching);
DeviceDataSet <int> deviceTest = new DeviceDataSet <int>(test);
FlattArray <byte> initialPopulation;
VectorReductionAccuracy acc = new VectorReductionAccuracy(context, deviceTeaching, deviceTest, popSize)
{
K = 5,
CountToPass = 3
};
VectorReductionFitness fitnessFunc =
new VectorReductionFitness(context, acc, popSize, deviceTeaching.length)
{
Alpha = 0.7f
};
//Drop3 drop = new Drop3();
//drop.CasheSize = 5;
//drop.K = 3;
//Profiler.Start("Drop3");
//var indexesToStay = drop.Apply(teaching, context);
//Profiler.Stop("Drop3");
//byte[] parrent = new byte[teaching.Vectors.GetLength(0)];
//foreach (var item in indexesToStay)
//{
// parrent[item] = 1;
//}
initialPopulation = CreateRandomPopulation(popSize, deviceTeaching.length);
//CreatePopulationBasedOnParent(parrent, popSize, 0.2f, 0.05f);
var d = new Evolutionary2(context, fitnessFunc, initialPopulation)
{
Elitism = 0.001f,
MutationRate = 0.001f
};
for (int i = 0; i < 30; i++)
{
Profiler.Start("iteration");
d.CreateNewPopulation();
Profiler.Stop("iteration");
}
var best = d.FindFitest();
Console.WriteLine(acc.GenAccuracy(best.index) / (float)deviceTest.length);
Console.WriteLine(fitnessFunc.GenLength(best.index));
Profiler.Print();
}
}
static void ReduceDimension(CudaDataSet <int> teaching, CudaDataSet <int> test)
{
using (CudaContext context = new CudaContext())
{
int popSize = 20;
int genLength = teaching.Vectors.GetLength(1);
DeviceDataSet <int> deviceTeaching = new DeviceDataSet <int>(teaching);
DeviceDataSet <int> deviceTest = new DeviceDataSet <int>(test);
FlattArray <byte> initialPopulation;
DimensionReductionAccuracy accuracy = new DimensionReductionAccuracy(context, deviceTeaching, deviceTest, popSize)
{
K = 3,
CountToPass = 2
};
DimensionReductionFitness fitnessFunc =
new DimensionReductionFitness(context, accuracy, popSize, genLength)
{
Alpha = 1f
};
Console.WriteLine(accuracy.BaseAccuracy());
Console.WriteLine();
initialPopulation = CreateRandomPopulation(popSize, deviceTeaching.attributeCount);
var d = new Evolutionary2(context, fitnessFunc, initialPopulation)
{
CrossOverRate = 0.5f,
MutationRate = 0.02f,
Elitism = 0.2f
};
for (int i = 0; i < 200; i++)
{
Profiler.Start("iteration");
d.CreateNewPopulation();
Profiler.Stop("iteration");
}
var best = d.FindFitest();
// Console.WriteLine(best.fitness);
var acc = accuracy.GenAccuracy(best.index);
var len = fitnessFunc.GenLength(best.index);
var gen = d.genGen(best.index);
foreach (var item in gen)
{
Console.Write(item + " ");
}
Console.WriteLine($"accuracy: {acc / (float)deviceTest.length} length: {len / (float)genLength}");
Console.WriteLine();
// Profiler.Print();
}
}
static void ReduceVectorsRegresion(CudaDataSet <float> teaching, CudaDataSet <float> test)
{
using (CudaContext context = new CudaContext())
{
int popSize = 100;
var deviceTeaching = new DeviceDataSet <float>(teaching);
var deviceTest = new DeviceDataSet <float>(test);
FlattArray <byte> initialPopulation;
var acc = new VectorReductionAccuracyRegresion(context, deviceTeaching, deviceTest, popSize)
{
K = 3,
CountToPass = 2
};
VectorReductionFitness fitnessFunc =
new VectorReductionFitness(context, acc, popSize, deviceTeaching.length)
{
Alpha = 0.7f
};
initialPopulation = CreateRandomPopulation(popSize, deviceTeaching.length);
var d = new Evolutionary2(context, fitnessFunc, initialPopulation)
{
Elitism = 0.1f,
MutationRate = 0.05f
};
for (int i = 0; i < 10; i++)
{
Profiler.Start("iteration");
d.CreateNewPopulation();
Profiler.Stop("iteration");
}
Console.WriteLine(acc.BaseAccuracy());
var best = d.FindFitest();
Console.WriteLine(acc.GenAccuracy(best.index));
Console.WriteLine(fitnessFunc.GenLength(best.index) / (float)deviceTeaching.length);
Profiler.Print();
}
}
public DimensionReductionAccuracy(
CudaContext context,
DeviceDataSet <int> teaching,
DeviceDataSet <int> test,
int popSize
)
{
this.popSize = popSize;
this.teaching = teaching;
this.test = test;
this.context = context;
accuracyKernel = context.LoadKernel
(
"kernels/dimensionsReductions.ptx",
"geneticKnn"
);
accuracyKernel.GridDimensions = new dim3()
{
x = (uint)(test.vectors.Size / ThreadsPerBlock) + 1,
y = (uint)popSize,
z = 1
};
accuracyKernel.BlockDimensions = ThreadsPerBlock;
K = 3;
CountToPass = 2;
accuracyKernel.SetConstantVariable("atributeCount", test.attributeCount);
accuracyKernel.SetConstantVariable("teachingVectorsCount", teaching.length);
accuracyKernel.SetConstantVariable("testVectorsCount", test.length);
accuracyKernel.SetConstantVariable("popSize", popSize);
accuracyKernel.DynamicSharedMemory = (uint)(test.attributeCount * sizeof(float));
saveCasheKernel = context.LoadKernel(
"kernels/dimensionsReductions.ptx",
"saveToCashe"
);
saveCasheKernel.GridDimensions = (popSize * 32) / ThreadsPerBlock + 1;
saveCasheKernel.BlockDimensions = ThreadsPerBlock;
saveCasheKernel.SetConstantVariable("atributeCount", teaching.attributeCount);
saveCasheKernel.SetConstantVariable("popSize", teaching.attributeCount);
readCasheKernel = context.LoadKernel(
"kernels/dimensionsReductions.ptx",
"readCashe"
);
readCasheKernel.GridDimensions = 1;
readCasheKernel.BlockDimensions = popSize;
readCasheKernel.SetConstantVariable("atributeCount", teaching.attributeCount);
casheTreeRoot = new Node()
{
mutex = 0,
one = (IntPtr)0,
zero = (IntPtr)0,
};
isInCashe = new CudaDeviceVariable <byte>(popSize);
accuracy = new CudaDeviceVariable <float>(popSize);
}
