忍者ブログ

Memeplexes

プログラミング、3DCGとその他いろいろについて

GPUでニューラルネットワーク更新 (OpenCL)

ニューロン更新

Deep Learningに備えて色々やってみます。
Deep Learningは重いことがわかったので、GPUで計算をしましょう。

using Cloo;
using System.Linq;

class Program
{
    static void Main()
    {
        ComputePlatform platform = ComputePlatform.Platforms[0];
        ComputeDevice[] devices = platform
            .Devices
            .Where(d => d.Type == ComputeDeviceTypes.Gpu)
            .ToArray();
        ComputeContext context = new ComputeContext(
            devices,
            new ComputeContextPropertyList(platform),
            null,
            System.IntPtr.Zero
            );
        ComputeCommandQueue commandQueue = new ComputeCommandQueue(
            context,
            devices[0],
            ComputeCommandQueueFlags.None
            );
        ComputeProgram program = new ComputeProgram(
            context,
            System.IO.File.ReadAllText("myKernelProgram.cl")
            );
        try
        {
            program.Build(devices, null, null, System.IntPtr.Zero);
        }
        catch
        {
            System.Console.WriteLine(program.GetBuildLog(devices[0]));
        }

        ComputeKernel kernel = program.CreateKernel("updateVisibleNeurons");

        int visibleNeuronCount = 3;
        int hiddenNeuronCount = 2;
        ComputeBuffer<float> visibleNeuronValues = new ComputeBuffer<float>(
            context,
            ComputeMemoryFlags.ReadWrite,
            visibleNeuronCount
            );
        ComputeBuffer<float> visibleNeuronBiases = new ComputeBuffer<float>(
            context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            new float[visibleNeuronCount]
            );
        ComputeBuffer<float> weights = new ComputeBuffer<float>(
            context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            Enumerable.Range(0, visibleNeuronCount * hiddenNeuronCount).Select(i => (float)i).ToArray()
            );
        ComputeBuffer<float> hiddenNeuronValues = new ComputeBuffer<float>(
            context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            Enumerable.Range(0, hiddenNeuronCount).Select(i => (float)i).ToArray()
            );
        kernel.SetMemoryArgument(0, visibleNeuronValues);
        kernel.SetMemoryArgument(1, visibleNeuronBiases);
        kernel.SetMemoryArgument(2, weights);
        kernel.SetMemoryArgument(3, hiddenNeuronValues);
        kernel.SetValueArgument(4, visibleNeuronCount);
        kernel.SetValueArgument(5, hiddenNeuronCount);


        commandQueue.Execute(
            kernel,
            null,
            new long[] { visibleNeuronCount },
            new long[] { 1 },
            null
            );
        float[] resultVisibleNeuronValues = new float[visibleNeuronCount];
        commandQueue.ReadFromBuffer(visibleNeuronValues, ref resultVisibleNeuronValues, true, null);

        foreach (var number in resultVisibleNeuronValues)
        {
            System.Console.WriteLine(number);
        }

        hiddenNeuronValues.Dispose();
        weights.Dispose();
        visibleNeuronBiases.Dispose();
        visibleNeuronValues.Dispose();

        kernel.Dispose();
        commandQueue.Dispose();
        program.Dispose();
        context.Dispose();
    }
}

GPUがわで動くプログラムはこちら:

myKernelProgram.cl

float sigmoid(float x)
{
	return 1.0f / (1.0f + exp(-x));
}

__kernel void updateVisibleNeurons(
	__global float *resultVisibleNeuronValues,
	__global float *visibleNeuronBiases, 
	__global float *weights, 
	__global float *hiddenNeuronValues,
	int visibleNeuronCount,
	int hiddenNeuronCount)
{
	int visibleNeuronIndex = get_global_id(0);
	float sum = 0;

	for(int i = 0; i < hiddenNeuronCount;i++)
	{
		sum += weights[visibleNeuronIndex * hiddenNeuronCount + i] * hiddenNeuronValues[i];
	}

	resultVisibleNeuronValues[visibleNeuronIndex] 
		= sigmoid(sum + visibleNeuronBiases[visibleNeuronIndex]);
}

このプログラムを実行すると、次のようになります:

0.7310586
0.9525741
0.9933071

上手く動いているようですね。
※sigmoid(1)は0.73...です。

拍手[1回]

PR