基于CNTK/C#实现MNIST【附源码】

---

前言

本文实现基于CNTK实现MNIST，并对其实现CNN以及MLP方法进行测试

---

一、环境搭建

因为本次使用C#，我们只需要对C#的一些依赖库进行整理放入到项目中即可，这里我把所需要的所有的dll都整理到win64文件夹，在C# 的项目工程上只需要引入Cntk.Core.Managed-2.7.dll、netstandard.dll即可。
如图：

---

二、MNIST代码解析

这里我们直接上代码并解释每行代码的意思和构建的思路和步骤

1.GPU/CPU的设置

//GPU设置
var device = DeviceDescriptor.GPUDevice(0);

2.参数变量的设置

//使用cnn还是使用mlp
bool useConvolution = true;
//模型内的名称
var featureStreamName = "features";
var labelsStreamName = "labels";
var classifierName = "classifierOutput";
Function classifierOutput;

 //输入和输出定义
int[] imageDim = useConvolution ? new int[] { 28, 28, 1 } : new int[] { 784 };
int imageSize = 28 * 28;
int numClasses = 10;

 //设置模型的保存名称
string modelFile = useConvolution ? "./MNIST_CNN.model" : "./MNIST_MLP.model";

3.关联模型构造

 //配置网络层的对应关系：features -> 28 * 28    labelsStreamName -> numClasses
IList<StreamConfiguration> streamConfigurations = new StreamConfiguration[]
{ new StreamConfiguration(featureStreamName, imageSize), new StreamConfiguration(labelsStreamName, numClasses) };
//定义输入变量，输出变量
var input = CNTKLib.InputVariable(imageDim, DataType.Float, featureStreamName);
var labels = CNTKLib.InputVariable(new int[] { numClasses }, DataType.Float, labelsStreamName);

4.模型构建

if (useConvolution)  //cnn
{
    var scaledInput = CNTKLib.ElementTimes(Constant.Scalar<float>(0.00390625f, device), input);
    //构建CNN网络结构
    classifierOutput = CreateConvolutionalNeuralNetwork(scaledInput, numClasses, device, classifierName);
}
else  //mlp
{
    int hiddenLayerDim = 200;   //mlp的隐藏节点
    //构建MLP结构
    var scaledInput = CNTKLib.ElementTimes(Constant.Scalar<float>(0.00390625f, device), input);
    classifierOutput = CreateMLPClassifier(device, numClasses, hiddenLayerDim, scaledInput, classifierName);
}

static Function CreateConvolutionalNeuralNetwork(Variable features, int outDims, DeviceDescriptor device, string classifierName)
{
    //CNN网络结构的构建

    //初始化卷积层的参数
    int kernelWidth1 = 3, kernelHeight1 = 3, numInputChannels1 = 1, outFeatureMapCount1 = 4;
    int hStride1 = 2, vStride1 = 2;
    int poolingWindowWidth1 = 3, poolingWindowHeight1 = 3;
    // 28x28x1 -> 14x14x4    卷积+激活函数
    Function pooling1 = ConvolutionWithMaxPooling(features, device, kernelWidth1, kernelHeight1,
        numInputChannels1, outFeatureMapCount1, hStride1, vStride1, poolingWindowWidth1, poolingWindowHeight1);

    //初始化卷积层的参数
    int kernelWidth2 = 3, kernelHeight2 = 3, numInputChannels2 = outFeatureMapCount1, outFeatureMapCount2 = 8;
    int hStride2 = 2, vStride2 = 2;
    int poolingWindowWidth2 = 3, poolingWindowHeight2 = 3;
    // 14x14x4 -> 7x7x8    卷积+激活函数
    Function pooling2 = ConvolutionWithMaxPooling(pooling1, device, kernelWidth2, kernelHeight2,
        numInputChannels2, outFeatureMapCount2, hStride2, vStride2, poolingWindowWidth2, poolingWindowHeight2);

    //Dense层设计
    Function denseLayer = Dense(pooling2, outDims, device, Activation.None, classifierName);
    return denseLayer;
}

5.评价指标

//出模型后求损失
var trainingLoss = CNTKLib.CrossEntropyWithSoftmax(new Variable(classifierOutput), labels, "lossFunction");
//出模型后求准确率
var prediction = CNTKLib.ClassificationError(new Variable(classifierOutput), labels, "classificationError");

6.数据集加载

//读取文件，数据集的读取，并确定好输入输出
var minibatchSource = MinibatchSource.TextFormatMinibatchSource("./mnist_data/MNIST_Train_cntk_text.txt", streamConfigurations, MinibatchSource.InfinitelyRepeat);
var featureStreamInfo = minibatchSource.StreamInfo(featureStreamName);
var labelStreamInfo = minibatchSource.StreamInfo(labelsStreamName);

7.学习率设置

//学习率的设置：SGD
TrainingParameterScheduleDouble learningRatePerSample = new TrainingParameterScheduleDouble(0.003125, 1);
IList<Learner> parameterLearners = new List<Learner>() { Learner.SGDLearner(classifierOutput.Parameters(), learningRatePerSample) };

8.获取模型训练器

var trainer = Trainer.CreateTrainer(classifierOutput, trainingLoss, prediction, parameterLearners);

9.模型训练

//开始训练
const uint minibatchSize = 64;
int outputFrequencyInMinibatches = 20, i = 0;
int epochs = 10000;
while (epochs > 0)
{
    var minibatchData = minibatchSource.GetNextMinibatch(minibatchSize, device);
    //定义好输入数据
    var arguments = new Dictionary<Variable, MinibatchData>
    {
        { input, minibatchData[featureStreamInfo] },
        { labels, minibatchData[labelStreamInfo] }
    };
    //训练入口
    trainer.TrainMinibatch(arguments, device);
    //输出训练结果
    PrintTrainingProgress(trainer, i++, outputFrequencyInMinibatches);

    if (MiniBatchDataIsSweepEnd(minibatchData.Values))
    {
        epochs--;
    }
}

10.模型保存

classifierOutput.Save(modelFile);

11.模型验证

 // 验证模型
var minibatchSourceNewModel = MinibatchSource.TextFormatMinibatchSource("./mnist_data/MNIST_Test_cntk_text.txt", streamConfigurations, MinibatchSource.FullDataSweep);
ValidateModelWithMinibatchSource(modelFile, minibatchSourceNewModel, imageDim, numClasses, featureStreamName, labelsStreamName, classifierName, device);

---

三、效果展示

---

四、源码下载

源码下载