CNTK学习笔记 - 经典的卷积神经网络

2024-06-15

字数统计: 790字 | 阅读时长: 4分

卷积神经网络是深度学习中一种高效的方法，已经成为众多科学领域的研究热点之一，特别是在模式识别领域，由于该网络可以直接输入原始图像，避免对图像进行复杂前期预处理，因而得到更为广泛的应用。本文以经典的卷积神经网络为例(使用MNIST数据集，手写数字识别)，来展示如何在CNTK中构建卷积网络，关于卷积神经网络的基础知识这里不再赘述。

在下面的代码中，我们首先构建一个包含2个卷积层、2个池化层和一个全连接层的神经网络，然后创建一个Reader和一个Trainer，并进行训练，最后，对训练好的模型进行验证。

public class MNISTClassifier
{
    public static void Main(string[] args)
    {
        /*1.build model*/
        var input = Var.Input(ND.Shape(28, 28, 1), "features");
        var scaledInput = CNTKLib.ElementTimes(Var.Scalar(0.00390625f), input);
        var labels = Var.Input(ND.Shape(10), "labels");

        /*28x28x1 -> 14x14x4*/
        var initializer = CNTKLib.GlorotUniformInitializer(0.26, -1, 2);
        var convParams = Var.Parameter(ND.Shape(3, 3, 1, 4), initializer);
        var conv1 = CNTKLib.Convolution(convParams, scaledInput, ND.Shape(1, 1, 1));
        var relu1 = CNTKLib.ReLU(conv1);
        var pooling1 = CNTKLib.Pooling(relu1, PoolingType.Max, ND.Shape(3, 3), ND.Shape(2, 2), Var.Vector(true));

        /*14x14x4 -> 7x7x8*/
        initializer = CNTKLib.GlorotUniformInitializer(0.26, -1, 2);
        convParams = Var.Parameter(ND.Shape(3, 3, 4, 8), initializer);
        var conv2 = CNTKLib.Convolution(convParams, pooling1, ND.Shape(1, 1, 4));
        var relu2 = CNTKLib.ReLU(conv2);
        var pooling2 = CNTKLib.Pooling(relu2, PoolingType.Max, ND.Shape(3, 3), ND.Shape(2, 2), Var.Vector(true));

        /*7x7x8=392 -> 10*/
        initializer = CNTKLib.GlorotUniformInitializer(CNTKLib.DefaultParamInitScale, CNTKLib.SentinelValueForInferParamInitRank, CNTKLib.SentinelValueForInferParamInitRank, 1);
        var w = Var.Parameter(ND.Shape(10, 7, 7, 8), initializer);
        var b = Var.Parameter(ND.Shape(10), 0f);
        var output = w * pooling2 + b;

        var trainingLoss = CNTKLib.CrossEntropyWithSoftmax(new Variable(output), labels, "lossFunction");
        var prediction = CNTKLib.ClassificationError(new Variable(output), labels, "classificationError");

        output.Save("data/MNIST.model");
        /*2.build reader and trainer*/
        var streamConfigurations = new[] { new StreamConfiguration("features", 28 * 28 * 1), new StreamConfiguration("labels", 10) };

        var trainFile = @"D:\cntk\Examples\Image\DataSets\MNIST\Train_cntk_text.txt";
        var minibatchSource = MinibatchSource.TextFormatMinibatchSource(trainFile, streamConfigurations, MinibatchSource.InfinitelyRepeat);
        var featureStreamInfo = minibatchSource.StreamInfo("features");
        var labelStreamInfo = minibatchSource.StreamInfo("labels");

        var learningRatePerSample = new TrainingParameterScheduleDouble(0.003125, 1);
        var parameterLearners = new[] { Learner.SGDLearner(output.Parameters(), learningRatePerSample) };
        var trainer = Trainer.CreateTrainer(output, trainingLoss, prediction, parameterLearners);

        /*3.train*/
        const uint minibatchSize = 64;
        int outputFrequencyInMinibatches = 20, i = 0;
        int epochs = 5;
        while (epochs > 0)
        {
            var minibatchData = minibatchSource.GetNextMinibatch(minibatchSize, Data.DefaultDevice);
            var arguments = new Dictionary<Variable, MinibatchData> {
                { input,minibatchData[featureStreamInfo]},
                { labels,minibatchData[labelStreamInfo]}
            };

            trainer.TrainMinibatch(arguments, Data.DefaultDevice);
            Log.TrainingProgress(trainer, i++, outputFrequencyInMinibatches);

            if (minibatchData.Values.Any(a => a.sweepEnd))
            {
                output.Save($"data/MNIST_{epochs}.model");
                epochs--;
            }
        }

        output.Save($"data/MNIST.model");
        Console.WriteLine("Model train complete!\n");

        /*4.validate*/
        var testFile = @"D:\cntk\Examples\Image\DataSets\MNIST\Test_cntk_text.txt";
        var testMinibatchSource = MinibatchSource.TextFormatMinibatchSource(testFile, streamConfigurations, MinibatchSource.FullDataSweep);
        featureStreamInfo = testMinibatchSource.StreamInfo("features");
        labelStreamInfo = testMinibatchSource.StreamInfo("labels");

        int batchSize = 50;
        int miscountTotal = 0, totalCount = 0;
        while (totalCount <= 1000)
        {
            var minibatchData = testMinibatchSource.GetNextMinibatch((uint)batchSize, Data.DefaultDevice);
            if (minibatchData == null || minibatchData.Count == 0) break;
            totalCount += (int)minibatchData[featureStreamInfo].numberOfSamples;

            var labelData = minibatchData[labelStreamInfo].data.GetDenseData<float>(output);
            var expectedLabels = labelData.Select(l => l.IndexOf(l.Max())).ToList();

            var inputDataMap = Data.Add(input, minibatchData[featureStreamInfo].data);
            var outputDataMap = Data.Add(output);

            output.Evaluate(inputDataMap, outputDataMap);
            var outputData = outputDataMap[output].GetDenseData<float>(output);
            var actualLabels = outputData.Select(l => l.IndexOf(l.Max())).ToList();

            int misMatches = actualLabels.Zip(expectedLabels, (_a, _b) => _a.Equals(_b) ? 0 : 1).Sum();

            miscountTotal += misMatches;
            Console.WriteLine($"Validating Model: Total Samples = {totalCount}, Misclassify Count = {miscountTotal}");
        }

        float errorRate = 1.0F * miscountTotal / totalCount;
        Console.WriteLine($"Model Validation Error = {errorRate}");

        Console.ReadLine();
    }
}

截图：

本文作者： killf
本文链接： http://www.killf.info/机器学习/CNTK/cntk_1/
版权声明： 本博客所有文章除特别声明外，均采用 Apache License 2.0 许可协议。转载请注明出处！