Load and train handwritten digits

gui/NeuroUI/netlearner.cpp

@@ -10,70 +10,64 @@ void NetLearner::run()
     {
         QElapsedTimer timer;
 
+        emit logMessage("Loading training data...");
+        emit progress(0.0);
+
         TrainingDataLoader dataLoader;
         dataLoader.addSamples("../NeuroUI/training data/mnist_train0.jpg", 0);
+        emit progress(0.1);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train1.jpg", 1);
+        emit progress(0.2);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train2.jpg", 2);
+        emit progress(0.3);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train3.jpg", 3);
+        emit progress(0.4);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train4.jpg", 4);
+        emit progress(0.5);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train5.jpg", 5);
+        emit progress(0.6);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train6.jpg", 6);
+        emit progress(0.7);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train7.jpg", 7);
+        emit progress(0.8);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train8.jpg", 8);
+        emit progress(0.9);
+        dataLoader.addSamples("../NeuroUI/training data/mnist_train9.jpg", 9);
+        emit progress(1.0);
 
-        Net myNet;
+        emit logMessage("done");
-        try
+        emit progress(0.0);
-        {
-            myNet.load("mynet.nnet");
-        }
-        catch (...)
-        {
-            myNet.initialize({2, 3, 1});
-        }
 
-        size_t batchSize = 5000;
+        Net digitClassifier({32*32, 16*16, 32, 1});
-        size_t batchIndex = 0;
-        double batchMaxError = 0.0;
-        double batchMeanError = 0.0;
 
         timer.start();
 
-        size_t numIterations = 2000000;
+        size_t numIterations = 10000;
         for (size_t iteration = 0; iteration < numIterations; ++iteration)
         {
-            std::vector<double> inputValues =
+            const TrainingDataLoader::Sample &trainingSample = dataLoader.getRandomSample();
-            {
-                std::rand() / (double)RAND_MAX,
-                std::rand() / (double)RAND_MAX
-            };
 
             std::vector<double> targetValues =
             {
-                (inputValues[0] + inputValues[1]) / 2.0
+                trainingSample.first / 10.0
             };
 
-            myNet.feedForward(inputValues);
+            digitClassifier.feedForward(trainingSample.second);
 
-            std::vector<double> outputValues = myNet.getOutput();
+            std::vector<double> outputValues = digitClassifier.getOutput();
 
             double error = outputValues[0] - targetValues[0];
 
-            batchMeanError += error;
+            QString logString;
-            batchMaxError = std::max<double>(batchMaxError, error);
 
-            if (batchIndex++ == batchSize)
+            logString.append("Error: ");
-            {
+            logString.append(QString::number(std::abs(error)));
-                QString logString;
 
-                logString.append("Batch error (");
+            emit logMessage(logString);
-                logString.append(QString::number(batchSize));
+            emit currentNetError(error);
-                logString.append(" iterations, max/mean): ");
+            emit progress((double)iteration / (double)numIterations);
-                logString.append(QString::number(std::abs(batchMaxError)));
-                logString.append(" / ");
-                logString.append(QString::number(std::abs(batchMeanError / batchSize)));
 
-                emit logMessage(logString);
+            digitClassifier.backProp(targetValues);
-                emit currentNetError(batchMaxError);
-                emit progress((double)iteration / (double)numIterations);
-
-                batchIndex = 0;
-                batchMaxError = 0.0;
-                batchMeanError = 0.0;
-            }
-
-            myNet.backProp(targetValues);
         }
 
         QString timerLogString;
@@ -83,7 +77,7 @@ void NetLearner::run()
 
         emit logMessage(timerLogString);
 
-        myNet.save("mynet.nnet");
+        digitClassifier.save("DigitClassifier.nnet");
     }
     catch (std::exception &ex)
     {

gui/NeuroUI/trainingdataloader.cpp

@@ -1,5 +1,7 @@
 #include "trainingdataloader.h"
 
+#include <sstream>
+
 #include <QImage>
 #include <QColor>
 
@@ -11,19 +13,53 @@ TrainingDataLoader::TrainingDataLoader()
 void TrainingDataLoader::addSamples(const QString &sourceFile, TrainingDataLoader::SampleId sampleId)
 {
     QImage sourceImage;
-    sourceImage.load(sourceFile);
+    if (sourceImage.load(sourceFile) == false)
-
-    Sample sample;
-    sample.first = sampleId;
-
-    for (unsigned int y = 0; y < 8; ++y)
     {
-        for (unsigned int x = 0; x < 8; ++x)
+        std::ostringstream errorString;
-        {
+        errorString << "error loading " << sourceFile.toStdString();
-            sample.second[x + y * 8] = qGray(sourceImage.pixel(x, y)) / 255.0;
+
-        }
+        throw std::runtime_error(errorString.str());
     }
 
-    m_samples.push_back(sample);
+    QSize scanWindow(32, 32);
+    QPoint scanPosition(0, 0);
+
+    while (scanPosition.y() + scanWindow.height() < sourceImage.height())
+    {
+        scanPosition.setX(0);
+
+        while (scanPosition.x() + scanWindow.width() < sourceImage.width())
+        {
+            Sample sample;
+            sample.first = sampleId;
+
+            for (int y = 0; y < scanWindow.height(); ++y)
+            {
+                for (int x = 0; x < scanWindow.width(); ++x)
+                {
+                    QRgb color = sourceImage.pixel(scanPosition.x() + x, scanPosition.y() + y);
+                    sample.second[x + y * scanWindow.height()] = qGray(color) / 255.0;
+                }
+            }
+
+            m_samples.push_back(sample);
+
+            scanPosition.rx() += scanWindow.width();
+        }
+
+        scanPosition.ry() += scanWindow.height();
+    }
+}
+
+const TrainingDataLoader::Sample &TrainingDataLoader::getRandomSample() const
+{
+    size_t sampleIndex = (std::rand() * m_samples.size()) / RAND_MAX;
+
+    auto it = m_samples.cbegin();
+    for (size_t index = 0; index < sampleIndex; ++index)
+    {
+        it++;
+    }
+    return *it;
 }
 

gui/NeuroUI/trainingdataloader.h

@@ -10,7 +10,7 @@
 class TrainingDataLoader
 {
 public:
-    using SampleData = double[64];
+    using SampleData = double[32*32];
     using SampleId = unsigned int;
     using Sample = std::pair<SampleId, SampleData>;
 
@@ -21,6 +21,8 @@ public:
     TrainingDataLoader();
 
     void addSamples(const QString &sourceFile, SampleId sampleId);
+
+    const Sample &getRandomSample() const;
 };
 
 #endif // TRAININGDATALOADER_H