Merge pull request #1 from mandlm/digits

Digits

Layer.cpp

@@ -21,7 +21,7 @@ void Layer::setOutputValues(const std::vector<double> & outputValues)
 	for (const double &value : outputValues)
 	{
 		(neuronIt++)->setOutputValue(value);
-	}
+    }
 }
 
 void Layer::feedForward(const Layer &inputLayer)
@@ -54,7 +54,7 @@ void Layer::connectTo(const Layer & nextLayer)
 
 void Layer::updateInputWeights(Layer & prevLayer)
 {
-	static const double trainingRate = 0.3;
+    static const double trainingRate = 0.2;
 
 	for (size_t targetLayerIndex = 0; targetLayerIndex < sizeWithoutBiasNeuron(); ++targetLayerIndex)
 	{

Layer.h

@@ -13,9 +13,10 @@ public:
 	Layer(size_t numNeurons);
 
 	void setOutputValues(const std::vector<double> & outputValues);
+
 	void feedForward(const Layer &inputLayer);
     double getWeightedSum(size_t outputNeuron) const;
-	void connectTo(const Layer & nextLayer);
+    void connectTo(const Layer &nextLayer);
 
 	void updateInputWeights(Layer &prevLayer);
 

Net.cpp

@@ -63,7 +63,7 @@ void Net::feedForward(const std::vector<double> &inputValues)
 		Layer &nextLayer = *(layerIt + 1);
 
 		nextLayer.feedForward(currentLayer);
-	}
+    }
 }
 
 std::vector<double> Net::getOutput()

gui/NeuroUI/NeuroUI.pro

@@ -18,14 +18,16 @@ SOURCES += main.cpp\
     ../../Net.cpp \
     ../../Neuron.cpp \
     netlearner.cpp \
-    errorplotter.cpp
+    errorplotter.cpp \
+    mnistloader.cpp
 
 HEADERS  += neuroui.h \
     ../../Layer.h \
     ../../Net.h \
     ../../Neuron.h \
     netlearner.h \
-    errorplotter.h
+    errorplotter.h \
+    mnistloader.h
 
 FORMS    += neuroui.ui
 

gui/NeuroUI/mnistloader.cpp

@@ -0,0 +1,97 @@
+#include "mnistloader.h"
+
+#include <fstream>
+
+void MnistLoader::load(const std::string &databaseFileName, const std::string &labelsFileName)
+{
+    loadDatabase(databaseFileName);
+    loadLabels(labelsFileName);
+}
+
+const MnistLoader::MnistSample &MnistLoader::getRandomSample() const
+{
+    size_t sampleIndex = (std::rand() * (samples.size() - 1)) / RAND_MAX;
+
+    return *(samples[sampleIndex].get());
+}
+
+void MnistLoader::loadDatabase(const std::string &fileName)
+{
+    std::ifstream databaseFile;
+    databaseFile.open(fileName, std::ios::binary);
+
+    if (!databaseFile.is_open())
+    {
+        throw std::runtime_error("unable to open MNIST database file");
+    }
+
+    int32_t magicNumber = readInt32(databaseFile);
+    if (magicNumber != DatabaseFileMagicNumber)
+    {
+        throw std::runtime_error("unexpected data reading MNIST database file");
+    }
+
+    int32_t sampleCount = readInt32(databaseFile);
+    int32_t sampleWidth = readInt32(databaseFile);
+    int32_t sampleHeight = readInt32(databaseFile);
+
+    if (sampleWidth != SampleWidth || sampleHeight != SampleHeight)
+    {
+        throw std::runtime_error("unexpected sample size loading MNIST database");
+    }
+
+    samples.reserve(samples.size() + sampleCount);
+
+    for (int32_t sampleIndex = 0; sampleIndex < sampleCount; ++sampleIndex)
+    {
+        std::unique_ptr<MnistSample> sample = std::make_unique<MnistSample>();
+
+        databaseFile.read(reinterpret_cast<char *>(sample->data), sampleWidth * sampleHeight);
+
+        samples.push_back(std::move(sample));
+    }
+}
+
+void MnistLoader::loadLabels(const std::string &fileName)
+{
+    std::ifstream labelFile;
+    labelFile.open(fileName, std::ios::binary);
+
+    if (!labelFile.is_open())
+    {
+        throw std::runtime_error("unable to open MNIST label file");
+    }
+
+    int32_t magicNumber = readInt32(labelFile);
+    if (magicNumber != LabelFileMagicNumber)
+    {
+        throw std::runtime_error("unexpected data reading MNIST label file");
+    }
+
+    int32_t labelCount = readInt32(labelFile);
+    if (labelCount != static_cast<int32_t>(samples.size()))
+    {
+        throw std::runtime_error("MNIST database and label files don't match in size");
+    }
+
+    auto sampleIt = samples.begin();
+    for (int32_t labelIndex = 0; labelIndex < labelCount; ++labelIndex)
+    {
+        (*sampleIt++)->label = readInt8(labelFile);
+    }
+}
+
+int8_t MnistLoader::readInt8(std::ifstream &file)
+{
+    int8_t buf8;
+    file.read(reinterpret_cast<char *>(&buf8), sizeof(buf8));
+    return buf8;
+}
+
+int32_t MnistLoader::readInt32(std::ifstream &file)
+{
+    int32_t buf32;
+    file.read(reinterpret_cast<char *>(&buf32), sizeof(buf32));
+    return _byteswap_ulong(buf32);
+}
+

gui/NeuroUI/mnistloader.h

@@ -0,0 +1,44 @@
+#ifndef MNISTLOADER_H
+#define MNISTLOADER_H
+
+#include <string>
+#include <vector>
+#include <memory>
+#include <inttypes.h>
+
+class MnistLoader
+{
+private:
+    static const uint32_t DatabaseFileMagicNumber = 2051;
+    static const uint32_t LabelFileMagicNumber = 2049;
+    static const size_t SampleWidth = 28;
+    static const size_t SampleHeight = 28;
+
+public:
+    template<size_t SAMPLE_WIDTH, size_t SAMPLE_HEIGHT>
+    class Sample
+    {
+    public:
+        uint8_t label;
+        uint8_t data[SAMPLE_WIDTH * SAMPLE_HEIGHT];
+    };
+
+    using MnistSample = Sample<SampleWidth, SampleHeight>;
+
+private:
+    std::vector<std::unique_ptr<MnistSample>> samples;
+
+public:
+    void load(const std::string &databaseFileName, const std::string &labelsFileName);
+
+    const MnistSample &getRandomSample() const;
+
+private:
+    void loadDatabase(const std::string &fileName);
+    void loadLabels(const std::string &fileName);
+
+    static int8_t readInt8(std::ifstream &file);
+    static int32_t readInt32(std::ifstream &file);
+};
+
+#endif // MNISTLOADER_H

gui/NeuroUI/netlearner.cpp

@@ -1,7 +1,9 @@
 #include "netlearner.h"
 #include "../../Net.h"
+#include "mnistloader.h"
 
 #include <QElapsedTimer>
+#include <QImage>
 
 void NetLearner::run()
 {
@@ -9,67 +11,54 @@ void NetLearner::run()
     {
         QElapsedTimer timer;
 
-        Net myNet;
-        try
-        {
-            myNet.load("mynet.nnet");
-        }
-        catch (...)
-        {
-            myNet.initialize({2, 3, 1});
-        }
+        emit logMessage("Loading training data...");
 
-        size_t batchSize = 5000;
-        size_t batchIndex = 0;
-        double batchMaxError = 0.0;
-        double batchMeanError = 0.0;
+        MnistLoader mnistLoader;
+        mnistLoader.load("../NeuroUI/MNIST Database/train-images.idx3-ubyte",
+                         "../NeuroUI/MNIST Database/train-labels.idx1-ubyte");
+
+        emit logMessage("done");
+
+        Net digitClassifier({28*28, 256, 1});
 
         timer.start();
 
-        size_t numIterations = 1000000;
+        size_t numIterations = 100000;
         for (size_t iteration = 0; iteration < numIterations; ++iteration)
         {
-            std::vector<double> inputValues =
-            {
-                std::rand() / (double)RAND_MAX,
-                std::rand() / (double)RAND_MAX
-            };
+            auto trainingSample = mnistLoader.getRandomSample();
+
+            QImage trainingImage(trainingSample.data, 28, 28, QImage::Format_Grayscale8);
+            emit sampleImageLoaded(trainingImage);
 
             std::vector<double> targetValues =
             {
-                (inputValues[0] + inputValues[1]) / 2.0
+                trainingSample.label / 10.0
             };
 
-            myNet.feedForward(inputValues);
+            std::vector<double> trainingData;
+            trainingData.reserve(28*28);
+            for (const uint8_t &val : trainingSample.data)
+            {
+                trainingData.push_back(val / 255.0);
+            }
 
-            std::vector<double> outputValues = myNet.getOutput();
+            digitClassifier.feedForward(trainingData);
+
+            std::vector<double> outputValues = digitClassifier.getOutput();
 
             double error = outputValues[0] - targetValues[0];
 
-            batchMeanError += error;
-            batchMaxError = std::max<double>(batchMaxError, error);
+            QString logString;
 
-            if (batchIndex++ == batchSize)
-            {
-                QString logString;
+            logString.append("Error: ");
+            logString.append(QString::number(std::abs(error)));
 
-                logString.append("Batch error (");
-                logString.append(QString::number(batchSize));
-                logString.append(" iterations, max/mean): ");
-                logString.append(QString::number(std::abs(batchMaxError)));
-                logString.append(" / ");
-                logString.append(QString::number(std::abs(batchMeanError / batchSize)));
+            emit logMessage(logString);
+            emit currentNetError(error);
+            emit progress((double)iteration / (double)numIterations);
 
-                emit logMessage(logString);
-                emit currentNetError(batchMaxError);
-                emit progress((double)iteration / (double)numIterations);
-
-                batchIndex = 0;
-                batchMaxError = 0.0;
-                batchMeanError = 0.0;
-            }
-
-            myNet.backProp(targetValues);
+            digitClassifier.backProp(targetValues);
         }
 
         QString timerLogString;
@@ -79,7 +68,7 @@ void NetLearner::run()
 
         emit logMessage(timerLogString);
 
-        myNet.save("mynet.nnet");
+        digitClassifier.save("DigitClassifier.nnet");
     }
     catch (std::exception &ex)
     {

gui/NeuroUI/netlearner.h

@@ -14,6 +14,7 @@ signals:
     void logMessage(const QString &logMessage);
     void progress(double progress);
     void currentNetError(double error);
+    void sampleImageLoaded(const QImage &image);
 };
 
 #endif // NETLEARNER_H

gui/NeuroUI/neuroui.cpp

@@ -31,6 +31,8 @@ void NeuroUI::on_runButton_clicked()
         connect(m_netLearner.get(), &NetLearner::finished, this, &NeuroUI::netLearnerFinished);
 
         connect(m_netLearner.get(), &NetLearner::currentNetError, ui->errorPlotter, &ErrorPlotter::addErrorValue);
+
+        connect(m_netLearner.get(), &NetLearner::sampleImageLoaded, this, &NeuroUI::setImage);
     }
 
     m_netLearner->start();
@@ -61,3 +63,10 @@ void NeuroUI::progress(double progress)
 
     ui->progressBar->setValue(value);
 }
+
+void NeuroUI::setImage(const QImage &image)
+{
+    QPixmap pixmap;
+    pixmap.convertFromImage(image);
+    ui->label->setPixmap(pixmap);
+}

gui/NeuroUI/neuroui.h

@@ -28,6 +28,7 @@ private slots:
     void netLearnerStarted();
     void netLearnerFinished();
     void progress(double progress);
+    void setImage(const QImage &image);
 
 private:
     Ui::NeuroUI *ui;

gui/NeuroUI/neuroui.ui

@@ -20,11 +20,37 @@
   <widget class="QWidget" name="centralWidget">
    <layout class="QVBoxLayout" name="verticalLayout_2">
     <item>
-     <widget class="QListWidget" name="logView">
-      <property name="uniformItemSizes">
-       <bool>true</bool>
-      </property>
-     </widget>
+     <layout class="QHBoxLayout" name="horizontalLayout_2">
+      <item>
+       <widget class="QListWidget" name="logView">
+        <property name="uniformItemSizes">
+         <bool>true</bool>
+        </property>
+       </widget>
+      </item>
+      <item>
+       <widget class="QLabel" name="label">
+        <property name="sizePolicy">
+         <sizepolicy hsizetype="Fixed" vsizetype="Preferred">
+          <horstretch>0</horstretch>
+          <verstretch>0</verstretch>
+         </sizepolicy>
+        </property>
+        <property name="minimumSize">
+         <size>
+          <width>128</width>
+          <height>0</height>
+         </size>
+        </property>
+        <property name="text">
+         <string/>
+        </property>
+        <property name="alignment">
+         <set>Qt::AlignCenter</set>
+        </property>
+       </widget>
+      </item>
+     </layout>
     </item>
     <item>
      <widget class="ErrorPlotter" name="errorPlotter" native="true">