/// <summary>
/// Train the network for 1 epoch
/// </summary>
/// <param name="data">The dataset</param>
/// <param name="batchSize">How many mini batch size before applying trained weights</param>
/// <param name="learningRate">The learning speed, 0.005f is a good value</param>
/// <param name="logTrainingData">Prints training data to the console</param>
public void Train(ILabeledData data, int batchSize, float learningRate, bool logTrainingData)
{
if (data.GetDataCount() != data.GetLabelCount())
{
throw new System.Exception("Data and labels are not matched");
}
Random rand = new Random();
List <int> shuffledData = new List <int>(data.GetDataCount());
for (int i = 0; i < data.GetDataCount(); i++)
{
shuffledData.Add(i);
}
float cost = 0;
int dataCount = 0;
int completeness = 1;
while (shuffledData.Count > 0)
{
for (int j = 0; j < batchSize; j++)
{
if (shuffledData.Count == 0)
{
break;
}
int randomIndex = rand.Next(0, shuffledData.Count);
int dataIndex = shuffledData[randomIndex];
shuffledData.RemoveAt(randomIndex);
m_inputLayer.FeedInput(data.GetInput(dataIndex));
m_inputLayer.Predict();
m_outputLayer.Backpropagate(data.GetTarget(dataIndex), learningRate);
cost += m_outputLayer.CalculateCost(data.GetTarget(dataIndex));
dataCount++;
}
ApplyChanges(learningRate);
float percentComplete = (((float)(data.GetDataCount() - shuffledData.Count)) / data.GetDataCount()) * 100f;
if (logTrainingData && percentComplete / 10.0f >= completeness)
{
completeness++;
Console.WriteLine(string.Format("{0,4:F2}% Done. Current Cost: {1,6:F4}", percentComplete, cost / dataCount));
cost = 0;
dataCount = 0;
}
}
}
// Fit to a dataset
public Network(ILabeledData data)
{
m_inputLayer = new InputLayer(data.GetDataDimension());
m_outputLayer = new OutputLayer(data.GetLabelDimension());
}
static void Main(string[] args)
{
float learningStep = 0.0025f;
int epochCount = 200;
int batchCount = 32;
ILabeledData trainingSet = MNISTDataset.LoadDataset("dataset/train-images.idx3-ubyte", "dataset/train-labels.idx1-ubyte");
ILabeledData testingSet = MNISTDataset.LoadDataset("dataset/t10k-images.idx3-ubyte", "dataset/t10k-labels.idx1-ubyte");
Console.WriteLine(string.Format("Training Dataset : \n{0}", (MNISTDataset)trainingSet));
Console.WriteLine(string.Format("\nTesting Dataset : \n{0}", (MNISTDataset)testingSet));
Network network = new Network(trainingSet);
network.AddLayer(new Layer(16, Activation.Relu));
network.AddLayer(new Layer(16, Activation.Relu));
network.Build();
Console.WriteLine(string.Format("Starting training\n\nepoch count : {0}\nbatch size : {1}\nlearning step : {2,4:F2}", epochCount, batchCount, learningStep));
for (int epoch = 0; epoch < epochCount; epoch++)
{
Console.WriteLine(string.Format("Starting epoch {0}", epoch));
network.Train(trainingSet, batchCount, learningStep, true);
}
Console.WriteLine("Finished training. Evaluating performance");
int total = testingSet.GetDataCount();
int correctPrediction = 0;
float cost = 0;
for (int i = 0; i < testingSet.GetDataCount(); i++)
{
float[] output = network.Predict(testingSet, i);
cost += network.CalculateCost(testingSet.GetTarget(i));
if (((MNISTDataset)testingSet).CheckPrediction(output, i))
{
correctPrediction++;
}
}
Console.WriteLine(string.Format("Cost : {0,4:F2}\nCorrect percentage : {1,4:F2}", cost / total, ((float)correctPrediction / total) * 100f));
Console.WriteLine(string.Format("Press any key to continue.."));
Console.ReadKey(true);
Random rand = new Random();
while (true)
{
int id = rand.Next(0, trainingSet.GetDataCount());
float[] input = trainingSet.GetInput(id);
Console.WriteLine(((MNISTDataset)trainingSet).GetImage(id));
Console.WriteLine("Number is : " + ((MNISTDataset)trainingSet).GetLabel(id));
float[] prediction = network.Predict(trainingSet, id);
for (int i = 0; i < prediction.Length; i++)
{
Console.Write(string.Format("{0} - {1,4:F2}%; ", i, prediction[i] * 100));
}
Console.WriteLine(string.Format("\nPredicted number is : {0}\n", ((MNISTDataset)trainingSet).PredictedNumber(prediction)));
Console.Write("\nTest again? (y/n)");
ConsoleKeyInfo key = Console.ReadKey();
Console.WriteLine("");
while (key.KeyChar != 'y' && key.KeyChar != 'n')
{
Console.Write(string.Format("Test again? (y/n)"));
key = Console.ReadKey();
Console.WriteLine("");
}
if (key.KeyChar == 'n')
{
break;
}
}
}
