private static TrainAndTest[] AddTrainOrTest(string fullPath)
{
string[] lines = File.ReadAllLines(fullPath);
//TRAIN or TEST units
TrainAndTest[] Data = new TrainAndTest[lines.Length / 2];
//Training and test files are .txt, where non-odd lines are INPUT values and odd lines are OUTPUT values
uint countOfUnit = 0;
for (uint i = 1; i <= lines.Length; ++i)
{
if (i % 2 != 0)
{
string[] line = lines[i - 1].Split(' ');
Data[countOfUnit].IN = new double[line.Length];
for (uint j = 0; j < line.Length; ++j) //Every INPUT value in the line must be readen
{
Data[countOfUnit].IN[j] = Convert.ToDouble(line[j]);
}
}
else
{
string[] line = lines[i - 1].Split(' ');
Data[countOfUnit].OUT = new double[line.Length];
for (int j = 0; j < line.Length; ++j) //Every OUTPUT value in the line must be readen
{
Data[countOfUnit].OUT[j] = Convert.ToDouble(line[j]);
}
++countOfUnit;
}
}
return(Data);
// return Data;
}
private static void Main(String[] args)
{
TrainAndTest trainAndTest = new TrainAndTest();
// 訓練
Matrix <Double> trainMatrix = trainAndTest.GetTrainMatrix();
Matrix <Double> testMatrix = trainAndTest.GetTestMatrix();
// 測試
Matrix <Double> resultMatrix = trainAndTest.GetResultMatrix();
Matrix <Double> resultTestMatrix = trainAndTest.GetResultTestMatrix();
/////////////////////////
/// 超參數
/////////////////////////
// 訓練次數
Int32 iterNum = 1 << 14; // 16384
// 學習樣本總數
Int32 trainNum = trainMatrix.ColumnCount;
// 批次內數量
Int32 batchNum = 1 << 7; // 128
// 學習速率
Double learnRate = 0.1;
var perEpoch = Max(iterNum / batchNum, 1);
var trainLossList = new List <Double>();
var ranM = new Func <Int32, Int32, Matrix <Double> >(Matrix <Double> .Build.Random);
var ranV = new Func <Int32, Vector <Double> >(Vector <Double> .Build.Random);
var size = new {
input = trainMatrix.RowCount,
hidden1 = (trainMatrix.RowCount * 2 + testMatrix.RowCount) / 3,
hidden2 = (trainMatrix.RowCount + testMatrix.RowCount * 2) / 3,
output = testMatrix.RowCount
};
var mat1 = ranM(size.input, size.hidden1);
var vct1 = ranV(size.hidden1);
var mat2 = ranM(size.hidden1, size.hidden2);
var vct2 = ranV(size.hidden2);
var mat3 = ranM(size.hidden2, size.output);
var vct3 = ranV(size.output);
var hiddenLayers = new List <IHiddenLayer <Double> > {
new AffineDouble(mat1, vct1),
new ReluDouble(),
new AffineDouble(mat2, vct2),
new AffineDouble(mat3, vct3)
};
var network = new NetworkDouble(hiddenLayers, new SoftmaxWithLossDouble());
var ran = new Random((1 << 13) - 1);
// 訓練回合
for (int i = 0; i < iterNum; i++)
{
var trainBatch = Matrix <Double> .Build.Dense(trainNum, batchNum);
var testBatch = Matrix <Double> .Build.Dense(trainNum, batchNum);
for (int j = 0; j < batchNum; j++)
{
var rNum = ran.Next(trainNum);
trainBatch.SetColumn(j, trainMatrix.Column(rNum));
testBatch.SetColumn(j, testMatrix.Column(rNum));
}
var grad = network.Gradient(trainBatch, testBatch);
network.Update(grad, learnRate);
var loss = network.Loss(trainBatch, testBatch);
trainLossList.Add(loss);
// 列印結果評估
if (!(i % perEpoch is 0))
{
continue;
}
var trainAccuracy = network.Accuracy(trainMatrix, testMatrix);
var testAccuracy = network.Accuracy(resultMatrix, resultTestMatrix);
Console.WriteLine(i.ToString(CultureInfo.InvariantCulture));
Console.WriteLine(nameof(trainAccuracy) + ": " + trainAccuracy.ToString(CultureInfo.InvariantCulture));
Console.WriteLine(nameof(testAccuracy) + ": " + testAccuracy.ToString(CultureInfo.InvariantCulture));
}
}
