static void Main(string[] args)//TODO BGD
{
var epochNum = 100;
var program = new Program();
var sTrainData = new SinTrainData(1000);
var initialWeights = new double[1];//tmp
// program.ManualPropSin(sTrainData, epochNum, false, InitialWeights: out var initialWeights,
// FinalWeights: out var finalWeights, debug: false);
// Console.WriteLine("");
program.RegressionUsingSGD(sTrainData, true, epochNum, initialWeights: ref initialWeights,
finalWeights: out var finalWeights);
}
public void ManualPropSin(SinTrainData sTrainData, int numMaxEpochs, bool shuffle_flag,
out double[] InitialWeights, out double[] FinalWeights, bool debug = false)
{
Console.WriteLine("NN Regression for predicting sin(x) begin:");
// create training data
// var sTrainData = new SinTrainData(500);
System.Console.WriteLine("\nTraining data:");
Helper.ShowMatrix(sTrainData.DataSet, 5, 4, true);
// create NN
const int numInput = 1;
var numHidden = 12;
var numOutput = 1; // regression
var rndSeed = 0;
System.Console.WriteLine("\n Creating a " + numInput + "-" + numHidden + "-" + numOutput +
" regression neural network");
var nn = new nn(numInput, numHidden, numOutput, rndSeed);
// const int numMaxEpochs = 1000;
const double lr = 0.005;
const double momentum = 0.001; // need more test
Console.WriteLine("\nSetting maxEpochs = " + numMaxEpochs);
Console.WriteLine("Setting learnRate = " + lr.ToString("F4"));
Console.WriteLine("Setting momentum = " + momentum.ToString("F4"));
// train NN
InitialWeights = new double[37]; //just for test
InitialWeights = nn.GetWeights();
Console.WriteLine("Initial weights");
Helper.ShowVector(InitialWeights, 6, 8, true);
FinalWeights = nn.Train(sTrainData, numMaxEpochs, lr, momentum, shuffle_flag, ref InitialWeights, debug);
if (debug)
{
Console.WriteLine("final weights after training");
Helper.ShowVector(FinalWeights);
}
// show model
// System.Console.WriteLine("Final model weights:");
// Helper.ShowVector(weights, 4, 4, true);
// Test Example
// var y = nn.Forward(new[] {Math.PI});
// Console.WriteLine("\nActual sin(PI) = 0.0 Predicted = " + y[0].ToString("F6"));
// y = nn.Forward(new[] {Math.PI / 2});
// Console.WriteLine("\nActual sin(PI / 2) = 1.0 Predicted = " + y[0].ToString("F6"));
// y = nn.Forward(new[] {3 * Math.PI / 2.0});
// Console.WriteLine("\nActual sin(3*PI / 2) = -1.0 Predicted = " + y[0].ToString("F6"));
}
public void RegressionUsingSGD(SinTrainData sinTrainData, bool shuffle_flag, int numEpochs,
ref double[] initialWeights, out double[] finalWeights, bool provideInitialWeights = false)
{
var inputLayer = new Linear(1, 13);
var activationLayer1TanH = new HyperTan();
var hiddenLayer1 = new Linear(13, 13);
var activationLayer2TanH = new HyperTan();
var hiddenLayer2 = new Linear(13, 1);
var model = new Model(new Layer[] { inputLayer, activationLayer1TanH, hiddenLayer1, activationLayer2TanH, hiddenLayer2 });
// var model = new Model(new Layer[] {inputLayer, activationLayer1TanH, hiddenLayer2});
if (provideInitialWeights)
{
model.SetWeights(initialWeights);
}
const double learning_rate = 0.005;
const double momentum = 0.001; // need more test
// generate dataSet
// var sinTrainData = new SinTrainData(); // Inherit from DataSet class:generate 1000 items default
//generate trainData, testData
Helper.SplitTrainTest(sinTrainData, 0.8, 1, trainData: out var trainSet, testData: out var testSet);
var optimizer = new SGD(model, learning_rate, momentum); // train via sgd
var dataLoader = new DataLoader(trainSet, 1, shuffle_flag);
//train using SGD
var epoch = 0;
var epochInternal = 10;
while (epoch < numEpochs)
{
epoch++;
if (epoch % epochInternal == 0 && epoch > 0) //print epoch & error info
{
var mse = Evaluate.MSE(model, testSet);
Console.WriteLine("epoch = " + epoch + " acc = " + (1 - mse).ToString("F4"));
// Console.WriteLine(" MSE: " + mse.ToString());
}
for (int i = 0; i < dataLoader.loopTime; i++)
{
var data = dataLoader.Enumerate(); //[1][x,y]
// Console.WriteLine(i);
Helper.SplitInputOutput(data, out var inputData, out var outputData);
var yPred = model.Forward(inputData[0]); //xValue
// Console.WriteLine("framework output:" + yPred.PredictedLinearValues[0]);
//compute and print loss
var loss = new RegressionLoss(yPred, outputData[0]); //tValue
// Console.WriteLine(loss.Item()); // print loss
optimizer.Zero_grad(); // refresh buffer before .backward()
loss.Backward(); // calculate grads
optimizer.Step(); // update weights
}
}
finalWeights = model.GetWeights();
// test
var y = model.Forward(new[] { Math.PI }).PredictedLinearValues;
Console.WriteLine("\nActual sin(PI) = 0.0 Predicted = " + y[0].ToString("F6"));
}