public void ClassificationUsingSoftmax()
{
var inputLayer = new Linear(1, 12);
var activationLayer1TanH = new HyperTan();
var hiddenLayer = new Linear(12, 1);
var softmaxLayer = new SoftMax();
var model = new Model(new Layer[] { inputLayer, activationLayer1TanH, hiddenLayer, softmaxLayer });
// implement yourself~~
}
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"));
}
