|
public TrainOnData ( |
||
| data | ||
| maxEpochs | uint | |
| epochsBetweenReports | uint | |
| desiredError | float | |
| Résultat | void | |
static void Main()
{
const uint num_layers = 3;
const uint num_neurons_hidden = 96;
const float desired_error = 0.001F;
using (TrainingData trainData = new TrainingData("..\\..\\..\\datasets\\robot.train"))
using (TrainingData testData = new TrainingData("..\\..\\..\\datasets\\robot.test"))
{
for (float momentum = 0.0F; momentum < 0.7F; momentum += 0.1F)
{
Console.WriteLine("============= momentum = {0} =============\n", momentum);
using (NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, trainData.InputCount, num_neurons_hidden, trainData.OutputCount))
{
net.TrainingAlgorithm = TrainingAlgorithm.TRAIN_INCREMENTAL;
net.LearningMomentum = momentum;
net.TrainOnData(trainData, 20000, 5000, desired_error);
Console.WriteLine("MSE error on train data: {0}", net.TestData(trainData));
Console.WriteLine("MSE error on test data: {0}", net.TestData(testData));
}
}
}
Console.ReadKey();
}
static void Main()
{
const uint num_input = 3;
const uint num_output = 1;
const uint num_layers = 4;
const uint num_neurons_hidden = 5;
const float desired_error = 0.0001F;
const uint max_epochs = 5000;
const uint epochs_between_reports = 1000;
using(NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, num_input, num_neurons_hidden, num_neurons_hidden, num_output))
{
net.ActivationFunctionHidden = ActivationFunction.SIGMOID_SYMMETRIC;
net.ActivationFunctionOutput = ActivationFunction.LINEAR;
net.TrainingAlgorithm = TrainingAlgorithm.TRAIN_RPROP;
using (TrainingData data = new TrainingData("..\\..\\..\\datasets\\scaling.data"))
{
net.SetScalingParams(data, -1, 1, -1, 1);
net.ScaleTrain(data);
net.TrainOnData(data, max_epochs, epochs_between_reports, desired_error);
net.Save("..\\..\\..\\datasets\\scaling.net");
Console.ReadKey();
}
}
}
static void Main(string[] args)
{
DataType[] calc_out;
const uint num_input = 2;
const uint num_output = 1;
const uint num_layers = 3;
const uint num_neurons_hidden = 3;
const float desired_error = 0;
const uint max_epochs = 1000;
const uint epochs_between_reports = 10;
int decimal_point;
Console.WriteLine("Creating network.");
using (NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, num_input, num_neurons_hidden, num_output))
using (TrainingData data = new TrainingData("..\\..\\..\\examples\\xor.data"))
{
net.ActivationFunctionHidden = ActivationFunction.SIGMOID_SYMMETRIC;
net.ActivationFunctionOutput = ActivationFunction.SIGMOID_SYMMETRIC;
net.TrainStopFunction = StopFunction.STOPFUNC_BIT;
net.BitFailLimit = 0.01F;
net.TrainingAlgorithm = TrainingAlgorithm.TRAIN_RPROP;
net.InitWeights(data);
Console.WriteLine("Training network.");
net.TrainOnData(data, max_epochs, epochs_between_reports, desired_error);
Console.WriteLine("Testing network");
// Keep a copy of the inputs and outputs so that we don't call TrainingData.Input
// and TrainingData.Output multiple times causing a copy of all the data on each
// call. An alternative is to use the Input/OutputAccessors which are fast with
// repeated calls to get data and can be cast to arrays with the Array property
DataType[][] input = data.Input;
DataType[][] output = data.Output;
for (int i = 0; i < data.TrainDataLength; i++)
{
calc_out = net.Run(input[i]);
Console.WriteLine("XOR test ({0},{1}) -> {2}, should be {3}, difference={4}",
input[i][0], input[i][1], calc_out[0], output[i][0],
FannAbs(calc_out[0] - output[i][0]));
}
Console.WriteLine("Saving network.\n");
net.Save("..\\..\\..\\examples\\xor_float.net");
decimal_point = net.SaveToFixed("..\\..\\..\\examples\\xor_fixed.net");
data.SaveTrainToFixed("..\\..\\..\\examples\\xor_fixed.data", (uint)decimal_point);
Console.ReadKey();
}
}
static void Main()
{
const uint num_layers = 3;
const uint num_neurons_hidden = 32;
const float desired_error = 0.0001F;
const uint max_epochs = 300;
const uint epochs_between_reports = 10;
Console.WriteLine("Creating network.");
using (TrainingData data = new TrainingData("..\\..\\..\\datasets\\mushroom.train"))
using (NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, data.InputCount, num_neurons_hidden, data.OutputCount))
{
Console.WriteLine("Training network.");
net.ActivationFunctionHidden = ActivationFunction.SIGMOID_SYMMETRIC;
net.ActivationFunctionOutput = ActivationFunction.SIGMOID;
net.TrainOnData(data, max_epochs, epochs_between_reports, desired_error);
Console.WriteLine("Testing network.");
using (TrainingData testData = new TrainingData())
{
testData.ReadTrainFromFile("..\\..\\..\\datasets\\mushroom.test");
net.ResetMSE();
for (int i = 0; i < testData.TrainDataLength; i++)
{
// The difference between calling GetTrain[Input|Output] and calling
// the Input and Output properties is huge in terms of speed
net.Test(testData.GetTrainInput((uint)i).Array, testData.GetTrainOutput((uint)i).Array);
}
Console.WriteLine("MSE error on test data {0}", net.MSE);
Console.WriteLine("Saving network.");
net.Save("..\\..\\..\\examples\\mushroom_float.net");
Console.ReadKey();
}
}
}
static void Main()
{
const uint num_layers = 3;
const uint num_neurons_hidden = 96;
const float desired_error = 0.001F;
Console.WriteLine("Creating network.");
using (TrainingData data = new TrainingData("..\\..\\..\\datasets\\robot.train"))
using (NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, data.InputCount, num_neurons_hidden, data.OutputCount))
using (TrainingData testData = new TrainingData())
{
Console.WriteLine("Training network.");
net.TrainingAlgorithm = TrainingAlgorithm.TRAIN_INCREMENTAL;
net.LearningMomentum = 0.4F;
net.TrainOnData(data, 3000, 10, desired_error);
Console.WriteLine("Testing network.");
testData.ReadTrainFromFile("..\\..\\..\\datasets\\robot.test");
try
{
net.ResetMSE();
for (int i = 0; i < testData.TrainDataLength; i++)
{
net.Test(testData.GetTrainInput((uint)i).Array, testData.GetTrainOutput((uint)i).Array);
}
Console.WriteLine("MSE error on test data: {0}", net.MSE);
Console.WriteLine("Saving network.");
net.Save("..\\..\\..\\datasets\\robot_float.net");
}
catch (Exception e)
{
Console.WriteLine("Exception: {0}", e.Message);
}
Console.ReadKey();
}
}
static void Main()
{
const uint num_layers = 3;
const uint num_neurons_hidden = 96;
const float desired_error = 0.00007F;
using (TrainingData trainData = new TrainingData())
using (TrainingData testData = new TrainingData())
{
trainData.CreateTrainFromCallback(374, 48, 3, TrainingDataCallback);
testData.CreateTrainFromCallback(594, 48, 3, TestDataCallback);
// Test Accessor classes
for (int i = 0; i < trainData.TrainDataLength; i++)
{
Console.Write("Input {0}: ", i);
for (int j = 0; j < trainData.InputCount; j++)
{
Console.Write("{0}, ", trainData.InputAccessor[i][j]);
}
Console.Write("\nOutput {0}: ", i);
for (int j = 0; j < trainData.OutputCount; j++)
{
Console.Write("{0}, ", trainData.OutputAccessor[i][j]);
}
Console.WriteLine("");
}
for (float momentum = 0.0F; momentum < 0.7F; momentum += 0.1F)
{
Console.WriteLine("============= momentum = {0} =============\n", momentum);
using (NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, trainData.InputCount, num_neurons_hidden, trainData.OutputCount))
{
net.SetCallback(TrainingCallback, "Hello!");
net.TrainingAlgorithm = TrainingAlgorithm.TRAIN_INCREMENTAL;
net.LearningMomentum = momentum;
net.TrainOnData(trainData, 20000, 500, desired_error);
Console.WriteLine("MSE error on train data: {0}", net.TestData(trainData));
Console.WriteLine("MSE error on test data: {0}", net.TestData(testData));
}
}
}
Console.ReadKey();
}
static void XorTest()
{
Console.WriteLine("\nXOR test started.");
const float learning_rate = 0.7f;
const uint num_layers = 3;
const uint num_input = 2;
const uint num_hidden = 3;
const uint num_output = 1;
const float desired_error = 0.001f;
const uint max_iterations = 300000;
const uint iterations_between_reports = 1000;
Console.WriteLine("\nCreating network.");
using (NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, num_input, num_hidden, num_output))
{
net.LearningRate = learning_rate;
net.ActivationSteepnessHidden = 1.0F;
net.ActivationSteepnessOutput = 1.0F;
net.ActivationFunctionHidden = ActivationFunction.SIGMOID_SYMMETRIC_STEPWISE;
net.ActivationFunctionOutput = ActivationFunction.SIGMOID_SYMMETRIC_STEPWISE;
// Output network type and parameters
Console.Write("\nNetworkType : ");
switch (net.NetworkType)
{
case NetworkType.LAYER:
Console.WriteLine("LAYER");
break;
case NetworkType.SHORTCUT:
Console.WriteLine("SHORTCUT");
break;
default:
Console.WriteLine("UNKNOWN");
break;
}
net.PrintParameters();
Console.WriteLine("\nTraining network.");
using (TrainingData data = new TrainingData())
{
if (data.ReadTrainFromFile("..\\..\\..\\examples\\xor.data"))
{
// Initialize and train the network with the data
net.InitWeights(data);
Console.WriteLine("Max Epochs " + String.Format("{0:D}", max_iterations).PadLeft(8) + ". Desired Error: " + String.Format("{0:F}", desired_error).PadRight(8));
net.SetCallback(PrintCallback, null);
net.TrainOnData(data, max_iterations, iterations_between_reports, desired_error);
Console.WriteLine("\nTesting network.");
for (uint i = 0; i < data.TrainDataLength; i++)
{
// Run the network on the test data
DataType[] calc_out = net.Run(data.Input[i]);
Console.WriteLine("XOR test ({0}, {1}) -> {2}, should be {3}, difference = {4}",
data.InputAccessor[(int)i][0].ToString("+#;-#"),
data.InputAccessor[(int)i][1].ToString("+#;-#"),
calc_out[0] == 0 ? 0.ToString() : calc_out[0].ToString("+#.#####;-#.#####"),
data.OutputAccessor[(int)i][0].ToString("+#;-#"),
FannAbs(calc_out[0] - data.Output[i][0]));
}
Console.WriteLine("\nSaving network.");
// Save the network in floating point and fixed point
net.Save("..\\..\\..\\examples\\xor_float.net");
uint decimal_point = (uint)net.SaveToFixed("..\\..\\..\\examples\\xor_fixed.net");
data.SaveTrainToFixed("..\\..\\..\\examples\\xor_fixed.data", decimal_point);
Console.WriteLine("\nXOR test completed.");
}
}
}
}
static void Main()
{
const uint num_layers = 3;
const uint num_neurons_hidden = 32;
const float desired_error = 0.0001F;
const uint max_epochs = 300;
const uint epochs_between_reports = 10;
long before;
Console.WriteLine("Creating network.");
using (TrainingData data = new TrainingData("..\\..\\..\\datasets\\mushroom.train"))
using (NeuralNet net = new NeuralNet(NetworkType.LAYER, num_layers, data.InputCount, num_neurons_hidden, data.OutputCount))
{
Console.WriteLine("Training network.");
net.ActivationFunctionHidden = ActivationFunction.SIGMOID_SYMMETRIC;
net.ActivationFunctionOutput = ActivationFunction.SIGMOID;
net.TrainOnData(data, max_epochs, epochs_between_reports, desired_error);
Console.WriteLine("Testing network.");
using (TrainingData testData = new TrainingData())
{
testData.ReadTrainFromFile("..\\..\\..\\datasets\\mushroom.test");
before = Environment.TickCount;
for (int i = 0; i < 5; i++)
{
DataType[] input = new DataType[testData.InputCount];
DataType[] output = new DataType[testData.OutputCount];
for(int j = 0; j < testData.InputCount; j++)
{
input[j] = testData.InputAccessor[i][j];
}
for (int j = 0; j < testData.OutputCount; j++)
{
output[j] = testData.OutputAccessor[i][j];
}
net.Test(input, output);
}
Console.WriteLine("Data Accessor ticks {0}", Environment.TickCount - before);
before = Environment.TickCount;
for (int i = 0; i < 5; i++)
{
DataType[] input = new DataType[testData.InputCount];
DataType[] output = new DataType[testData.OutputCount];
for (int j = 0; j < testData.InputCount; j++)
{
input[j] = testData.Input[i][j];
}
for (int j = 0; j < testData.OutputCount; j++)
{
output[j] = testData.Output[i][j];
}
net.Test(input, output);
}
Console.WriteLine("Array ticks {0}", Environment.TickCount - before);
}
Console.WriteLine("MSE error on test data {0}", net.MSE);
Console.WriteLine("Saving network.");
net.Save("..\\..\\..\\examples\\mushroom_float.net");
Console.ReadKey();
}
}