public ClassifierRecurrentNeuralNetwork(int input, int output, int hidden, int units, float learningRate, float std, EnumOptimizerFunction opt)
{
encoder = new RecurrentCellModel();
decoder = new DenseModel();
int x = input;
for (int i = 0; i < units; i++)
{
encoder.AddLayer(BuildedModels.RecurrentTanh(x, hidden, learningRate, std, opt));
x = hidden;
}
decoder.AddLayer(BuildedModels.DenseSoftMax(hidden, output, learningRate, std, opt));
decoder.SetLossFunction(new CrossEntropyLossFunction());
input_size = input;
hidden_size = hidden;
output_size = output;
recurrentUnits = units;
learning_rate = learningRate;
}
private static void Main(string[] args)
{
Console.Clear();
var watch = Stopwatch.StartNew();
ProcessingDevice.Device = DeviceType.CPU_Parallel;
watch.Stop();
Console.WriteLine($"Device Time: {watch.ElapsedMilliseconds}ms");
var model = new DenseModel();
model.AddLayer(BuildedModels.DenseLeakRelu(784, 100, 1e-4f, 2e-1f, EnumOptimizerFunction.RmsProp));
model.AddLayer(BuildedModels.DenseLeakRelu(100, 30, 1e-4f, 2e-1f, EnumOptimizerFunction.RmsProp));
model.AddLayer(BuildedModels.DenseSoftMax(30, 10, 1e-4f, 2e-1f, EnumOptimizerFunction.RmsProp));
model.SetLossFunction(new CrossEntropyLossFunction());
watch = Stopwatch.StartNew();
watch.Stop();
Console.WriteLine($"Sinapse Time: {watch.ElapsedMilliseconds}ms");
MnistLoader.DataPath = path;
var trainingValues = MnistLoader.OpenMnist();
int cont = 0;
int sizeTrain = trainingValues.Count;
var err = 100f;
var e = 0f;
while (true)
{
watch = Stopwatch.StartNew();
e = 0f;
int ct = 0;
for (int i = 0; i < sizeTrain; i++)
{
var index = i;
var inputs = new FloatArray(ArrayMethods.ByteToArray(trainingValues[index].pixels, 28, 28));
var target = new FloatArray(ArrayMethods.ByteToArray(trainingValues[index].label, 10));
// Learning
e += model.Learn(inputs, target);
// Sample
if (ct % 1000 == 0)
{
Write(trainingValues[index].ToString());
Write(ArrayMethods.PrintArray(model.Output(inputs), 10));
Write("\n");
}
ct++;
}
err = 0.999f * err + 0.001f * e;
cont++;
watch.Stop();
var time = watch.ElapsedMilliseconds;
Console.Title =
$"Error: {err} --- TSPS (Training Sample per Second): {Math.Ceiling(1000d / ((double)time / (double)sizeTrain))}";
}
}