internal void SetOptimizer(Optimizer.Optimizer optimizer)
{
foreach (var connection in _connections)
{
connection.Optimizer = optimizer.Clone();
}
}
public static void Training(Network network, DataSet.DataSet data, Optimizer.Optimizer optimizer = null,
float limitError = 1e-5f, bool printLog = false)
{
var error = float.MaxValue;
var it = 0;
var usedOptimizer = optimizer?.Clone() ?? DefaultOptimizer.Clone();
usedOptimizer.LearningRate /= data.TrainingDataSize;
network.SetOptimizer(usedOptimizer);
while (error > limitError)
{
if (printLog)
{
WriteLine($"Epoch {it++}");
}
error = 0.0f;
network.ClearDeltaW();
foreach (var datum in data.TrainingData())
{
network.SetInputs(datum.Input);
network.ForwardPropagation();
error += network.Error(datum.Output);
network.BackPropagation(datum.Output);
}
network.UpdateWeights();
error /= data.TrainingDataSize;
if (printLog)
{
WriteLine($"Error : {error: ##0.00000;-##0.00000}");
}
}
network.FinishLearning();
}
public static void PreTraining(Network network, DataSet.DataSet data, Optimizer.Optimizer optimizer = null,
int epoch = 1, int batchSize = 1, bool printLog = false)
{
var trainingDataSize = data.TrainingDataSize;
var iterationSize = trainingDataSize / batchSize;
var usedOptimizer = optimizer?.Clone() ?? DefaultOptimizer.Clone();
usedOptimizer.LearningRate /= batchSize;
network.SetOptimizer(usedOptimizer);
Network.Factory.DockIn(network);
for (var hidLayerSize = 1; hidLayerSize < network.LayersCount - 1; hidLayerSize++)
{
if (printLog)
{
WriteLine($"Building auto encoder {hidLayerSize}.");
}
var partial = Network.Factory.PartialNetwork(hidLayerSize);
partial.SetOptimizer(usedOptimizer);
for (var i = 0; i < epoch; i++)
{
if (printLog)
{
WriteLine($"Epoch {i}");
Write("Error : ");
}
for (var j = 0; j < iterationSize; j++)
{
partial.ClearDeltaW();
partial.CopyWeightValues();
var error = 0.0f;
foreach (var datum in data.MiniBatch(batchSize))
{
partial.SetInputs(datum.Input);
partial.ForwardPropagation();
if (printLog)
{
error += partial.Error();
}
partial.PartialBackPropagation();
}
partial.UpdatePartialWeights();
if (printLog)
{
Write($"\rError : {error / batchSize: ##0.00000;-##0.00000}");
}
}
if (printLog)
{
WriteLine();
}
}
partial.FinishLearning();
}
Network.Factory.OutOfDock();
}
public static void Training(Network network, DataSet.DataSet data, Optimizer.Optimizer optimizer = null,
int epoch = 1, int batchSize = 1, bool printLog = false)
{
var trainingDataSize = data.TrainingDataSize;
var iterationSize = trainingDataSize / batchSize;
var usedOptimizer = optimizer?.Clone() ?? DefaultOptimizer.Clone();
usedOptimizer.LearningRate /= batchSize;
network.SetOptimizer(usedOptimizer);
for (var i = 0; i < epoch; i++)
{
WriteLine($"Epoch {i}");
if (printLog)
{
Write("Error : ");
}
for (var j = 0; j < iterationSize; j++)
{
network.ClearDeltaW();
var error = 0.0f;
foreach (var datum in data.MiniBatch(batchSize))
{
network.SetInputs(datum.Input);
network.ForwardPropagation();
if (printLog)
{
error += network.Error(datum.Output);
}
network.BackPropagation(datum.Output);
}
network.UpdateWeights();
if (printLog)
{
Write($"\rError : {error / batchSize: ##0.00000;-##0.00000}");
}
}
if (printLog)
{
WriteLine();
}
}
network.FinishLearning();
}