static void Main(string[] args)
{
// Network initialize:
ServiceCNNImages evaNetwork = new ServiceCNNImages();
NetworkStructure netStructure = new NetworkStructure
{
InputVectorLength = 192,
NeuronsByLayers = new[] { 200, 200, 1 }
};
evaNetwork.CreateNetwork(netStructure, 3, "testDatasets.txt");
// Main training:
TrainConfiguration trainConfig = new TrainConfiguration
{
StartIteration = 26000000,
EndIteration = 29110000,
InputDatasetFilename = "inputSets.txt",
OutputDatasetFilename = "outputSets.txt",
SourceFolderName = "images",
MemoryFolder = ""
};
evaNetwork.Train(trainConfig, 3110000);
// TODO: Temp
DirectoryInfo dirInfo = new DirectoryInfo("testImages");
double[] sssss = evaNetwork.Handle(dirInfo.GetFiles()[0].FullName);
Console.ReadKey();
}
public void Train(TrainConfiguration trainConfiguration, int iterationToPause = 100)
{
if (_networkTeacher.CheckMemory(trainConfiguration.MemoryFolder))
{
_networkTeacher.TrainNets(trainConfiguration, iterationToPause);
_networkTeacher.PrintLearnStatistic(trainConfiguration, true);
if (_networkTeacher.CheckMemory(trainConfiguration.MemoryFolder))
{
_networkTeacher.BackupMemory();
}
}
else
{
Console.WriteLine("Train failed! Invalid memory!");
}
}
private List <TrainConfiguration> InitializeTrainConfigs(TrainConfiguration trainConfig, int iterationsToPause)
{
List <TrainConfiguration> trainConfigs = new List <TrainConfiguration>();
int currentIterPosition = trainConfig.StartIteration;
while (true)
{
if (trainConfig.EndIteration - currentIterPosition - 1 >= iterationsToPause)
{
var trainConfigItem = new TrainConfiguration()
{
StartIteration = currentIterPosition,
EndIteration = currentIterPosition + iterationsToPause,
InputDatasetFilename = trainConfig.InputDatasetFilename,
OutputDatasetFilename = trainConfig.OutputDatasetFilename
};
trainConfigs.Add(trainConfigItem);
currentIterPosition += iterationsToPause;
}
else
{
var trainConfigItem = new TrainConfiguration()
{
StartIteration = currentIterPosition,
EndIteration = trainConfig.EndIteration,
InputDatasetFilename = trainConfig.InputDatasetFilename,
OutputDatasetFilename = trainConfig.OutputDatasetFilename
};
trainConfigs.Add(trainConfigItem);
break;
}
}
Console.WriteLine("Train configuration object created!");
return(trainConfigs);
}
public void CreateTrainSets(Extractor extractor, ImageLoader imageLoader, TrainConfiguration trainConfig)
{
Console.WriteLine("Start creating train sets...");
List <double[]> inputDataSets = new List <double[]>();
List <double[]> outputDataSets = new List <double[]>();
// Загрузка файлов из папки
string[] outputClassesFolders = Directory.GetDirectories(trainConfig.SourceFolderName);
for (int i = 0; i < outputClassesFolders.Length; i++)
{
// Получение информации о файлах в текущей папке:
string[] fileInfo = Directory.GetFiles(outputClassesFolders[i]);
for (int k = 0; k < fileInfo.Length; k++)
{
// Создание входного вектора:
List <double[, ]> imageMatrix = imageLoader.LoadImageDataRGB(fileInfo[k]);
inputDataSets.Add(extractor.Extract(imageMatrix));
// Создание выходного вектора:
double[] outputVector = new double[outputClassesFolders.Length];
outputVector[i] = 1;
outputDataSets.Add(outputVector);
}
}
Console.WriteLine("Creating completed!");
Console.WriteLine("Saving train sets data...");
//bool isEqual = Enumerable.SequenceEqual(inputDataSets[0], inputDataSets[1]);
SaveSets(inputDataSets, outputDataSets, trainConfig.InputDatasetFilename, trainConfig.OutputDatasetFilename);
Console.WriteLine("Save train sets completed!");
}
/// <summary>
/// Обучение сети
/// </summary>
/// <param name="startIteration"></param>
/// <param name="withSort"></param>
public void TrainNets(TrainConfiguration trainConfig, int iterationsToPause)
{
Iteration = trainConfig.EndIteration;
// 1. Создание обучающих сетов из указанных папок:
_fileManager.CreateTrainSets(_extractor, _imageLoader, trainConfig);
#region Load data from file
Console.WriteLine("Load train sets...");
List <double[]> inputDataSets;
List <double[]> outputDataSets;
try
{
inputDataSets = _fileManager.LoadSingleDataset(trainConfig.InputDatasetFilename);
outputDataSets = _fileManager.LoadSingleDataset(trainConfig.OutputDatasetFilename);
}
catch (Exception ex)
{
_logger.LogError(ErrorType.SetMissing, ex);
return;
}
#endregion
Console.WriteLine("Training nets...");
try
{
SingleNetworkTeacher[] netTeachers = new SingleNetworkTeacher[_netsList.Count];
List <TrainConfiguration> trainConfigs = InitializeTrainConfigs(trainConfig, iterationsToPause);
// Initialize teachers:
for (int i = 0; i < netTeachers.Length; i++)
{
netTeachers[i] = new SingleNetworkTeacher
{
Id = i,
Network = _netsList[i],
TrainConfiguration = trainConfig,
InputDatasets = inputDataSets,
OutputDatasets = outputDataSets,
Logger = _logger
};
}
List <Task> tasks;
// Iteration multithreading train:
for (int j = 0; j < trainConfigs.Count; j++)
{
tasks = new List <Task>();
for (int i = 0; i < netTeachers.Length; i++)
{
var teacherNumber = i;
var task = Task.Run(() => { netTeachers[teacherNumber].Train(); });
tasks.Add(task);
}
;
Task.WaitAll(tasks.ToArray());
if (j != trainConfigs.Count - 1)
{
Console.WriteLine("Iterations already finished: " + iterationsToPause * (j + 1));
}
else
{
Console.WriteLine("Iterations already finished: " + trainConfig.EndIteration);
}
CommonTestColorized();
}
Console.WriteLine("Training success!");
}
catch (Exception ex)
{
_logger.LogError(ErrorType.TrainError, ex);
}
}
public void PrintLearnStatistic(TrainConfiguration trainConfig, bool withLogging = false)
{
Console.WriteLine("Start calculating statistic...");
int testPassed = 0;
int testFailed = 0;
int testFailed_lowActivationCause = 0;
#region Load data from file
List <double[]> inputDataSets;
List <double[]> outputDataSets;
try
{
inputDataSets = _fileManager.LoadSingleDataset(trainConfig.InputDatasetFilename);
outputDataSets = _fileManager.LoadSingleDataset(trainConfig.OutputDatasetFilename);
}
catch (Exception ex)
{
_logger.LogError(ErrorType.SetMissing, ex);
return;
}
#endregion
for (int i = 0; i < inputDataSets.Count; i++)
{
List <double> netResults = new List <double>();
for (int k = 0; k < _netsList.Count; k++)
{
// Получение ответа:
netResults.Add(_netsList[k].Handle(inputDataSets[i])[0]);
}
// Поиск максимально активирующейся сети (класса) с заданным порогом активации:
int maxIndex = FindMaxIndex(netResults, 0.8);
if (maxIndex == -1)
{
testFailed++;
testFailed_lowActivationCause++;
}
else
{
if (outputDataSets[i][maxIndex] != 1)
{
testFailed++;
}
else
{
testPassed++;
}
}
}
// Logging (optional):
if (withLogging)
{
_logger.LogTrainResults(testPassed, testFailed, testFailed_lowActivationCause, Iteration);
}
Console.WriteLine("Test passed: {0}\nTest failed: {1}\n - Low activation causes: {2}\nPercent learned: {3:f2}", testPassed,
testFailed,
testFailed_lowActivationCause,
(double)testPassed * 100 / (testPassed + testFailed));
}
