public static void loadFromSaved()
{
DirectoryInfo experimentsDirectory = new DirectoryInfo("C://nnlib/experiments");
if (!experimentsDirectory.Exists)
{
experimentsDirectory.Create();
}
else
{
DirectoryInfo[] directories = experimentsDirectory.GetDirectories();
foreach (DirectoryInfo directory in directories)
{
try {
string info = File.ReadAllText(directory.FullName + "/info.json");
if (info.Equals(""))
{
return;
}
ExperimentLog experiment = JsonConvert.DeserializeObject <ExperimentLog>(info);
experiments.Add(experiment.id, experiment);
} catch (Exception e) {
Console.WriteLine("Failed to get experiment from {0}", directory.FullName);
}
}
}
}
public static Experiment?getExperiment(string id)
{
ExperimentLog log = experiments.First(experiment => experiment.Key == id).Value;
return(new Experiment {
id = log.id,
name = log.title,
description = log.description,
amountOfParallelNetworks = log.amountOfParallelNetworks,
duration = (long)Math.Round(log.duration.TotalMilliseconds),
initDateTime = log.initDateTime.Ticks / 1000,
finishDateTime = log.finishDateTime.Ticks / 1000,
phases = log.phases.Select(phase => new ExperimentPhase {
title = phase.title,
configuration = phase.configuration,
iterations = phase.iterations,
duration = (long)Math.Round(phase.duration.TotalMilliseconds),
startDateTime = phase.startDateTime.Ticks / 1000,
finishDateTime = phase.finishDateTime.Ticks / 1000,
}).ToList(),
topologyLayerRecords = log.topologyLayerRecords.Select(record => new LayerTopologyRecord {
type = record.type,
numberOfUnits = record.numberOfNeurons
}).ToList()
});
}
public static void endLogAndWriteToFile(ExperimentLog log, string rootPath)
{
createRootResultsDirectoryIfNotExists(rootPath);
log.end();
File.WriteAllText($"{rootPath}/log.txt", log.getString());
File.WriteAllText($"{rootPath}/info.json", log.getJsonString());
}
public static void train(List <NeuralNetwork> nns, int iterations, ExperimentLog log = null)
{
FileStream trainImages = new FileStream("data/train_imgs", FileMode.Open);
FileStream trainLabels = new FileStream("data/train_lbls", FileMode.Open);
trainImages.Read(new byte[4 * 4], 0, 4 * 4);
trainLabels.Read(new byte[4 * 2], 0, 4 * 2);
List <double> answer = new List <double>();
for (int i = 0; i < 10; i++)
{
answer.Add(0);
}
List <double> input = new List <double>();
for (int i = 0; i < 784; i++)
{
input.Add(0);
}
log?.startPhase("Training",
$"MNIST dataset, learning step {learningStep}, correct answer as 1, wrong answer as 0",
iterations * 1000);
for (int h = 0; h < iterations; h++)
{
for (int i = 0; i < 1000; i++)
{
int digit = Utils.getNextByte(trainLabels);
byte[] byteInput = new byte[784];
trainImages.Read(byteInput, 0, 784);
for (int j = 0; j < 784; j++)
{
input[j] = byteInput[j] / 255d;
}
foreach (NeuralNetwork nn in nns)
{
nn.putData(input);
nn.run();
}
answer[digit] = 1;
foreach (NeuralNetwork nn in nns)
{
nn.backpropagate(answer, learningStep);
}
answer[digit] = 0;
}
Console.Write(".");
}
log?.endPhase();
}
public static void run()
{
Console.Write("Hello. Enter experiments title: ");
string experimentTitle = Console.ReadLine();
Console.Write("Enter experiments description: ");
string experimentDescription = Console.ReadLine();
string rootPath = $"C://nnlib/experiments/{experimentTitle}";
ExperimentLog log = new ExperimentLog(experimentTitle, experimentDescription, parallelNns);
Console.Write("Initialisation of NNs...");
List <NeuralNetwork> nns = setupNNs(log);
Console.WriteLine(" Done.");
Console.WriteLine("Training started!");
Console.Write("In progress");
train(nns, trainSize, log);
Console.WriteLine(" Done.");
Console.WriteLine("Testing started!");
Console.Write("In progress");
NNsTestResults <int> testResults = test(nns, testSize, log);
Console.WriteLine(" Done.");
Console.Write("Saving log... ");
Utils.endLogAndWriteToFile(log, rootPath);
Console.WriteLine(" Done.");
Console.Write("Writing results to excel... ");
Utils.writeToExcel(testResults, rootPath);
Console.WriteLine(" Done.");
Console.Write("Writing NNs configurations to files... ");
Utils.writeNNsToFiles(nns, rootPath);
Console.WriteLine(" Done.");
Console.WriteLine("Process is completed and result is saved in excel file.");
Console.ReadKey();
}
public static void run()
{
Console.Write("Hello. Enter experiments title: ");
string experimentTitle = Console.ReadLine();
Console.Write("Enter experiments description: ");
string experimentDescription = Console.ReadLine();
string rootPath = $"./experiments/{experimentTitle}";
ExperimentLog log = new ExperimentLog(experimentTitle, experimentDescription, 1);
Console.Write("Enter config files name: ./experiments/loads/");
string loadName = Console.ReadLine();
Console.Write("Deserialization of NN...");
NeuralNetwork nn = deserialize($"./experiments/loads/{loadName}", log);
Console.WriteLine(" Done.");
Console.WriteLine("Testing started!");
Console.Write("In progress");
NNsTestResults <int> testResults = Mnist.test(new List <NeuralNetwork> {
nn
}, testSize, log);
Console.WriteLine(" Done.");
Console.Write("Saving log... ");
Utils.endLogAndWriteToFile(log, rootPath);
Console.WriteLine(" Done.");
Console.Write("Writing data to excel... ");
Utils.writeToExcel(testResults, rootPath);
Console.WriteLine(" Done.");
Console.WriteLine("Process is completed and result is saved in excel file.");
Console.ReadKey();
}
public static NNsTestResults <int> test(List <NeuralNetwork> nns, int iterations, ExperimentLog log = null)
{
FileStream testImages = new FileStream("data/test_imgs", FileMode.Open);
FileStream testLabels = new FileStream("data/test_lbls", FileMode.Open);
testImages.Read(new byte[4 * 4], 0, 4 * 4);
testLabels.Read(new byte[4 * 2], 0, 4 * 2);
List <double> input = new List <double>();
for (int i = 0; i < 784; i++)
{
input.Add(0);
}
List <int> correctAnswersAmount = nns.Select(nn => 0).ToList();
List <int> correctAnswers = new List <int>(iterations * 1000);
List <List <int> > answers = nns.Select(nn => new List <int>()).ToList();
log?.startPhase("Testing",
$"MNIST dataset",
iterations * 1000);
for (int h = 0; h < iterations; h++)
{
for (int i = 0; i < 1000; i++)
{
int digit = Utils.getNextByte(testLabels);
correctAnswers.Add(digit);
byte[] byteInput = new byte[784];
testImages.Read(byteInput, 0, 784);
for (int j = 0; j < 784; j++)
{
input[j] = byteInput[j] / 255d;
}
for (int nnIndex = 0; nnIndex < nns.Count; nnIndex++)
{
NeuralNetwork nn = nns[nnIndex];
nn.putData(input);
nn.run();
int answer = nn.getMaxIndexInOutput();
if (answer == digit)
{
correctAnswersAmount[nnIndex]++;
}
answers[nnIndex].Add(answer);
}
}
Console.Write(".");
}
log?.endPhase();
return(new NNsTestResults <int>(correctAnswersAmount, correctAnswers, answers));
}
