public void TrainNetwork(SomNetwork network, Vector input)
{
var bmu = BestMatchingUnit(network, input);
UpdateNetwork(network, bmu, input);
network.Time++;
}
public Bmu DrawInput(SomNetwork network, DataPoint data)
{
var bmu = BestMatchingUnit(network, data.Input);
var color = new Color();
switch (data.Output.AbsoluteMaximumIndex())
{
case 0:
color = Color.Blue;
break;
case 1:
color = Color.BlueViolet;
break;
case 2:
color = Color.Turquoise;
break;
case 3:
color = Color.Green;
break;
case 4:
color = Color.DeepSkyBlue;
break;
case 5:
color = Color.Yellow;
break;
case 6:
color = Color.Orange;
break;
case 7:
color = Color.DeepPink;
break;
case 8:
color = Color.Red;
break;
case 9:
color = Color.SaddleBrown;
break;
default:
color = Color.White;
break;
}
bmu.Color = color;
return(bmu);
}
public void UpdateNetwork(SomNetwork network, Bmu bmu, Vector input)
{
var radius = network.Radius * SomUtilities.Shrink(network.Time, network.T0);
foreach (SomNode node in network.Weights)
{
if (Distance.Euclidean(bmu.Position, node.Position) < radius)
{
Vector delta = (Vector)(
SomUtilities.Shrink(network.Time, network.T0) *
SomUtilities.Shrink(Distance.Euclidean(bmu.Position, node.Position), network.L0) *
network.LearningRate * (input - node.Weight));
node.Weight = (Vector)(node.Weight + delta);
}
}
}
public Bmu BestMatchingUnit(SomNetwork network, Vector input)
{
//Inelegant.
var bmu = new Bmu()
{
Distance = int.MaxValue
};
foreach (SomNode node in network.Weights)
{
if (Distance.Euclidean(node.Weight, input) < bmu.Distance)
{
bmu.Position = node.Position;
bmu.Distance = Distance.Euclidean(node.Weight, input);
}
}
return(bmu);
}
public Bitmap Illustrate(SomNetwork network)
{
//var lengths = (from somNode1 in network.Weights from somNode2 in network.Weights where network.Weights.IndexOf(somNode2) < network.Weights.IndexOf(somNode1) select Distance.Euclidean(somNode2.Weight, somNode1.Weight)).ToList();
var lengths =
network.Weights.Select(
parentNode => network.Neighbors(parentNode, 1.5).Average(childNode => Distance.Euclidean(childNode.Weight, parentNode.Weight))).ToList();
//var lengths = network.Weights.Select(node => node.Weight.L2Norm()).ToList();
var mean = lengths.Average();
var quan = lengths.Quantile(0.02);
var standardDeviation = lengths.StandardDeviation();
var min = lengths.Min();
var bitmap = new Bitmap(network.Width, network.Height);
Graphics g = Graphics.FromImage(bitmap);
foreach (var node in network.Weights)
{
int i = Convert.ToInt32(node.Position[0]);
int j = Convert.ToInt32(node.Position[1]);
var neighbors = network.Neighbors(node, 1.5);
var dist = 35 * (neighbors.Average(n => Distance.Euclidean(n.Weight, node.Weight)) - quan) / standardDeviation;
var strength = Convert.ToInt32(dist);
if (strength > 255)
{
strength = 255;
}
if (strength < 0)
{
strength = 0;
}
var color = Color.FromArgb(strength, Color.Black);
bitmap.SetPixel(j, i, color);
}
return(bitmap);
}
static void Main(string[] args)
{
//NeuralNetwork network = new NeuralNetwork(new List<int> {784, 30, 10}, stepSize:0.5,regularizationParameter:0.0001);
////TODO: take the paths as inputs
var trainingData = DataHelper.LoadData("../../../MNIST/train-images.idx3-ubyte", "../../../MNIST/train-labels.idx1-ubyte", numberOfImages: 50000);
var testData = DataHelper.LoadData("../../../MNIST/t10k-images.idx3-ubyte", "../../../MNIST/t10k-labels.idx1-ubyte", numberOfImages: 10000);
NetworkUtilities networkUtilities = new NetworkUtilities();
NetworkHelper networkHelper = new NetworkHelper(networkUtilities);
//int batchSize = 50;
//int epochs = 30;
//int length = trainingData.Count;
//int testLength = testData.Count;
//var network = new NeuralNetwork(new List<int> {784, 30, 10}, 0.1, 0.001);
//for (int k = 0; k < epochs; k++)
//{
// trainingData.Shuffle();
// for (int j = 0; j < length / batchSize; j++)
// {
// var batch = trainingData.GetRange(j * batchSize, batchSize);
// networkHelper.TrainNetwork(network, batch);
// }
// Console.WriteLine("Epoch: " + (k + 1));
// Console.WriteLine(testData.Sum(x => x.Output[networkHelper.Evaluate(x.Input, network).AbsoluteMaximumIndex()]) + " / " + testLength);
// DatabaseHelper.SaveNetworkToDatabase(network, k + 1);
//}
//Console.ReadKey();
//trainingData.Shuffle();
SomNetwork somNetwork = new SomNetwork(height: 50, width: 50, t0: 2500,
learningRate: 0.08, l0: 100, inputDimension: 784);
trainingData.Shuffle();
trainingData = trainingData.Take(10000).ToList();
SomHelper helper = new SomHelper();
NetworkIllustrator illustrator = new NetworkIllustrator();
for (int k = 0; k < trainingData.Count(); k++)
{
helper.TrainNetwork(somNetwork, trainingData[k].Input);
}
var bitmap = illustrator.Illustrate(somNetwork);
//Illustrate the first test data points
for (int k = 0; k < 800; k++)
{
var bmu = helper.DrawInput(somNetwork, testData[k]);
int i = Convert.ToInt32(bmu.Position[0]);
int j = Convert.ToInt32(bmu.Position[1]);
bitmap.SetPixel(j, i, bmu.Color);
}
bitmap.Save("Network.bmp");
}
