//Neuron Network on twitter
private static void LearnImage()
{
//Bitmap img = new Bitmap(@"C:\SK\Programming\VSProjects\Stuff\Files\icon_smile2.bmp");
//Bitmap img = new Bitmap(@"C:\SK\Programming\VSProjects\Stuff\Files\twitter.png");
Bitmap img = new Bitmap(@"C:\SK\Programming\VSProjects\Stuff\Files\chrome.png");
//var ntwk = new Network(2, 1024*32, 3);
//var ntwk = new Network(2, 50, 30, 15, 6, 3);
//var ntwk = new Network(2, 1024, 1024, 1024, 3);
var ntwk = new Network(2, 200, 100, 50, 25, 15, 10, 3);
//Load existing network
//if (File.Exists(ntwkPath))
//{
// using (Stream stream = File.Open(ntwkPath, FileMode.Open))
// {
// var binaryFormatter = new System.Runtime.Serialization.Formatters.Binary.BinaryFormatter();
// ntwk = Serialization.DeSerialize((float[])binaryFormatter.Deserialize(stream));
// }
//}
//var cuda = new CudaGo(Enumerable.Range(0, ntwk.Layers.Length - 1).Max(i => ntwk.Layers[i].Length * ntwk.Layers[i + 1].Length));
//cuda.PushToGpu(ntwk, img);
var xx = EnumerateImgCoords(img).ToArray();
var yy = EnumerateImgColors(img).ToArray();
var time = new Stopwatch();
time.Start();
int iter = 0;
while(++iter > 0)
{
ntwk.Train(xx, yy, iter);
//cuda.Learn(alpha);
Console.SetCursorPosition(0, Console.CursorTop);
Console.Write("It " + iter + ", " + (time.ElapsedMilliseconds / iter) + " mill/it. Alpha=" + Hyperparams.Alpha(iter));
if (iter % 100 == 0) //Test
{
//ntwk = cuda.PeekFromGpu();
//Guess
var resImg = new Bitmap(img.Width, img.Height);
var avgError = 0.0f;
for (int w = 0; w < img.Width; ++w)
{
for (int h = 0; h < img.Height; ++h)
{
var v = ntwk.Test(new float[] {
((float)w) / img.Width,
((float)h) / img.Height });
var r = Math.Max(0, Math.Min(255, (int)(v[0] * 255.0)));
var g = Math.Max(0, Math.Min(255, (int)(v[1] * 255.0)));
var b = Math.Max(0, Math.Min(255, (int)(v[2] * 255.0)));
var c = Color.FromArgb(r, g, b);
var p = img.GetPixel(w, h);
avgError += Math.Abs(p.R - r);
avgError += Math.Abs(p.G - g);
avgError += Math.Abs(p.B - b);
resImg.SetPixel(w, h, c);
}
}
avgError /= img.Width * img.Height * 3;
resImg.Save(savePath(iter) + ".bmp");
Console.SetCursorPosition(0, Console.CursorTop);
Console.WriteLine("It " + iter + ", " + (time.ElapsedMilliseconds / iter) + " mill/it. Alpha=" + Hyperparams.Alpha(iter) + ", err=" + avgError);
//int name = 0;
//foreach (var layer in ntwk.Layers.Skip(1).Take(ntwk.Layers.Count - 2))
//{
// foreach (var neur in layer)
// {
// for (int w = 0; w < img.Width; ++w)
// {
// for (int h = 0; h < img.Height; ++h)
// {
// var v = ntwk.Test(new float[] { ((float)w) / img.Width, ((float)h) / img.Height });
// var a = neur.Activation;
// var c = Color.FromArgb(
// Math.Max(0, Math.Min(255, (int)(a * 255.0))),
// Math.Max(0, Math.Min(255, (int)(a * 255.0))),
// Math.Max(0, Math.Min(255, (int)(a * 255.0))));
// resImg.SetPixel(w, h, c);
// }
// }
// resImg.Save(savePath(iter) + name++ + ".bmp");
// }
//}
//Save network
//{
// using (Stream stream = File.Open(ntwkPath, FileMode.Create))
// {
// var binaryFormatter = new System.Runtime.Serialization.Formatters.Binary.BinaryFormatter();
// binaryFormatter.Serialize(stream, Serialization.Serialize(ntwk));
// }
//}
}
}
//Console.ReadKey();
}
//Neuron Network on numbers
private static void LearnMNIST()
{
Console.WriteLine("Loading files");
var trainX = Matrix.Build.DenseOfRows(MNIST.LoadImages(@"C:\SK\Programming\VSProjects\Stuff\Files\data\common\train-images-idx3-ubyte")).Map(x => x / 255.0);
var trainY = Vector.Build.Dense(MNIST.LoadLabels(@"C:\SK\Programming\VSProjects\Stuff\Files\data\common\train-labels-idx1-ubyte"));
var testX = Matrix.Build.DenseOfRows(MNIST.LoadImages(@"C:\SK\Programming\VSProjects\Stuff\Files\data\common\t10k-images-idx3-ubyte")).Map(x => x / 255.0);
var testY = Vector.Build.Dense(MNIST.LoadLabels(@"C:\SK\Programming\VSProjects\Stuff\Files\data\common\t10k-labels-idx1-ubyte"));
Console.WriteLine("Learning");
var ntwk = new Network(trainX.ColumnCount, 1024, 10);
//Load existing network
//if (File.Exists(ntwkPath))
//{
// using (Stream stream = File.Open(ntwkPath, FileMode.Open))
// {
// var binaryFormatter = new System.Runtime.Serialization.Formatters.Binary.BinaryFormatter();
// ntwk = Serialization.DeSerialize((float[])binaryFormatter.Deserialize(stream));
// }
//}
var xx = trainX.EnumerateRows().Select(x => x.Select(y => (float)y).ToArray())/*.Take(1000)*/.ToArray();
var yy = trainY.Select(y => new[] { (float)y })/*.Take(1000)*/.ToArray();
var testXx = testX.EnumerateRows().Select(r => r.Select(t => (float)t).ToArray()).ToArray();
var time = new Stopwatch();
time.Start();
int iter = 0;
while (++iter > 0)
{
ntwk.TrainSoftMax(xx, yy, iter);
//Save network
//{
// using (Stream stream = File.Open(ntwkPath, FileMode.Create))
// {
// var binaryFormatter = new System.Runtime.Serialization.Formatters.Binary.BinaryFormatter();
// binaryFormatter.Serialize(stream, Serialization.Serialize(ntwk));
// }
//}
//Console.SetCursorPosition(0, Console.CursorTop);
if (iter % 1 == 0)
{
Console.WriteLine("It " + iter + ", " + (time.ElapsedMilliseconds / iter) + " mill/it. Alpha=" + Hyperparams.Alpha(iter));
int howManyToTest = 1000/*testY.Count*/;
var resImg = new Bitmap(28, 28);
int maxShow = 10;
int rightCount = 0;
for (int testI = 0; testI < howManyToTest; ++testI)
{
var x = testXx[testI];
var v = ntwk.Test(x);
var guess = Vector<float>.Build.DenseOfArray(v).MaximumIndex();
var right = testY[testI];
if (guess == right)
{
++rightCount;
//resImg.Save(savePath(iter) + (testI) + "RIGHT" + guess + ".bmp");
}
else if(maxShow-- > 0)
{
for (int h = 0; h < 28; ++h)
{
for (int w = 0; w < 28; ++w)
{
var c = Color.FromArgb(
(int)(x[h * 28 + w] * 255.0),
(int)(x[h * 28 + w] * 255.0),
(int)(x[h * 28 + w] * 255.0));
resImg.SetPixel(w, h, c);
}
}
resImg.Save(savePath(iter) + (testI) + "WRONG" + guess + ".bmp");
}
}
Console.WriteLine("Hit ratio " + rightCount + "/" + howManyToTest + " (" + ((float)rightCount / howManyToTest) + ")");
}
}
}