private void MainForm_Load(object sender, EventArgs e)
{
loading = true;
Brain = new Brain(28 * 28, 10, 3, 16, -0.5F, 0.5F, false, Neuron.ReLU);
Brain = new Brain(new FileStream(@"network.brainStream", FileMode.Open), Neuron.ReLU);//new Brain(File.ReadAllText(@"network.brain"),28 * 28, 10, 3, 16);
Thread loadThread = new Thread(new ThreadStart(loadData));
loadThread.Start();
trainingActive = true;
void loadData()
{
if (trainingActive)
{
trainingAmount = 60000;
Images = ParseDatabase.ParseImages(@"train-images.idx3-ubyte", trainingAmount);
Labels = ParseDatabase.ParseLabels(@"train-labels.idx1-ubyte", trainingAmount);
}
else
{
trainingAmount = 10000;
Images = ParseDatabase.ParseImages(@"t10k-images.idx3-ubyte", trainingAmount);
Labels = ParseDatabase.ParseLabels(@"t10k-labels.idx1-ubyte", trainingAmount);
}
loading = false;
}
}
private void MainForm_Paint(object sender, PaintEventArgs e)
{
if (!loading)
{
e.Graphics.InterpolationMode = System.Drawing.Drawing2D.InterpolationMode.NearestNeighbor;
Bitmap bmp = ParseDatabase.ByteToBitmap(Images[count], ImageSize, ImageSize);
Rectangle rec = new Rectangle(10, 10, 200, 200);
e.Graphics.DrawImage(bmp, rec);
bmp.Dispose();
}
}
private void MainForm_Load(object sender, EventArgs e)
{
loading = true;
Brain = new Brain(28 * 28, 10, 2, 16);
Thread loadThread = new Thread(new ThreadStart(loadData));
loadThread.Start();
void loadData()
{
Images = ParseDatabase.ParseImages(@"train-images-idx3-ubyte");
Labels = ParseDatabase.ParseLabels(@"train-labels-idx1-ubyte");
loading = false;
}
}
private void train_button_Click(object sender, EventArgs e)
{
count++;
if (trainThread == null)
{
trainThread = new Thread(new ThreadStart(training));
trainThread.Start();
}
float[] output = Brain.Think(ParseDatabase.ByteToFloat(Images[count]));
outputNum = Training.OutputNumber(output);
if (Labels[count] == outputNum)
{
good++;
}
float accuracy = good / (count + 1);
Console.Write(" accuracy: " + accuracy + "\r");
outputNum_label.Text = outputNum + " accuracy: " + accuracy;
show_output(output);
this.Refresh();
}
private void training()
{
float tweakAmount = 0.0001F;
count = 0;
bool Active = true;
bool adjustTweak = true;
while (Active)
{
float[] output = Brain.Think(ParseDatabase.ByteToFloat(Images[count]));
outputNum = Training.OutputNumber(output);
float cost = Training.CalculateCost(output, Labels[count]);
Training.Backpropagate(Brain, tweakAmount, outputNum, Labels[count]);
int expected = Labels[count];
// Normalerweise trennt man die Prüfdaten von den Trainingsdaten,
// um zu gucken ob das Netz nicht nur auswendig lernt, sondern auch generalisiert
if (count % 1000 == 0)
{
good = 0; //reset statistic every 1000 steps
}
if (expected == outputNum)
{
good++;
}
double accuracy = good / (count % 1000 + 1.0);
int percent = (int)(100 * accuracy);
if (count % 100 == 90)
{
Console.Write("accuracy: " + percent + "% cost: " + cost + " \r");
if (percent > 50 && adjustTweak)
{
tweakAmount = tweakAmount / 10;
adjustTweak = false;// only once
}
}
//this.Invoke(new MethodInvoker(delegate {
// this.Refresh();
// outputNum_label.Text =outputNum + " accuracy: " + percent+ "%";
// //show_output(output);
//}));
count++;
if (count >= 59999)
{
count = 0;
string path = @"save.brain";
File.Create(path).Close();
File.WriteAllText(path, Brain.BrainStructure);
DialogResult res = MessageBox.Show("Finished", "Training", MessageBoxButtons.OKCancel);
if (res.HasFlag(DialogResult.Cancel))
{
Active = false;
}
}
}
}
private void training()
{
float tweakAmount = 0.0005F;
count = 0;
bool Active = true;
while (Active)
{
float[] output = Brain.Think(ParseDatabase.ByteToFloat(Images[count]), Neuron.ReLU);
outputNum = Training.OutputNumber(output);
float cost = Training.CalculateCost(output, Labels[count]);
Training.Backpropagate(Brain, tweakAmount, outputNum, Labels[count]);
int expected = Labels[count];
// Normalerweise trennt man die Prüfdaten von den Trainingsdaten,
// um zu gucken ob das Netz nicht nur auswendig lernt, sondern auch generalisiert
if (count % 1000 == 0)
{
good = 0; //reset statistic every 1000 steps
}
if (expected == outputNum)
{
totalSuccess++;
good++;
}
double accuracy = good / (count % 1000 + 1.0);
int percent = (int)(100 * accuracy);
int totalAccuracy = Convert.ToInt32(totalSuccess * 100 / (count + 1));
Console.Write("accuracy: " + percent + "% cost: " + cost + "total:" + totalAccuracy + "% \r");
if (display)
{
this.Invoke(new MethodInvoker(delegate {
this.Refresh();
outputNum_label.Text = outputNum + " accuracy: " + percent + "%";
show_output(output);
}));
}
count++;
if (count >= trainingAmount - 1)
{
count = 0;
totalSuccess = 0;
string path = @"save.brain";
File.Create(path).Close();
File.WriteAllText(path, Brain.BrainStructureString);
DialogResult res = MessageBox.Show("Finished", "Training", MessageBoxButtons.OKCancel);
if (res.HasFlag(DialogResult.Cancel))
{
Active = false;
}
}
}
}
