public static void Training()
{
for (int j = 0; j < SaveState; j++)
{
List <NN> __ = new List <NN>();
var tasks = new Task[batchsize];
for (int ii = 0; ii < batchsize; ii++)
{
//B/c ii may change and this code can't let that happen
int iterator = ii;
double[,] image = Reader.ReadNextImage();
int correct = Reader.ReadNextLabel();
__.Add(new NN());
tasks[iterator] = Task.Run(() => __[iterator].Backprop(image, correct));
}
Task.WaitAll(tasks);
//Syncronously descend
foreach (NN nn in __)
{
nn.Descend();
nn.Dispose();
}
//Updating the weights with the avg gradients
NN.Descend(batchsize);
UserValidation();
}
//Save weights and biases
D.WriteWeightBias();
}
//Reset (re-initialize) weights and biases of the neural network
public static void reset()
{
NN nn = new NN();
nn.initialize();
D.WriteWeightBias();
}
public static void UserValidation()
{
NN nn = new NN(); double[,] image; int correct;
//Some user validation code
image = Reader.ReadNextImage();
correct = Reader.ReadNextLabel();
//Backprop again for averaging
nn.Calculate(image);
//Print out various things
int guess = 0; double certainty = 0;
//yeah, yeah, this SHOULD be a NN method, not public and used here but I'm tired right now
//Is a calculation of r^2 error
for (int i = 0; i < 10; i++)
{
if (nn.OutputValues[i] > certainty)
{
certainty = nn.OutputValues[i]; guess = i;
}
}
//Calculate the moving average of the percentage of trials correct of those written to console
double error = 0;
for (int i = 0; i < NN.OutputCount; i++)
{
error += ((i == correct ? 1d : 0d) - nn.OutputValues[i]) * ((i == correct ? 1d : 0d) - nn.OutputValues[i]);
}
iterator++;
avgerror = ((iterator / (iterator + 1)) * avgerror) + ((1 / iterator) * error);
avg = (avg * (iterator / (iterator + 1))) + ((guess == correct) ? (1 / iterator) : 0d);
//Some safety code which is currently disabled
//if (avgerror > maxavg && iterator > 300) { maxavg = avgerror; }
//if (avgerror > maxavg * 10 && iterator > 300) { finished = true; }
//Print various things to the console for verification that things are nominal
Console.WriteLine("Correct: " + correct + " Guess: " + guess + " Correct? " + (guess == correct ? "1 " : "0 ") + "Certainty: " + Math.Round(certainty, 5).ToString().PadRight(7)
+ " %Correct: " + Math.Round(avg, 5).ToString().PadRight(7) + " Avg error: " + Math.Round(avgerror, 5).ToString().PadRight(8) + " Avg gradient: " + NN.AvgGradient, 15);
//Dispose of the neural network (may not be necessary)
nn.Dispose();
//Reset the console data every few iterations to ensure up to date numbers
if (iterator > 1000)
{
iterator = 100; NN.Epoch++;
}
}
public static void Testing()
{
NN nn = new NN();
while (iterator < 9000)
{
iterator++;
nn.Calculate(Reader.ReadNextImage());
int correct = Reader.ReadNextLabel();
double certainty = -99d; int guess = -1;
for (int i = 0; i < 10; i++)
{
if (nn.OutputValues[i] > certainty)
{
certainty = nn.OutputValues[i]; guess = i;
}
}
avg = (avg * (iterator / (iterator + 1))) + ((guess == correct) ? (1 / iterator) : 0d);
Console.WriteLine("Correct: " + correct + " Correct? " + (guess == correct ? "1 " : "0 ") + " %Correct: " + Math.Round(avg, 10).ToString().PadRight(12) + " Certainty " + Math.Round(certainty, 10));
nn.Dispose();
}
}