public void TrainNetwork()
{
Console.WriteLine("Training network...");
Console.WriteLine();
Random rnd = new Random();
Parents potentialParents = new Parents();
while (true)
{
if (shouldPause)
{
return;
}
bool lastSuccess = false;
int currParentInd = 0;
if (currentParents.parentKernels.Count > 0)
{
//if(currentParents.parentKernels.Length == convolution.kernels.Length)
// convolution.SetKernels(currentParents.parentKernels);
currParentInd = (int)(rnd.NextDouble() * currentParents.parentKernels.Count);
convolution.SetKernels(currentParents.parentKernels[currParentInd].kernels);
}
convolution.MutateKernel(settings.mutationRate);
float fitness = TestConnections(convolution, rnd);
runningAvg.Enqueue(fitness);
potentialParents.parentKernels.Add(new Parents.Parent(fitness, convolution.DeepCopyKernels(), gens > 0?currentParents.parentKernels[currParentInd].species + " " + (iters - gens * settings.childrenPerGen):iters.ToString()));
//Genetic Selection
if (iters % settings.childrenPerGen == settings.childrenPerGen - 1)
{
currentParents = new Parents();
potentialParents.parentKernels.Sort(delegate(Parents.Parent x, Parents.Parent y)
{
return(x.fitness.CompareTo(y.fitness));
});
for (int x = 0; x < settings.parentsPerGeneration; x++)
{
//float rndval = (float)rnd.NextDouble();
//float newTarget = Map(potentialParents.parentKernels.Last().fitness, potentialParents.parentKernels.First().fitness, rndval);
//Parents.Parent newParent = potentialParents.parentKernels.First(y => y.fitness>=newTarget);
Parents.Parent newParent = potentialParents.parentKernels[potentialParents.parentKernels.Count - x - 1];
currentParents.parentKernels.Add(newParent);
//potentialParents.parentKernels.Remove(newParent);
}
potentialParents.parentKernels.Clear();
//Retest the previous best//This is tricky now, just assume all parents die after reproducing
/*if (currentParents.parentKernels.Length == convolution.kernels.Length)
* {
* convolution.SetKernels(currentParents.parentKernels);
* potentialParents.parentFitness = TestConnections(convolution, rnd);
* }*/
//TEMP
try
{
System.IO.File.WriteAllBytes("genOut.png", convolution.dstImg.ToBytes());
System.IO.File.WriteAllBytes("genIn.png", convolution.srcImgs[lastIndex].ToBytes());
//System.IO.File.WriteAllText(inpath, JsonConvert.SerializeObject(networkConnections));
}
catch (Exception e)
{
Console.WriteLine(e.Message);
}
gens++;
}
if (runningAvg.Count > settings.overrallFitIters)
{
runningAvg.Dequeue();
}
//GUI
if (iters % 2 == 0)
{
float overallFit = 0;
foreach (float i in runningAvg)
{
overallFit += i;
}
overallFit /= runningAvg.Count;
Console.Write("[" + (DateTime.Now - startTime).ToString(@"h\:mm\:ss") + "] Training... Current fitness: " + Math.Round(fitness)
+ "%; Overall fitness: " + Math.Round(overallFit)
+ "% \n[" + (DateTime.Now - startTime).ToString(@"h\:mm\:ss") + "] Best fitness (gen): " + Math.Round(potentialParents.parentKernels.FirstOrDefault().fitness)
+ "%; Iterations: " + iters
+ "; Generations: " + gens + " "
+ "\n[" + (DateTime.Now - startTime).ToString(@"h\:mm\:ss") + "] List Parents: " + " \n");
for (int i = 0; i < currentParents.parentKernels.Count; i++)
{
Console.WriteLine("Parent: " + currentParents.parentKernels[i].species.Substring(Math.Max(currentParents.parentKernels[i].species.Length - 20, 0)) + " Fitness: " + currentParents.parentKernels[i].fitness + " ");
}
Console.SetCursorPosition(0, Console.CursorTop - (3 + currentParents.parentKernels.Count));
}
if (lastSuccess)
{
successIters++;
}
iters++;
}
}
