// Might not have to load statically, but it's easy for now.
public static bool LoadNet(ref NeuralNet bestNet, string path)
{
if (File.Exists(path))
{
/* 0 - float - Best score
* 1 - int - Amount of layers in net
* 2 - int - Amount of neurons in layer
* 3 - int - amount of floats in set for neuron
* 4/5 - float - Weight/slope data
* f i i i fffffff i ffffffff etc
* 0 1 2 3 45454545 3 45454545 3 454545 2 3 45454545 etc
*/
float best = 0;
using (BinaryReader reader = new BinaryReader(File.Open(path, FileMode.Open)))
{
best = reader.ReadSingle();
int layers = reader.ReadInt32();
bestNet = new NeuralNet(layers);
bestNet.bestScore = best;
for (int i = layers - 1; i >= 0; i--)
{
int neurons = reader.ReadInt32();
//bestNet.Neurons = neurons; // Temp
if (i == layers - 1) // Output layer
{
bestNet.array[i] = new NeuronList(null, neurons);
}
else
{
bestNet.array[i] = new NeuronList(bestNet.array[i + 1], neurons);
}
float[] weights;
float[] slopes;
for (int n = 0; n < neurons; n++)
{
int size = reader.ReadInt32();
weights = new float[size / 2];
slopes = new float[size / 2];
for (int d = 0; d < size; d += 2)
{
weights[d / 2] = reader.ReadSingle();
slopes[d / 2] = reader.ReadSingle();
}
bestNet.array[i].array[n] = new Neuron(weights, slopes);
}
}
}
return(true);
}
return(false);
}
public NeuralNet Copy()
{
NeuronList connections = null;
NeuralNet ret = new NeuralNet(array.Length);
//ret.bestScore = bestScore;
ret.count = count;
//ret.Neurons = Neurons;
for (int i = array.Length - 1; i >= 0; i--)
{
ret.array[i] = connections = array[i].Copy(connections);
}
return(ret);
}
public void LearnSessionNoOutput(float[][] lessonplan, int bestof)
{
bool Kill = false;
//int practice = 100; // Doesn't matter what it is, usually doesn't go above 20 anyway.
NeuralNet[] neuralNets = new NeuralNet[TASKS];
float[] curScore = new float[TASKS];
Console.WriteLine("[Ctrl+D] to stop learning.");
Task Listen = Task.Factory.StartNew(() =>
{
ConsoleKeyInfo key;
while (!Kill)
{
if (Console.KeyAvailable)
{
key = Console.ReadKey(true);
if (key.Key == ConsoleKey.D && key.Modifiers == ConsoleModifiers.Control)
{
Kill = true;
}
}
}
});
for (int h = 0; h < bestof; h++)
{
int curBest = -1;
int currentLeft = Console.CursorLeft;
int currentTop = Console.CursorTop;
//int tryNew = new Random().Next(24);
NeuralNet randNet = null;
if (h > 0 && h % 10 == 0)
{
Random rand = new Random(DateTime.Now.Millisecond);
randNet = new NeuralNet(rand.Next(Math.Max(3, array.Length - 2), array.Length + 3), 6, rand.Next(Math.Max(1, array[1].array.Length - 2), array[1].array.Length + 3), 1, rand);
}
for (int ts = 0; ts < taskArray.Length; ts++)
{
int t = ts;
taskArray[ts] = Task.Factory.StartNew(() =>
{
Random rand = new Random(DateTime.Now.Millisecond + ((int)Task.CurrentId * 478));
float lcurScore = 0, llcurScore = -1, lllcurScore = -2, llllcurScore = -3;
if (randNet != null)
{
neuralNets[t] = randNet.MutateCopy(rand);//new NeuralNet(rand.Next(Math.Max(3, bestNet.array.Length - 2), bestNet.array.Length + 3), 6, rand.Next(Math.Max(1, bestNet.array[1].array.Length - 2), bestNet.array[1].array.Length + 3), 1, rand);
}
else
{
neuralNets[t] = MutateCopy(rand);
}
for (int i = 0; i < 33; i++)
{
float[] res = new float[lessonplan.Length / 2];
for (int j = 0; j < lessonplan.Length; j += 2)
{
res[j / 2] = neuralNets[t].Learn(lessonplan[j], lessonplan[j + 1]);
}
if (Kill)
{
curScore = new float[TASKS];
break;
}
learnLen = res.Length;
curScore[t] = 0;
for (int j = 0; j < res.Length; j++)
{
float accu = res[j];
if (accu < ACC_THRESHOLD - 0.1f)
{
curScore[t] += 1f;
}
curScore[t] += 1f - accu;
}
if ((curScore[t] == llcurScore && curScore[t] == lllcurScore && curScore[t] == llllcurScore && curScore[t] == lcurScore) || i == 32)
{ // Probably a better way to do this. Checks if we're not getting anywhere
break;
}
if (curScore[t] < bestScore / 2.035) //(res.Length / 7.0f))
// "res.length/?" lowered will allow more offshoots to potentially grow. Slower but maybe better results? Probably not.
{
break;
}
llllcurScore = llcurScore;
lllcurScore = lcurScore;
llcurScore = lcurScore;
lcurScore = curScore[t];
}
});
}
Task.WaitAll(taskArray);
for (int i = 0; i < curScore.Length; i++)
{
if (curScore[i] > bestScore)
{
bestScore = curScore[i];
curBest = i;
}
}
if (curBest >= 0)
{
CopyFrom(neuralNets[curBest]);
}
if (bestScore >= (learnLen * 2.0f) - 0.1f || Kill)
{
break;
}
}
Kill = true;
Console.WriteLine("Done Test");
}
public void LearnSession(float[][] lessonplan, int bestof)
{
bool Kill = false;
//int practice = 100; // Doesn't matter what it is, usually doesn't go above 20 anyway.
Console.WriteLine();
NeuralNet[] neuralNets = new NeuralNet[TASKS];
float[] curScore = new float[TASKS];
Console.WriteLine("[Ctrl+D] to stop learning.\n\nTIME\t\tNUM\tBEST/{0}\t\tCAP\t%ACC", TASKS);
Task Listen = Task.Factory.StartNew(() =>
{
ConsoleKeyInfo key;
while (!Kill)
{
if (Console.KeyAvailable)
{
key = Console.ReadKey(true);
if (key.Key == ConsoleKey.D && key.Modifiers == ConsoleModifiers.Control)
{
Kill = true;
}
}
}
});
for (int h = 0; h < bestof; h++)
{
int curBest = -1;
int currentLeft = Console.CursorLeft;
int currentTop = Console.CursorTop;
//int tryNew = new Random().Next(24);
NeuralNet randNet = null;
if (h > 0 && h % 10 == 0)
{
Random rand = new Random(DateTime.Now.Millisecond);
randNet = new NeuralNet(rand.Next(Math.Max(3, array.Length - 2), array.Length + 3), 6, rand.Next(Math.Max(1, array[1].array.Length - 2), array[1].array.Length + 3), 1, rand);
}
for (int ts = 0; ts < taskArray.Length; ts++)
{
int t = ts;
taskArray[ts] = Task.Factory.StartNew(() =>
{
Random rand = new Random(DateTime.Now.Millisecond + ((int)Task.CurrentId * 478));
float lcurScore = 0, llcurScore = -1, lllcurScore = -2, llllcurScore = -3;
if (randNet != null)
{
neuralNets[t] = randNet.MutateCopy(rand);//new NeuralNet(rand.Next(Math.Max(3, bestNet.array.Length - 2), bestNet.array.Length + 3), 6, rand.Next(Math.Max(1, bestNet.array[1].array.Length - 2), bestNet.array[1].array.Length + 3), 1, rand);
}
else
{
neuralNets[t] = MutateCopy(rand);
}
lock (_lock)
{
Console.SetCursorPosition(currentLeft + 94 + TASKS, currentTop);
Console.Write(neuralNets[t].GetID().PadRight(16));
Console.SetCursorPosition(currentLeft, currentTop);
}
for (int i = 0; i < 33; i++)
{
if (t == 0)
{
lock (_lock)
{
Console.SetCursorPosition(currentLeft + 60, currentTop);
Console.Write(" " + i.ToString().PadRight(2));
Console.SetCursorPosition(currentLeft, currentTop);
}
}
float[] res = new float[lessonplan.Length / 2];
for (int j = 0; j < lessonplan.Length; j += 2)
{
res[j / 2] = neuralNets[t].Learn(lessonplan[j], lessonplan[j + 1]);
if (t == 0)
{
lock (_lock)
{
int conloc = (int)((float)j / lessonplan.Length * 30);
Console.SetCursorPosition(currentLeft + 60 + conloc, currentTop);
Console.Write(conloc % 2 == 0 ? '-' : '=');
Console.SetCursorPosition(currentLeft, currentTop);
}
}
}
if (Kill)
{
curScore = new float[TASKS];
break;
}
learnLen = res.Length;
curScore[t] = 0;
for (int j = 0; j < res.Length; j++)
{
float accu = res[j];
if (accu < ACC_THRESHOLD - 0.1f)
{
curScore[t] += 1f;
}
curScore[t] += 1f - accu;
}
if ((curScore[t] == llcurScore && curScore[t] == lllcurScore && curScore[t] == llllcurScore && curScore[t] == lcurScore) || i == 32)
{ // Probably a better way to do this. Checks if we're not getting anywhere
break;
}
if (curScore[t] < bestScore / 2.035) //(res.Length / 7.0f))
// "res.length/?" lowered will allow more offshoots to potentially grow. Slower but maybe better results? Probably not.
{
break;
}
llllcurScore = llcurScore;
lllcurScore = lcurScore;
llcurScore = lcurScore;
lcurScore = curScore[t];
}
lock (_lock)
{
Console.SetCursorPosition(currentLeft + 93 + t, currentTop);
Console.Write(curScore[t] > bestScore ? 'o' : '.');
Console.SetCursorPosition(currentLeft, currentTop);
}
});
}
lock (_lock)
{
Console.SetCursorPosition(currentLeft, currentTop);
Console.WriteLine("[" + DateTime.Now.ToLongTimeString() + "]\t" + h.ToString("d4") + "\t" + "WORKING" + "\t\t" + learnLen * 2 + "\t" + "WORKING".PadRight(78));
Console.SetCursorPosition(currentLeft, currentTop);
}
Task.WaitAll(taskArray);
for (int i = 0; i < curScore.Length; i++)
{
if (curScore[i] > bestScore)
{
bestScore = curScore[i];
curBest = i;
}
}
if (curBest >= 0)
{
CopyFrom(neuralNets[curBest]);
Console.WriteLine("[" + DateTime.Now.ToLongTimeString() + "]\t" + h.ToString("d4") + "\t" + bestScore.ToString("f5").PadRight(8) + "\t" + learnLen * 2 + "\t" + ((bestScore / (learnLen * 2)) * 100).ToString("f2") + "\t");
}
if (bestScore >= (learnLen * 2.0f) - 0.1f || Kill)
{
break;
}
}
Kill = true;
Console.WriteLine("[" + DateTime.Now.ToLongTimeString() + "]\tDone Test\t\t\t\t\t\t\t\n");
}
public void CopyFrom(NeuralNet net)
{
array = (NeuronList[])net.array.Clone();
count = net.count;
}