static void TrainMelModel(Set sourceFiles,
Matrix Model, Action <double> SetLoss, Func <bool> HasCtrlBreak)
{
if (Model == null)
{
Console.WriteLine("Model not loaded.");
return;
}
Thread[] threads = new Thread[Environment.ProcessorCount * 2];
int numberOfThreads = 0,
verbOut = 0;
for (var t = 0; t < threads.Length; t++)
{
threads[t] = new Thread(() => {
string[] Shuffle = ((IEnumerable <string>)sourceFiles).ToArray();
Random.Shuffle(
Shuffle,
Shuffle.Length);
Interlocked.Increment(ref numberOfThreads);
try {
for (int iter = 0; iter < GENS; iter++)
{
if (HasCtrlBreak != null && HasCtrlBreak())
{
break;
}
foreach (string file in Shuffle)
{
if (HasCtrlBreak != null && HasCtrlBreak())
{
return;
}
try {
Matrix Data = null;
Data = System.Ai.Model.LoadFromFile(file, SIZE, out string fmt, out CBOW.DIMS);
Debug.Assert(fmt == "MIDI");
double loss = 0,
cc = 0;
var wo = Model["a"];
if (wo == null)
{
Interlocked.Increment(ref verbOut);
continue;
}
foreach (var it in Data)
{
double score = 0,
err = 0;
double[] X
= new double[it.Axis.Length];
for (int j = 0; j < X.Length; j++)
{
X[j] = it.Axis[j].Re;
}
bool label = it.Id == wo.Id;
if (!label)
{
for (int j = 0; j < X.Length; j++)
{
X[j] = ((Random.Next() & 0xFFFF) / (65536f));
if (X[j] < 0 || X[j] > 0.7)
{
X[j] = 0;
}
}
if (((Random.Next() & 0xFFFF) / (65536f)) > 0.7)
{
for (int j = 0; j < X.Length; j++)
{
X[j] = 0;
}
}
}
score = CBOW.binaryLogistic(
X,
null,
wo.Axis,
label ? 1.0 : 0.0,
CBOW.lr);
if (label)
{
err = -System.Math.Log(score);
}
else
{
err = -System.Math.Log(1.0 - score);
}
if (double.IsNaN(err) || double.IsInfinity(err))
{
Console.WriteLine("NaN detected...");
}
else
{
loss += err;
cc++;
if (err == 0 || (verbOut >= 0 && ((verbOut % CBOW.VERBOSITY) == 0)))
{
Console.Write($"[{Thread.CurrentThread.ManagedThreadId}.{verbOut}] " +
$" Cost: {loss / cc}\r\n");
}
Interlocked.Increment(ref verbOut);
}
}
Thread.Sleep(3000 + Random.Next(3000));
} finally {
}
}
}
} finally {
Interlocked.Decrement(ref numberOfThreads);
}
Console.Write($"[{Thread.CurrentThread.ManagedThreadId}] stopped...\r\n");
});
}
foreach (var t in threads)
{
t.Start();
}
foreach (var t in threads)
{
t.Join();
}
Debug.Assert(numberOfThreads == 0);
}
public static Vector[] RunFullCosineSort(IOrthography lex, Matrix Model, string Q, int max)
{
if (Model == null || string.IsNullOrWhiteSpace(Q))
{
Console.ForegroundColor = ConsoleColor.Yellow;
Console.WriteLine("Model not loaded.\r\n");
Console.ResetColor();
Console.WriteLine("See '--load' command for more info...\r\n");
return(null);
}
double[] Re = new double[CBOW.DIMS];
double norm = 0;
var sign = +1;
foreach (var tok in PlainText.ForEach(Q, 0, Q.Length, 0))
{
string wi = lex.GetKey(tok.TextFragment.Substring(tok.StartIndex, tok.Length));
if (wi == "+")
{
sign = +1;
}
else if (wi == "-")
{
sign = -1;
}
else
{
var vec = Model[wi];
if (vec != null)
{
Debug.Assert(vec.Axis.Length == Re.Length);
for (var j = 0; j < Re.Length; j++)
{
Re[j] += sign * vec.Axis[j].Re;
}
norm++;
}
else
{
Console.ForegroundColor = ConsoleColor.Yellow;
Console.WriteLine($"'{wi}' not found.");
Console.ResetColor();
}
}
}
if (norm > 0)
{
for (var j = 0; j < Re.Length; j++)
{
Re[j] /= (double)norm;
}
}
Vector[] output = CBOW.Predict(Model, Re, max);
Array.Sort(output,
(a, b) => Scalar.CompareTo(a, b));
Console.WriteLine();
Console.WriteLine(" [" + string.Join(",", Re.Select(re => Math.Round(re, 4)).Take(7)) + "...]");
Console.WriteLine();
int len = 0;
for (int i = output.Length - 1; i >= 0; i--)
{
Vector n = output[i];
if (n != null)
{
string str = n.Id;
var it = Model[n.Id];
if (it != null)
{
// if (it.Count > 0) {
// var best = it.ArgMax();
// if (best != null) {
// str = best.Id;
// }
// }
}
if (len + str.Length > 37 /* break like if does not fit */)
{
Console.WriteLine(
output.Length <= 31
? $" {str} : {n.Score}"
: $" {str}");
len = 0;
}
else
{
Console.Write(
output.Length <= 31
? $" {str} : {n.Score}"
: $" {str}");
len += str.Length;
}
}
}
Console.WriteLine();
return(output);
}