private IDictionary <string, float[]> GetPredictions(ILearner <TSample> learner)
{
var predictions = new ConcurrentDictionary <string, float[]>();
var sampleCount = _positiveSamples.Length + _negativeSamples.Length;
var samplesProcessed = 0;
foreach (var config in learner.AllPossibleConfigurations())
{
if (config.IndexOfAny(new[] { '>', '?', ':', '\r', '\n' }) > 0)
{
throw new FormatException("Learner configuration cannot contain any of '>', '?', ':', or a line break.");
}
predictions.TryAdd(config, new float[sampleCount]);
}
Parallel.ForEach(_positiveSamples.Concat(_negativeSamples), s =>
{
var localLearner = learner;
localLearner.SetSample(s.Sample);
foreach (var config in learner.AllPossibleConfigurations())
{
localLearner = localLearner.WithConfiguration(config);
predictions[config][s.Index] = learner.Classify();
}
if (_cancellation.IsCancellationRequested)
{
return;
}
if (_progress != null)
{
if (Monitor.TryEnter(_progress))
{
try
{
samplesProcessed++;
_progress.Report(Tuple.Create("Running learner \"" + learner.UniqueId + "\"", samplesProcessed * 100 / sampleCount));
}
finally
{
Monitor.Exit(_progress);
}
}
}
});
return(predictions);
}
private LayerHolder OptimizeCoefficients(ILearner <TSample> learner, float[] outputs)
{
var positiveWeights = _positiveSamples.Select(s => s.Weight).ToArray();
var negativeWeights = _negativeSamples.Select(s => s.Weight).ToArray();
var positiveCorrect = new bool[_positiveSamples.Length];
var negativeCorrect = new bool[_negativeSamples.Length];
#if DEBUG
var perfect = true;
foreach (var s in _positiveSamples.Concat(_negativeSamples))
{
learner.SetSample(s.Sample);
// ReSharper disable CompareOfFloatsByEqualityOperator
if (learner.Classify() != outputs[s.Index])
{
throw new Exception();
}
perfect &= s.Actual == outputs[s.Index];
// ReSharper restore CompareOfFloatsByEqualityOperator
}
#endif
//Find best bias point, adding one sample at a time to the predicted target set and updating the cost.
//Note that -cost is the cost of the inverse of the current hypothesis i.e. "all samples are in the target".
var bestThres = float.PositiveInfinity;
var bestCost = float.PositiveInfinity;
var bestPos = float.PositiveInfinity;
var bestNeg = float.NegativeInfinity;
var threshold = float.NegativeInfinity;
while (threshold < float.PositiveInfinity)
{
{
var i = 0;
foreach (var s in _positiveSamples)
{
positiveCorrect[i++] = outputs[s.Index] > threshold;
}
i = 0;
foreach (var s in _negativeSamples)
{
negativeCorrect[i++] = outputs[s.Index] <= threshold;
}
}
double sumWeightsPosCorrect = 0;
double sumWeightsPosIncorrect = 0;
double sumWeightsNegCorrect = 0;
double sumWeightsNegIncorrect = 0;
SumWeights(positiveWeights, positiveCorrect, ref sumWeightsPosCorrect, ref sumWeightsPosIncorrect);
SumWeights(negativeWeights, negativeCorrect, ref sumWeightsNegCorrect, ref sumWeightsNegIncorrect);
var coefPos = (float)Math.Log(sumWeightsPosCorrect / sumWeightsNegIncorrect) / 2;
var coefNeg = (float)Math.Log(sumWeightsNegCorrect / sumWeightsPosIncorrect) / -2;
coefPos = float.IsNaN(coefPos) ? 0 : coefPos;
coefNeg = float.IsNaN(coefNeg) ? 0 : coefNeg;
var loss = (float)(sumWeightsPosCorrect * Math.Exp(-coefPos) + sumWeightsNegIncorrect * Math.Exp(coefPos));
loss += (float)(sumWeightsNegCorrect * Math.Exp(coefNeg) + sumWeightsPosIncorrect * Math.Exp(-coefNeg));
#if DEBUG
if (!ApproxEqual((float)(sumWeightsNegCorrect + sumWeightsNegIncorrect + sumWeightsPosCorrect + sumWeightsPosIncorrect), 1))
{
// ReSharper disable once EmptyStatement
;
}
if (!perfect && (float.IsInfinity(coefNeg) || float.IsInfinity(coefPos)))
{
throw new Exception("Unexpected coefficients");
}
//float calcLoss = 0;
//calcLoss += getLoss(coefPos, coefNeg, positiveSamples.Select(s => s.confidence).Zip(positiveCorrect, (s, c) => new Tuple<float, bool>(s, c)), 1f);
//calcLoss += getLoss(coefNeg, coefPos, negativeSamples.Select(s => s.confidence).Zip(negativeCorrect, (s, c) => new Tuple<float, bool>(s, c)), -1f);
#endif
if (!float.IsNaN(loss) && loss < bestCost)
{
bestCost = loss;
bestPos = coefPos;
bestNeg = coefNeg;
bestThres = threshold;
}
var next = outputs.Aggregate(float.PositiveInfinity, (min, i) => (i > threshold && i < min) ? i : min);
threshold = next;
}
return(new LayerHolder(bestCost, new Layer <TSample>(learner, bestPos, bestNeg, bestThres), outputs));
}
internal float Classify(TSample s)
{
_learner.SetSample(s);
return(_learner.Classify() > Threshold ? CoefPos : CoefNeg);
}
