/// <summary>
/// Trains the model to recognise the specified input as the specified symbol
/// </summary>
public void Train(WrappedBitmap inputData, TSymbol representedValue, uint rounds)
{
var initialGeneration = Enumerable.Range(0, (int)rounds)
.Select(i => new Scorer(new TuningParams.Scoring())) // TODO nested constructors is smelly
.Select(s => new ScorerScorePair(s, s.Score(inputData)))
.AsParallel()
.OrderByDescending(s => s.Score);
// Having acquired the initial set of sampling pixels via sheer randomness, we'll work with
// the best we've got and try and massage them into something better.
var mutParams = new TuningParams.Mutation();
var initialBestCandidates = initialGeneration.Take(100); // TODO parameterise magic value (which was decided totally arbitrarily anyway)
var initialBestCandidatesWithMutatedVariants = from p in initialBestCandidates
let mutatedOffspring = p.Scorer.MutateMany(mutParams)
let mutatedScores = from m in mutatedOffspring
let mutatedScore = m.Score(inputData)
select new ScorerScorePair(m, mutatedScore)
select new ParentChildScoreCollection(p, mutatedScores);
var bestOftheMutantStrains = initialBestCandidatesWithMutatedVariants
.SelectMany(x => x.Children)
.OrderByDescending(s => s.Score)
.First();
// Compute some stats for these mutations. These stats are for interest only, they don't affect any computation.
Benchmarking.Mutation.AddDataPoints(initialBestCandidatesWithMutatedVariants);
var rec = new Recogniser <TSymbol>(representedValue, bestOftheMutantStrains.Scorer);
RegisterRecogniser(rec);
}
/// <summary>
/// Randomises the sampled pixels, but keeps them relatively close to their existing locations.
/// </summary>
public IScorer Mutate(TuningParams.Mutation tuningParams)
{
var newPixels = new List <Point>(_pixelsToSample);
int n = _randomNumGenerator.Next(tuningParams.PixelRandomChurn);
for (int i = 0; i < n; i++)
{
// Remove one of the pixels at random
newPixels.RemoveAt(_randomNumGenerator.Next(TuningParams.ImageSize));
}
// Skew some of the remaining ones
n = _randomNumGenerator.Next(tuningParams.PixelRandomSkewNumber);
for (int i = 0; i < n; i++)
{
// TODO the same pixels could be skewed more than once (does that even matter?)
int xOffset = _randomNumGenerator.Next(tuningParams.PixelRandomSkewOffset) - tuningParams.PixelRandomSkewOffset;
int yOffset = _randomNumGenerator.Next(tuningParams.PixelRandomSkewOffset) - tuningParams.PixelRandomSkewOffset;
int pixIndex = _randomNumGenerator.Next(newPixels.Count - 1);
newPixels[pixIndex] = ConstrainPixelToImage(
newPixels[pixIndex].X + xOffset,
newPixels[pixIndex].Y + yOffset);
}
// Add some more random new ones (approx the same number to replace those that were removed or "churned")
n = _randomNumGenerator.Next(tuningParams.PixelRandomChurn);
for (int i = 0; i < n; i++)
{
// NOTE - it's possible to get duplicates, in which case the same pixel is counted twice
// TODO check for off-by-ones (can we hit the image edges?)
newPixels.Add(
new Point(
_randomNumGenerator.Next(TuningParams.ImageSize),
_randomNumGenerator.Next(TuningParams.ImageSize)
)
);
}
// modify the negative/positive offset randomly
//_positiveScoreOffset = (byte)(_positiveScoreOffset + _randomNumGenerator.Next(tuningParams.PositiveRandOffset) - TuningParams.PositiveRandOffset / 2);
//_negativeScoreOffset = (byte)(_negativeScoreOffset + _randomNumGenerator.Next(TuningParams.Scoring.NegativeRandOffset) - TuningParams.NegativeRandOffset / 2);
// TODO keeping these params the same as our parent for now, and only randomoising on the sampled pixels. the randomisation of varying parameters could go on forever!!!
return(new Scorer(_tuningParams, newPixels));
}
public IEnumerable <IScorer> MutateMany(TuningParams.Mutation tuningParams)
{
return(Enumerable.Range(0, tuningParams.SpawnedDescendants)
.Select(i => Mutate(tuningParams)));
}