public float Distance(SparseVectorF x, float[] y)
{
return(x.EuclideanDistance(y, 0));
}
public float Distance(SparseVectorF x, float[] y)
{
return(x.ManhattanDistance(y, 0));
}
public (List <string> classes, KMeans kmeans, OneVsAllSupportVectorMachine svm) FinishLearning(
int vectorLength,
CancellationToken cancellationToken)
{
// count classes
List <string> classes = new List <string>(this.features.Select(x => x.truth).ToLookup(x => x).Select(x => x.Key));
if (classes.Count < 2)
{
throw new ArgumentException();
}
classes.Sort();
// count vectors
int numberOfVectors = this.features.Sum(x => x.features.Count);
// copy vectors
Dictionary <IVector <float>, float> vectors = new Dictionary <IVector <float>, float>(numberOfVectors);
for (int i = 0, ii = this.features.Count; i < ii; i++)
{
FeatureDetectors.Features f = this.features[i].features;
for (int j = 0, jj = f.Count, len = f.Length, off = 0; j < jj; j++, off += len)
{
////DenseVectorF vector = new DenseVectorF(len, f.X, off);
SparseVectorF vector = SparseVectorF.FromDense(len, f.X, off);
vectors[vector] = vectors.TryGetValue(vector, out float weight) ? weight + 1.0f : 1.0f;
}
}
cancellationToken.ThrowIfCancellationRequested();
// learn k-means
KMeans kmeans = KMeans.Learn(
vectorLength,
KMeansSeeding.Random,
2,
default(EuclideanDistance),
vectors.Keys.ToList(),
vectors.Values.ToList(),
cancellationToken);
cancellationToken.ThrowIfCancellationRequested();
// learn svm
Dictionary <string, int> classesLookup = classes.ToDictionary((x, i) => x, (x, i) => i);
SequentualMinimalOptimization smo = new SequentualMinimalOptimization(new ChiSquare())
{
Algorithm = SMOAlgorithm.LibSVM,
Tolerance = 0.01f,
};
List <float[]> svmx = new List <float[]>(this.features.Count);
List <int> svmy = new List <int>(this.features.Count);
for (int i = 0, ii = this.features.Count; i < ii; i++)
{
(FeatureDetectors.Features features, string truth) = this.features[i];
svmx.Add(PointsOfInterestClassifier.PrepareVector(kmeans, features, cancellationToken));
svmy.Add(classesLookup[truth]);
}
cancellationToken.ThrowIfCancellationRequested();
OneVsAllSupportVectorMachine svm = OneVsAllSupportVectorMachine.Learn(
smo,
classes.Count,
svmx,
svmy,
null,
cancellationToken);
cancellationToken.ThrowIfCancellationRequested();
return(classes, kmeans, svm);
}
public void XXXTest()
{
#if false
const int Count = 10000000;
Stopwatch stopwatch = new Stopwatch();
RandomGenerator random = new RandomGenerator();
int length = 128;
float[] dense1 = random.Generate(length);
float[] dense2 = new float[length];
for (int i = 0; i < 16; i++)
{
dense2[(int)random.Generate(0, length)] = random.Generate();
}
SparseVectorF sparse = SparseVectorF.FromDense(length, dense2, 0);
stopwatch.Restart();
for (int i = 0; i < Count; i++)
{
Math32f.EuclideanDistance(length, dense1, 0, dense2, 0);
}
stopwatch.Stop();
Console.WriteLine("{0:F4} ms", stopwatch.ElapsedMilliseconds /* / Count*/);
stopwatch.Restart();
for (int i = 0; i < Count; i++)
{
sparse.EuclideanDistance(dense1, 0);
}
stopwatch.Stop();
Console.WriteLine("{0:F4} ms", stopwatch.ElapsedMilliseconds /* / Count*/);
#else
const int Count = 5;
Stopwatch stopwatch = new Stopwatch();
////Image image = new Image(2000, 3000, 8, 200, 200);
////image.Randomize();
////foreach ((Image image, _, _) in Image.FromFile(@"C:\DNN\dnn\test.jpg"))
////foreach ((Image image, _, _) in Image.FromFile(@"C:\DNN\dnn\363978.tif"))
foreach ((Image image, _, _) in Imaging.Image.FromFile(@"L:\FormXtra\Receipts\Concur\Full\jpg\Work\08240EB488FE324E83C7C54042D9813C.jpg"))
{
stopwatch.Restart();
Image xxx = image.ConvertTo(null, 8);
for (int i = 0; i < Count; i++)
{
xxx.Erode(xxx, StructuringElement.Square(2), 1, BorderType.BorderRepl, 0);
xxx.ScaleByDownsampling2(xxx);
////image.Binarize(null, 0, 0, 0, 0, true, 0, 0);
////IntegralImage.FromImage(image);
////long power = image.Power();
////Histogram hyst = image.GrayHistogram();
/*Image workImage = image
* .ConvertTo(null, 8)
* .Scale(null, 100.0 / image.HorizontalResolution, 100.0 / image.VerticalResolution, ScalingOptions.None)
* ////.Binarize(null)
* .Convert8To1(null, 128)
* .CleanOverscan(0.5f, 0.5f)
* .Deskew(null)
* .Despeckle(null);
*
* ISet<ConnectedComponent> components = workImage.FindConnectedComponents(8);
* workImage.RemoveConnectedComponents(components);*/
/*workImage = workImage.Scale(100.0 / image.HorizontalResolution, 100.0 / image.VerticalResolution, Imaging.ScalingOptions.None);
* workImage = workImage.Binarize();
* ////workImage = workImage.Convert8To1(128);
* workImage = workImage.Dilate(StructuringElement.Rectangle(5, 1), 1);
* workImage = workImage.Dilate(StructuringElement.Rectangle(1, 5), 1);
* workImage = workImage.Convert1To8();
*
* DenseVectorPackF vectors = workImage.HOG(8, 2, 1, 9, 0.2f);
*
* vectors = DenseVectorPackF.Pack(vectors.Unpack().Where(x => x.Sum() != 0.0f).ToList());*/
////Image temp = image.Binarize();
}
stopwatch.Stop();
}
Console.WriteLine("{0:F4} ms", stopwatch.ElapsedMilliseconds / Count);
#endif
}
