public override LearningImage Project(LearningImage image)
{
LearningImage o = new LearningImage(image);
double[] maxs = new double[image.Plane];
for (int h = 0; h < o.Height; h++)
{
for (int w = 0; w < o.Width; w++)
{
for (int p = 0; p < o.Plane; p++)
{
double v = o.Data[(o.Width * h + w) * o.Plane + p];
if (maxs[p] < Math.Abs(v)) maxs[p] = Math.Abs(v);
}
}
}
for (int h = 0; h < o.Height; h++)
{
for (int w = 0; w < o.Width; w++)
{
for (int p = 0; p < o.Plane; p++)
{
if (maxs[p] == 0) continue;
o.Data[(o.Width * h + w) * o.Plane + p] /= maxs[p];
}
}
}
return o;
}
public override LearningImage BackProject(LearningImage i)
{
List<double> list = new List<double>(i.Data);
LearningImage image = new LearningImage(Height, Width, Plane);
LearningImage tmp = new LearningImage(Height, Width, Plane);
for (int m = 0; m < MainMax; m++)
{
// double length = LearningImage.EuclideanLength(_MainImages[m]); 本来はlist[m]*length
LearningImage.Sacle(_MainImages[m], tmp, list[m]);
LearningImage.Add(image, tmp, image);
}
return image;
}
public override LearningImage BackProject(LearningImage image)
{
int step = FrameIn.Height;
int height = image.Height * step;
int width = image.Width * step;
LearningImage backed = new LearningImage(height, width, image.Plane);
for (int h = 0; h < height; h++)
{
for (int w = 0; w < width; w++)
{
var p = image.GetPlane((double)h / (double)step, (double)w / (double)step);
backed.SetPlane(h, w, p);
}
}
return backed;
}
public override LearningImage Project(LearningImage image)
{
int step = FrameIn.Height;
int height = image.Height / step;
int width = image.Width / step;
LearningImage projected = new LearningImage(height, width, image.Plane);
for (int h = 0; h < image.Height; h += step)
{
for(int w = 0; w < image.Width; w += step)
{
var list = image.GetPlanes(new Rectangle(w, h, step, step));
var best = list.OrderByDescending(x => x.SpecialEuclidean()).First();
projected.SetPlane(h / step, w / step, best);
}
}
return projected;
}
public override LearningImage Project(LearningImage i)
{
List<double> results = new List<double>();
int scaledH = 0;
int scaledW = 0;
for (int h = -Height / 2; h < i.Height - Height / 2; h += this.Scale)
{
scaledW = 0;
for (int w = -Width / 2; w < i.Width - Width / 2; w += this.Scale)
{
var trimed = i.Trim(new Rectangle(w, h, Width, Height));
var projected = base.Project(trimed);
results.AddRange(projected.Data);
scaledW++;
}
scaledH++;
}
return new LearningImage(scaledH, scaledW, this.FrameOut.Plane, results.ToArray());
}
public override LearningImage BackProject(LearningImage i)
{
int Scale = this.Scale;
ListImage li = new ListImage(i.Height * Scale + Height, i.Width * Scale + Width);
for (int h = 0; h < i.Height; h++)
{
for (int w = 0; w < i.Width; w++)
{
LearningImage trimed = i.Trim(new Rectangle(w, h, 1, 1));
LearningImage pasting = base.BackProject(trimed);
for (int hh = 0; hh < pasting.Height; hh++)
for (int ww = 0; ww < pasting.Width; ww++)
li.Add(h * Scale + hh, w * Scale + ww, pasting.GetPlane(hh, ww));
}
}
LearningImage o = new LearningImage(i.Height * Scale, i.Width * Scale, this.FrameIn.Plane);
for (int h = 0; h < o.Height; h++)
for (int w = 0; w < o.Width; w++)
o.SetPlane(h, w, li.Median(h + Height / 2, w + Width / 2));
return o;
}
public void Paste(int x, int y, LearningImage image)
{
for (int h = 0; h < image.Height; h++)
{
for (int w = 0; w < image.Width; w++)
{
if (h + y >= this.Height || w + x >= this.Width) continue;
int pt = (this.Width * (h + y) + (w + x)) * Plane;
int pi = (image.Width * h + w) * Plane;
for (int l = 0; l < Plane; l++)
this.Data[pt + l] = image.Data[pi + l]; // デフォルト
}
}
}
public static double DotProduct(LearningImage a, LearningImage b)
{
return Matrix.InnerProduct(a.Data, b.Data);
}
public static double EuclideanLength(LearningImage a)
{
return Norm.Euclidean(a.Data);
}
public override void Initialize()
{
_FrameNow = 0;
_MainImages = new LearningImage[MainMax];
_TmpImages = new LearningImage[MainMax];
for (int m = 0; m < MainMax; m++)
{
_MainImages[m] = new LearningImage(Height, Width, Plane);
_TmpImages[m] = new LearningImage(Height, Width, Plane);
}
}
public static void Add(LearningImage a, LearningImage b, LearningImage o)
{
for (int l = 0; l < o.Length; l++) o.Data[l] = a.Data[l] + b.Data[l];
}
public static List<double> HighLow(LearningImage a)
{
return new List<double>() { a.Data.Max(), a.Data.Min() };
}
private LearningImage ForecastUnit(int index, LearningImage i)
{
if(index >= _Units.Count) return i;
var c = _Units[index].Project(i);
//c.SavePngAdjusted("../c" + index + ".png");
var o = ForecastUnit(index + 1, c);
//var e = _Units[index].BackProject(o);
//e.SavePngAdjusted("../e" + index + ".png");
return o;
}
public static double ParticularEuclideanLength(LearningImage a)
{
double amount = 0;
for (int i = 1; i < a.Data.Length; i++) amount += a.Data[i] * a.Data[i];
return Math.Sqrt(amount);
}
public static LearningImage LoadBin(string path)
{
BinaryReader reader = null;
try
{
if (!File.Exists(path)) return null;
byte[] bytes = File.ReadAllBytes(path);
int shift = 0;
int height = BitConverter.ToInt32(bytes, shift); shift += sizeof(int);
int width = BitConverter.ToInt32(bytes, shift); shift += sizeof(int);
int plane = BitConverter.ToInt32(bytes, shift); shift += sizeof(int);
LearningImage image = new LearningImage(height, width, plane);
for (int l = 0; l < image.Length; l++) image.Data[l] = BitConverter.ToDouble(bytes, l * sizeof(double) + shift);
return image;
}
catch(Exception ex)
{
Log.Instance.Error(ex.Message + "@" + ex.StackTrace);
return null;
}
finally
{
if (reader != null) reader.Close();
reader = null;
}
}
// 正規化
public LearningImage Normalize()
{
double amount_x = 0;
double amount_xx = 0;
for(int j = 0; j < this.Length; j++)
{
double x = this.Data[j];
amount_x += x;
amount_xx += x * x;
}
double average = amount_x / this.Length;
double deviation = Math.Sqrt(amount_xx / this.Length - average * average);
if (deviation == 0) deviation = 1.0;
LearningImage i = new LearningImage(this);
for (int j = 0; j < Length; j++) i.Data[j] = (i.Data[j] - average) / deviation;
return i;
}
// ドロップアウト
public LearningImage DropOut(double rate)
{
LearningImage i = new LearningImage(this);
Random random = new Random();
for(int j = 0; j < i.Length; j++)
{
if (random.NextDouble() < rate) i.Data[j] = 0;
}
return i;
}
public void Blend(int x, int y, LearningImage image, double rate = 0.75)
{
for (int h = 0; h < image.Height; h++)
{
for (int w = 0; w < image.Width; w++)
{
if (h + y >= this.Height || w + x >= this.Width) continue;
int pt = (this.Width * (h + y) + (w + x)) * Plane;
int pi = (image.Width * h + w) * Plane;
for (int l = 0; l < Plane; l++) this.Data[pt + l] = this.Data[pt + l] * (1 - rate) + image.Data[pi + l] * rate;
}
}
}
// ノイズ追加、
public LearningImage AddNoise(double range)
{
LearningImage i = new LearningImage(this);
Random random = new Random();
for (int j = 0; j < i.Length; j++)
{
i.Data[j] += (random.NextDouble() - 0.5) * range;
}
return i;
}
public static void Sacle(LearningImage i, LearningImage o, double scale = 1, double bias = 0)
{
for (int l = 0; l < o.Length; l++) o.Data[l] = i.Data[l] * scale + bias;
}
public virtual LearningImage Forecast(LearningImage image)
{
return image;
}
public virtual LearningImage Project(LearningImage image)
{
return image;
}
public LearningImage(LearningImage image)
: this(image._Frame, image.Data)
{
}
public static unsafe LearningImage FromBitmap(Bitmap src, ColorType color = ColorType.Color)
{
try
{
PixelFormat format = PixelFormat.Format24bppRgb;
int plane = 3;
if (color == ColorType.Gray)
{
format = PixelFormat.Format8bppIndexed;
plane = 1;
}
BitmapData srcData = src.LockBits(new Rectangle(Point.Empty, src.Size), ImageLockMode.ReadOnly, format);
LearningImage i = new LearningImage(src.Height, src.Width, plane);
for (int h = 0; h < srcData.Height; h++)
{
byte* ps = (byte*)srcData.Scan0 + srcData.Stride * h;
for (int w = 0; w < srcData.Width; w++, ps += i.Plane)
{
int postion = (i.Width * h + w) * i.Plane;
for (int l = 0; l < i.Plane; l++)
i.Data[postion + l] = (double)ps[l] / 255;
}
}
src.UnlockBits(srcData);
return i;
}
catch(Exception ex)
{
Log.Instance.Error(ex.Message + "@" + ex.StackTrace);
return null;
}
}
public virtual void ParallelProject(List<LearningImage> inputs, out List<LearningImage> outputs)
{
LearningImage[] array = new LearningImage[inputs.Count];
Parallel.For(0, inputs.Count, GetParallelOptions(), i => array[i] = Project(inputs[i]));
outputs = new List<LearningImage>(array);
}
public LearningImage Trim(Rectangle r)
{
LearningImage i = new LearningImage(r.Height, r.Width, Plane);
for(int h = r.Top; h < r.Bottom; h++)
{
for(int w = r.Left; w < r.Right; w++)
{
int sh = h;
int sw = w;
if (sh < 0) sh = 0;
if (sh >= this.Height) sh = this.Height - 1;
if (sw < 0) sw = 0;
if (sw >= this.Width) sw = this.Width - 1;
int d = i.Width * (h - r.Top) + (w - r.Left);
int s = sh * this.Width + sw;
Array.Copy(this.Data, s * Plane, i.Data, d * Plane, Plane);
}
}
return i;
}
public override LearningImage Project(LearningImage image)
{
double[] data = _Network.Compute(image.Homogenize());
return new LearningImage(FrameOut, data);
}
// 0から1へ値を調整
public LearningImage ZeroToOne()
{
LearningImage i = new LearningImage(this);
for (int j = 0; j < Length; j++) i.Data[j] = (i.Data[j] + 1.0) * 0.5;
return i;
}
protected LearningImage MakeOutimageByDirectory(int height, int width, string path)
{
string dir = Path.GetDirectoryName(path);
int index = dir.LastIndexOf('\\');
string group = dir.Substring(index + 1);
index = group.IndexOf('-');
if (index == 0) return null; // 頭に-がついているフォルダは無視
if (index > 0) group = group.Substring(0, index);
int number = 0;
if (!int.TryParse(group, out number)) return null;
LearningImage result = new LearningImage(height, width, 1);
if (0 <= number && number < height * width) result.Data[number] = 1;
return result;
}
public LearningImagePair(LearningImage i, LearningImage o)
{
In = i; Out = o;
}
