public virtual CNN.Filter Configure(int?outputfieldsize)
{
if (outputfieldsize != null)
{
this.outputfieldsize = outputfieldsize.Value;
}
// 0.
int c = 0;
for (int i = 0; i < sources.Count; i++)
{
c += sources[i].Columns * sources[i].Rows;
}
// 1. set target
target = new fMap();
target.Configure(c, 1);
// 2. set target elements
for (int i = 0; i < c; i++)
{
target.SetElement(i, 0, new Node(new double?[this.outputfieldsize]));
}
return(this);
}
public override Image Configure(int rows, int cols, double?min, double?max)
{
Foundation.Node node;
// 0. set range
this.max = (max == null ? +1.0 : max.Value);
this.min = (min == null ? -1.0 : min.Value);
this.cols = cols;
this.rows = rows;
// 1. set delta
dq = (max.Value - min.Value) / 255;
// 2. initialize fmap
fMap fmap = (fMap) new fMap()
.Configure(rows, cols);
for (int i = 0; i < rows; i++)
{
for (int j = 0; j < cols; j++)
{
node = new Foundation.Node(new double?[] { null, null });
fmap.SetElement(i, j, node);
}
}
fmaps.Add(fmap);
return(this);
}
public override Model.Unit Configure(string configuration)
{
base.Configure(configuration);
Foundation.Node node;
// 1. set delta
dq = (max.Value - min.Value) / 255;
// 2. initialize fmap
fMap fmap = (fMap) new fMap()
.Configure(rows, cols);
for (int i = 0; i < rows; i++)
{
for (int j = 0; j < cols; j++)
{
node = new Foundation.Node(new double?[] { null, null });
fmap.SetElement(i, j, node);
}
}
fmaps.Add(fmap);
return(this);
}
/// <summary>
/// configures connected layer
/// </summary>
/// <param name="configuration">neu=outputperceptron; act=sigmoid(a=-1, b=2, c=6); nodes=6; fieldsize=2(def:2)</param>
/// <returns></returns>
public override Model.Unit Configure(string configuration)
{
string cfg = ANN.Global.Parser.RemoveWhiteSpaces(configuration);
string[] a = ANN.Global.Parser.Split(cfg, ";"), b;
int nofN = ANN.Global.Parser.Extract <int>(a, new string[] { "nodes" }, ANN.Global.Parser.Option.None, out b);
string neu = ANN.Global.Parser.Extract <string>(b, new string[] { "neu", "neuron" }, ANN.Global.Parser.Option.None, out a);
cfg = ANN.Global.Parser.Build(a, 0, ";");
neurons = new ANN.Neuron[nofN];
for (int i = 0; i < nofN; i++)
{
switch (neu)
{
case "hiddenperceptron":
neurons[i] = new ANN.Neurons.Perceptron.Hidden(cfg);
break;
case "outputperceptron":
neurons[i] = new ANN.Neurons.Perceptron.Output(cfg);
break;
}
neurons[i].SetElement(0.0);
}
ANN.Neuron n;
fmaps.Clear();
fMap sfmap = Input[0];
fMap tfmap = (fMap) new fMap()
.Configure(neurons.Length, 1);
for (int i = 0; i < neurons.Length; i++)
{
n = neurons[i];
// 1. connect sources
for (int j = 0; j < sfmap.Rows; j++)
{
n.Source = sfmap.GetElement(j, 0);
}
// 2. set target into map
tfmap.SetElement(i, 0, n.Output);
}
fmaps.Add(tfmap);
return(this);
}
public Kernel(fMap source, int?padding, int?size)
: base(source, padding, size)
{
synapse = new Synapse[size.Value, size.Value];
for (int i = 0; i < size; i++)
{
for (int j = 0; j < size; j++)
{
synapse[i, j] = new Synapse(null, Math.Daemon.NextGaussianDouble(0.0, 0.004)[0], 0.0);
}
}
}
public override CNN.Filter Configure(string configuration)
{
bias = (fMap) new fMap()
.Configure(1, 1);
bias.SetElement(0, 0, new Foundation.Node(new double?[] { 1.0, null }));
string cfg = Global.Parser.RemoveWhiteSpaces(configuration);
string[] a = Global.Parser.Split(cfg, ";"), b;
padding = Global.Parser.Extract <int?>(a, new string[] { "kernelSize" }, Global.Parser.Option.StripDefaultToken, out b) / 2;
kernel.Add(new Kernel(bias, 0, 1));
base.Configure(cfg);
return(this);
}
protected override void ConstructTargetMap()
{
int nr = ((rows.Value - kernelSize.Value) / stride.Value) + 1;
int nc = ((cols.Value - kernelSize.Value) / stride.Value) + 1;
target = new fMap();
target.Configure(nr, nc);
for (int i = 0; i < nr; i++)
{
for (int j = 0; j < nc; j++)
{
target.SetElement(i, j, new Foundation.Node(new double?[outputfieldSize]));
}
}
}
public override CNN.Filter Configure <T>(double?[] funcparams, int kernelSize, int?stride, int?outputfieldSize)
{
bias = (fMap) new fMap()
.Configure(1, 1);
bias.SetElement(0, 0, new Foundation.Node(new double?[] { 1.0, null }));
padding = kernelSize / 2;
Kernel k = new Kernel(bias, 0, 1);
k.WeightCorrection = new double?[] { null };
kernel.Add(k);
base.Configure <T>(funcparams, kernelSize, stride, outputfieldSize);
return(this);
}
public virtual Filter Configure <T>(double?[] funcparams, int kernelsize, int?kernelstride, int?kernelpadding,
int?kernelweightfieldsize, int?outputfieldsize)
where T : Function, new()
{
// 0. construct bias
bias = (fMap) new fMap()
.Configure(1, 1);
bias.SetElement(0, 0, new Foundation.Node(new double?[] { 1.0, null }));
Kernel kernel = new Kernel();
kernel.Configure(1, 1, 0, 0, kernelweightfieldsize == null ? 2 : kernelweightfieldsize.Value, true);
kernel.Source = bias;
kernels.Add(kernel);
// 1. construct kernels
for (int i = 0; i < sources.Count; i++)
{
kernel = new Kernel();
kernel.Configure(kernelsize, kernelsize, kernelstride.Value, kernelpadding.Value, kernelweightfieldsize.Value);
kernel.Source = sources[i];
kernels.Add(kernel);
}
// 2. construct function
function = new T();
function.Configure(funcparams);
function.Input = new double[kernels.Count];
// 3. construct target
int rows = ((sources[0].Rows - kernel.Rows + (2 * kernel.Padding)) / kernel.Stride) + 1;
int cols = ((sources[0].Columns - kernel.Columns + (2 * kernel.Padding)) / kernel.Stride) + 1;
target = new fMap();
target.Configure <double?>(rows, cols, outputfieldsize == null ? 2 : outputfieldsize.Value);
grad = new double?[target.Rows][];
for (int i = 0; i < target.Rows; i++)
{
grad[i] = new double?[target.Columns];
}
SetElement(grad);
return(this);
}
public virtual Filter Configure <T>(int kernelsize, int?kernelstride, int?kernelpadding, int?outputfieldsize)
where T : CNN.Kernel, new()
{
Kernel kernel = new T();
kernel.Configure(kernelsize, kernelsize, kernelstride, kernelpadding);
kernel.Source = sources[0];
kernels.Add(kernel);
this.outputfieldsize = outputfieldsize == null ? 2 : outputfieldsize.Value;
// size = (((src.Size - kernel.Size) + (2 * kernel.Padding)) / kernel.Stride) + 1;
int rows = ((sources[0].Rows - kernel.Rows + (2 * kernel.Padding)) / kernel.Stride) + 1;
int cols = ((sources[0].Columns - kernel.Columns + (2 * kernel.Padding)) / kernel.Stride) + 1;
target = new fMap();
target.Configure <double?>(rows, cols, this.outputfieldsize);
return(this);
}
public override CNN.Filter Configure(string configuration)
{
string cfg = ANN.Global.Parser.RemoveWhiteSpaces(configuration);
string[] a = ANN.Global.Parser.Split(cfg, "=");
switch (a[0])
{
case "outputfieldsize":
outputfieldsize = int.Parse(a[1]);
break;
case "void":
break;
default:
throw new Exception();
}
// 0.
int c = 0;
for (int i = 0; i < sources.Count; i++)
{
c += sources[i].Columns * sources[i].Rows;
}
// 1. set target
target = new fMap();
target.Configure(c, 1);
// 2. set target elements
for (int i = 0; i < c; i++)
{
target.SetElement(i, 0, new Node(new double?[outputfieldsize]));
}
return(this);
}
protected override void ConstructTargetMap()
{
int nr = ((rows.Value - kernelSize.Value + (2 * padding.Value)) / stride.Value) + 1;
int nc = ((cols.Value - kernelSize.Value + (2 * padding.Value)) / stride.Value) + 1;
double?[][] gradient = new double?[nr][];
for (int i = 0; i < nr; i++)
{
gradient[i] = new double?[nc];
}
Gradient = gradient;
target = new fMap();
target.Configure(nr, nc);
for (int i = 0; i < nr; i++)
{
for (int j = 0; j < nc; j++)
{
target.SetElement(i, j, new Foundation.Node(new double?[outputfieldSize]));
}
}
}
public override CNN.Filter Configure(string configuration)
{
bias = (fMap) new fMap()
.Configure(1, 1);
bias.SetElement(0, 0, new Foundation.Node(new double?[] { 1.0, null }));
string cfg = Global.Parser.RemoveWhiteSpaces(configuration);
string[] a = Global.Parser.Split(cfg, ";");
cfg = Global.Parser.Extract <string>(a, new string[] { "ker", "kernel" }, ANN.Global.Parser.Option.None, out string[] b);
a = Global.Parser.Split(cfg, "(", ")");
if (a[0].CompareTo("convolution") != 0)
{
throw new Exception();
}
a = Global.Parser.Split(a[1], ",");
cfg = Global.Parser.Extract <string>(a, new string[] { "weightfieldsize" }, Global.Parser.Option.None, out b);
Kernel kernel = new Kernel();
kernel.Configure(1, 1, 0, 0, int.Parse(cfg));
kernel.Source = bias;
kernels.Add(kernel);
base.Configure(configuration);
function.Input = new double[kernels.Count];
grad = new double?[target.Rows][];
for (int i = 0; i < target.Rows; i++)
{
grad[i] = new double?[target.Columns];
}
SetElement(grad);
return(this);
}
