static Kernel FormHypothesis(int BlockSize, float[] w_block, float[] b_block)
{
var kernel = new AffineKernel(BlockSize);
for (int j = 0; j < BlockSize; j++)
{
kernel.threshold = .01f + .5f * (float)rnd.NextDouble();
//kernel.threshold = 1;
//kernel.w[j] = (float)rnd.NextDouble();
//kernel.b[j] = (float)rnd.NextDouble();
kernel.w[j] = w_block[j] - b_block[j];
kernel.b[j] = b_block[j];
//kernel.w[j] = w_block[j];
//kernel.b[j] = 0;
//kernel.w[j] = 1;
//kernel.b[j] = b_block[j];
}
Normalize(kernel.w);
return(kernel);
}
static void Test()
{
return;
UtilTest.Run();
Console.WriteLine(s(new float[, ] {
{ 1, 2 }, { 4, 5 }
}));
var k = new AffineKernel(new float[] { 1, 2, 3 }, new float[] { 0, 0, 0 });
var input = new float[] { 2.1f, 4.1f, 6.1f };
Normalize(k.w);
var d = k.Apply(input);
Console.WriteLine(d);
var m = k.ArgMinInverseError(input);
Console.WriteLine(m);
var result = Min(x => (float)Math.Cos(x), threshold: .000001f);
Console.WriteLine(result);
Console.ReadLine();
}
static Kernel[] LearnLevel(int Rows, float[][,,] input, int Width, int Height, int Channels,
int NumKernels = 3, int BlockDim = 3)
{
int BlockSize = BlockDim * BlockDim * Channels;
// Unpack blocks.
int num_blocks = Rows * (Width - (BlockDim - 1)) * (Height - (BlockDim - 1));
var blocks = new float[num_blocks][];
var block_labels = new byte[num_blocks];
var block_sources = new int[num_blocks];
int block_index = 0;
for (int row = 0; row < Rows; row++)
{
var image = input[row];
for (int x = 0; x < Width - (BlockDim - 1); x++)
{
for (int y = 0; y < Height - (BlockDim - 1); y++)
{
var block = blocks[block_index] = new float[BlockSize];
block_sources[block_index] = row;
block_index++;
int i = 0;
for (int _x = x; _x < x + BlockDim; _x++)
{
for (int _y = y; _y < y + BlockDim; _y++)
{
for (int _c = 0; _c < Channels; _c++)
{
block[i++] = (float)image[_x, _y, _c];
}
}
}
}
}
}
Debug.Assert(block_index == blocks.Length);
// Find kernels.
var Kernels = new AffineKernel[NumKernels];
for (int i = 0; i < NumKernels; i++)
{
Console.WriteLine($"Searching for kernel {i + 1} of {NumKernels}");
var kernel = Kernels[i] = FindKernel(blocks, BlockSize);
//kernel.Plot();
var remainder = SetRemainder(blocks, kernel, GroupThreshold, labels: block_labels);
blocks = remainder;
}
return(Kernels);
}
static AffineKernel FindKernel(float[][] blocks, int BlockSize, int[] sources = null)
{
float best_score = 0;
AffineKernel best_kernel = null;
for (int i = 0; i < 500000; i++)
{
var kernel = new AffineKernel(BlockSize);
int w_index, b_index;
w_index = rnd.Next(0, blocks.Length - 1);
b_index = rnd.Next(0, blocks.Length - 1);
for (int j = 0; j < BlockSize; j++)
{
//kernel.w[j] = (float)rnd.NextDouble();
//kernel.b[j] = (float)rnd.NextDouble();
kernel.w[j] = blocks[w_index][j] - blocks[b_index][j];
kernel.b[j] = blocks[b_index][j];
kernel.threshold = .01f + .5f * (float)rnd.NextDouble();
//kernel.w[j] = blocks[w_index][j];
//kernel.b[j] = 0;
}
Normalize(kernel.w);
var score = TestKernel(blocks, kernel,
GroupThreshold, step: 599);
if (score > best_score)
{
best_score = score;
best_kernel = kernel;
Console.WriteLine($"(Iteration {i}) New best with a score of {best_score}.");
}
}
return(best_kernel);
}
static List <Tuple <AffineKernel, float> > Learn_SupervisedKernelHg(int Rows, LabeledImage[] images, int InLabel,
float InLabelMinRatio = 0.05f, int SearchLength = 1000000, int preferred_source = -1)
{
const int BlockDim = 5;
const int Channels = 3;
const int BlockSize = BlockDim * BlockDim * Channels;
var blocked_images = GetBlocks(images, Rows, Width, Height, Channels: 3, BlockDim: BlockDim, AddMirrors: true);
var list = new List <Tuple <AffineKernel, float> >();
Console.WriteLine("\n\nSearching for kernels via supervised hypothesis generation\n\n");
float best_score = 0;
AffineKernel best_kernel = null;
for (int i = 0; i < SearchLength; i++)
{
int w_index, b_index;
if (preferred_source >= 0)
{
w_index = b_index = preferred_source;
}
else
{
do
{
w_index = rnd.Next(0, blocked_images.Length - 1);
} while (images[w_index].Label != InLabel);
do
{
b_index = rnd.Next(0, blocked_images.Length - 1);
} while (images[b_index].Label != InLabel);
}
var w_blocks = blocked_images[w_index].Blocks;
var b_blocks = blocked_images[b_index].Blocks;
var w_block = w_blocks[rnd.Next(0, w_blocks.Length)];
var b_block = b_blocks[rnd.Next(0, b_blocks.Length)];
var kernel = (AffineKernel)FormHypothesis(BlockSize, w_block, b_block);
var score = OptimizeKernelCutoff(Rows, blocked_images, kernel,
InLabel: InLabel, InLabelMinRatio: InLabelMinRatio);
if (score.InHitRatio > InLabelMinRatio)
{
list.Add(new Tuple <AffineKernel, float>(kernel, score.InOutHitRatio));
if (score.InOutHitRatio > best_score)
{
best_score = score.InOutHitRatio;
best_kernel = kernel;
Console.WriteLine($"(Iteration {i}) New best with a score of {best_score}, in label ratio of {score.InHitRatio} with a threshold of {kernel.ClassificationCutoff}.");
best_kernel.Plot();
}
}
}
return(list);
}