private Pixel GetAbsMax(float[,] xImage, float[,] bMap, float lambda, float alpha)
{
var currentMax = new Pixel(0, -1, -1);
for (int y = 0; y < imageSection.YExtent(); y++)
{
for (int x = 0; x < imageSection.XExtent(); x++)
{
var currentA = aMap[y, x];
var old = xImage[y, x];
var xTmp = ElasticNet.ProximalOperator(old * currentA + bMap[y, x], currentA, lambda, alpha);
var xDiff = xTmp - old;
if (currentMax.MaxDiff < Math.Abs(xDiff))
{
var yGlobal = y + imageSection.Y;
var xGlobal = x + imageSection.X;
currentMax = new Pixel(xDiff, yGlobal, xGlobal);
}
}
}
var maxPixelGlobal = comm.Allreduce(currentMax, (aC, bC) => aC.MaxDiff > bC.MaxDiff ? aC : bC);
return(maxPixelGlobal);
}
public static Tuple <int, int, double>[] GetMaxBlocks(float[,] gExplore, float[,] gCorrection, float[,] xExplore, float stepSize, float theta, float lambda, float alpha, int yBlockSize, int xBlockSize, int tau)
{
var yBlocks = gExplore.GetLength(0) / yBlockSize;
var xBlocks = gExplore.GetLength(1) / xBlockSize;
var tmp = new List <Tuple <int, int, double> >(yBlocks * xBlocks);
for (int i = 0; i < yBlocks; i++)
{
for (int j = 0; j < xBlocks; j++)
{
int yIdx = i * yBlockSize;
int xIdx = j * xBlockSize;
var sum = 0.0;
for (int y = yIdx; y < yIdx + yBlockSize; y++)
{
for (int x = xIdx; x < xIdx + xBlockSize; x++)
{
var update = theta * gCorrection[y, x] + gExplore[y, x] + xExplore[y, x] * stepSize;
var shrink = ElasticNet.ProximalOperator(update, stepSize, lambda, alpha) - xExplore[y, x];
sum += Math.Abs(shrink);
}
}
tmp.Add(new Tuple <int, int, double>(i, j, sum));
}
}
tmp.Sort((x, y) => x.Item3.CompareTo(y.Item3));
var output = new Tuple <int, int, double> [tau];
for (int i = 0; i < tau; i++)
{
output[i] = tmp[tmp.Count - i - 1];
}
return(output);
}
public void Deconvolve()
{
var update = 0.0f;
var blockCount = shared.XExpl.Length;
float eta = 1.0f / blockCount;
DiffMax = 0.0f;
var beta = CalcESO(shared.ProcessorCount, shared.DegreeOfSeperability, blockCount);
var continueAsync = Thread.VolatileRead(ref shared.AsyncFinished) == 0;
for (int inner = 0; inner < shared.MaxConcurrentIterations & continueAsync; inner++)
{
continueAsync = Thread.VolatileRead(ref shared.AsyncFinished) == 0;
var stepFactor = (float)beta * Theta / shared.Theta0;
var theta2 = Theta * Theta;
var blockIdx = GetPseudoRandomBlock(stepFactor, theta2);
var blockSample = shared.ActiveSet[blockIdx];
var yPixel = blockSample.Item1;
var xPixel = blockSample.Item2;
var step = shared.AMap[yPixel, xPixel] * stepFactor;
var correctionFactor = -(1.0f - Theta / shared.Theta0) / theta2;
var xExpl = Thread.VolatileRead(ref shared.XExpl[yPixel, xPixel]);
update = theta2 * Thread.VolatileRead(ref shared.GCorr[yPixel, xPixel]) + Thread.VolatileRead(ref shared.GExpl[yPixel, xPixel]) + xExpl * step;
update = ElasticNet.ProximalOperator(update, step, shared.Lambda, shared.Alpha) - xExpl;
DiffMax = Math.Max(DiffMax, Math.Abs(update));
//update gradients
if (0.0f != Math.Abs(update))
{
UpdateGradientsApprox(shared.GExpl, shared.GCorr, shared.Psf2, updateCache, yPixel, xPixel, correctionFactor, update);
var oldExplore = shared.XExpl[yPixel, xPixel]; //does not need to be volatile, this index is blocked until this process is finished, and we already made sure with a volatile read that the latest value is in the cache
var oldXCorr = Thread.VolatileRead(ref shared.XCorr[yPixel, xPixel]);
var newXExplore = oldExplore + update;
Thread.VolatileWrite(ref shared.XExpl[yPixel, xPixel], shared.XExpl[yPixel, xPixel] + update);
Thread.VolatileWrite(ref shared.XCorr[yPixel, xPixel], shared.XCorr[yPixel, xPixel] + update * correctionFactor);
//not 100% sure this is the correct generalization from single pixel thread rule to block rule
var testRestartUpdate = (update) * (newXExplore - (theta2 * oldXCorr + oldExplore));
ConcurrentUpdateTestRestart(ref shared.TestRestart, eta, testRestartUpdate);
}
AsyncIterations++;
//unlockBlock
Thread.VolatileWrite(ref shared.BlockLock[blockIdx], 0);
Theta = (float)(Math.Sqrt((theta2 * theta2) + 4 * (theta2)) - theta2) / 2.0f;
}
Thread.VolatileWrite(ref shared.AsyncFinished, 1);
}
private static float GetMaxAbsPixelValue(SharedData shared, Tuple <int, int> block, float stepFactor)
{
var yPixel = block.Item1;
var xPixel = block.Item2;
var step = shared.AMap[yPixel, xPixel] * stepFactor;
var xExpl = Thread.VolatileRead(ref shared.XExpl[yPixel, xPixel]);
var update = Thread.VolatileRead(ref shared.GExpl[yPixel, xPixel]) + xExpl * step;
update = ElasticNet.ProximalOperator(update, step, shared.Lambda, shared.Alpha) - xExpl;
return(Math.Abs(update));
}
public static List <Tuple <int, int> > GetActiveSet(float[,] xExplore, float[,] gExplore, int yBlockSize, int xBlockSize, float lambda, float alpha, float[,] lipschitzMap)
{
var debug = new float[xExplore.GetLength(0), xExplore.GetLength(1)];
var output = new List <Tuple <int, int> >();
for (int i = 0; i < xExplore.GetLength(0) / yBlockSize; i++)
{
for (int j = 0; j < xExplore.GetLength(1) / xBlockSize; j++)
{
var yPixel = i * yBlockSize;
var xPixel = j * xBlockSize;
var nonZero = false;
for (int y = yPixel; y < yPixel + yBlockSize; y++)
{
for (int x = xPixel; x < xPixel + xBlockSize; x++)
{
var lipschitz = lipschitzMap[y, x];
var tmp = gExplore[y, x] + xExplore[y, x] * lipschitz;
tmp = ElasticNet.ProximalOperator(tmp, lipschitz, lambda, alpha);
if (0.0f < Math.Abs(tmp - xExplore[y, x]))
{
nonZero = true;
}
}
}
if (nonZero)
{
output.Add(new Tuple <int, int>(i, j));
for (int y = yPixel; y < yPixel + yBlockSize; y++)
{
for (int x = xPixel; x < xPixel + xBlockSize; x++)
{
debug[y, x] = 1.0f;
}
}
}
}
}
FitsIO.Write(debug, "activeSet.fits");
return(output);
}
private static List <Tuple <int, int> > GetActiveSet(float[,] xExplore, float[,] gExplore, float lambda, float alpha, float[,] lipschitzMap)
{
var output = new List <Tuple <int, int> >();
for (int y = 0; y < xExplore.GetLength(0); y++)
{
for (int x = 0; x < xExplore.GetLength(1); x++)
{
var lipschitz = lipschitzMap[y, x];
var tmp = gExplore[y, x] + xExplore[y, x] * lipschitz;
tmp = ElasticNet.ProximalOperator(tmp, lipschitz, lambda, alpha);
if (0.0f < Math.Abs(tmp - xExplore[y, x]))
{
output.Add(new Tuple <int, int>(y, x));
}
}
}
return(output);
}
private List <Tuple <int, int> > GetActiveSet(float[,] xExplore, float[,] gExplore, float lambda, float alpha, float lipschitz)
{
var debug = new float[xExplore.GetLength(0), xExplore.GetLength(1)];
var output = new List <Tuple <int, int> >();
for (int i = 0; i < xExplore.GetLength(0) / yBlockSize; i++)
{
for (int j = 0; j < xExplore.GetLength(1) / xBlockSize; j++)
{
var yPixel = i * yBlockSize;
var xPixel = j * xBlockSize;
var nonZero = false;
for (int y = yPixel; y < yPixel + yBlockSize; y++)
{
for (int x = xPixel; x < xPixel + xBlockSize; x++)
{
var tmp = gExplore[y, x] + xExplore[y, x] * lipschitz;
tmp = ElasticNet.ProximalOperator(tmp, lipschitz, lambda, alpha);
if (ACTIVE_SET_CUTOFF < Math.Abs(tmp - xExplore[y, x]))
{
nonZero = true;
}
}
}
if (nonZero)
{
output.Add(new Tuple <int, int>(i, j));
for (int y = yPixel; y < yPixel + yBlockSize; y++)
{
for (int x = xPixel; x < xPixel + xBlockSize; x++)
{
debug[y, x] = 1.0f;
}
}
}
}
}
//FitsIO.Write(debug, "activeSet.fits");
//can write max change for convergence purposes
return(output);
}
private Pixel GetAbsMaxSingle(Rectangle subpatch, float[,] xImage, float[,] gradients, float lambda, float alpha)
{
var maxPixels = new Pixel[subpatch.YExtent()];
for (int y = subpatch.Y; y < subpatch.YEnd; y++)
{
var yLocal = y;
var currentMax = new Pixel(-1, -1, 0, 0);
for (int x = subpatch.X; x < subpatch.XEnd; x++)
{
var xLocal = x;
var currentA = aMap[yLocal, xLocal];
var old = xImage[yLocal, xLocal];
//var xTmp = old + bMap[y, x] / currentA;
//xTmp = ShrinkElasticNet(xTmp, lambda, alpha);
var xTmp = ElasticNet.ProximalOperator(old * currentA + gradients[y, x], currentA, lambda, alpha);
var xDiff = old - xTmp;
if (currentMax.PixelMaxDiff < Math.Abs(xDiff))
{
currentMax.Y = y;
currentMax.X = x;
currentMax.PixelMaxDiff = Math.Abs(xDiff);
currentMax.PixelNew = xTmp;
}
}
maxPixels[yLocal] = currentMax;
}
var maxPixel = new Pixel(-1, -1, 0, 0);
for (int i = 0; i < maxPixels.Length; i++)
{
if (maxPixel.PixelMaxDiff < maxPixels[i].PixelMaxDiff)
{
maxPixel = maxPixels[i];
}
}
return(maxPixel);
}
private static float GetMaxAbsBlockValue(SharedData shared, Tuple <int, int> block, float stepFactor, float theta2)
{
var yOffset = block.Item1 * shared.YBlockSize;
var xOffset = block.Item2 * shared.XBlockSize;
var blockLipschitz = GetBlockLipschitz(shared.AMap, yOffset, xOffset, shared.YBlockSize, shared.XBlockSize);
var step = blockLipschitz * stepFactor;
var updateAbsSum = 0.0f;
for (int y = yOffset; y < (yOffset + shared.YBlockSize); y++)
{
for (int x = xOffset; x < (xOffset + shared.XBlockSize); x++)
{
var xExpl = Thread.VolatileRead(ref shared.XExpl[y, x]);
var update = theta2 * Thread.VolatileRead(ref shared.GCorr[y, x]) + Thread.VolatileRead(ref shared.GExpl[y, x]) + xExpl * step;
update = ElasticNet.ProximalOperator(update, step, shared.Lambda, shared.Alpha) - xExpl;
updateAbsSum += Math.Abs(update);
}
}
return(updateAbsSum);
}
public static float GetAbsMax(float[,] xImage, float[,] bMap, float[,] aMap, float lambda, float alpha)
{
var maxPixels = new float[xImage.GetLength(0)];
Parallel.For(0, xImage.GetLength(0), (y) =>
{
var yLocal = y;
var currentMax = 0.0f;
for (int x = 0; x < xImage.GetLength(1); x++)
{
var xLocal = x;
var currentA = aMap[yLocal, xLocal];
var old = xImage[yLocal, xLocal];
//var xTmp = old + bMap[y, x] / currentA;
//xTmp = ShrinkElasticNet(xTmp, lambda, alpha);
var xTmp = ElasticNet.ProximalOperator(old * currentA + bMap[y, x], currentA, lambda, alpha);
var xDiff = old - xTmp;
if (currentMax < Math.Abs(xDiff))
{
currentMax = Math.Abs(xDiff);
}
}
maxPixels[yLocal] = currentMax;
});
var maxPixel = 0.0f;
for (int i = 0; i < maxPixels.Length; i++)
{
if (maxPixel < maxPixels[i])
{
maxPixel = maxPixels[i];
}
}
return(maxPixel);
}
public void ISTAStep(float[,] xImage, float[,] residuals, float[,] psf, float lambda, float alpha)
{
var xOld = Copy(xImage);
var corrKernel = PSF.CalcPaddedFourierCorrelation(psf, new Rectangle(0, 0, residuals.GetLength(0), residuals.GetLength(1)));
var gradients = Residuals.CalcGradientMap(residuals, corrKernel, new Rectangle(0, 0, psf.GetLength(0), psf.GetLength(1)));
var lipschitz = (float)PSF.CalcMaxLipschitz(psf) * xImage.Length;
for (int i = 0; i < xImage.GetLength(0); i++)
{
for (int j = 0; j < xImage.GetLength(1); j++)
{
var tmp = gradients[i, j] + xImage[i, j] * lipschitz;
tmp = ElasticNet.ProximalOperator(tmp, lipschitz, lambda, alpha);
xImage[i, j] = tmp;
}
}
//update residuals
for (int i = 0; i < xImage.GetLength(0); i++)
{
for (int j = 0; j < xImage.GetLength(1); j++)
{
xOld[i, j] = xImage[i, j] - xOld[i, j];
}
}
var convKernel = PSF.CalcPaddedFourierConvolution(psf, new Rectangle(0, 0, residuals.GetLength(0), residuals.GetLength(1)));
var residualsCalculator = new PaddedConvolver(convKernel, new Rectangle(0, 0, psf.GetLength(0), psf.GetLength(1)));
residualsCalculator.ConvolveInPlace(xOld);
for (int i = 0; i < xImage.GetLength(0); i++)
{
for (int j = 0; j < xImage.GetLength(1); j++)
{
residuals[i, j] -= xOld[i, j];
}
}
}
public void Deconvolve()
{
var update = 0.0f;
var blockCount = shared.XExpl.Length;
DiffMax = 0.0f;
var beta = CalcESO(shared.ProcessorCount, shared.DegreeOfSeperability, blockCount);
var continueAsync = Thread.VolatileRead(ref shared.AsyncFinished) == 0;
for (int inner = 0; inner < shared.MaxConcurrentIterations & continueAsync; inner++)
{
continueAsync = Thread.VolatileRead(ref shared.AsyncFinished) == 0;
var blockIdx = GetPseudoRandomBlock((float)beta);
var blockSample = shared.ActiveSet[blockIdx];
var yPixel = blockSample.Item1;
var xPixel = blockSample.Item2;
var step = shared.AMap[yPixel, xPixel] * (float)beta;
var xExpl = Thread.VolatileRead(ref shared.XExpl[yPixel, xPixel]);
update = Thread.VolatileRead(ref shared.GExpl[yPixel, xPixel]) + xExpl * step;
update = ElasticNet.ProximalOperator(update, step, shared.Lambda, shared.Alpha) - xExpl;
DiffMax = Math.Max(DiffMax, Math.Abs(update));
//update gradients
if (0.0f != Math.Abs(update))
{
UpdateGradients(shared.GExpl, shared.Psf2, yPixel, xPixel, update);
Thread.VolatileWrite(ref shared.XExpl[yPixel, xPixel], shared.XExpl[yPixel, xPixel] + update);
}
AsyncIterations++;
//unlockBlock
Thread.VolatileWrite(ref shared.BlockLock[blockIdx], 0);
}
Thread.VolatileWrite(ref shared.AsyncFinished, 1);
}
public void Deconvolve()
{
var blockCount = shared.XExpl.Length / (shared.YBlockSize * shared.XBlockSize);
//var blockCount = shared.ActiveSet.Count;
float eta = 1.0f / blockCount;
var beta = CalcESO(shared.ProcessorCount, shared.DegreeOfSeperability, blockCount);
xDiffMax = 0.0f;
var continueAsync = Thread.VolatileRead(ref shared.asyncFinished) == 0;
for (int inner = 0; inner < shared.MaxConcurrentIterations & continueAsync; inner++)
{
continueAsync = Thread.VolatileRead(ref shared.asyncFinished) == 0;
var blockIdx = GetRandomBlockIdx(random, id, shared.BlockLock);
var blockSample = shared.ActiveSet[blockIdx];
var yOffset = blockSample.Item1 * shared.YBlockSize;
var xOffset = blockSample.Item2 * shared.XBlockSize;
var blockLipschitz = GetBlockLipschitz(shared.AMap, yOffset, xOffset, shared.YBlockSize, shared.XBlockSize);
var step = blockLipschitz * (float)beta * Theta / shared.theta0;
var theta2 = Theta * Theta;
var correctionFactor = -(1.0f - Theta / shared.theta0) / theta2;
var updateSum = 0.0f;
var updateAbsSum = 0.0f;
for (int y = yOffset; y < (yOffset + shared.YBlockSize); y++)
{
for (int x = xOffset; x < (xOffset + shared.XBlockSize); x++)
{
var xExpl = Thread.VolatileRead(ref shared.XExpl[y, x]);
var update = theta2 * Thread.VolatileRead(ref shared.GCorr[y, x]) + Thread.VolatileRead(ref shared.GExpl[y, x]) + xExpl * step;
update = ElasticNet.ProximalOperator(update, step, shared.Lambda, shared.Alpha) - xExpl;
blockUpdate[y - yOffset, x - xOffset] = update;
updateSum = update;
updateAbsSum += Math.Abs(update);
}
}
//update gradients
if (0.0f != updateAbsSum)
{
xDiffMax = Math.Max(xDiffMax, updateAbsSum);
UpdateBMaps(blockUpdate, blockSample.Item1, blockSample.Item2, shared.Psf2, shared.GExpl, shared.GCorr, correctionFactor);
var newXExplore = 0.0f;
var oldXExplore = 0.0f;
var oldXCorr = 0.0f;
for (int y = yOffset; y < (yOffset + shared.YBlockSize); y++)
{
for (int x = xOffset; x < (xOffset + shared.XBlockSize); x++)
{
var update = blockUpdate[y - yOffset, x - xOffset];
var oldExplore = shared.XExpl[y, x]; //does not need to be volatile, this index is blocked until this process is finished, and we already made sure with a volatile read that the latest value is in the cache
oldXExplore += shared.XExpl[y, x];
oldXCorr += Thread.VolatileRead(ref shared.XCorr[y, x]);
newXExplore += oldExplore + update;
Thread.VolatileWrite(ref shared.XExpl[y, x], shared.XExpl[y, x] + update);
Thread.VolatileWrite(ref shared.XCorr[y, x], shared.XCorr[y, x] + update * correctionFactor);
}
}
//not 100% sure this is the correct generalization from single pixel thread rule to block rule
var testRestartUpdate = (updateSum) * (newXExplore - (theta2 * oldXCorr + oldXExplore));
ConcurrentUpdateTestRestart(ref shared.testRestart, eta, testRestartUpdate);
}
//unlockBlock
Thread.VolatileWrite(ref shared.BlockLock[blockIdx], 0);
if (useAcceleration)
{
Theta = (float)(Math.Sqrt((theta2 * theta2) + 4 * (theta2)) - theta2) / 2.0f;
}
}
Thread.VolatileWrite(ref shared.asyncFinished, 1);
}
public float DeconvolveAccelerated(float[,] xExplore, float[,] xCorrection, float[,] gExplore, float[,] gCorrection, float[,] psf2, ref List <Tuple <int, int> > activeSet, float maxLipschitz, float lambda, float alpha, Random random, int maxIteration, float epsilon)
{
var blockCount = activeSet.Count;
var beta = CalcESO(tau, degreeOfSeperability, blockCount);
var lipschitz = maxLipschitz * yBlockSize * xBlockSize;
lipschitz *= (float)beta;
var theta = tau / (float)blockCount;
var theta0 = theta;
float eta = 1.0f / blockCount;
var testRestart = 0.0f;
var iter = 0;
var converged = false;
Console.WriteLine("Starting Active Set iterations with " + activeSet.Count + " blocks");
while (iter < maxIteration & !converged)
{
var xDiffMax = new float[tau];
var innerIterCount = Math.Min(activeSet.Count / tau, MAX_ACTIVESET_ITER / tau);
for (int inner = 0; inner < innerIterCount; inner++)
{
var stepSize = lipschitz * theta / theta0;
var theta2 = theta * theta;
var correctionFactor = -(1.0f - theta / theta0) / theta2;
var samples = activeSet.Shuffle(random).Take(tau).ToList();
Parallel.For(0, tau, (i) =>
{
var blockSample = samples[i];
var yOffset = blockSample.Item1 * yBlockSize;
var xOffset = blockSample.Item2 * xBlockSize;
var blockUpdate = new float[yBlockSize, xBlockSize];
var updateSum = 0.0f;
var updateAbsSum = 0.0f;
for (int y = yOffset; y < (yOffset + yBlockSize); y++)
{
for (int x = xOffset; x < (xOffset + xBlockSize); x++)
{
var update = theta2 * gCorrection[y, x] + gExplore[y, x] + xExplore[y, x] * stepSize;
update = ElasticNet.ProximalOperator(update, stepSize, lambda, alpha) - xExplore[y, x];
blockUpdate[y - yOffset, x - xOffset] = update;
updateSum = update;
updateAbsSum += Math.Abs(update);
}
}
//update gradients
if (0.0f != updateAbsSum)
{
xDiffMax[i] = Math.Max(xDiffMax[i], updateAbsSum);
UpdateBMaps(blockUpdate, blockSample.Item1, blockSample.Item2, psf2, gExplore, gCorrection, correctionFactor);
var newXExplore = 0.0f;
var oldXExplore = 0.0f;
var oldXCorr = 0.0f;
for (int y = yOffset; y < (yOffset + yBlockSize); y++)
{
for (int x = xOffset; x < (xOffset + xBlockSize); x++)
{
var update = blockUpdate[y - yOffset, x - xOffset];
var oldExplore = xExplore[y, x];
var oldCorrection = xCorrection[y, x];
oldXExplore += xExplore[y, x];
oldXCorr += xCorrection[y, x];
newXExplore += oldExplore + update;
xExplore[y, x] += update;
xCorrection[y, x] += update * correctionFactor;
}
}
//not 100% sure this is the correct generalization from single pixel/single thread rule to block/parallel rule
var testRestartUpdate = (updateSum) * (newXExplore - (theta2 * oldXCorr + oldXExplore));
ConcurrentUpdateTestRestart(ref testRestart, eta, testRestartUpdate);
}
});
theta = (float)(Math.Sqrt((theta2 * theta2) + 4 * (theta2)) - theta2) / 2.0f;
}
if (testRestart > 0)
{
//restart acceleration
var tmpTheta = theta < 1.0f ? ((theta * theta) / (1.0f - theta)) : theta0;
for (int y = 0; y < xExplore.GetLength(0); y++)
{
for (int x = 0; x < xExplore.GetLength(1); x++)
{
xExplore[y, x] += tmpTheta * xCorrection[y, x];
xCorrection[y, x] = 0;
gExplore[y, x] += tmpTheta * gCorrection[y, x];
gCorrection[y, x] = 0;
}
}
Console.WriteLine("restarting");
//new active set
activeSet = GetActiveSet(xExplore, gExplore, lambda, alpha, maxLipschitz);
blockCount = activeSet.Count;
theta = tau / (float)blockCount;
theta0 = theta;
beta = CalcESO(tau, degreeOfSeperability, blockCount);
lipschitz = maxLipschitz * yBlockSize * xBlockSize;
lipschitz *= (float)beta;
}
if (xDiffMax.Sum() < epsilon)
{
converged = true;
}
Console.WriteLine("Done Active Set iteration " + iter);
iter++;
}
return(theta);
}
public float DeconvolveRandomActiveSet(float[,] xExplore, float[,] xCorrection, float[,] gExplore, float[,] gCorrection, float[,] psf2, ref List <Tuple <int, int> > activeSet, float maxLipschitz, float lambda, float alpha, Random random, int maxIteration, float epsilon)
{
var blockCount = activeSet.Count;
var beta = CalcESO(tau, degreeOfSeperability, blockCount);
var lipschitz = maxLipschitz * yBlockSize * xBlockSize;
lipschitz *= (float)beta;
var theta = tau / (float)blockCount;
var theta0 = theta;
float eta = 1.0f / blockCount;
var testRestart = 0.0;
var iter = 0;
var blocks = new float[tau, yBlockSize, xBlockSize];
var containsNonZero = new bool[tau];
var converged = false;
Console.WriteLine("Starting Active Set iterations with " + activeSet.Count + " blocks");
while (iter < maxIteration & !converged)
{
var xDiffMax = 0.0f;
for (int inner = 0; inner < Math.Min(activeSet.Count / tau, 10000 / tau); inner++)
{
var stepSize = lipschitz * theta / theta0;
var theta2 = theta * theta;
var samples = CreateSamples(blockCount, tau, random);
//minimize blocks
for (int i = 0; i < samples.Length; i++)
{
var blockSample = activeSet[samples[i]];
var yOffset = blockSample.Item1 * yBlockSize;
var xOffset = blockSample.Item2 * xBlockSize;
containsNonZero[i] = false;
for (int y = yOffset; y < (yOffset + yBlockSize); y++)
{
for (int x = xOffset; x < (xOffset + xBlockSize); x++)
{
var update = theta2 * gCorrection[y, x] + gExplore[y, x] + xExplore[y, x] * stepSize;
update = ElasticNet.ProximalOperator(update, stepSize, lambda, alpha) - xExplore[y, x];
blocks[i, y - yOffset, x - xOffset] = update;
if (update != 0.0)
{
containsNonZero[i] = true;
}
}
}
}
//update bMaps
var correctionFactor = -(1.0f - theta / theta0) / (theta * theta);
for (int i = 0; i < samples.Length; i++)
{
if (containsNonZero[i])
{
var blockSample = activeSet[samples[i]];
var yBlock = blockSample.Item1;
var xBlock = blockSample.Item2;
UpdateBMaps(i, blocks, yBlock, xBlock, psf2, gExplore, gCorrection, correctionFactor);
//FitsIO.Write(gExplore, "gExplore2.fits");
//FitsIO.Write(gCorrection, "gCorr2.fits");
var currentDiff = 0.0f;
//update reconstructed image
var yOffset = yBlock * yBlockSize;
var xOffset = xBlock * xBlockSize;
for (int y = 0; y < blocks.GetLength(1); y++)
{
for (int x = 0; x < blocks.GetLength(2); x++)
{
var update = blocks[i, y, x];
var oldExplore = xExplore[yOffset + y, xOffset + x];
var oldCorrection = xCorrection[yOffset + y, xOffset + x];
var newValue = oldExplore + update;
testRestart = (1.0 - eta) * testRestart - eta * (update) * (newValue - (theta * theta * oldCorrection + oldExplore));
currentDiff += Math.Abs(update);
xExplore[yOffset + y, xOffset + x] += update;
xCorrection[yOffset + y, xOffset + x] += update * correctionFactor;
}
}
xDiffMax = Math.Max(xDiffMax, currentDiff);
}
}
theta = (float)(Math.Sqrt((theta * theta * theta * theta) + 4 * (theta * theta)) - theta * theta) / 2.0f;
}
if (testRestart > 0)
{
//restart acceleration
var tmpTheta = theta < 1.0f ? ((theta * theta) / (1.0f - theta)) : theta0;
for (int y = 0; y < xExplore.GetLength(0); y++)
{
for (int x = 0; x < xExplore.GetLength(1); x++)
{
xExplore[y, x] += tmpTheta * xCorrection[y, x];
xCorrection[y, x] = 0;
gExplore[y, x] += tmpTheta * gCorrection[y, x];
gCorrection[y, x] = 0;
}
}
Console.WriteLine("restarting");
//new active set
activeSet = GetActiveSet(xExplore, gExplore, lambda, alpha, maxLipschitz);
blockCount = activeSet.Count;
theta = tau / (float)blockCount;
theta0 = theta;
beta = CalcESO(tau, degreeOfSeperability, blockCount);
lipschitz = maxLipschitz * yBlockSize * xBlockSize;
lipschitz *= (float)beta;
}
if (xDiffMax < epsilon)
{
//converged = true;
}
Console.WriteLine("Done Active Set iteration " + iter);
iter++;
}
return(theta);
}