// В качестве радиальной функции чаще всего применяется функция Гаусса. При размещении ее центра в точке она может быть определена в сокращенной форме как:
public double Gaussian(double[] input)
{
Func <int, double> func = j =>
{
var numerator = input[j] - Centroids[j];
numerator *= numerator;
var denominator = Widths[j] * Widths[j];
return(numerator / denominator);
};
//для каждого центра считаем функцию гаусса, потом суммируем результаты
var sum = Centroids.Select((c, j) => func(j)).Sum();
return(Math.Pow(Math.E, -0.5 * sum));
}
public float4[] Match(Image volume, float diameterAngstrom, float threshold, Func <int3, int, string, bool> progressCallback)
{
float PixelSizeBN = PixelSize;
int3 DimsRegionBN = CubeNet.BoxDimensionsPredict;
int3 DimsRegionValidBN = CubeNet.BoxDimensionsValidPredict;
int BorderBN = (DimsRegionBN.X - DimsRegionValidBN.X) / 2;
int3 DimsBN = volume.Dims;
GPU.Normalize(volume.GetDevice(Intent.Read),
volume.GetDevice(Intent.Write),
(uint)volume.ElementsReal,
1);
volume.FreeDevice();
float[] Predictions = new float[DimsBN.Elements()];
float[] Mask = new float[DimsBN.Elements()];
{
int3 DimsPositions = (DimsBN + DimsRegionValidBN - 1) / DimsRegionValidBN;
float3 PositionStep = new float3(DimsBN - DimsRegionBN) / new float3(Math.Max(DimsPositions.X - 1, 1),
Math.Max(DimsPositions.Y - 1, 1),
Math.Max(DimsPositions.Z - 1, 1));
int NPositions = (int)DimsPositions.Elements();
int3[] Positions = new int3[NPositions];
for (int p = 0; p < NPositions; p++)
{
int Z = p / (DimsPositions.X * DimsPositions.Y);
int Y = (p - Z * DimsPositions.X * DimsPositions.Y) / DimsPositions.X;
int X = p % DimsPositions.X;
Positions[p] = new int3((int)(X * PositionStep.X + DimsRegionBN.X / 2),
(int)(Y * PositionStep.Y + DimsRegionBN.Y / 2),
(int)(Z * PositionStep.Z + DimsRegionBN.Z / 2));
}
float[][] PredictionTiles = Helper.ArrayOfFunction(i => new float[DimsRegionBN.Elements()], NPositions);
int NGPUThreads = MaxThreads;
Image[] Extracted = new Image[NGPUThreads];
int DeviceID = GPU.GetDevice();
int BatchesDone = 0;
Helper.ForCPU(0, NPositions, NGPUThreads,
threadID =>
{
GPU.SetDevice(DeviceID);
Extracted[threadID] = new Image(IntPtr.Zero, DimsRegionBN);
},
(b, threadID) =>
{
int GPUID = threadID / NGPUThreads;
int GPUThreadID = threadID % NGPUThreads;
#region Extract and normalize windows
GPU.Extract(volume.GetDevice(Intent.Read),
Extracted[threadID].GetDevice(Intent.Write),
volume.Dims,
DimsRegionBN,
new int[] { Positions[b].X - DimsRegionBN.X / 2, Positions[b].Y - DimsRegionBN.Y / 2, Positions[b].Z - DimsRegionBN.Z / 2 },
1);
//GPU.Normalize(Extracted[threadID].GetDevice(Intent.Read),
// Extracted[threadID].GetDevice(Intent.Write),
// (uint)Extracted[threadID].ElementsReal,
// 1);
#endregion
//Extracted[threadID].WriteMRC("d_extracted.mrc", true);
#region Predict
long[] BatchArgMax;
float[] BatchProbability;
Predict(Extracted[threadID].GetDevice(Intent.Read),
GPUThreadID,
out BatchArgMax,
out BatchProbability);
//new Image(BatchArgMax.Select(v => (float)v).ToArray(), new int3(DimsRegionBN)).WriteMRC("d_labels.mrc", true);
for (int i = 0; i < BatchArgMax.Length; i++)
{
int Label = (int)BatchArgMax[i];
float Probability = BatchProbability[i * NClasses + Label];
PredictionTiles[b][i] = (Label >= 1 && Probability >= threshold ? Probability : 0);
}
#endregion
lock (volume)
progressCallback?.Invoke(new int3(NPositions, 1, 1), ++BatchesDone, "");
},
threadID =>
{
int GPUID = threadID / NGPUThreads;
int GPUThreadID = threadID % NGPUThreads;
Extracted[threadID].Dispose();
});
for (int z = 0; z < DimsBN.Z; z++)
{
for (int y = 0; y < DimsBN.Y; y++)
{
for (int x = 0; x < DimsBN.X; x++)
{
int ClosestX = (int)Math.Max(0, Math.Min(DimsPositions.X - 1, (int)(((float)x - DimsRegionBN.X / 2) / PositionStep.X + 0.5f)));
int ClosestY = (int)Math.Max(0, Math.Min(DimsPositions.Y - 1, (int)(((float)y - DimsRegionBN.Y / 2) / PositionStep.Y + 0.5f)));
int ClosestZ = (int)Math.Max(0, Math.Min(DimsPositions.Z - 1, (int)(((float)z - DimsRegionBN.Z / 2) / PositionStep.Z + 0.5f)));
int ClosestID = (ClosestZ * DimsPositions.Y + ClosestY) * DimsPositions.X + ClosestX;
int3 Position = Positions[ClosestID];
int LocalX = Math.Max(0, Math.Min(DimsRegionBN.X - 1, x - Position.X + DimsRegionBN.X / 2));
int LocalY = Math.Max(0, Math.Min(DimsRegionBN.Y - 1, y - Position.Y + DimsRegionBN.Y / 2));
int LocalZ = Math.Max(0, Math.Min(DimsRegionBN.Z - 1, z - Position.Z + DimsRegionBN.Z / 2));
Predictions[(z * DimsBN.Y + y) * DimsBN.X + x] = PredictionTiles[ClosestID][(LocalZ * DimsRegionBN.Y + LocalY) * DimsRegionBN.X + LocalX];
}
}
}
volume.FreeDevice();
}
#region Apply Gaussian and find peaks
Image PredictionsImage = new Image(Predictions, new int3(DimsBN));
PredictionsImage.WriteMRC("d_predictions.mrc", true);
//Image PredictionsConvolved = PredictionsImage.AsConvolvedGaussian((float)options.ExpectedDiameter / PixelSizeBN / 6);
//PredictionsConvolved.Multiply(PredictionsImage);
//PredictionsImage.Dispose();
//PredictionsImage.WriteMRC(MatchingDir + RootName + "_boxnet.mrc", PixelSizeBN, true);
int3[] Peaks = PredictionsImage.GetLocalPeaks((int)(diameterAngstrom / PixelSizeBN / 4 + 0.5f), 1e-6f);
PredictionsImage.Dispose();
int BorderDist = (int)(diameterAngstrom / PixelSizeBN * 0.8f + 0.5f);
Peaks = Peaks.Where(p => p.X > BorderDist &&
p.Y > BorderDist &&
p.Z > BorderDist &&
p.X < DimsBN.X - BorderDist &&
p.Y < DimsBN.Y - BorderDist &&
p.Z < DimsBN.Z - BorderDist).ToArray();
#endregion
#region Label connected components and get centroids
List <float3> Centroids;
List <int> Extents;
{
List <List <int3> > Components = new List <List <int3> >();
int[] PixelLabels = Helper.ArrayOfConstant(-1, Predictions.Length);
foreach (var peak in Peaks)
{
if (PixelLabels[DimsBN.ElementFromPosition(peak)] >= 0)
{
continue;
}
List <int3> Component = new List <int3>()
{
peak
};
int CN = Components.Count;
PixelLabels[DimsBN.ElementFromPosition(peak)] = CN;
Queue <int3> Expansion = new Queue <int3>(100);
Expansion.Enqueue(peak);
while (Expansion.Count > 0)
{
int3 pos = Expansion.Dequeue();
int PosElement = (int)DimsBN.ElementFromPosition(pos);
if (pos.X > 0 && Predictions[PosElement - 1] > 0 && PixelLabels[PosElement - 1] < 0)
{
PixelLabels[PosElement - 1] = CN;
Component.Add(pos + new int3(-1, 0, 0));
Expansion.Enqueue(pos + new int3(-1, 0, 0));
}
if (pos.X < DimsBN.X - 1 && Predictions[PosElement + 1] > 0 && PixelLabels[PosElement + 1] < 0)
{
PixelLabels[PosElement + 1] = CN;
Component.Add(pos + new int3(1, 0, 0));
Expansion.Enqueue(pos + new int3(1, 0, 0));
}
if (pos.Y > 0 && Predictions[PosElement - DimsBN.X] > 0 && PixelLabels[PosElement - DimsBN.X] < 0)
{
PixelLabels[PosElement - DimsBN.X] = CN;
Component.Add(pos + new int3(0, -1, 0));
Expansion.Enqueue(pos + new int3(0, -1, 0));
}
if (pos.Y < DimsBN.Y - 1 && Predictions[PosElement + DimsBN.X] > 0 && PixelLabels[PosElement + DimsBN.X] < 0)
{
PixelLabels[PosElement + DimsBN.X] = CN;
Component.Add(pos + new int3(0, 1, 0));
Expansion.Enqueue(pos + new int3(0, 1, 0));
}
if (pos.Z > 0 && Predictions[PosElement - DimsBN.X * DimsBN.Y] > 0 && PixelLabels[PosElement - DimsBN.X * DimsBN.Y] < 0)
{
PixelLabels[PosElement - DimsBN.X * DimsBN.Y] = CN;
Component.Add(pos + new int3(0, 0, -1));
Expansion.Enqueue(pos + new int3(0, 0, -1));
}
if (pos.Z < DimsBN.Z - 1 && Predictions[PosElement + DimsBN.X * DimsBN.Y] > 0 && PixelLabels[PosElement + DimsBN.X * DimsBN.Y] < 0)
{
PixelLabels[PosElement + DimsBN.X * DimsBN.Y] = CN;
Component.Add(pos + new int3(0, 0, 1));
Expansion.Enqueue(pos + new int3(0, 0, 1));
}
}
Components.Add(Component);
}
Centroids = Components.Select(c => MathHelper.Mean(c.Select(v => new float3(v)))).ToList();
Extents = Components.Select(c => c.Count).ToList();
}
List <int> ToDelete = new List <int>();
// Hit test with crap mask
//for (int c1 = 0; c1 < Centroids.Count; c1++)
//{
// float2 P1 = Centroids[c1];
// if (MaskHitTest[(int)P1.Y * DimsBN.X + (int)P1.X] == 0)
// {
// ToDelete.Add(c1);
// continue;
// }
//}
for (int c1 = 0; c1 < Centroids.Count - 1; c1++)
{
float3 P1 = Centroids[c1];
for (int c2 = c1 + 1; c2 < Centroids.Count; c2++)
{
if ((P1 - Centroids[c2]).Length() < diameterAngstrom / PixelSizeBN / 1.5f)
{
int D = Extents[c1] < Extents[c2] ? c1 : c2;
if (!ToDelete.Contains(D))
{
ToDelete.Add(D);
}
}
}
}
ToDelete.Sort();
for (int i = ToDelete.Count - 1; i >= 0; i--)
{
Centroids.RemoveAt(ToDelete[i]);
Extents.RemoveAt(ToDelete[i]);
}
#endregion
//new Image(Predictions, DimsBN).WriteMRC("d_predictions.mrc", true);
#region Write peak positions and angles into table
return(Centroids.Select((c, i) => new float4(c.X, c.Y, c.Z, Extents[i])).ToArray());
#endregion
}