public void TestRegion_ABBCBB()
{
int regionIterationsToLearnExpected = 28;
int repeatSequencesToRun = 20;
this.DefaultSynapseValues();
var imageFiles = new string[]
{
"../CLA.Tests/images/ABBCBBA/image-A.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp",
"../CLA.Tests/images/ABBCBBA/image-C.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp"
//"../CLA.Tests/images/ABBCBBA/image-A.bmp",
};
string outDir = "../CLA.Tests/images/ABBCBBA/";
var data = new int[imageFiles.Length][,];
for (int i = 0; i < imageFiles.Length; ++i)
{
data[i] = this.GetNextPictureFromFile(imageFiles[i]);
}
int[,] sample = this.GetNextPictureFromFile(imageFiles[0]);
var inputSize = new Size(sample.GetLength(0), sample.GetLength(1));
int localityRadius = 0;
int cellsPerCol = 4;
int segmentActiveThreshold = 4;
int newSynapses = 5;
var region = new Region(0, null, inputSize, 1.0f, 1.0f, 1, localityRadius, cellsPerCol,
segmentActiveThreshold, newSynapses);
Global.TemporalLearning = true;
int iters = 0;
for (int iter = 0; iter < repeatSequencesToRun; ++iter)
{
for (int img = 0; img < imageFiles.Length; ++img)
{
iters++;
region.SetInput(data[img]);
region.NextTimeStep();
// Examine region accuracy
float columnActivationAccuracy = region.Statistics.ColumnActivationAccuracy;
float columnPredictionAccuracy = region.Statistics.ColumnPredictionAccuracy;
#if (DEBUG)
Console.Write("\niter=" + iter + " img=" + img + " accuracy: " + columnActivationAccuracy + " " + columnPredictionAccuracy);
Console.Write(" nc: " + region.Statistics.NumberActiveColumns);
#endif
int[,] outData = region.GetPredictingColumnsByTimeStep();
int n1 = 0, n2 = 0, n3 = 0;
foreach (var outVal in outData)
{
n1 += outVal == 1 ? 1 : 0;
n2 += outVal == 2 ? 1 : 0;
n3 += outVal == 3 ? 1 : 0;
}
#if (DEBUG)
Console.Write(" np:" + n1 + " " + n2 + " " + n3);
#endif
this.WritePredictionsToFile(outData, outDir + "prediction-pass-" + iter + "-image-" + (img + 1) + ".bmp");
if (iter > regionIterationsToLearnExpected)
{
// we expect 100% accuracy
Assert.Equal(1.0f, region.Statistics.ColumnActivationAccuracy, 5);
Assert.Equal(1.0f, region.Statistics.ColumnPredictionAccuracy, 5);
}
else
{
//before that we expect 0% accuracy
//Assert.Equal(0.0f, region.ColumnActivationAccuracy, 5);
//Assert.Equal(0.0f, region.ColumnPredictionAccuracy, 5);
}
}
}
}
public void TestRegion_ABBCBB()
{
int regionIterationsToLearnExpected = 28;
int repeatSequencesToRun = 20;
this.DefaultSynapseValues();
var imageFiles = new string[]
{
"../CLA.Tests/images/ABBCBBA/image-A.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp",
"../CLA.Tests/images/ABBCBBA/image-C.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp",
"../CLA.Tests/images/ABBCBBA/image-B.bmp"
//"../CLA.Tests/images/ABBCBBA/image-A.bmp",
};
string outDir = "../CLA.Tests/images/ABBCBBA/";
var data = new int[imageFiles.Length][, ];
for (int i = 0; i < imageFiles.Length; ++i)
{
data[i] = this.GetNextPictureFromFile(imageFiles[i]);
}
int[,] sample = this.GetNextPictureFromFile(imageFiles[0]);
var inputSize = new Size(sample.GetLength(0), sample.GetLength(1));
int localityRadius = 0;
int cellsPerCol = 4;
int segmentActiveThreshold = 4;
int newSynapses = 5;
var region = new Region(0, null, inputSize, 1.0f, 1.0f, 1, localityRadius, cellsPerCol,
segmentActiveThreshold, newSynapses);
Global.TemporalLearning = true;
int iters = 0;
for (int iter = 0; iter < repeatSequencesToRun; ++iter)
{
for (int img = 0; img < imageFiles.Length; ++img)
{
iters++;
region.SetInput(data[img]);
region.NextTimeStep();
// Examine region accuracy
float columnActivationAccuracy = region.Statistics.ColumnActivationAccuracy;
float columnPredictionAccuracy = region.Statistics.ColumnPredictionAccuracy;
#if (DEBUG)
Console.Write("\niter=" + iter + " img=" + img + " accuracy: " + columnActivationAccuracy + " " + columnPredictionAccuracy);
Console.Write(" nc: " + region.Statistics.NumberActiveColumns);
#endif
int[,] outData = region.GetPredictingColumnsByTimeStep();
int n1 = 0, n2 = 0, n3 = 0;
foreach (var outVal in outData)
{
n1 += outVal == 1 ? 1 : 0;
n2 += outVal == 2 ? 1 : 0;
n3 += outVal == 3 ? 1 : 0;
}
#if (DEBUG)
Console.Write(" np:" + n1 + " " + n2 + " " + n3);
#endif
this.WritePredictionsToFile(outData, outDir + "prediction-pass-" + iter + "-image-" + (img + 1) + ".bmp");
if (iter > regionIterationsToLearnExpected)
{
// we expect 100% accuracy
Assert.Equal(1.0f, region.Statistics.ColumnActivationAccuracy, 5);
Assert.Equal(1.0f, region.Statistics.ColumnPredictionAccuracy, 5);
}
else
{
//before that we expect 0% accuracy
//Assert.Equal(0.0f, region.ColumnActivationAccuracy, 5);
//Assert.Equal(0.0f, region.ColumnPredictionAccuracy, 5);
}
}
}
}
public void testRegion_BasicSpatialTemporalPooling()
{
this.DefaultSynapseValues();
var inputSize = new Size(128, 128); //input data is size 128x128
var regionSize = new Size(32, 32); //region's column grid is size 32x32
float pctInputPerCol = 0.01f; //each column connects to 1% random input bits
float pctMinOverlap = 0.07f; //7% of column bits at minimum to be active
int localityRadius = 0; //columns can connect anywhere within input
float pctLocalActivity = 0.5f; //half of columns within radius inhibited
int cellsPerCol = 4;
int segActiveThreshold = 10;
int newSynapseCount = 10;
var region = new Region(0, null, regionSize, pctInputPerCol, pctMinOverlap,
localityRadius, pctLocalActivity, cellsPerCol, segActiveThreshold,
newSynapseCount, true);
Global.SpatialLearning = false;
Global.TemporalLearning = true;
int dataSize = inputSize.Width * inputSize.Height;
var data = new int[inputSize.Width,inputSize.Height];
region.SetInput(data);
int iters = 0;
int accCount = 0;
double accSum = 0.0;
for (int k = 0; k < 10; ++k)
{
for (int i = 0; i < 10; ++i)
{
iters++;
//data will contain a 128x128 bit representation
for (int di = 0; di < inputSize.Width; ++di) //reset all data to 0
{
for (int dj = 0; dj < inputSize.Height; ++dj)
{
data[di, dj] = 0;
}
}
for (int j = 0; j < dataSize / 10; ++j) //assign next 10% set to 1's
{
int index = (i * (dataSize / 10)) + j;
int di = index == 0 ? 0 : index % inputSize.Width;
int dj = index == 0 ? 0 : index / inputSize.Width;
data[di, dj] = 1;
}
region.NextTimeStep();
// Expect 15-25 active columns per time step
Assert.InRange(region.Statistics.NumberActiveColumns, 15, 25);
float columnActivationAccuracy = region.Statistics.ColumnActivationAccuracy;
float columnPredictionAccuracy = region.Statistics.ColumnPredictionAccuracy;
#if (DEBUG)
Console.Write("\niter" + iters + " Acc: " + columnActivationAccuracy + " " + columnPredictionAccuracy);
Console.Write(" nc:" + region.Statistics.NumberActiveColumns);
#endif
//find the max sequence segment count across the Region; should be 1
int maxSeg = 0;
float meanSeg = 0.0f;
float totalSeg = 0;
foreach (var col in region.Columns)
{
foreach (var cell in col.Cells)
{
int nseg = 0;
foreach (var seg in cell.DistalSegments)
{
if (seg.NumberPredictionSteps == 1)
{
nseg++;
}
}
if (nseg > maxSeg)
{
maxSeg = nseg;
}
meanSeg += nseg;
totalSeg += 1;
}
}
meanSeg /= totalSeg;
#if (DEBUG)
Console.Write(" maxSeg: " + maxSeg);
#endif
if (iters > 1)
{
Assert.Equal(maxSeg, 1);
}
if (k > 0 && i > 0)
{
Console.Write(" ");
}
// Get the current column predictions. outData is size 32x32 to match the
// column grid. each value represents whether the column is predicted to
// happen soon. a value of 1 indicates the column is predicted to be active
// in t+1, value of 2 for t+2, etc. value of 0 indicates column is not
// being predicted any time soon.
int[,] outData = region.GetPredictingColumnsByTimeStep();
int n1 = 0, n2 = 0, n3 = 0;
foreach (var outVal in outData)
{
n1 += outVal == 1 ? 1 : 0;
n2 += outVal == 2 ? 1 : 0;
n3 += outVal == 3 ? 1 : 0;
}
#if (DEBUG)
Console.Write(" np:" + n1 + " " + n2 + " " + n3);
#endif
if (k > 1 || (k == 1 && i >= 1))
{
//after 1 full sequence presented, we expect 100% prediction accuracy.
//Activation accuracy may be slightly less than 100% if some columns share
//activation states amongst different inputs (can happen based on random
//initial connections in the spatial pooler).
Assert.InRange(region.Statistics.ColumnActivationAccuracy, 0.9f, 1.0f);
Assert.Equal(1.0f, region.Statistics.ColumnPredictionAccuracy, 5);
accCount += 1;
accSum += region.Statistics.ColumnActivationAccuracy;
//we also expect predicting columns to match previous active columns
//each successive time step we expect farther-out predictions
Assert.InRange(n1, 15, 25);
if (k > 1)
{
Assert.InRange(n2, 15, 25);
if (k > 2)
{
Assert.InRange(n3, 15, 25);
}
}
}
else
{
//before that we expect 0% accuracy and 0 predicting columns
Assert.Equal(0.0f, region.Statistics.ColumnActivationAccuracy, 5);
Assert.Equal(0.0f, region.Statistics.ColumnPredictionAccuracy, 5);
if (k == 0)
{
Assert.Equal(0, n1);
Assert.Equal(0, n2);
Assert.Equal(0, n3);
}
}
}
#if (DEBUG)
Console.Write("\n");
#endif
}
double meanAccuracy = accSum / accCount;
#if (DEBUG)
Console.WriteLine("total iters = " + iters);
Console.WriteLine("meanAcc: " + meanAccuracy);
#endif
Assert.InRange(meanAccuracy, 0.99, 1.0); //at least 99% average activation accuracy
}
public void testRegion_BasicSpatialTemporalPooling()
{
this.DefaultSynapseValues();
var inputSize = new Size(128, 128); //input data is size 128x128
var regionSize = new Size(32, 32); //region's column grid is size 32x32
float pctInputPerCol = 0.01f; //each column connects to 1% random input bits
float pctMinOverlap = 0.07f; //7% of column bits at minimum to be active
int localityRadius = 0; //columns can connect anywhere within input
float pctLocalActivity = 0.5f; //half of columns within radius inhibited
int cellsPerCol = 4;
int segActiveThreshold = 10;
int newSynapseCount = 10;
var region = new Region(0, null, regionSize, pctInputPerCol, pctMinOverlap,
localityRadius, pctLocalActivity, cellsPerCol, segActiveThreshold,
newSynapseCount, true);
Global.SpatialLearning = false;
Global.TemporalLearning = true;
int dataSize = inputSize.Width * inputSize.Height;
var data = new int[inputSize.Width, inputSize.Height];
region.SetInput(data);
int iters = 0;
int accCount = 0;
double accSum = 0.0;
for (int k = 0; k < 10; ++k)
{
for (int i = 0; i < 10; ++i)
{
iters++;
//data will contain a 128x128 bit representation
for (int di = 0; di < inputSize.Width; ++di) //reset all data to 0
{
for (int dj = 0; dj < inputSize.Height; ++dj)
{
data[di, dj] = 0;
}
}
for (int j = 0; j < dataSize / 10; ++j) //assign next 10% set to 1's
{
int index = (i * (dataSize / 10)) + j;
int di = index == 0 ? 0 : index % inputSize.Width;
int dj = index == 0 ? 0 : index / inputSize.Width;
data[di, dj] = 1;
}
region.NextTimeStep();
// Expect 15-25 active columns per time step
Assert.InRange(region.Statistics.NumberActiveColumns, 15, 25);
float columnActivationAccuracy = region.Statistics.ColumnActivationAccuracy;
float columnPredictionAccuracy = region.Statistics.ColumnPredictionAccuracy;
#if (DEBUG)
Console.Write("\niter" + iters + " Acc: " + columnActivationAccuracy + " " + columnPredictionAccuracy);
Console.Write(" nc:" + region.Statistics.NumberActiveColumns);
#endif
//find the max sequence segment count across the Region; should be 1
int maxSeg = 0;
float meanSeg = 0.0f;
float totalSeg = 0;
foreach (var col in region.Columns)
{
foreach (var cell in col.Cells)
{
int nseg = 0;
foreach (var seg in cell.DistalSegments)
{
if (seg.NumberPredictionSteps == 1)
{
nseg++;
}
}
if (nseg > maxSeg)
{
maxSeg = nseg;
}
meanSeg += nseg;
totalSeg += 1;
}
}
meanSeg /= totalSeg;
#if (DEBUG)
Console.Write(" maxSeg: " + maxSeg);
#endif
if (iters > 1)
{
Assert.Equal(maxSeg, 1);
}
if (k > 0 && i > 0)
{
Console.Write(" ");
}
// Get the current column predictions. outData is size 32x32 to match the
// column grid. each value represents whether the column is predicted to
// happen soon. a value of 1 indicates the column is predicted to be active
// in t+1, value of 2 for t+2, etc. value of 0 indicates column is not
// being predicted any time soon.
int[,] outData = region.GetPredictingColumnsByTimeStep();
int n1 = 0, n2 = 0, n3 = 0;
foreach (var outVal in outData)
{
n1 += outVal == 1 ? 1 : 0;
n2 += outVal == 2 ? 1 : 0;
n3 += outVal == 3 ? 1 : 0;
}
#if (DEBUG)
Console.Write(" np:" + n1 + " " + n2 + " " + n3);
#endif
if (k > 1 || (k == 1 && i >= 1))
{
//after 1 full sequence presented, we expect 100% prediction accuracy.
//Activation accuracy may be slightly less than 100% if some columns share
//activation states amongst different inputs (can happen based on random
//initial connections in the spatial pooler).
Assert.InRange(region.Statistics.ColumnActivationAccuracy, 0.9f, 1.0f);
Assert.Equal(1.0f, region.Statistics.ColumnPredictionAccuracy, 5);
accCount += 1;
accSum += region.Statistics.ColumnActivationAccuracy;
//we also expect predicting columns to match previous active columns
//each successive time step we expect farther-out predictions
Assert.InRange(n1, 15, 25);
if (k > 1)
{
Assert.InRange(n2, 15, 25);
if (k > 2)
{
Assert.InRange(n3, 15, 25);
}
}
}
else
{
//before that we expect 0% accuracy and 0 predicting columns
Assert.Equal(0.0f, region.Statistics.ColumnActivationAccuracy, 5);
Assert.Equal(0.0f, region.Statistics.ColumnPredictionAccuracy, 5);
if (k == 0)
{
Assert.Equal(0, n1);
Assert.Equal(0, n2);
Assert.Equal(0, n3);
}
}
}
#if (DEBUG)
Console.Write("\n");
#endif
}
double meanAccuracy = accSum / accCount;
#if (DEBUG)
Console.WriteLine("total iters = " + iters);
Console.WriteLine("meanAcc: " + meanAccuracy);
#endif
Assert.InRange(meanAccuracy, 0.99, 1.0); //at least 99% average activation accuracy
}
