def testActivateCorrectlyPredictiveCells(self):
tm = TemporalMemory()
prevPredictiveCells = set([tm.layer.columns[0].cells[0],
tm.layer.columns[7].cells[13],
tm.layer.columns[32].cells[2],
tm.layer.columns[823].cells[1],
tm.layer.columns[823].cells[3],
tm.layer.columns[1735].cells[16]])
activeColumns = set([tm.layer.columns[32],
tm.layer.columns[47],
tm.layer.columns[823]])
prevMatchingCells = set()
(activeCells,
winnerCells,
predictedColumns,
predictedInactiveCells) = tm.activateCorrectlyPredictiveCells(prevPredictiveCells,
prevMatchingCells,
activeColumns)
self.assertEqual(activeCells, set([tm.layer.columns[32].cells[2],
tm.layer.columns[823].cells[1],
tm.layer.columns[823].cells[3]]))
self.assertEqual(winnerCells, set([tm.layer.columns[32].cells[2],
tm.layer.columns[823].cells[1],
tm.layer.columns[823].cells[3]]))
self.assertEqual(predictedColumns, set([tm.layer.columns[32],
tm.layer.columns[823]]))
self.assertEqual(predictedInactiveCells, set())
def testBurstColumns(self):
tm = TemporalMemory(
layer=Layer(numCellsPerColumn=4),
connectedPermanence=0.50,
minThreshold=1,
seed=42
)
segment1 = tm.layer.columns[0].cells[0].createSegment()
segment1.createSynapse(presynapticCell=tm.layer.columns[5].cells[3], permanence=0.6)
segment1.createSynapse(presynapticCell=tm.layer.columns[9].cells[1], permanence=0.4)
segment1.createSynapse(presynapticCell=tm.layer.columns[119].cells[1], permanence=0.9)
segment2 = tm.layer.columns[0].cells[0].createSegment()
segment2.createSynapse(presynapticCell=tm.layer.columns[12].cells[1], permanence=0.9)
segment2.createSynapse(presynapticCell=tm.layer.columns[0].cells[3], permanence=0.8)
segment3 = tm.layer.columns[0].cells[1].createSegment()
segment3.createSynapse(presynapticCell=tm.layer.columns[183].cells[1], permanence=0.7)
segment4 = tm.layer.columns[27].cells[0].createSegment()
segment4.createSynapse(presynapticCell=tm.layer.columns[121].cells[2], permanence=0.9)
activeColumns = set([tm.layer.columns[0],
tm.layer.columns[1],
tm.layer.columns[26]])
predictedColumns = set([tm.layer.columns[26]])
prevActiveCells = set([tm.layer.columns[5].cells[3],
tm.layer.columns[9].cells[1],
tm.layer.columns[12].cells[1],
tm.layer.columns[183].cells[1]])
prevWinnerCells = set([tm.layer.columns[5].cells[3],
tm.layer.columns[9].cells[1],
tm.layer.columns[12].cells[1],
tm.layer.columns[183].cells[1]])
(activeCells,
winnerCells,
learningSegments) = tm.burstColumns(activeColumns,
predictedColumns,
prevActiveCells,
prevWinnerCells)
# 5 is the new segment was added to winner cell (6) in column 1
segment5 = list(tm.layer.columns[1].cells[2].segments)[0]
self.assertEqual(activeCells, set([tm.layer.columns[0].cells[0],
tm.layer.columns[0].cells[1],
tm.layer.columns[0].cells[2],
tm.layer.columns[0].cells[3],
tm.layer.columns[1].cells[0],
tm.layer.columns[1].cells[1],
tm.layer.columns[1].cells[2],
tm.layer.columns[1].cells[3]]))
self.assertEqual(winnerCells, set([tm.layer.columns[0].cells[0],
tm.layer.columns[1].cells[2]])) # 6 is randomly chosen cell
self.assertEqual(learningSegments, set([segment1, segment5]))
def testPickCellsToLearnOnAvoidDuplicates(self):
tm = TemporalMemory(seed=42)
segment1 = tm.layer.columns[0].cells[0].createSegment()
segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.6)
winnerCells = set([tm.layer.columns[0].cells[23]])
# Ensure that no additional (duplicate) cells were picked
self.assertEqual(tm.pickCellsToLearnOn(2, segment1, winnerCells),
set())
def testAdaptSegmentToMin(self):
tm = TemporalMemory()
segment1 = tm.layer.columns[0].cells[0].createSegment()
synapse1 = segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.1)
tm.adaptSegment(segment1, set(),
tm.permanenceIncrement,
tm.permanenceDecrement)
synapses = segment1.synapses
self.assertFalse(synapse1 in synapses)
def testLearnOnSegments(self):
tm = TemporalMemory(maxNewSynapseCount=2)
segment1 = tm.layer.columns[0].cells[0].createSegment()
synapse1 = segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.6)
synapse2 = segment1.createSynapse(presynapticCell=tm.layer.columns[1].cells[5], permanence=0.4)
synapse3 = segment1.createSynapse(presynapticCell=tm.layer.columns[14].cells[29], permanence=0.9)
segment2 = tm.layer.columns[0].cells[1].createSegment()
synapse4 = segment2.createSynapse(presynapticCell=tm.layer.columns[22].cells[29], permanence=0.7)
segment3 = tm.layer.columns[0].cells[8].createSegment()
synapse5 = segment3.createSynapse(presynapticCell=tm.layer.columns[15].cells[6], permanence=0.9)
segment4 = tm.layer.columns[3].cells[4].createSegment()
prevActiveSegments = set([segment1, segment3])
learningSegments = set([segment2, segment4])
prevActiveCells = set([tm.layer.columns[0].cells[23],
tm.layer.columns[1].cells[5],
tm.layer.columns[22].cells[29]])
winnerCells = set([tm.layer.columns[0].cells[0]])
prevWinnerCells = set([tm.layer.columns[0].cells[10],
tm.layer.columns[0].cells[11],
tm.layer.columns[0].cells[12],
tm.layer.columns[0].cells[13],
tm.layer.columns[0].cells[14]])
predictedInactiveCells = set()
prevMatchingSegments = set()
tm.learnOnSegments(prevActiveSegments,
learningSegments,
prevActiveCells,
winnerCells,
prevWinnerCells,
predictedInactiveCells,
prevMatchingSegments)
# Check segment 1
self.assertAlmostEqual(synapse1.permanence, 0.7)
self.assertAlmostEqual(synapse2.permanence, 0.5)
self.assertAlmostEqual(synapse3.permanence, 0.8)
# Check segment 2
self.assertAlmostEqual(synapse4.permanence, 0.8)
self.assertEqual(len(segment2.synapses), 2)
# Check segment 3
self.assertAlmostEqual(synapse5.permanence, 0.9)
self.assertEqual(len(segment3.synapses), 1)
# Check segment 4
self.assertEqual(len(segment4.synapses), 2)
def testLeastUsedCell(self):
tm = TemporalMemory(
layer=Layer(columnDimensions=[2], numCellsPerColumn=2),
seed=42
)
segment1 = tm.layer.columns[0].cells[0].createSegment()
segment1.createSynapse(presynapticCell=tm.layer.columns[1].cells[1], permanence=0.3)
for _ in range(100):
# Never pick cell 0, always pick cell 1
self.assertEqual(tm.leastUsedCell(tm.layer.columns[0].cells),
tm.layer.columns[0].cells[1])
def testAdaptSegmentToMax(self):
tm = TemporalMemory()
segment1 = tm.layer.columns[0].cells[0].createSegment()
synapse1 = segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.9)
tm.adaptSegment(segment1, set([synapse1]),
tm.permanenceIncrement,
tm.permanenceDecrement)
self.assertAlmostEqual(synapse1.permanence, 1.0)
# Now permanence should be at max
tm.adaptSegment(segment1, set([synapse1]),
tm.permanenceIncrement,
tm.permanenceDecrement)
self.assertAlmostEqual(synapse1.permanence, 1.0)
def testAdaptSegment(self):
tm = TemporalMemory(
layer=Layer(columnDimensions=[2048], numCellsPerColumn=32)
)
segment1 = tm.layer.columns[0].cells[0].createSegment()
synapse1 = segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.6)
synapse2 = segment1.createSynapse(presynapticCell=tm.layer.columns[1].cells[5], permanence=0.4)
synapse3 = segment1.createSynapse(presynapticCell=tm.layer.columns[14].cells[29], permanence=0.9)
tm.adaptSegment(segment1, set([synapse1, synapse2]),
tm.permanenceIncrement,
tm.permanenceDecrement)
self.assertAlmostEqual(synapse1.permanence, 0.7)
self.assertAlmostEqual(synapse2.permanence, 0.5)
self.assertAlmostEqual(synapse3.permanence, 0.8)
def testBestMatchingCellFewestSegments(self):
tm = TemporalMemory(
layer=Layer(columnDimensions=[2], numCellsPerColumn=2),
connectedPermanence=0.50,
minThreshold=1,
seed=42
)
segment1 = tm.layer.columns[0].cells[0].createSegment()
segment1.createSynapse(presynapticCell=tm.layer.columns[1].cells[1], permanence=0.3)
activeSynapsesForSegment = set([])
for _ in range(100):
# Never pick cell 0, always pick cell 1
(cell, _) = tm.bestMatchingCell(tm.layer.columns[0].cells,
activeSynapsesForSegment)
self.assertEqual(cell, tm.layer.columns[0].cells[1])
def testBurstColumnsEmpty(self):
tm = TemporalMemory()
activeColumns = set()
predictedColumns = set()
prevActiveCells = set()
prevWinnerCells = set()
(activeCells,
winnerCells,
learningSegments) = tm.burstColumns(activeColumns,
predictedColumns,
prevActiveCells,
prevWinnerCells)
self.assertEqual(activeCells, set())
self.assertEqual(winnerCells, set())
self.assertEqual(learningSegments, set())
def testBestMatchingSegment(self):
tm = TemporalMemory(
layer=Layer(columnDimensions=[2048], numCellsPerColumn=32),
connectedPermanence=0.50,
minThreshold=1
)
segment1 = tm.layer.columns[0].cells[0].createSegment()
segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.6)
segment1.createSynapse(presynapticCell=tm.layer.columns[1].cells[5], permanence=0.4)
segment1.createSynapse(presynapticCell=tm.layer.columns[14].cells[29], permanence=0.9)
segment2 = tm.layer.columns[0].cells[0].createSegment()
segment2.createSynapse(presynapticCell=tm.layer.columns[1].cells[17], permanence=0.9)
segment2.createSynapse(presynapticCell=tm.layer.columns[0].cells[3], permanence=0.8)
segment3 = tm.layer.columns[0].cells[1].createSegment()
segment3.createSynapse(presynapticCell=tm.layer.columns[22].cells[29], permanence=0.7)
segment4 = tm.layer.columns[0].cells[8].createSegment()
segment4.createSynapse(presynapticCell=tm.layer.columns[15].cells[6], permanence=0.9)
activeCells = set([tm.layer.columns[0].cells[23],
tm.layer.columns[1].cells[5],
tm.layer.columns[1].cells[17],
tm.layer.columns[22].cells[29]])
self.assertEqual(tm.bestMatchingSegment(tm.layer.columns[0].cells[0],
activeCells),
(segment1, 2))
self.assertEqual(tm.bestMatchingSegment(tm.layer.columns[0].cells[1],
activeCells),
(segment3, 1))
self.assertEqual(tm.bestMatchingSegment(tm.layer.columns[0].cells[8],
activeCells),
(None, None))
self.assertEqual(tm.bestMatchingSegment(tm.layer.columns[3].cells[4],
activeCells),
(None, None))
def testPickCellsToLearnOn(self):
tm = TemporalMemory(seed=42)
segment1 = tm.layer.columns[0].cells[0].createSegment()
winnerCells = set([tm.layer.columns[0].cells[4],
tm.layer.columns[1].cells[15],
tm.layer.columns[1].cells[26],
tm.layer.columns[2].cells[29]])
self.assertEqual(tm.pickCellsToLearnOn(2, segment1, winnerCells),
set([tm.layer.columns[0].cells[4],
tm.layer.columns[1].cells[26]])) # randomly picked
self.assertEqual(tm.pickCellsToLearnOn(100, segment1, winnerCells),
set([tm.layer.columns[0].cells[4],
tm.layer.columns[1].cells[15],
tm.layer.columns[1].cells[26],
tm.layer.columns[2].cells[29]]))
self.assertEqual(tm.pickCellsToLearnOn(0, segment1, winnerCells),
set())
def testBestMatchingCell(self):
tm = TemporalMemory(
connectedPermanence=0.50,
minThreshold=1,
seed=42
)
segment1 = tm.layer.columns[0].cells[0].createSegment()
segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.6)
segment1.createSynapse(presynapticCell=tm.layer.columns[1].cells[5], permanence=0.4)
segment1.createSynapse(presynapticCell=tm.layer.columns[14].cells[29], permanence=0.9)
segment2 = tm.layer.columns[0].cells[0].createSegment()
segment2.createSynapse(presynapticCell=tm.layer.columns[1].cells[17], permanence=0.9)
segment2.createSynapse(presynapticCell=tm.layer.columns[0].cells[3], permanence=0.8)
segment3 = tm.layer.columns[0].cells[1].createSegment()
segment3.createSynapse(presynapticCell=tm.layer.columns[22].cells[29], permanence=0.7)
segment4 = tm.layer.columns[3].cells[12].createSegment()
segment4.createSynapse(presynapticCell=tm.layer.columns[15].cells[6], permanence=0.9)
activeCells = set([tm.layer.columns[0].cells[23],
tm.layer.columns[1].cells[5],
tm.layer.columns[1].cells[17],
tm.layer.columns[22].cells[29]])
self.assertEqual(tm.bestMatchingCell(tm.layer.columns[0].cells,
activeCells),
(tm.layer.columns[0].cells[0], segment1))
self.assertEqual(tm.bestMatchingCell(tm.layer.columns[3].cells, # column containing cell 108
activeCells),
(tm.layer.columns[3].cells[20], None)) # Random cell from column
self.assertEqual(tm.bestMatchingCell(tm.layer.columns[999].cells,
activeCells),
(tm.layer.columns[999].cells[19], None)) # Random cell from column
def testComputePredictiveCells(self):
tm = TemporalMemory(
layer=Layer(columnDimensions=[2048], numCellsPerColumn=32),
activationThreshold=2,
minThreshold=2,
predictedSegmentDecrement=0.004)
segment1 = tm.layer.columns[0].cells[0].createSegment()
segment1.createSynapse(presynapticCell=tm.layer.columns[0].cells[23], permanence=0.6)
segment1.createSynapse(presynapticCell=tm.layer.columns[1].cells[5], permanence=0.5)
segment1.createSynapse(presynapticCell=tm.layer.columns[14].cells[29], permanence=0.9)
segment2 = tm.layer.columns[0].cells[1].createSegment()
segment2.createSynapse(presynapticCell=tm.layer.columns[22].cells[29], permanence=0.7)
segment2.createSynapse(presynapticCell=tm.layer.columns[22].cells[29], permanence=0.4)
segment3 = tm.layer.columns[0].cells[1].createSegment()
segment3.createSynapse(presynapticCell=tm.layer.columns[30].cells[14], permanence=0.9)
segment4 = tm.layer.columns[0].cells[8].createSegment()
segment4.createSynapse(presynapticCell=tm.layer.columns[15].cells[6], permanence=0.9)
segment5 = tm.layer.columns[3].cells[4].createSegment()
activeCells = set([tm.layer.columns[0].cells[23],
tm.layer.columns[1].cells[5],
tm.layer.columns[22].cells[29],
tm.layer.columns[30].cells[14]])
(activeSegments,
predictiveCells,
matchingSegments,
matchingCells) = tm.computePredictiveCells(activeCells)
self.assertEqual(activeSegments, set([segment1]))
self.assertEqual(predictiveCells, set([tm.layer.columns[0].cells[0]]))
self.assertEqual(matchingSegments, set([segment1, segment2]))
self.assertEqual(matchingCells, set([tm.layer.columns[0].cells[0],
tm.layer.columns[0].cells[1]]))
def testActivateCorrectlyPredictiveCellsOrphan(self):
tm = TemporalMemory()
tm.predictedSegmentDecrement = 0.001
prevPredictiveCells = set([])
activeColumns = set([tm.layer.columns[32],
tm.layer.columns[47],
tm.layer.columns[823]])
prevMatchingCells = set([tm.layer.columns[1].cells[0],
tm.layer.columns[1].cells[15]])
(activeCells,
winnerCells,
predictedColumns,
predictedInactiveCells) = tm.activateCorrectlyPredictiveCells(prevPredictiveCells,
prevMatchingCells,
activeColumns)
self.assertEqual(activeCells, set([]))
self.assertEqual(winnerCells, set([]))
self.assertEqual(predictedColumns, set([]))
self.assertEqual(predictedInactiveCells, set([tm.layer.columns[1].cells[0],
tm.layer.columns[1].cells[15]]))
