def test_getPotentialSynapsePos(self):
'''
Test the teensorflow overlap calculators getPotentialSynapsePos
function.
'''
potWidth = 3
potHeight = 3
centerPotSynapses = 1
numColumnRows = 3
numColumnCols = 3
connectedPerm = 0.3
minOverlap = 3
wrapInput = 0
numInputRows = 3
numInputCols = 3
numInputs = 1
numPotSyn = potWidth * potHeight
numColumns = numColumnRows * numColumnCols
newInputMat = np.random.randint(2,
size=(numInputs, numInputRows,
numInputCols))
# Create an array representing the permanences of colums synapses
colSynPerm = np.random.rand(numColumns, numPotSyn)
# Create an instance of the overlap calculation class
overlapCalc = tf_overlap.OverlapCalculator(
potWidth, potHeight, numColumnCols, numColumnRows, numInputCols,
numInputRows, centerPotSynapses, connectedPerm, minOverlap,
wrapInput)
#import ipdb; ipdb.set_trace()
columnPotSynPositions = overlapCalc.getPotentialSynapsePos(
numInputCols, numInputRows)
colOverlaps, colPotInputs = overlapCalc.calculateOverlap(
colSynPerm, newInputMat)
print("columnPotSynPositions = \n", columnPotSynPositions)
result = (np.array([[-1., -1., -1., -1., -0., 0., -1., 1., 1.],
[-1., -1., -1., -0., 0., 0., 1., 1., 1.],
[-1., -1., -1., 0., 0., -1., 1., 1., -1.],
[-1., -0., 0., -1., 1., 1., -1., 2., 2.],
[-0., 0., 0., 1., 1., 1., 2., 2., 2.],
[0., 0., -1., 1., 1., -1., 2., 2., -1.],
[-1., 1., 1., -1., 2., 2., -1., -1., -1.],
[1., 1., 1., 2., 2., 2., -1., -1., -1.],
[1., 1., -1., 2., 2., -1., -1., -1., -1.]]),
np.array([[2., 2., 2., 2., 0., 1., 2., 0., 1.],
[2., 2., 2., 0., 1., 2., 0., 1., 2.],
[2., 2., 2., 1., 2., 2., 1., 2., 2.],
[2., 0., 1., 2., 0., 1., 2., 0., 1.],
[0., 1., 2., 0., 1., 2., 0., 1., 2.],
[1., 2., 2., 1., 2., 2., 1., 2., 2.],
[2., 0., 1., 2., 0., 1., 2., 2., 2.],
[0., 1., 2., 0., 1., 2., 2., 2., 2.],
[1., 2., 2., 1., 2., 2., 2., 2., 2.]]))
np.array_equal(columnPotSynPositions, result)
def test_uncenteredCase1(self):
'''
Test the tf overlap calculator with a case where
each column calculates the overlap with that columns
potential synpases begining from the top right. The
potential synpases are not cenetered around the column.
'''
potWidth = 2
potHeight = 2
centerPotSynapses = 0
numColumnRows = 5
numColumnCols = 4
connectedPerm = 0.3
minOverlap = 3
wrapInput = 0
# The below colsynPerm needs to have potWidth * potHeight number of columns
# and needs to have numColumnCols * numColumnRows number of rows.
colSynPerm = np.array([[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0], [1, 1, 1, 1],
[1, 1, 1, 1], [1, 1, 1, 1], [1, 1, 1, 1],
[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0]])
# Needs to have numColumnCols number of columns and
# numColumnRows number of rows for it to be valid.
newInputMat = np.array([[[1, 1, 1, 1], [0, 0, 0, 0], [1, 1, 1, 1],
[0, 0, 0, 0]]])
numInputCols = 4
numInputRows = 4
# Create an instance of the overlap calculation class
overlapCalc = tf_overlap.OverlapCalculator(
potWidth, potHeight, numColumnCols, numColumnRows, numInputCols,
numInputRows, centerPotSynapses, connectedPerm, minOverlap,
wrapInput)
# Return both the overlap values and the inputs from
# the potential synapses to all columns.
colOverlaps, colPotInputs = overlapCalc.calculateOverlap(
colSynPerm, newInputMat)
#import ipdb; ipdb.set_trace()
assert np.sum(colOverlaps) == 0
def setupCalculators(self):
# Setup the theano calculator classes used to calculate
# efficiently the spatial, temporal and sequence pooling.
self.overlapCalc = overlap.OverlapCalculator(
self.potentialWidth, self.potentialHeight, self.width, self.height,
self.inputWidth, self.inputHeight, self.centerPotSynapses,
self.connectPermanence, self.minOverlap, self.wrapInput)
# Get the output tensors and link them to the input tensors of the next calculators
# This creates a continuous tensorflow graph so tensors can flow through it with having
# to be converted back to numpy arrays.
overlapTensor = self.overlapCalc.getOverlapTensor()
potOverlapTensor = self.overlapCalc.getPotOverlapTensor()
self.inhibCalc = inhibition.inhibitionCalculator(
self.width, self.height, self.inhibitionWidth,
self.inhibitionHeight, self.desiredLocalActivity, self.minOverlap,
self.centerPotSynapses, overlapTensor, potOverlapTensor)
def test_inputSizes(self):
'''
Test the tensorflow overlap calculator with a range of input sizes
'''
potWidth = 4
potHeight = 4
centerPotSynapses = 1
numColumnRows = 7
numColumnCols = 5
connectedPerm = 0.3
minOverlap = 3
wrapInput = 0
numPotSyn = potWidth * potHeight
numColumns = numColumnRows * numColumnCols
numInputs = 1
# Create an array representing the permanences of colums synapses
colSynPerm = np.random.rand(numColumns, numPotSyn)
for i in range(4, 100, 3):
numInputRows = i
for j in range(4, 100, 7):
numInputCols = j
print "NEW TEST ROUND"
print "numInputRows, numInputCols = %s, %s " % (numInputRows,
numInputCols)
newInputMat = np.random.randint(2,
size=(numInputs, numInputRows,
numInputCols))
# Create an instance of the overlap calculation class
overlapCalc = tf_overlap.OverlapCalculator(
potWidth, potHeight, numColumnCols, numColumnRows,
numInputCols, numInputRows, centerPotSynapses,
connectedPerm, minOverlap, wrapInput, numInputs)
# Return both the overlap values and the inputs from
# the potential synapses to all columns.
colOverlaps, colPotInputs = overlapCalc.calculateOverlap(
colSynPerm, newInputMat)
columnPotSynPositions = overlapCalc.getPotentialSynapsePos(
numInputCols, numInputRows)
assert len(colOverlaps) == numColumns
def test_minOverlap(self):
'''
Test the tf overlap calculator with a case where their is no
columns with an overlap value larger then the min overlap value.
'''
potWidth = 2
potHeight = 2
centerPotSynapses = 1
numColumnRows = 4
numColumnCols = 4
connectedPerm = 0.3
minOverlap = 3
wrapInput = 0
colSynPerm = np.array([[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0], [1, 1, 1, 1],
[1, 1, 1, 1], [1, 1, 1, 1], [1, 1, 1, 1],
[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 0]])
newInputMat = np.array([[[1, 1, 1, 1], [0, 0, 0, 0], [1, 1, 1, 1],
[0, 0, 0, 0]]])
numInputCols = 4
numInputRows = 4
# Create an instance of the overlap calculation class
overlapCalc = tf_overlap.OverlapCalculator(
potWidth, potHeight, numColumnCols, numColumnRows, numInputCols,
numInputRows, centerPotSynapses, connectedPerm, minOverlap,
wrapInput)
# Return both the overlap values and the inputs from
# the potential synapses to all columns.
colOverlaps, colPotInputs = overlapCalc.calculateOverlap(
colSynPerm, newInputMat)
#import ipdb; ipdb.set_trace()
assert np.sum(colOverlaps) == 0
def test_smallNumCol(self):
'''
Test the tensorflow overlap calculator when there is very few columns.
'''
potWidth = 2
potHeight = 2
centerPotSynapses = 1
numColumnRows = 5
numColumnCols = 5
connectedPerm = 0.3
minOverlap = 0
wrapInput = 0
numInputRows = 2
numInputCols = 2
numInputs = 1
numPotSyn = potWidth * potHeight
numColumns = numColumnRows * numColumnCols
newInputMat = np.random.randint(2,
size=(numInputs, numInputRows,
numInputCols))
print("newInputMat = \n%s" % newInputMat)
# Create an array representing the permanences of colums synapses
colSynPerm = np.ones((numColumns, numPotSyn))
# Create an instance of the overlap calculation class
overlapCalc = tf_overlap.OverlapCalculator(
potWidth, potHeight, numColumnCols, numColumnRows, numInputCols,
numInputRows, centerPotSynapses, connectedPerm, minOverlap,
wrapInput)
#import ipdb; ipdb.set_trace()
#columnPotSynPositions = overlapCalc.getPotentialSynapsePos(numInputCols, numInputRows)
colOverlaps, colPotInputs = overlapCalc.calculateOverlap(
colSynPerm, newInputMat)
print("colOverlaps = \n", colOverlaps)
def test_smallPotSizes(self):
'''
Test the tf overlap calculator with a specific case.
Also test the overlap calculator remove small overlap values
with an edge case.
'''
potWidth = 2
potHeight = 2
centerPotSynapses = 0
numColumnRows = 5
numColumnCols = 4
connectedPerm = 0.3
minOverlap = 2
wrapInput = 0
# The below colsynPerm needs to have potWidth * potHeight number of columns
# and needs to have numColumnCols * numColumnRows number of rows.
colSynPerm = np.array([[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31],
[0.31, 0.31, 0.31, 0.31]])
# Needs to have numColumnCols number of columns and
# numColumnRows number of rows for it to be valid.
newInputMat = np.array([[[0, 0, 1, 0], [0, 0, 1, 0], [0, 0, 1, 0],
[0, 0, 1, 0]]])
numInputCols = 4
numInputRows = 4
# Create an instance of the overlap calculation class
overlapCalc = tf_overlap.OverlapCalculator(
potWidth, potHeight, numColumnCols, numColumnRows, numInputCols,
numInputRows, centerPotSynapses, connectedPerm, minOverlap,
wrapInput)
# Return both the overlap values and the inputs from
# the potential synapses to all columns.
colOverlaps, colPotInputs = overlapCalc.calculateOverlap(
colSynPerm, newInputMat)
print "colOverlaps = ", colOverlaps
#import ipdb; ipdb.set_trace()
# remove the tie breaker values from the overlap scores.
colOverlaps = np.floor(colOverlaps)
result = [0, 2, 2, 0, 0, 2, 2, 0, 0, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0]
assert np.array_equal(colOverlaps, result)
