def testSerialization(self):
params = {
'inputDimensions': [2, 4, 5],
'columnDimensions': [4, 3, 3],
'potentialRadius': 30,
'potentialPct': 0.7,
'globalInhibition': False,
'localAreaDensity': 0.23,
'numActiveColumnsPerInhArea': 0,
'stimulusThreshold': 2,
'synPermInactiveDec': 0.02,
'synPermActiveInc': 0.1,
'synPermConnected': 0.12,
'minPctOverlapDutyCycle': 0.011,
'dutyCyclePeriod': 25,
'boostStrength': 11.0,
'seed': 19,
'spVerbosity': 0
}
sp1 = CreateSP("py", params)
sp2 = pickle.loads(pickle.dumps(sp1))
self.compare(sp1, sp2)
sp1 = CreateSP("cpp", params)
sp2 = pickle.loads(pickle.dumps(sp1))
self.compare(sp1, sp2)
def testPermInit(self):
params = {
'inputDimensions' : [2,8,4],
'columnDimensions' : [9,3,6],
'potentialRadius' : 1000,
'potentialPct' : 0.8,
'globalInhibition' : True,
'localAreaDensity' : -1,
'numActiveColumnsPerInhArea' : 3,
'stimulusThreshold' : 0,
'synPermInactiveDec' : 0.001,
'synPermActiveInc' : 0.001,
'synPermConnected' : 0.3,
'minPctOverlapDutyCycle' : 0.001,
'minPctActiveDutyCycle' : 0.001,
'dutyCyclePeriod' : 1000,
'maxBoost' : 1.0,
'seed' : 1956,
'spVerbosity' : 0
}
def max_perm(sp):
numCols = sp.getNumColumns()
numInputs = sp.getNumInputs()
perm = numpy.zeros(numInputs)
max_perm = 0
for col in range(numCols):
sp.getPermanence(col, perm)
column_max = perm.max()
if column_max > max_perm:
max_perm = column_max
return max_perm
pySP = CreateSP("py", params)
cppSP = CreateSP("cpp", params)
self.assertAlmostEqual(max_perm(pySP), max_perm(cppSP))
def runSideBySide(self,
params,
seed=None,
learnMode=None,
convertEveryIteration=False):
"""
Run the PY and CPP implementations side by side on random inputs.
If seed is None a random seed will be chosen based on time, otherwise
the fixed seed will be used.
If learnMode is None learning will be randomly turned on and off.
If it is False or True then set it accordingly.
If convertEveryIteration is True, the CPP will be copied from the PY
instance on every iteration just before each compute.
"""
randomState = getNumpyRandomGenerator(seed)
cppSp = CreateSP("cpp", params)
pySp = CreateSP("py", params)
self.compare(pySp, cppSp)
numColumns = pySp.getNumColumns()
numInputs = pySp.getNumInputs()
threshold = 0.8
inputMatrix = (randomState.rand(numRecords, numInputs) >
threshold).astype(uintType)
# Run side by side for numRecords iterations
for i in range(numRecords):
if learnMode is None:
learn = (randomState.rand() > 0.5)
else:
learn = learnMode
if self.verbosity > 1:
print("Iteration:", i, "learn=", learn)
PyActiveArray = numpy.zeros(numColumns).astype(uintType)
CppActiveArray = numpy.zeros(numColumns).astype(uintType)
inputVector = inputMatrix[i, :]
pySp.compute(inputVector, learn, PyActiveArray)
cppSp.compute(inputVector, learn, CppActiveArray)
self.assertListEqual(list(PyActiveArray), list(CppActiveArray))
self.compare(pySp, cppSp)
# The boost factors were similar enough to get this far.
# Now make them completely equal so that small variations don't cause
# columns to have slightly higher boosted overlaps.
cppBoostFactors = numpy.zeros(numColumns, dtype=realType)
cppSp.getBoostFactors(cppBoostFactors)
pySp.setBoostFactors(cppBoostFactors)
# The permanence values for the two implementations drift ever so slowly
# over time due to numerical precision issues. This occasionally causes
# different permanences to be connected. By transferring the permanence
# values every so often, we can avoid this drift but still check that
# the logic is applied equally for both implementations.
if convertEveryIteration or ((i + 1) % 10 == 0):
convertPermanences(pySp, cppSp)
def basicComputeLoop(self, imp, params, inputSize, columnDimensions,
seed = None):
"""
Feed in some vectors and retrieve outputs. Ensure the right number of
columns win, that we always get binary outputs, and that nothing crashes.
"""
sp = CreateSP(imp,params)
# Create a set of input vectors as well as various numpy vectors we will
# need to retrieve data from the SP
numRecords = 100
randomState = getNumpyRandomGenerator(seed)
inputMatrix = (
randomState.rand(numRecords,inputSize) > 0.8).astype(uintType)
y = numpy.zeros(columnDimensions, dtype = uintType)
dutyCycles = numpy.zeros(columnDimensions, dtype = uintType)
# With learning on we should get the requested number of winners
for v in inputMatrix:
y.fill(0)
sp.compute(v, True, y)
self.assertEqual(sp.getNumActiveColumnsPerInhArea(),y.sum())
self.assertEqual(0,y.min())
self.assertEqual(1,y.max())
# With learning off and some prior training we should get the requested
# number of winners
for v in inputMatrix:
y.fill(0)
sp.compute(v, False, y)
self.assertEqual(sp.getNumActiveColumnsPerInhArea(),y.sum())
self.assertEqual(0,y.min())
self.assertEqual(1,y.max())
def testInhibitColumnsGlobal(self):
params = {
"inputDimensions": [512],
"columnDimensions": [512],
"globalInhibition": True,
"numActiveColumnsPerInhArea": 40,
"seed": 19
}
sp1 = CreateSP("py", params)
sp2 = CreateSP("cpp", params)
for _ in range(100):
overlaps = numpy.random.randint(10, size=512).astype(realType)
columns1 = sp1._inhibitColumns(overlaps)
columns2 = sp2._inhibitColumns(overlaps)
self.assertEqual(set(columns1), set(columns2))
def boostTestLoop(self, imp):
"""Main test loop."""
self.sp = CreateSP(imp, self.params)
self.spImplementation = imp
self.winningIteration.fill(0)
self.lastSDR = {}
self.boostTestPhase1()
self.boostTestPhase2()
self.boostTestPhase3()
self.boostTestPhase4()
def runSerialize(self, imp, params, seed=None):
randomState = getNumpyRandomGenerator(seed)
sp1 = CreateSP(imp, params)
numColumns = sp1.getNumColumns()
numInputs = sp1.getNumInputs()
threshold = 0.8
inputMatrix = (randomState.rand(numRecords, numInputs) >
threshold).astype(uintType)
for i in range(numRecords / 2):
activeArray = numpy.zeros(numColumns).astype(uintType)
inputVector = inputMatrix[i, :]
learn = (randomState.rand() > 0.5)
sp1.compute(inputVector, learn, activeArray)
sp2 = pickle.loads(pickle.dumps(sp1))
for i in range(numRecords / 2 + 1, numRecords):
activeArray1 = numpy.zeros(numColumns).astype(uintType)
activeArray2 = numpy.zeros(numColumns).astype(uintType)
inputVector = inputMatrix[i, :]
learn = (randomState.rand() > 0.5)
sp1.compute(inputVector, learn, activeArray1)
sp2.compute(inputVector, learn, activeArray2)
self.assertListEqual(list(activeArray1), list(activeArray2))