def __init__(self, numColumns, L2Params, L4Params, L6aParams, repeat, logCalls=False):
"""
Create a network consisting of multiple columns. Each column contains one L2,
one L4 and one L6a layers. In addition all the L2 columns are fully
connected to each other through their lateral inputs.
:param numColumns: Number of columns to create
:type numColumns: int
:param L2Params: constructor parameters for :class:`ColumnPoolerRegion`
:type L2Params: dict
:param L4Params: constructor parameters for :class:`ApicalTMPairRegion`
:type L4Params: dict
:param L6aParams: constructor parameters for :class:`GridCellLocationRegion`
:type L6aParams: dict
:param repeat: Number of times each pair should be seen to be learned
:type repeat: int
:param logCalls: If true, calls to main functions will be logged internally.
The log can then be saved with saveLogs(). This allows us
to recreate the complete network behavior using
rerunExperimentFromLogfile which is very useful for
debugging.
:type logCalls: bool
"""
# Handle logging - this has to be done first
self.logCalls = logCalls
self.numColumns = numColumns
self.repeat = repeat
network = Network()
self.network = createMultipleL246aLocationColumn(network=network,
numberOfColumns=self.numColumns,
L2Params=L2Params,
L4Params=L4Params,
L6aParams=L6aParams)
network.initialize()
self.sensorInput = []
self.motorInput = []
self.L2Regions = []
self.L4Regions = []
self.L6aRegions = []
for i in range(self.numColumns):
col = str(i)
self.sensorInput.append(network.getRegion("sensorInput_" + col))
self.motorInput.append(network.getRegion("motorInput_" + col))
self.L2Regions.append(network.getRegion("L2_" + col))
self.L4Regions.append(network.getRegion("L4_" + col))
self.L6aRegions.append(network.getRegion("L6a_" + col))
if L6aParams is not None and "dimensions" in L6aParams:
self.dimensions = L6aParams["dimensions"]
else:
self.dimensions = 2
self.sdrSize = L2Params["sdrSize"]
# will be populated during training
self.learnedObjects = {}
def testRun(self):
"""
A simple NetworkAPI example with three regions.
///////////////////////////////////////////////////////////////
//
// .------------------.
// | encoder |
// |(RDSEEncoderRegion)|<------ inital input
// | | (INPUT.begin)
// `-------------------'
// | --------------. sp.bottomUpIn
// |/ |
// .-----------------. |
// | sp | |
// | (SPRegion) | |
// | | |
// `-----------------' |
// | ^
// .-----------------. |
// | tm | |
// | (TMRegion) |-----' (tm.bottomUpOut)
// | |
// `-----------------'
//
//////////////////////////////////////////////////////////////////
"""
""" Creating network instance. """
config = """
{network: [
{addRegion: {name: "encoder", type: "RDSEEncoderRegion", params: {size: 1000, sparsity: 0.2, radius: 0.03, seed: 2019, noise: 0.01}, phase: [1]}},
{addRegion: {name: "sp", type: "SPRegion", params: {columnCount: 1000, globalInhibition: true}, phase: [2]}},
{addRegion: {name: "tm", type: "TMRegion", params: {cellsPerColumn: 8, orColumnOutputs: true}, phase: [2]}},
{addLink: {src: "INPUT.begin", dest: "encoder.values", dim: [1]}},
{addLink: {src: "encoder.encoded", dest: "sp.bottomUpIn", mode: "overwrite"}},
{addLink: {src: "sp.bottomUpOut", dest: "tm.bottomUpIn"}},
{addLink: {src: "tm.bottomUpOut", dest: "sp.bottomUpIn"}}
]}"""
net = Network()
net.configure(config)
net.initialize()
# for debugging: print(net.getExecutionMap())
""" Force initial data. """
net.setInputData("begin", np.array([10]))
""" Execute encoder once, the loop (sp and tm) 4 times """
net.run(1, [1]) # execute initial entry (phase 1)
net.run(4, [2]) # execute loop 4 times (phase 2)
# Here is how you access the buffers
sp_input_buffer = np.array(net.getRegion('sp').getInputArray('bottomUpIn'))
tn_output_buffer = np.array(net.getRegion('tm').getOutputArray('bottomUpOut'))
self.assertEqual(sp_input_buffer.size, tn_output_buffer.size)
def _create_network(self, L4Params, L6aParams):
"""
Constructor.
"""
network = Network()
# Create network
network = createL4L6aLocationColumn(network=network,
L4Params=L4Params,
L6aParams=L6aParams,
inverseReadoutResolution=None,
baselineCellsPerAxis=6,
suffix="")
network.initialize()
return network
def testCreateL4L6aLocationColumn(self):
"""
Test 'createL4L6aLocationColumn' by inferring a set of hand crafted objects
"""
scale = []
orientation = []
# Initialize L6a location region with 5 modules varying scale by sqrt(2) and
# 4 different random orientations for each scale
for i in range(5):
for _ in range(4):
angle = np.radians(random.gauss(7.5, 7.5))
orientation.append(random.choice([angle, -angle]))
scale.append(10.0 * (math.sqrt(2)**i))
net = Network()
createL4L6aLocationColumn(net,
L4Params={
"columnCount": NUM_OF_COLUMNS,
"cellsPerColumn": CELLS_PER_COLUMN,
"activationThreshold": 15,
"minThreshold": 15,
"initialPermanence": 1.0,
"implementation": "ApicalTiebreak",
"maxSynapsesPerSegment": -1
},
L6aParams={
"moduleCount": len(scale),
"scale": scale,
"orientation": orientation,
"anchorInputSize": NUM_OF_CELLS,
"activationThreshold": 8,
"initialPermanence": 1.0,
"connectedPermanence": 0.5,
"learningThreshold": 8,
"sampleSize": 10,
"permanenceIncrement": 0.1,
"permanenceDecrement": 0.0,
"bumpOverlapMethod": "probabilistic"
},
inverseReadoutResolution=8)
net.initialize()
L6a = net.getRegion('L6a')
sensor = net.getRegion('sensorInput')
motor = net.getRegion('motorInput')
# Keeps a list of learned objects
learnedRepresentations = defaultdict(list)
# Learn Objects
self._setLearning(net, True)
for objectDescription in OBJECTS:
reset = True
previousLocation = None
L6a.executeCommand("activateRandomLocation")
for iFeature, feature in enumerate(objectDescription["features"]):
# Move the sensor to the center of the object
locationOnObject = np.array([
feature["top"] + feature["height"] / 2.,
feature["left"] + feature["width"] / 2.
])
# Calculate displacement from previous location
if previousLocation is not None:
motor.executeCommand(
'addDataToQueue',
list(locationOnObject - previousLocation))
else:
motor.executeCommand('addDataToQueue', [0, 0])
previousLocation = locationOnObject
# Sense feature at location
sensor.executeCommand('addDataToQueue',
FEATURE_ACTIVE_COLUMNS[feature["name"]],
reset, 0)
net.run(1)
reset = False
# Save learned representations
representation = L6a.getOutputArray("sensoryAssociatedCells")
representation = np.array(representation).nonzero()[0]
learnedRepresentations[(objectDescription["name"],
iFeature)] = representation
# Infer objects
self._setLearning(net, False)
for objectDescription in OBJECTS:
reset = True
previousLocation = None
inferred = False
features = objectDescription["features"]
touchSequence = list(range(len(features)))
random.shuffle(touchSequence)
for iFeature in touchSequence:
feature = features[iFeature]
# Move the sensor to the center of the object
locationOnObject = np.array([
feature["top"] + feature["height"] / 2.,
feature["left"] + feature["width"] / 2.
])
# Calculate displacement from previous location
if previousLocation is not None:
motor.executeCommand('addDataToQueue',
locationOnObject - previousLocation)
else:
motor.executeCommand('addDataToQueue', [0, 0])
previousLocation = locationOnObject
# Sense feature at location
sensor.executeCommand('addDataToQueue',
FEATURE_ACTIVE_COLUMNS[feature["name"]],
reset, 0)
net.run(1)
reset = False
representation = L6a.getOutputArray("sensoryAssociatedCells")
representation = np.array(representation).nonzero()[0]
target_representations = set(
learnedRepresentations[(objectDescription["name"],
iFeature)])
inferred = (set(representation) <= target_representations)
if inferred:
break
self.assertTrue(inferred)
def _createNetwork(inverseReadoutResolution, anchorInputSize, dualPhase=False):
"""
Create a simple network connecting sensor and motor inputs to the location
region. Use :meth:`RawSensor.addDataToQueue` to add sensor input and growth
candidates. Use :meth:`RawValues.addDataToQueue` to add motor input.
::
+----------+
[ sensor* ] --> | | --> [ activeCells ]
[ candidates* ] --> | location | --> [ learnableCells ]
[ motor ] --> | | --> [ sensoryAssociatedCells ]
+----------+
:param inverseReadoutResolution:
Specifies the diameter of the circle of phases in the rhombus encoded by a
bump.
:type inverseReadoutResolution: int
:type anchorInputSize: int
:param anchorInputSize:
The number of input bits in the anchor input.
.. note::
(*) This function will only add the 'sensor' and 'candidates' regions when
'anchorInputSize' is greater than zero. This is useful if you would like to
compute locations ignoring sensor input
.. seealso::
- :py:func:`htmresearch.frameworks.location.path_integration_union_narrowing.createRatModuleFromReadoutResolution`
"""
net = Network()
# Create simple region to pass motor commands as displacement vectors (dx, dy)
net.addRegion("motor", "py.RawValues", json.dumps({"outputWidth": 2}))
if anchorInputSize > 0:
# Create simple region to pass growth candidates
net.addRegion("candidates", "py.RawSensor",
json.dumps({"outputWidth": anchorInputSize}))
# Create simple region to pass sensor input
net.addRegion("sensor", "py.RawSensor",
json.dumps({"outputWidth": anchorInputSize}))
# Initialize region with 5 modules varying scale by sqrt(2) and 4 different
# random orientations for each scale
scale = []
orientation = []
for i in range(5):
for _ in range(4):
angle = np.radians(random.gauss(7.5, 7.5))
orientation.append(random.choice([angle, -angle]))
scale.append(10.0 * (math.sqrt(2)**i))
# Create location region
params = computeRatModuleParametersFromReadoutResolution(
inverseReadoutResolution)
params.update({
"moduleCount": len(scale),
"scale": scale,
"orientation": orientation,
"anchorInputSize": anchorInputSize,
"activationThreshold": 8,
"initialPermanence": 1.0,
"connectedPermanence": 0.5,
"learningThreshold": 8,
"sampleSize": 10,
"permanenceIncrement": 0.1,
"permanenceDecrement": 0.0,
"dualPhase": dualPhase,
"bumpOverlapMethod": "probabilistic"
})
net.addRegion("location", "py.GridCellLocationRegion", json.dumps(params))
if anchorInputSize > 0:
# Link sensor
net.link("sensor",
"location",
"UniformLink",
"",
srcOutput="dataOut",
destInput="anchorInput")
net.link("sensor",
"location",
"UniformLink",
"",
srcOutput="resetOut",
destInput="resetIn")
net.link("candidates",
"location",
"UniformLink",
"",
srcOutput="dataOut",
destInput="anchorGrowthCandidates")
# Link motor input
net.link("motor",
"location",
"UniformLink",
"",
srcOutput="dataOut",
destInput="displacement")
# Initialize network objects
net.initialize()
return net
def main(parameters=default_parameters, argv=None, verbose=True):
if verbose:
import pprint
print("Parameters:")
pprint.pprint(parameters, indent=4)
print("")
# Read the input file.
records = []
with open(_INPUT_FILE_PATH, "r") as fin:
reader = csv.reader(fin)
headers = next(reader)
next(reader)
next(reader)
for record in reader:
records.append(record)
net = Network()
# Make the Encoders. These will convert input data into binary representations.
dateEncoderRegion = net.addRegion(
'dateEncoder', 'DateEncoderRegion',
str(
dict(timeOfDay_width=parameters["enc"]["time"]["timeOfDay"][0],
timeOfDay_radius=parameters["enc"]["time"]["timeOfDay"][1],
weekend_width=parameters["enc"]["time"]["weekend"])))
valueEncoderRegion = net.addRegion(
'valueEncoder', 'RDSEEncoderRegion',
str(
dict(size=parameters["enc"]["value"]["size"],
sparsity=parameters["enc"]["value"]["sparsity"],
resolution=parameters["enc"]["value"]["resolution"])))
# Make the HTM. SpatialPooler & TemporalMemory & associated tools.
spParams = parameters["sp"]
spRegion = net.addRegion(
'sp',
'SPRegion',
str(
dict(
columnCount=spParams['columnCount'],
potentialPct=spParams["potentialPct"],
potentialRadius=0, # 0 is auto assign as inputWith
globalInhibition=True,
localAreaDensity=spParams["localAreaDensity"],
synPermInactiveDec=spParams["synPermInactiveDec"],
synPermActiveInc=spParams["synPermActiveInc"],
synPermConnected=spParams["synPermConnected"],
boostStrength=spParams["boostStrength"],
wrapAround=True)))
tmParams = parameters["tm"]
tmRegion = net.addRegion(
'tm', 'TMRegion',
str(
dict(columnCount=spParams['columnCount'],
cellsPerColumn=tmParams["cellsPerColumn"],
activationThreshold=tmParams["activationThreshold"],
initialPermanence=tmParams["initialPerm"],
connectedPermanence=spParams["synPermConnected"],
minThreshold=tmParams["minThreshold"],
maxNewSynapseCount=tmParams["newSynapseCount"],
permanenceIncrement=tmParams["permanenceInc"],
permanenceDecrement=tmParams["permanenceDec"],
predictedSegmentDecrement=0.0,
maxSegmentsPerCell=tmParams["maxSegmentsPerCell"],
maxSynapsesPerSegment=tmParams["maxSynapsesPerSegment"])))
net.link('dateEncoder', 'sp', '', '', 'encoded', 'bottomUpIn')
net.link('valueEncoder', 'sp', '', '', 'encoded', 'bottomUpIn')
net.link('sp', 'tm', '', '', 'bottomUpOut', 'bottomUpIn')
net.initialize()
# Iterate through every datum in the dataset, record the inputs & outputs.
inputs = []
anomaly = []
for count, record in enumerate(records):
# Convert date string into Python date object.
dateString = datetime.datetime.strptime(record[0], "%m/%d/%y %H:%M")
# Convert data value string into float.
consumption = float(record[1])
inputs.append(consumption)
# Call the encoders to create bit representations for each value. These are SDR objects.
dateEncoderRegion.setParameterInt64('sensedTime',
int(dateString.timestamp()))
valueEncoderRegion.setParameterReal64('sensedValue', consumption)
net.run(1)
anomaly.append(np.array(tmRegion.getOutputArray("anomaly"))[0])
try:
import matplotlib.pyplot as plt
except:
print("WARNING: failed to import matplotlib, plots cannot be shown.")
return
plt.title("Anomaly Score")
plt.xlabel("Time")
plt.ylabel("Power Consumption")
inputs = np.array(inputs) / max(inputs)
plt.plot(
np.arange(len(inputs)),
inputs,
'red',
np.arange(len(inputs)),
anomaly,
'blue',
)
plt.legend(labels=('Input', 'Anomaly Score'))
plt.show()
return