def __init__(self,
# Constructor arguments for FDRSpatial and FDRTemporal are
# picked up automatically. There is no need to add them anywhere
# in CLARegion, unless you need to do something special with them.
# See docstring above.
# These args are used by CLARegion only or need special handling
disableSpatial=False,
disableTemporal=False,
orColumnOutputs=False,
nCellsPerCol=1,
trainingStep = 'temporal',
cellsSavePath='',
statelessMode=False,
storeDenseOutput=False, #DEPRECATED
outputCloningWidth=0,
outputCloningHeight=0,
saveMasterCoincImages = 0,
temporalImp='py', #'py', 'simple' or 'cpp'
spatialImp=getDefaultSPImp(), #'py', 'cpp', or 'oldpy'
computeTopDown = 0,
nMultiStepPrediction = 0,
# We have separate seeds for spatial and temporal
spSeed=-1,
tpSeed=-1,
# Needed for vision framework
bottomUpOut=None,
**kwargs):
#if disableSpatial and disableTemporal:
# raise RuntimeError("Disable both the spatial and temporal components? "
# "That would make it too easy.")
# Make sure our tuple arguments are integers
for name in ['coincidencesShape', 'inputShape']:
if name in kwargs:
(height, width) = kwargs[name]
kwargs[name] = (int(height), int(width))
# Which FDR Temporal implementation?
FDRCSpatialClass = getSPClass(spatialImp)
FDRTemporalClass = _getTPClass(temporalImp)
# Pull out the spatial and temporal arguments automatically
# These calls whittle down kwargs and create instance variables of CLARegion
sArgTuples = _buildArgs(FDRCSpatialClass.__init__, self, kwargs)
tArgTuples = _buildArgs(FDRTemporalClass.__init__, self, kwargs)
# Make a list of automatic spatial and temporal arg names for later use
self._spatialArgNames = [t[0] for t in sArgTuples]
self._temporalArgNames = [t[0] for t in tArgTuples]
# Start out in stage learn
self.learningMode = True
self.inferenceMode = False
PyRegion.__init__(self, **kwargs)
# Initialize all non-persistent base members, as well as give
# derived class an opportunity to do the same.
self._loaded = False
self._initialize()
# Debugging support, used in _conditionalBreak
self.breakPdb = False
self.breakKomodo = False
# CLARegion only, or special handling
self.disableSpatial = disableSpatial
self.saveMasterCoincImages = saveMasterCoincImages
self.disableTemporal = disableTemporal
self.orColumnOutputs = orColumnOutputs
self.nCellsPerCol = nCellsPerCol # Modified in initInNetwork
self.coincidenceCount = self.coincidencesShape[0] * self.coincidencesShape[1]
self.temporalImp = temporalImp
self.spatialImp = spatialImp
self.computeTopDown = computeTopDown
self.nMultiStepPrediction = nMultiStepPrediction
# Handle -1 for cloning sizes, which essentially just means no cloning...
# ...also handle 0, since that's the new default...
if outputCloningWidth in (0, -1):
outputCloningWidth = self.coincidencesShape[1]
if outputCloningHeight in (0, -1):
outputCloningHeight = self.coincidencesShape[0]
self.outputCloningWidth = outputCloningWidth
self.outputCloningHeight = outputCloningHeight
# Make the clone map, which is used by both spatial and temporal components.
self._cloneMap, self._numCloneMasters = self.makeCloneMap(
self.coincidencesShape, outputCloningWidth, outputCloningHeight
)
# Both FDRCSpatial and FDRTemporal
self.tpSeed = tpSeed
self.spSeed = spSeed
self.trainingStep = trainingStep
self.logPathSPInput = ''
self.logPathSP = ''
self.logPathSPDense = ''
self.logPathTP = ''
# Used to save TP cells data structure to auxiliary file
self.cellsSavePath = cellsSavePath
# Instructs node to ignore past temporal state when operating in
# inference mode; i.e., tells node to ignore the actual resetSignal
# input and treat it as if the resetSignal was *always* set (in
# inference mode only)
self.statelessMode = statelessMode
self._hasRunInference = False
# Variables set up in initInNetwork()
self._sfdr = None # FDRCSpatial instance
self._tfdr = None # FDRTemporal instance
self._numOutputs = None # Number of outputs allocated per node
self._spatialPoolerOutput = None # Hang on to this output for debugging
self._tpSeqOutput = None # Hang on to this for supporting the
# tpSeqOutputNonZeros parameter
self._spatialPoolerInput = None # Hang on to this for supporting the
# spInputNonZeros parameter
self._rfOutput = None # Hang on to this for supporting the
# tpOutputNonZeros parameter
# Read-only node parameters
self.activeOutputCount = None
self.cppOutput = None
self.file = None
# For inspector usage
#from dbgp.client import brk; brk(port=9019)
self._spatialSpec, self._temporalSpec, self._otherSpec = \
_getAdditionalSpecs(spatialImp=self.spatialImp, temporalImp=self.temporalImp)
def _getAdditionalSpecs(spatialImp, temporalImp, kwargs={}):
"""Build the additional specs in three groups (for the inspector)
Use the type of the default argument to set the Spec type, defaulting
to 'Byte' for None and complex types
Determines the spatial parameters based on the selected implementation.
It defaults to FDRCSpatial.
Determines the temporal parameters based on the temporalImp
"""
typeNames = {int: 'UInt32', float: 'Real32', str: 'Byte', bool: 'bool', tuple: 'tuple'}
def getArgType(arg):
t = typeNames.get(type(arg), 'Byte')
count = 0 if t == 'Byte' else 1
if t == 'tuple':
t = typeNames.get(type(arg[0]), 'Byte')
count = len(arg)
if t == 'bool':
t = 'UInt32'
return (t, count)
def getConstraints(arg):
t = typeNames.get(type(arg), 'Byte')
if t == 'Byte':
return 'multiple'
elif t == 'bool':
return 'bool'
else:
return ''
FDRSpatialClass = getSPClass(spatialImp)
FDRTemporalClass = _getTPClass(temporalImp)
spatialSpec = {}
sArgTuples = _buildArgs(FDRSpatialClass.__init__)
tArgTuples = _buildArgs(FDRTemporalClass.__init__)
for argTuple in sArgTuples:
d = dict(
description=argTuple[1],
accessMode='ReadWrite',
dataType=getArgType(argTuple[2])[0],
count=getArgType(argTuple[2])[1],
constraints=getConstraints(argTuple[2]))
spatialSpec[argTuple[0]] = d
temporalSpec = {}
for argTuple in tArgTuples:
d = dict(
description=argTuple[1],
accessMode='ReadWrite',
dataType=getArgType(argTuple[2])[0],
count=getArgType(argTuple[2])[1],
constraints=getConstraints(argTuple[2]))
temporalSpec[argTuple[0]] = d
# Add special parameters that weren't handled automatically
# Spatial parameters only!
spatialSpec.update(dict(
disableSpatial=dict(
description='Disable the spatial FDR (become a temporal nregion)',
accessMode='ReadWrite',
dataType='UInt32',
count=1,
constraints='bool'),
coincidenceCount=dict(
description='Total number of coincidences.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
spInputNonZeros=dict(
description='The indices of the non-zero inputs to the spatial pooler',
accessMode='Read',
dataType='UInt32',
count=0,
constraints=''),
spOutputNonZeros=dict(
description='The indices of the non-zero outputs from the spatial pooler',
accessMode='Read',
dataType='UInt32',
count=0,
constraints=''),
spOverlapDistribution=dict(
description="""The overlaps between the active output coincidences
and the input. The overlap amounts for each coincidence are sorted
from highest to lowest. """,
accessMode='Read',
dataType='Real32',
count=0,
constraints=''),
sparseCoincidenceMatrix=dict(
description='The coincidences, as a SparseMatrix',
accessMode='Read',
dataType='Byte',
count=0,
constraints=''),
spatialPoolerInput=dict(
description='Input to the spatial pooler.',
accessMode='Read',
dataType='Real32',
count=0,
constraints=''),
spatialPoolerOutput=dict(
description='Output of the spatial pooler.',
accessMode='Read',
dataType='Real32',
count=0,
constraints=''),
spNumActiveOutputs=dict(
description='Number of active spatial pooler outputs',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
denseOutput=dict(
description='Score for each coincidence.',
accessMode='Read',
dataType='Real32',
count=0,
constraints=''),
outputCloningWidth=dict(
description="""The number of columns you'd have to move horizontally
(or vertically) before you get back to the same the same
clone that you started with.""",
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
outputCloningHeight=dict(
description="""If non-negative, can be used to make rectangular
(instead of square) cloning fields.""",
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
saveMasterCoincImages=dict(
description="""If non-zero, saves an image of the master coincidences
every N iterations.""",
accessMode='Read',
dataType='UInt32',
count=1,
constraints='',
defaultValue=0),
spLearningStatsStr=dict(
description="""String representation of dictionary containing a number
of statistics related to learning.""",
accessMode='Read',
dataType='Byte',
count=0,
constraints='handle'),
))
# Add temporal parameters that weren't handled automatically
temporalSpec.update(dict(
disableTemporal=dict(
description='Disable the temporal FDR (become a spatial region)',
accessMode='Read',
dataType='UInt32',
count=1,
constraints='bool'),
orColumnOutputs=dict(
description="""OR together the cell outputs from each column to produce
the temporal pooler output. When this mode is enabled, the number of
cells per column must also be specified (via the 'nCellsPerCol'
creation parameter) and the output size of the node should be set the
same as the number of spatial pooler coincidences
('coincidenceCount')""",
accessMode='Read',
dataType='UInt32',
count=1,
constraints='bool'),
nCellsPerCol=dict(
description="""The number of cells, or states, allocated per column
(the number of columns is the same as the number of coincidences).""",
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
maxSegmentsPerCell=dict(
description='Max allowed number of segments per cellf.',
accessMode='ReadWrite',
dataType='Int32',
count=1,
constraints=''),
maxSynapsesPerSegment=dict(
description='Max allowed number of segments per cellf.',
accessMode='ReadWrite',
dataType='Int32',
count=1,
constraints=''),
tpOutputNonZeros=dict(
description='The indices of the non-zero outputs from the temporal pooler',
accessMode='Read',
dataType='UInt32',
count=0,
constraints=''),
tpNumSynapses=dict(
description='Total number of synapses learned in the temporal pooler.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
tpNumSynapsesPerSegmentMax=dict(
description='Max number of synapses found in any one dendritic segment of the temporal pooler.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
tpNumSynapsesPerSegmentAvg=dict(
description='Average number of synapses learned per dendritic segment.',
accessMode='Read',
dataType='Real32',
count=1,
constraints=''),
tpNumSegments=dict(
description='Total number of dendritic segments present in the temporal pooler.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
tpNumCells=dict(
description='Total number of cells present in the temporal pooler.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
tpNumActiveCells=dict(
description='Total number of active cells in the temporal pooler output.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
tpNumActiveColumns=dict(
description='Total number of active columns in the temporal pooler output.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
tpNumBurstingColumns=dict(
description='Total number of bursting columns in the temporal pooler output.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
))
# The last group is for parameters that aren't strictly spatial or temporal
otherSpec = dict(
learningMode=dict(
description='1 if the node is learning (default 1).',
accessMode='ReadWrite',
dataType='UInt32',
count=1,
constraints='bool'),
trainingStep=dict(
description="""Controls which pooler is trained when learningMode is true.
Set to 'spatial' to train spatial pooler, 'temporal'
to train temporal pooler, or 'both' to train both
simultaneously.""",
accessMode='ReadWrite',
dataType='Byte',
count=0,
constraints='enum: spatial,temporal,both'),
inferenceMode=dict(
description='1 if the node is inferring (default 0).',
accessMode='ReadWrite',
dataType='UInt32',
count=1,
constraints='bool'),
activeOutputCount=dict(
description='Number of active elements in bottomUpOut output.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
spSeed=dict(
description='Random seed for the spatial pooler.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
tpSeed=dict(
description='Random seed for the temporal pooler.',
accessMode='Read',
dataType='UInt32',
count=1,
constraints=''),
logPathSPInput=dict(
description='Optional name of spatial pooler input log file.',
accessMode='ReadWrite',
dataType='Byte',
count=0,
constraints=''),
logPathSP=dict(
description='Optional name of spatial pooler output log file.',
accessMode='ReadWrite',
dataType='Byte',
count=0,
constraints=''),
logPathSPDense=dict(
description='Optional name of spatial pooler dense output log file.',
accessMode='ReadWrite',
dataType='Byte',
count=0,
constraints=''),
logPathTP=dict(
description='Optional name of temporal pooler output log file.',
accessMode='ReadWrite',
dataType='Byte',
count=0,
constraints=''),
cellsSavePath=dict(
description="""Optional path to file in which large temporal pooler cells
data structure is to be saved.""",
accessMode='ReadWrite',
dataType='Byte',
count=0,
constraints=''),
statelessMode=dict(
description='Ignores temporal state when in inference mode.',
accessMode='ReadWrite',
dataType='UInt32',
count=1,
constraints='bool'),
temporalImp=dict(
description="""Which temporal pooler implementation to use. Set to either
'simple', 'py' or 'cpp'. The 'simple' implementation supports only 1 cell
per column and is typically only used for vision spatial invariance
applications. The 'cpp' implementation is optimized for speed in C++.
The 'trivial' implementation makes random or zeroth order predictions.""",
accessMode='ReadWrite',
dataType='Byte',
count=0,
constraints='enum: simple, full, cpp, trivial'),
computeTopDown=dict(
description='1 to compute and output the topDownOut output',
accessMode='ReadWrite',
dataType='UInt32',
count=1,
constraints='bool'),
nMultiStepPrediction=dict(
description="""The number of time steps to be predicted in future.""",
accessMode='ReadWrite',
dataType='UInt32',
count=1,
constraints=''),
)
return spatialSpec, temporalSpec, otherSpec
