Browse files

python-only modification to support multiple predicted fields. works …

…by ignoring the underlying record sensor, as it seems that _handleSDRClassifierMultiStep fetches the bucketIdx and actual values, so we no longer need it.
  • Loading branch information...
jclevesque committed Oct 19, 2017
1 parent 1aea72a commit 6d8e629d5eaf5dcf38e3d475432a41098a512fd2
Showing with 74 additions and 56 deletions.
  1. +74 −56 src/nupic/frameworks/opf/
@@ -223,7 +223,6 @@ def __init__(self,
self._hasTP = tmEnable
self._hasCL = clEnable

self._classifierInputEncoder = None
self._predictedFieldIdx = None
self._predictedFieldName = None
self._numFields = None
@@ -323,8 +322,9 @@ def setFieldStatistics(self, fieldStats):
def enableInference(self, inferenceArgs=None):
super(HTMPredictionModel, self).enableInference(inferenceArgs)
if inferenceArgs is not None and "predictedField" in inferenceArgs:
# self._getSensorRegion().setParameter("predictedField",
# str(inferenceArgs["predictedField"]))

def enableLearning(self):
@@ -464,7 +464,6 @@ def run(self, inputRecord):
if self.__logger.isEnabledFor(logging.DEBUG):
self.__logger.debug("inputRecord: %r, results: %r" % (inputRecord,

return results

@@ -591,13 +590,36 @@ def _multiStepCompute(self, rawInput):
"TM, SP, or Sensor regions")

inputTSRecordIdx = rawInput.get('_timestampRecordIdx')
return self._handleSDRClassifierMultiStep(

inferenceArgs = self.getInferenceArgs()
predictedField = inferenceArgs.get('predictedField', None)
if predictedField is None:
raise ValueError(
"No predicted field was enabled! Did you call enableInference()?"

if isinstance(predictedField, str):
return self._handleSDRClassifierMultiStep(classifier=self._getClassifierRegion(),
elif isinstance(predictedField, list):
# Now returns a nested structure, with one inference per predictedField
inferences = {}
# Multiple output prediction
for i, pf in enumerate(predictedField):
inf = self._handleSDRClassifierMultiStep(classifier=self._getClassifierRegion(i=i),
inferences[pf] = inf
return inferences

def _classificationCompute(self):
# This is called by the deprecated classification mode
inference = {}
classifier = self._getClassifierRegion()
classifier.setParameter('inferenceMode', True)
@@ -702,7 +724,10 @@ def _anomalyCompute(self):
return inferences

def _handleSDRClassifierMultiStep(self, patternNZ,
def _handleSDRClassifierMultiStep(self,
""" Handle the CLA Classifier compute logic when implementing multi-step
@@ -714,6 +739,9 @@ def _handleSDRClassifierMultiStep(self, patternNZ,
classifier: Classifier with which to compute predictions, there is one
classifier per predictedField.
predictedFieldName: Corresponding predictedField name for given classifier.
patternNZ: The input to the CLA Classifier as a list of active input indices
inputTSRecordIdx: The index of the record as computed from the timestamp
and aggregation interval. This normally increments by 1
@@ -722,32 +750,20 @@ def _handleSDRClassifierMultiStep(self, patternNZ,
rawInput: The raw input to the sensor, as a dict.
inferenceArgs = self.getInferenceArgs()
predictedFieldName = inferenceArgs.get('predictedField', None)
if predictedFieldName is None:
raise ValueError(
"No predicted field was enabled! Did you call enableInference()?"
self._predictedFieldName = predictedFieldName

classifier = self._getClassifierRegion()
classifier._predictedFieldName = predictedFieldName
if not self._hasCL or classifier is None:
# No classifier so return an empty dict for inferences.
return {}

sensor = self._getSensorRegion()
minLikelihoodThreshold = self._minLikelihoodThreshold
maxPredictionsPerStep = self._maxPredictionsPerStep
needLearning = self.isLearningEnabled()
inferences = {}

# Get the classifier input encoder, if we don't have it already
if self._classifierInputEncoder is None:
if predictedFieldName is None:
raise RuntimeError("This experiment description is missing "
"the 'predictedField' in its config, which is required "
"for multi-step prediction inference.")

if getattr(classifier, "_inputEncoder", None) is None:
encoderList = sensor.getSelf().encoder.getEncoderList()
self._numFields = len(encoderList)

@@ -768,15 +784,14 @@ def _handleSDRClassifierMultiStep(self, patternNZ,
encoderList = []
if len(encoderList) >= 1:
fieldNames = sensor.getSelf().disabledEncoder.getScalarNames()
self._classifierInputEncoder = encoderList[fieldNames.index(
classifier._inputEncoder = encoderList[fieldNames.index(
# Legacy multi-step networks don't have a separate encoder for the
# classifier, so use the one that goes into the bottom of the network
encoderList = sensor.getSelf().encoder.getEncoderList()
self._classifierInputEncoder = encoderList[self._predictedFieldIdx]

classifier._inputEncoder = encoderList[self._predictedFieldIdx]

# Get the actual value and the bucket index for this sample. The
# predicted field may not be enabled for input to the network, so we
@@ -787,11 +802,11 @@ def _handleSDRClassifierMultiStep(self, patternNZ,
"field configured for this model. Missing value for '%s'"
% predictedFieldName)
absoluteValue = rawInput[predictedFieldName]
bucketIdx = self._classifierInputEncoder.getBucketIndices(absoluteValue)[0]
bucketIdx = classifier._inputEncoder.getBucketIndices(absoluteValue)[0]

# Convert the absolute values to deltas if necessary
# The bucket index should be handled correctly by the underlying delta encoder
if isinstance(self._classifierInputEncoder, DeltaEncoder):
if isinstance(classifier._inputEncoder, DeltaEncoder):
# Make the delta before any values have been seen 0 so that we do not mess up the
# range for the adaptive scalar encoder.
if not hasattr(self,"_ms_prevVal"):
@@ -876,13 +891,13 @@ def _handleSDRClassifierMultiStep(self, patternNZ,
# calculate likelihood for each bucket
bucketLikelihood = {}
for k in likelihoodsDict.keys():
bucketLikelihood[self._classifierInputEncoder.getBucketIndices(k)[0]] = (
bucketLikelihood[classifier._inputEncoder.getBucketIndices(k)[0]] = (

# ---------------------------------------------------------------------
# If we have a delta encoder, we have to shift our predicted output value
# by the sum of the deltas
if isinstance(self._classifierInputEncoder, DeltaEncoder):
if isinstance(classifier._inputEncoder, DeltaEncoder):
# Get the prediction history for this number of timesteps.
# The prediction history is a store of the previous best predicted values.
# This is used to get the final shift from the current absolute value.
@@ -907,7 +922,7 @@ def _handleSDRClassifierMultiStep(self, patternNZ,
# calculate likelihood for each bucket
bucketLikelihoodOffset = {}
for k in offsetDict.keys():
bucketLikelihoodOffset[self._classifierInputEncoder.getBucketIndices(k)[0]] = (
bucketLikelihoodOffset[classifier._inputEncoder.getBucketIndices(k)[0]] = (

@@ -946,7 +961,6 @@ def _handleSDRClassifierMultiStep(self, patternNZ,
inferences[InferenceElement.multiStepBucketLikelihoods][steps] = (

return inferences

@@ -1049,13 +1063,14 @@ def _getSensorRegion(self):

def _getClassifierRegion(self):
def _getClassifierRegion(self, i=None):
Returns reference to the network's Classifier region
classifierName = "Classifier" + (str(i) if i is not None else "")
if ( is not None and
"Classifier" in
classifierName in
return None

@@ -1179,22 +1194,29 @@ def __createHTMNetwork(self, sensorParams, spEnable, spParams, tmEnable,
if clEnable and clParams is not None:
clParams = clParams.copy()
clRegionName = clParams.pop('regionName')
self.__logger.debug("Adding %s; clParams: %r" % (clRegionName,
n.addRegion("Classifier", "py.%s" % str(clRegionName), json.dumps(clParams))

# SDR Classifier-specific links
if str(clRegionName) == "SDRClassifierRegion":"sensor", "Classifier", "UniformLink", "", srcOutput="actValueOut",
destInput="actValueIn")"sensor", "Classifier", "UniformLink", "", srcOutput="bucketIdxOut",
clPredictedFields = clParams.pop('predictedFields', [""])

for i, pf in enumerate(clPredictedFields):
self.__logger.debug("Adding %s%i; clParams: %r" % (clRegionName, i,
clRawName = "Classifier"
if len(clPredictedFields) > 1:
clRawName += "%i" % (i)
n.addRegion(clRawName, "py.%s" % str(clRegionName), json.dumps(clParams))

# SDR Classifier-specific links
if str(clRegionName) == "SDRClassifierRegion":"sensor", clRawName, "UniformLink", "", srcOutput="actValueOut",
destInput="actValueIn")"sensor", clRawName, "UniformLink", "", srcOutput="bucketIdxOut",

# This applies to all (SDR and KNN) classifiers"sensor", clRawName, "UniformLink", "", srcOutput="categoryOut",
destInput="categoryIn"), clRawName, "UniformLink", "")

# This applies to all (SDR and KNN) classifiers"sensor", "Classifier", "UniformLink", "", srcOutput="categoryOut",
destInput="categoryIn"), "Classifier", "UniformLink", "")

if self.getInferenceType() == InferenceType.TemporalAnomaly:
anomalyClParams = dict(
@@ -1297,10 +1319,6 @@ def __setstate__(self, state):
self.__dict__.pop("_HTMPredictionModel__temporalNetInfo", None)

# This gets filled in during the first infer because it can only be
# determined at run-time
self._classifierInputEncoder = None

if not hasattr(self, '_minLikelihoodThreshold'):
self._minLikelihoodThreshold = DEFAULT_LIKELIHOOD_THRESHOLD

0 comments on commit 6d8e629

Please sign in to comment.