def learn(self, inputPattern, inputCategory, partitionId=None, isSparse=0,
rowID=None):
"""
Train the classifier to associate specified input pattern with a
particular category.
:param inputPattern: (list) The pattern to be assigned a category. If
isSparse is 0, this should be a dense array (both ON and OFF bits
present). Otherwise, if isSparse > 0, this should be a list of the
indices of the non-zero bits in sorted order
:param inputCategory: (int) The category to be associated to the training
pattern
:param partitionId: (int) partitionID allows you to associate an id with each
input vector. It can be used to associate input patterns stored in the
classifier with an external id. This can be useful for debugging or
visualizing. Another use case is to ignore vectors with a specific id
during inference (see description of infer() for details). There can be
at most one partitionId per stored pattern (i.e. if two patterns are
within distThreshold, only the first partitionId will be stored). This
is an optional parameter.
:param isSparse: (int) If 0, the input pattern is a dense representation. If
isSparse > 0, the input pattern is a list of non-zero indices and
isSparse is the length of the dense representation
:param rowID: (int) UNKNOWN
:returns: The number of patterns currently stored in the classifier
"""
if self.verbosity >= 1:
print("%s learn:" % g_debugPrefix)
print(" category:", int(inputCategory))
print(" active inputs:", _labeledInput(inputPattern,
cellsPerCol=self.cellsPerCol))
if isSparse > 0:
assert all(inputPattern[i] <= inputPattern[i+1]
for i in range(len(inputPattern)-1)), \
"Sparse inputPattern must be sorted."
assert all(bit < isSparse for bit in inputPattern), \
("Sparse inputPattern must not index outside the dense "
"representation's bounds.")
if rowID is None:
rowID = self._iterationIdx
# Dense vectors
if not self.useSparseMemory:
# Not supported
assert self.cellsPerCol == 0, "not implemented for dense vectors"
# If the input was given in sparse form, convert it to dense
if isSparse > 0:
denseInput = numpy.zeros(isSparse)
denseInput[inputPattern] = 1.0
inputPattern = denseInput
if self._specificIndexTraining and not self._nextTrainingIndices:
# Specific index mode without any index provided - skip training
return self._numPatterns
if self._Memory is None:
# Initialize memory with 100 rows and numPatterns = 0
inputWidth = len(inputPattern)
self._Memory = numpy.zeros((100,inputWidth))
self._numPatterns = 0
self._M = self._Memory[:self._numPatterns]
addRow = True
if self._vt is not None:
# Compute projection
inputPattern = numpy.dot(self._vt, inputPattern - self._mean)
if self.distThreshold > 0:
# Check if input is too close to an existing input to be accepted
dist = self._calcDistance(inputPattern)
minDist = dist.min()
addRow = (minDist >= self.distThreshold)
if addRow:
self._protoSizes = None # need to re-compute
if self._numPatterns == self._Memory.shape[0]:
# Double the size of the memory
self._doubleMemoryNumRows()
if not self._specificIndexTraining:
# Normal learning - append the new input vector
self._Memory[self._numPatterns] = inputPattern
self._numPatterns += 1
self._categoryList.append(int(inputCategory))
else:
# Specific index training mode - insert vector in specified slot
vectorIndex = self._nextTrainingIndices.pop(0)
while vectorIndex >= self._Memory.shape[0]:
self._doubleMemoryNumRows()
self._Memory[vectorIndex] = inputPattern
self._numPatterns = max(self._numPatterns, vectorIndex + 1)
if vectorIndex >= len(self._categoryList):
self._categoryList += [-1] * (vectorIndex -
len(self._categoryList) + 1)
self._categoryList[vectorIndex] = int(inputCategory)
# Set _M to the "active" part of _Memory
self._M = self._Memory[0:self._numPatterns]
self._addPartitionId(self._numPatterns-1, partitionId)
# Sparse vectors
else:
# If the input was given in sparse form, convert it to dense if necessary
if isSparse > 0 and (self._vt is not None or self.distThreshold > 0 \
or self.numSVDDims is not None or self.numSVDSamples is not None \
or self.numWinners > 0):
denseInput = numpy.zeros(isSparse)
denseInput[inputPattern] = 1.0
inputPattern = denseInput
isSparse = 0
# Get the input width
if isSparse > 0:
inputWidth = isSparse
else:
inputWidth = len(inputPattern)
# Allocate storage if this is the first training vector
if self._Memory is None:
self._Memory = NearestNeighbor(0, inputWidth)
# Support SVD if it is on
if self._vt is not None:
inputPattern = numpy.dot(self._vt, inputPattern - self._mean)
# Threshold the input, zeroing out entries that are too close to 0.
# This is only done if we are given a dense input.
if isSparse == 0:
thresholdedInput = self._sparsifyVector(inputPattern, True)
addRow = True
# If given the layout of the cells, then turn on the logic that stores
# only the start cell for bursting columns.
if self.cellsPerCol >= 1:
burstingCols = thresholdedInput.reshape(-1,
self.cellsPerCol).min(axis=1).nonzero()[0]
for col in burstingCols:
thresholdedInput[(col * self.cellsPerCol) + 1 :
(col * self.cellsPerCol) + self.cellsPerCol] = 0
# Don't learn entries that are too close to existing entries.
if self._Memory.nRows() > 0:
dist = None
# if this vector is a perfect match for one we already learned, then
# replace the category - it may have changed with online learning on.
if self.replaceDuplicates:
dist = self._calcDistance(thresholdedInput, distanceNorm=1)
if dist.min() == 0:
rowIdx = dist.argmin()
self._categoryList[rowIdx] = int(inputCategory)
if self.fixedCapacity:
self._categoryRecencyList[rowIdx] = rowID
addRow = False
# Don't add this vector if it matches closely with another we already
# added
if self.distThreshold > 0:
if dist is None or self.distanceNorm != 1:
dist = self._calcDistance(thresholdedInput)
minDist = dist.min()
addRow = (minDist >= self.distThreshold)
if not addRow:
if self.fixedCapacity:
rowIdx = dist.argmin()
self._categoryRecencyList[rowIdx] = rowID
# If sparsity is too low, we do not want to add this vector
if addRow and self.minSparsity > 0.0:
if isSparse==0:
sparsity = ( float(len(thresholdedInput.nonzero()[0])) /
len(thresholdedInput) )
else:
sparsity = float(len(inputPattern)) / isSparse
if sparsity < self.minSparsity:
addRow = False
# Add the new sparse vector to our storage
if addRow:
self._protoSizes = None # need to re-compute
if isSparse == 0:
self._Memory.addRow(thresholdedInput)
else:
self._Memory.addRowNZ(inputPattern, [1]*len(inputPattern))
self._numPatterns += 1
self._categoryList.append(int(inputCategory))
self._addPartitionId(self._numPatterns-1, partitionId)
if self.fixedCapacity:
self._categoryRecencyList.append(rowID)
if self._numPatterns > self.maxStoredPatterns and \
self.maxStoredPatterns > 0:
leastRecentlyUsedPattern = numpy.argmin(self._categoryRecencyList)
self._Memory.deleteRow(leastRecentlyUsedPattern)
self._categoryList.pop(leastRecentlyUsedPattern)
self._categoryRecencyList.pop(leastRecentlyUsedPattern)
self._numPatterns -= 1
if self.numSVDDims is not None and self.numSVDSamples is not None \
and self._numPatterns == self.numSVDSamples:
self.computeSVD()
return self._numPatterns
def learn(self,
inputPattern,
inputCategory,
partitionId=None,
isSparse=0,
rowID=None):
"""
Learn a new training presentation
Parameters:
------------------------------------------------------------------------
inputPattern: training pattern to learn. This should be a dense array if
isSparse==0 or a list of non-zero indices if
isSparse>0
inputCategory: category index of the training pattern.
partitionID: ??
isSparse: If >0, the input pattern is a list of non-zero indices and
isSparse is the length of the dense representation.
"""
if self.verbosity >= 1:
print "%s learn:" % (g_debugPrefix)
print " category:", int(inputCategory)
print " active inputs:", _labeledInput(
inputPattern, cellsPerCol=self.cellsPerCol)
if rowID is None:
rowID = self._iterationIdx
assert partitionId is None, \
"No documentation is available for partitionId, not sure how it works."
#---------------------------------------------------------------------------------
# Dense vectors
if not self.useSparseMemory:
# Not supported
assert self.cellsPerCol == 0, "not implemented for dense vectors"
# If the input was given in sparse form, convert it to dense
if isSparse > 0:
denseInput = numpy.zeros(isSparse)
denseInput[inputPattern] = 1.0
inputPattern = denseInput
if self._specificIndexTraining and not self._nextTrainingIndices:
# Specific index mode without any index provided - skip training
return self._numPatterns
if self._Memory is None:
# Initialize memory with 100 rows and numPatterns = 0
inputWidth = len(inputPattern)
self._Memory = numpy.zeros((100, inputWidth))
self._numPatterns = 0
self._M = self._Memory[:self._numPatterns]
addRow = True
if self._vt is not None:
# Compute projection
inputPattern = numpy.dot(self._vt, inputPattern - self._mean)
if self.distThreshold > 0:
# Check if input is too close to an existing input to be accepted
dist = self._calcDistance(inputPattern)
minDist = dist.min()
addRow = (minDist >= self.distThreshold)
if addRow:
self._protoSizes = None # need to re-compute
if self._numPatterns == self._Memory.shape[0]:
# Double the size of the memory
self._doubleMemoryNumRows()
if not self._specificIndexTraining:
# Normal learning - append the new input vector
self._Memory[self._numPatterns] = inputPattern
self._numPatterns += 1
self._categoryList.append(int(inputCategory))
else:
# Specific index training mode - insert vector in specified slot
vectorIndex = self._nextTrainingIndices.pop(0)
while vectorIndex >= self._Memory.shape[0]:
self._doubleMemoryNumRows()
self._Memory[vectorIndex] = inputPattern
self._numPatterns = max(self._numPatterns, vectorIndex + 1)
if vectorIndex >= len(self._categoryList):
self._categoryList += [-1] * (
vectorIndex - len(self._categoryList) + 1)
self._categoryList[vectorIndex] = int(inputCategory)
# Set _M to the "active" part of _Memory
self._M = self._Memory[0:self._numPatterns]
if partitionId is not None:
self._partitionIdList.append(partitionId)
#---------------------------------------------------------------------------------
# Sparse vectors
else:
# If the input was given in sparse form, convert it to dense if necessary
if isSparse > 0 and (self._vt is not None or self.distThreshold > 0 \
or self.numSVDDims is not None or self.numSVDSamples is not None \
or self.numWinners > 0):
denseInput = numpy.zeros(isSparse)
denseInput[inputPattern] = 1.0
inputPattern = denseInput
isSparse = 0
# Get the input width
if isSparse > 0:
inputWidth = isSparse
else:
inputWidth = len(inputPattern)
# Allocate storage if this is the first training vector
if self._Memory is None:
self._Memory = NearestNeighbor(0, inputWidth)
# Support SVD if it is on
if self._vt is not None:
inputPattern = numpy.dot(self._vt, inputPattern - self._mean)
# Threshold the input, zeroing out entries that are too close to 0.
# This is only done if we are given a dense input.
if isSparse == 0:
thresholdedInput = self._sparsifyVector(inputPattern, True)
addRow = True
# If given the layout of the cells, then turn on the logic that stores
# only the start cell for bursting columns.
if self.cellsPerCol >= 1:
numCols = thresholdedInput.size / self.cellsPerCol
burstingCols = thresholdedInput.reshape(
-1, self.cellsPerCol).min(axis=1).nonzero()[0]
for col in burstingCols:
thresholdedInput[(col * self.cellsPerCol) +
1:(col * self.cellsPerCol) +
self.cellsPerCol] = 0
# Don't learn entries that are too close to existing entries.
if self._Memory.nRows() > 0:
dist = None
# if this vector is a perfect match for one we already learned, then
# replace the category - it may have changed with online learning on.
if self.replaceDuplicates:
dist = self._calcDistance(thresholdedInput, distanceNorm=1)
if dist.min() == 0:
rowIdx = dist.argmin()
self._categoryList[rowIdx] = int(inputCategory)
if self.fixedCapacity:
self._categoryRecencyList[rowIdx] = rowID
addRow = False
# Don't add this vector if it matches closely with another we already
# added
if self.distThreshold > 0:
if dist is None or self.distanceNorm != 1:
dist = self._calcDistance(thresholdedInput)
minDist = dist.min()
addRow = (minDist >= self.distThreshold)
if not addRow:
if self.fixedCapacity:
rowIdx = dist.argmin()
self._categoryRecencyList[rowIdx] = rowID
# Add the new vector to our storage
if addRow:
self._protoSizes = None # need to re-compute
if isSparse == 0:
self._Memory.addRow(thresholdedInput)
else:
self._Memory.addRowNZ(inputPattern,
[1] * len(inputPattern))
self._numPatterns += 1
self._categoryList.append(int(inputCategory))
if partitionId is not None:
self._partitionIdList.append(partitionId)
if self.fixedCapacity:
self._categoryRecencyList.append(rowID)
if self._numPatterns > self.maxStoredPatterns and \
self.maxStoredPatterns > 0:
leastRecentlyUsedPattern = numpy.argmin(
self._categoryRecencyList)
self._Memory.deleteRow(leastRecentlyUsedPattern)
self._categoryList.pop(leastRecentlyUsedPattern)
self._categoryRecencyList.pop(leastRecentlyUsedPattern)
self._numPatterns -= 1
if self.numSVDDims is not None and self.numSVDSamples is not None \
and self._numPatterns == self.numSVDSamples:
self.computeSVD()
return self._numPatterns
def read(cls, proto):
if proto.version != KNNCLASSIFIER_VERSION:
raise RuntimeError("Invalid KNNClassifier Version")
knn = object.__new__(cls)
knn.version = proto.version
knn.k = proto.k
knn.exact = proto.exact
knn.distanceNorm = proto.distanceNorm
knn.distanceMethod = proto.distanceMethod
knn.distThreshold = proto.distThreshold
knn.doBinarization = proto.doBinarization
knn.binarizationThreshold = proto.binarizationThreshold
knn.useSparseMemory = proto.useSparseMemory
knn.sparseThreshold = proto.sparseThreshold
knn.relativeThreshold = proto.relativeThreshold
knn.numWinners = proto.numWinners
knn.numSVDSamples = proto.numSVDSamples
knn.numSVDDims = proto.numSVDDims
knn.fractionOfMax = proto.fractionOfMax
knn.verbosity = proto.verbosity
knn.maxStoredPatterns = proto.maxStoredPatterns
knn.replaceDuplicates = proto.replaceDuplicates
knn.cellsPerCol = proto.cellsPerCol
knn.minSparsity = proto.minSparsity
if knn.numSVDDims == "adaptive":
knn._adaptiveSVDDims = True
else:
knn._adaptiveSVDDims = False
# Read private state
knn.clear()
if proto.memory is not None:
which = proto.memory.which()
if which == "ndarray":
knn._Memory = numpy.array(proto.memory.ndarray, dtype=numpy.float64)
elif which == "nearestNeighbor":
knn._Memory = NearestNeighbor()
knn._Memory.read(proto.memory.nearestNeighbor)
knn._numPatterns = proto.numPatterns
if proto.m is not None:
knn._M = numpy.array(proto.m, dtype=numpy.float64)
if proto.categoryList is not None:
knn._categoryList = list(proto.categoryList)
if proto.partitionIdList is not None:
knn._partitionIdList = list(proto.partitionIdList)
knn._rebuildPartitionIdMap(knn._partitionIdList)
knn._iterationIdx = proto.iterationIdx
knn._finishedLearning = proto.finishedLearning
if proto.s is not None:
knn._s = numpy.array(proto.s, dtype=numpy.float32)
if proto.vt is not None:
knn._vt = numpy.array(proto.vt, dtype=numpy.float32)
if proto.mean is not None:
knn._mean = numpy.array(proto.mean, dtype=numpy.float32)
return knn
