def testBadDataset(self):
filename = _getTempFileName()
print 'Creating tempfile:', filename
# Write bad dataset with records going backwards in time
fields = [('timestamp', 'datetime', 'T')]
o = FileRecordStream(streamID=filename, write=True, fields=fields)
# Records
records = (
[datetime(day=3, month=3, year=2010)],
[datetime(day=2, month=3, year=2010)])
o.appendRecord(records[0])
o.appendRecord(records[1])
o.close()
# Write bad dataset with broken sequences
fields = [('sid', 'int', 'S')]
o = FileRecordStream(streamID=filename, write=True, fields=fields)
# Records
records = ([1], [2], [1])
o.appendRecord(records[0])
o.appendRecord(records[1])
self.assertRaises(Exception, o.appendRecord, (records[2],))
o.close()
def testBadDataset(self):
filename = _getTempFileName()
print 'Creating tempfile:', filename
# Write bad dataset with records going backwards in time
fields = [
FieldMetaInfo('timestamp', FieldMetaType.datetime,
FieldMetaSpecial.timestamp)
]
o = FileRecordStream(streamID=filename, write=True, fields=fields)
# Records
records = ([datetime(day=3, month=3, year=2010)],
[datetime(day=2, month=3, year=2010)])
o.appendRecord(records[0])
o.appendRecord(records[1])
o.close()
# Write bad dataset with broken sequences
fields = [
FieldMetaInfo('sid', FieldMetaType.integer,
FieldMetaSpecial.sequence)
]
o = FileRecordStream(streamID=filename, write=True, fields=fields)
# Records
records = ([1], [2], [1])
o.appendRecord(records[0])
o.appendRecord(records[1])
self.assertRaises(Exception, o.appendRecord, (records[2], ))
o.close()
def _generateScalar(filename="simple.csv",
numSequences=2,
elementsPerSeq=1,
numRepeats=10,
stepSize=0.1,
includeRandom=False):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences of scalar values.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
stepSize: how far apart each scalar is
includeRandom: if true, include another random field
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('classification', 'float', ''), ('field1', 'float', '')]
if includeRandom:
fields += [('randomData', 'float', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i * elementsPerSeq, (i + 1) * elementsPerSeq)]
sequences.append(seq)
random.seed(42)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
if includeRandom:
outFile.appendRecord([seqIdx, x * stepSize, random.random()])
else:
outFile.appendRecord([seqIdx, x * stepSize])
outFile.close()
def _generateScalar(
filename="simple.csv", numSequences=2, elementsPerSeq=1, numRepeats=10, stepSize=0.1, includeRandom=False
):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences of scalar values.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
stepSize: how far apart each scalar is
includeRandom: if true, include another random field
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, "datasets", filename)
print "Creating %s..." % (pathname)
fields = [("classification", "float", ""), ("field1", "float", "")]
if includeRandom:
fields += [("randomData", "float", "")]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i * elementsPerSeq, (i + 1) * elementsPerSeq)]
sequences.append(seq)
random.seed(42)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
if includeRandom:
outFile.appendRecord([seqIdx, x * stepSize, random.random()])
else:
outFile.appendRecord([seqIdx, x * stepSize])
outFile.close()
def _generateScalar(filename="simple.csv",
numSequences=2,
elementsPerSeq=1,
numRepeats=10,
stepSize=0.1,
resets=False):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences of scalar values.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
stepSize: how far apart each scalar is
resets: if True, turn on reset at start of each sequence
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print("Creating %s..." % (pathname))
fields = [('reset', 'int', 'R'), ('category', 'int', 'C'),
('field1', 'float', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i * elementsPerSeq, (i + 1) * elementsPerSeq)]
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += list(range(numSequences))
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
reset = int(resets)
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([reset, str(seqIdx), x * stepSize])
reset = 0
outFile.close()
def _generateScalar(filename="simple.csv", numSequences=2, elementsPerSeq=1,
numRepeats=10, stepSize=0.1, resets=False):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences of scalar values.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
stepSize: how far apart each scalar is
resets: if True, turn on reset at start of each sequence
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('reset', 'int', 'R'), ('category', 'int', 'C'),
('field1', 'float', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i*elementsPerSeq, (i+1)*elementsPerSeq)]
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
reset = int(resets)
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([reset, str(seqIdx), x*stepSize])
reset = 0
outFile.close()
def _generateCategory(filename="simple.csv",
numSequences=2,
elementsPerSeq=1,
numRepeats=10):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('classification', 'string', ''), ('field1', 'string', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i * elementsPerSeq, (i + 1) * elementsPerSeq)]
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([str(seqIdx), str(x)])
outFile.close()
def _generateCategory(filename="simple.csv", numSequences=2, elementsPerSeq=1,
numRepeats=10):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('classification', 'string', ''),
('field1', 'string', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i*elementsPerSeq, (i+1)*elementsPerSeq)]
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([str(seqIdx), str(x)])
outFile.close()
class StreamReader(RecordStreamIface):
"""
Implements a stream reader. This is a high level class that owns one or more
underlying implementations of a RecordStreamIFace. Each RecordStreamIFace
implements the raw reading of records from the record store (which could be a
file, hbase table or something else).
In the future, we will support joining of two or more RecordStreamIface's (
which is why the streamDef accepts a list of 'stream' elements), but for now
only 1 source is supported.
The class also implements aggregation of the (in the future) joined records
from the sources.
This module parses the stream definition (as defined in
/nupic/frameworks/opf/jsonschema/stream_def.json), creates the
RecordStreamIFace for each source ('stream's element) defined in the stream
def, performs aggregation, and returns each record in the correct format
according to the desired column names specified in the streamDef.
This class implements the RecordStreamIFace interface and thus can be used
in place of a raw record stream.
This is an example streamDef:
{
'version': 1
'info': 'test_hotgym',
'streams': [
{'columns': [u'*'],
'info': u'hotGym.csv',
'last_record': 4000,
'source': u'file://extra/hotgym/hotgym.csv'}.
],
'timeField': 'timestamp',
'aggregation': {
'hours': 1,
'fields': [
('timestamp', 'first'),
('gym', 'first'),
('consumption', 'sum')
],
}
}
"""
def __init__(self, streamDef, bookmark=None, saveOutput=False,
isBlocking=True, maxTimeout=0, eofOnTimeout=False):
""" Base class constructor, performs common initialization
Parameters:
----------------------------------------------------------------
streamDef: The stream definition, potentially containing multiple sources
(not supported yet). See
/nupic/frameworks/opf/jsonschema/stream_def.json for the format
of this dict
bookmark: Bookmark to start reading from. This overrides the first_record
field of the streamDef if provided.
saveOutput: If true, save the output to a csv file in a temp directory.
The path to the generated file can be found in the log
output.
isBlocking: should read operation block *forever* if the next row of data
is not available, but the stream is not marked as 'completed'
yet?
maxTimeout: if isBlocking is False, max seconds to wait for more data before
timing out; ignored when isBlocking is True.
eofOnTimeout: If True and we get a read timeout (isBlocking must be False
to get read timeouts), assume we've reached the end of the
input and produce the last aggregated record, if one can be
completed.
"""
# Call superclass constructor
super(StreamReader, self).__init__()
loggerPrefix = 'com.numenta.nupic.data.StreamReader'
self._logger = logging.getLogger(loggerPrefix)
jsonhelpers.validate(streamDef,
schemaPath=pkg_resources.resource_filename(
jsonschema.__name__, "stream_def.json"))
assert len(streamDef['streams']) == 1, "Only 1 source stream is supported"
# Save constructor args
sourceDict = streamDef['streams'][0]
self._recordCount = 0
self._eofOnTimeout = eofOnTimeout
self._logger.debug('Reading stream with the def: %s', sourceDict)
# Dictionary to store record statistics (min and max of scalars for now)
self._stats = None
# ---------------------------------------------------------------------
# Get the stream definition params
# Limiting window of the stream. It would not return any records until
# 'first_record' ID is read (or very first with the ID above that). The
# stream will return EOS once it reads record with ID 'last_record' or
# above (NOTE: the name 'lastRecord' is misleading because it is NOT
# inclusive).
firstRecordIdx = sourceDict.get('first_record', None)
self._sourceLastRecordIdx = sourceDict.get('last_record', None)
# If a bookmark was given, then override first_record from the stream
# definition.
if bookmark is not None:
firstRecordIdx = None
# Column names must be provided in the streamdef json
# Special case is ['*'], meaning all available names from the record stream
self._streamFieldNames = sourceDict.get('columns', None)
if self._streamFieldNames != None and self._streamFieldNames[0] == '*':
self._needFieldsFiltering = False
else:
self._needFieldsFiltering = True
# Types must be specified in streamdef json, or in case of the
# file_recod_stream types could be implicit from the file
streamFieldTypes = sourceDict.get('types', None)
self._logger.debug('Types from the def: %s', streamFieldTypes)
# Validate that all types are valid
if streamFieldTypes is not None:
for dataType in streamFieldTypes:
assert FieldMetaType.isValid(dataType)
# Reset, sequence and time fields might be provided by streamdef json
streamResetFieldName = streamDef.get('resetField', None)
streamTimeFieldName = streamDef.get('timeField', None)
streamSequenceFieldName = streamDef.get('sequenceIdField', None)
self._logger.debug('r, t, s fields: %s, %s, %s', streamResetFieldName,
streamTimeFieldName,
streamSequenceFieldName)
# =======================================================================
# Open up the underlying record store
dataUrl = sourceDict.get('source', None)
assert dataUrl is not None
self._recordStore = self._openStream(dataUrl, isBlocking, maxTimeout,
bookmark, firstRecordIdx)
assert self._recordStore is not None
# =======================================================================
# Prepare the data structures we need for returning just the fields
# the caller wants from each record
recordStoreFields = self._recordStore.getFields()
self._recordStoreFieldNames = self._recordStore.getFieldNames()
if not self._needFieldsFiltering:
self._streamFieldNames = self._recordStoreFieldNames
# Build up the field definitions for each field. This is a list of tuples
# of (name, type, special)
self._streamFields = []
for dstIdx, name in enumerate(self._streamFieldNames):
if name not in self._recordStoreFieldNames:
raise RuntimeError("The column '%s' from the stream definition "
"is not present in the underlying stream which has the following "
"columns: %s" % (name, self._recordStoreFieldNames))
fieldIdx = self._recordStoreFieldNames.index(name)
fieldType = recordStoreFields[fieldIdx].type
fieldSpecial = recordStoreFields[fieldIdx].special
# If the types or specials were defined in the stream definition,
# then override what was found in the record store
if streamFieldTypes is not None:
fieldType = streamFieldTypes[dstIdx]
if streamResetFieldName is not None and streamResetFieldName == name:
fieldSpecial = FieldMetaSpecial.reset
if streamTimeFieldName is not None and streamTimeFieldName == name:
fieldSpecial = FieldMetaSpecial.timestamp
if (streamSequenceFieldName is not None and
streamSequenceFieldName == name):
fieldSpecial = FieldMetaSpecial.sequence
self._streamFields.append(FieldMetaInfo(name, fieldType, fieldSpecial))
# ========================================================================
# Create the aggregator which will handle aggregation of records before
# returning them.
self._aggregator = Aggregator(
aggregationInfo=streamDef.get('aggregation', None),
inputFields=recordStoreFields,
timeFieldName=streamDef.get('timeField', None),
sequenceIdFieldName=streamDef.get('sequenceIdField', None),
resetFieldName=streamDef.get('resetField', None))
# We rely on the aggregator to tell us the bookmark of the last raw input
# that contributed to the aggregated record
self._aggBookmark = None
# Compute the aggregation period in terms of months and seconds
if 'aggregation' in streamDef:
self._aggMonthsAndSeconds = nupic.support.aggregationToMonthsSeconds(
streamDef.get('aggregation'))
else:
self._aggMonthsAndSeconds = None
# ========================================================================
# Are we saving the generated output to a csv?
if saveOutput:
tmpDir = tempfile.mkdtemp()
outFilename = os.path.join(tmpDir, "generated_output.csv")
self._logger.info("StreamReader: Saving generated records to: '%s'" %
outFilename)
self._writer = FileRecordStream(streamID=outFilename,
write=True,
fields=self._streamFields)
else:
self._writer = None
@staticmethod
def _openStream(dataUrl,
isBlocking, # pylint: disable=W0613
maxTimeout, # pylint: disable=W0613
bookmark,
firstRecordIdx):
"""Open the underlying file stream
This only supports 'file://' prefixed paths.
:returns: record stream instance
:rtype: FileRecordStream
"""
filePath = dataUrl[len(FILE_PREF):]
if not os.path.isabs(filePath):
filePath = os.path.join(os.getcwd(), filePath)
return FileRecordStream(streamID=filePath,
write=False,
bookmark=bookmark,
firstRecord=firstRecordIdx)
def close(self):
""" Close the stream
"""
return self._recordStore.close()
def getNextRecord(self):
""" Returns combined data from all sources (values only).
Returns None on EOF; empty sequence on timeout.
"""
# Keep reading from the raw input till we get enough for an aggregated
# record
while True:
# Reached EOF due to lastRow constraint?
if self._sourceLastRecordIdx is not None and \
self._recordStore.getNextRecordIdx() >= self._sourceLastRecordIdx:
preAggValues = None # indicates EOF
bookmark = self._recordStore.getBookmark()
else:
# Get the raw record and bookmark
preAggValues = self._recordStore.getNextRecord()
bookmark = self._recordStore.getBookmark()
if preAggValues == (): # means timeout error occurred
if self._eofOnTimeout:
preAggValues = None # act as if we got EOF
else:
return preAggValues # Timeout indicator
self._logger.debug('Read source record #%d: %r',
self._recordStore.getNextRecordIdx()-1, preAggValues)
# Perform aggregation
(fieldValues, aggBookmark) = self._aggregator.next(preAggValues, bookmark)
# Update the aggregated record bookmark if we got a real record back
if fieldValues is not None:
self._aggBookmark = aggBookmark
# Reached EOF?
if preAggValues is None and fieldValues is None:
return None
# Return it if we have a record
if fieldValues is not None:
break
# Do we need to re-order the fields in the record?
if self._needFieldsFiltering:
values = []
srcDict = dict(zip(self._recordStoreFieldNames, fieldValues))
for name in self._streamFieldNames:
values.append(srcDict[name])
fieldValues = values
# Write to debug output?
if self._writer is not None:
self._writer.appendRecord(fieldValues)
self._recordCount += 1
self._logger.debug('Returning aggregated record #%d from getNextRecord(): '
'%r. Bookmark: %r',
self._recordCount-1, fieldValues, self._aggBookmark)
return fieldValues
def getDataRowCount(self):
"""Iterates through stream to calculate total records after aggregation.
This will alter the bookmark state.
"""
inputRowCountAfterAggregation = 0
while True:
record = self.getNextRecord()
if record is None:
return inputRowCountAfterAggregation
inputRowCountAfterAggregation += 1
if inputRowCountAfterAggregation > 10000:
raise RuntimeError('No end of datastream found.')
def getLastRecords(self, numRecords):
"""Saves the record in the underlying storage."""
raise RuntimeError("Not implemented in StreamReader")
def getRecordsRange(self, bookmark=None, range=None):
""" Returns a range of records, starting from the bookmark. If 'bookmark'
is None, then records read from the first available. If 'range' is
None, all available records will be returned (caution: this could be
a lot of records and require a lot of memory).
"""
raise RuntimeError("Not implemented in StreamReader")
def getNextRecordIdx(self):
"""Returns the index of the record that will be read next from
getNextRecord()
"""
return self._recordCount
def recordsExistAfter(self, bookmark):
"""Returns True iff there are records left after the bookmark."""
return self._recordStore.recordsExistAfter(bookmark)
def getAggregationMonthsAndSeconds(self):
""" Returns the aggregation period of the record stream as a dict
containing 'months' and 'seconds'. The months is always an integer and
seconds is a floating point. Only one is allowed to be non-zero at a
time.
If there is no aggregation associated with the stream, returns None.
Typically, a raw file or hbase stream will NOT have any aggregation info,
but subclasses of RecordStreamIFace, like StreamReader, will and will
return the aggregation period from this call. This call is used by the
getNextRecordDict() method to assign a record number to a record given
its timestamp and the aggregation interval
Parameters:
------------------------------------------------------------------------
retval: aggregationPeriod (as a dict) or None
'months': number of months in aggregation period
'seconds': number of seconds in aggregation period (as a float)
"""
return self._aggMonthsAndSeconds
def appendRecord(self, record, inputRef=None):
"""Saves the record in the underlying storage."""
raise RuntimeError("Not implemented in StreamReader")
def appendRecords(self, records, inputRef=None, progressCB=None):
"""Saves multiple records in the underlying storage."""
raise RuntimeError("Not implemented in StreamReader")
def removeOldData(self):
raise RuntimeError("Not implemented in StreamReader")
def seekFromEnd(self, numRecords):
"""Seeks to numRecords from the end and returns a bookmark to the new
position.
"""
raise RuntimeError("Not implemented in StreamReader")
def getFieldNames(self):
""" Returns all fields in all inputs (list of plain names).
NOTE: currently, only one input is supported
"""
return [f.name for f in self._streamFields]
def getFields(self):
""" Returns a sequence of nupic.data.fieldmeta.FieldMetaInfo
name/type/special tuples for each field in the stream.
"""
return self._streamFields
def getBookmark(self):
""" Returns a bookmark to the current position
"""
return self._aggBookmark
def clearStats(self):
""" Resets stats collected so far.
"""
self._recordStore.clearStats()
def getStats(self):
""" Returns stats (like min and max values of the fields).
TODO: This method needs to be enhanced to get the stats on the *aggregated*
records.
"""
# The record store returns a dict of stats, each value in this dict is
# a list with one item per field of the record store
# {
# 'min' : [f1_min, f2_min, f3_min],
# 'max' : [f1_max, f2_max, f3_max]
# }
recordStoreStats = self._recordStore.getStats()
# We need to convert each item to represent the fields of the *stream*
streamStats = dict()
for (key, values) in recordStoreStats.items():
fieldStats = dict(zip(self._recordStoreFieldNames, values))
streamValues = []
for name in self._streamFieldNames:
streamValues.append(fieldStats[name])
streamStats[key] = streamValues
return streamStats
def getError(self):
""" Returns errors saved in the stream.
"""
return self._recordStore.getError()
def setError(self, error):
""" Saves specified error in the stream.
"""
self._recordStore.setError(error)
def isCompleted(self):
""" Returns True if all records have been read.
"""
return self._recordStore.isCompleted()
def setCompleted(self, completed=True):
""" Marks the stream completed (True or False)
"""
# CSV file is always considered completed, nothing to do
self._recordStore.setCompleted(completed)
def setTimeout(self, timeout):
""" Set the read timeout """
self._recordStore.setTimeout(timeout)
def flush(self):
""" Flush the file to disk """
raise RuntimeError("Not implemented in StreamReader")
class _BasicPredictionWriter(PredictionWriterIface):
""" This class defines the basic (file-based) implementation of
PredictionWriterIface, whose instances are returned by
BasicPredictionWriterFactory
"""
def __init__(self,
experimentDir,
label,
inferenceType,
fields,
metricNames=None,
checkpointSource=None):
""" Constructor
experimentDir:
experiment directory path that contains description.py
label: A label string to incorporate into the filename.
inferenceElements:
inferenceType:
An constant from opfutils.InferenceType for the
requested prediction writer
fields: a non-empty sequence of nupic.data.fieldmeta.FieldMetaInfo
representing fields that will be emitted to this prediction
writer
metricNames: OPTIONAL - A list of metric names that well be emiited by this
prediction writer
checkpointSource:
If not None, a File-like object containing the
previously-checkpointed predictions for setting the initial
contents of this PredictionOutputStream. Will be copied
before returning, if needed.
"""
#assert len(fields) > 0
self.__experimentDir = experimentDir
# opfutils.InferenceType kind value
self.__inferenceType = inferenceType
# A tuple of nupic.data.fieldmeta.FieldMetaInfo
self.__inputFieldsMeta = tuple(copy.deepcopy(fields))
self.__numInputFields = len(self.__inputFieldsMeta)
self.__label = label
if metricNames is not None:
metricNames.sort()
self.__metricNames = metricNames
# Define our output field meta info
self.__outputFieldsMeta = []
# The list of inputs that we include in the prediction output
self._rawInputNames = []
# Output dataset
self.__datasetPath = None
self.__dataset = None
# Save checkpoint data until we're ready to create the output dataset
self.__checkpointCache = None
if checkpointSource is not None:
checkpointSource.seek(0)
self.__checkpointCache = StringIO.StringIO()
shutil.copyfileobj(checkpointSource, self.__checkpointCache)
return
############################################################################
def __openDatafile(self, modelResult):
"""Open the data file and write the header row"""
# Write reset bit
resetFieldMeta = FieldMetaInfo(name="reset",
type=FieldMetaType.integer,
special=FieldMetaSpecial.reset)
self.__outputFieldsMeta.append(resetFieldMeta)
# -----------------------------------------------------------------------
# Write each of the raw inputs that go into the encoders
rawInput = modelResult.rawInput
rawFields = rawInput.keys()
rawFields.sort()
for field in rawFields:
if field.startswith('_') or field == 'reset':
continue
value = rawInput[field]
meta = FieldMetaInfo(name=field,
type=FieldMetaType.string,
special=FieldMetaSpecial.none)
self.__outputFieldsMeta.append(meta)
self._rawInputNames.append(field)
# -----------------------------------------------------------------------
# Handle each of the inference elements
for inferenceElement, value in modelResult.inferences.iteritems():
inferenceLabel = InferenceElement.getLabel(inferenceElement)
# TODO: Right now we assume list inferences are associated with
# The input field metadata
if type(value) in (list, tuple):
# Append input and prediction field meta-info
self.__outputFieldsMeta.extend(
self.__getListMetaInfo(inferenceElement))
elif isinstance(value, dict):
self.__outputFieldsMeta.extend(
self.__getDictMetaInfo(inferenceElement, value))
else:
if InferenceElement.getInputElement(inferenceElement):
self.__outputFieldsMeta.append(
FieldMetaInfo(name=inferenceLabel + ".actual",
type=FieldMetaType.string,
special=''))
self.__outputFieldsMeta.append(
FieldMetaInfo(name=inferenceLabel,
type=FieldMetaType.string,
special=''))
if self.__metricNames:
for metricName in self.__metricNames:
metricField = FieldMetaInfo(name=metricName,
type=FieldMetaType.float,
special=FieldMetaSpecial.none)
self.__outputFieldsMeta.append(metricField)
# Create the inference directory for our experiment
inferenceDir = _FileUtils.createExperimentInferenceDir(
self.__experimentDir)
# Consctruct the prediction dataset file path
filename = (self.__label + "." +
opfutils.InferenceType.getLabel(self.__inferenceType) +
".predictionLog.csv")
self.__datasetPath = os.path.join(inferenceDir, filename)
# Create the output dataset
print "OPENING OUTPUT FOR PREDICTION WRITER AT: %r" % self.__datasetPath
print "Prediction field-meta: %r" % (
[tuple(i) for i in self.__outputFieldsMeta], )
self.__dataset = FileRecordStream(streamID=self.__datasetPath,
write=True,
fields=self.__outputFieldsMeta)
# Copy data from checkpoint cache
if self.__checkpointCache is not None:
self.__checkpointCache.seek(0)
reader = csv.reader(self.__checkpointCache, dialect='excel')
# Skip header row
try:
header = reader.next()
except StopIteration:
print "Empty record checkpoint initializer for %r" % (
self.__datasetPath, )
else:
assert tuple(self.__dataset.getFieldNames()) == tuple(header), \
"dataset.getFieldNames(): %r; predictionCheckpointFieldNames: %r" % (
tuple(self.__dataset.getFieldNames()), tuple(header))
# Copy the rows from checkpoint
numRowsCopied = 0
while True:
try:
row = reader.next()
except StopIteration:
break
#print "DEBUG: restoring row from checkpoint: %r" % (row,)
self.__dataset.appendRecord(row)
numRowsCopied += 1
self.__dataset.flush()
print "Restored %d rows from checkpoint for %r" % (
numRowsCopied, self.__datasetPath)
# Dispose of our checkpoint cache
self.__checkpointCache.close()
self.__checkpointCache = None
return
############################################################################
def setLoggedMetrics(self, metricNames):
""" Tell the writer which metrics should be written
Parameters:
-----------------------------------------------------------------------
metricsNames: A list of metric lables to be written
"""
if metricNames is None:
self.__metricNames = set([])
else:
self.__metricNames = set(metricNames)
############################################################################
def close(self):
""" [virtual method override] Closes the writer (e.g., close the underlying
file)
"""
if self.__dataset:
self.__dataset.close()
self.__dataset = None
return
############################################################################
def __getListMetaInfo(self, inferenceElement):
""" Get field metadata information for inferences that are of list type
TODO: Right now we assume list inferences are associated with the input field
metadata
"""
fieldMetaInfo = []
inferenceLabel = InferenceElement.getLabel(inferenceElement)
for inputFieldMeta in self.__inputFieldsMeta:
if InferenceElement.getInputElement(inferenceElement):
outputFieldMeta = FieldMetaInfo(name=inputFieldMeta.name +
".actual",
type=inputFieldMeta.type,
special=inputFieldMeta.special)
predictionField = FieldMetaInfo(name=inputFieldMeta.name + "." +
inferenceLabel,
type=inputFieldMeta.type,
special=inputFieldMeta.special)
fieldMetaInfo.append(outputFieldMeta)
fieldMetaInfo.append(predictionField)
return fieldMetaInfo
############################################################################
def __getDictMetaInfo(self, inferenceElement, inferenceDict):
"""Get field metadate information for inferences that are of dict type"""
fieldMetaInfo = []
inferenceLabel = InferenceElement.getLabel(inferenceElement)
if InferenceElement.getInputElement(inferenceElement):
fieldMetaInfo.append(
FieldMetaInfo(name=inferenceLabel + ".actual",
type=FieldMetaType.string,
special=''))
keys = sorted(inferenceDict.keys())
for key in keys:
fieldMetaInfo.append(
FieldMetaInfo(name=inferenceLabel + "." + str(key),
type=FieldMetaType.string,
special=''))
return fieldMetaInfo
############################################################################
def append(self, modelResult):
""" [virtual method override] Emits a single prediction as input versus
predicted.
modelResult: An opfutils.ModelResult object that contains the model input
and output for the current timestep.
"""
#print "DEBUG: _BasicPredictionWriter: writing modelResult: %r" % (modelResult,)
# If there are no inferences, don't write anything
inferences = modelResult.inferences
hasInferences = False
if inferences is not None:
for value in inferences.itervalues():
hasInferences = hasInferences or (value is not None)
if not hasInferences:
return
if self.__dataset is None:
self.__openDatafile(modelResult)
inputData = modelResult.sensorInput
sequenceReset = int(bool(inputData.sequenceReset))
outputRow = [sequenceReset]
# -----------------------------------------------------------------------
# Write out the raw inputs
rawInput = modelResult.rawInput
for field in self._rawInputNames:
outputRow.append(str(rawInput[field]))
# -----------------------------------------------------------------------
# Write out the inference element info
for inferenceElement, outputVal in inferences.iteritems():
inputElement = InferenceElement.getInputElement(inferenceElement)
if inputElement:
inputVal = getattr(inputData, inputElement)
else:
inputVal = None
if type(outputVal) in (list, tuple):
assert type(inputVal) in (list, tuple, None)
for iv, ov in zip(inputVal, outputVal):
# Write actual
outputRow.append(str(iv))
# Write inferred
outputRow.append(str(ov))
elif isinstance(outputVal, dict):
if inputVal is not None:
# If we have a predicted field, include only that in the actuals
if modelResult.predictedFieldIdx is not None:
outputRow.append(
str(inputVal[modelResult.predictedFieldIdx]))
else:
outputRow.append(str(inputVal))
for key in sorted(outputVal.keys()):
outputRow.append(str(outputVal[key]))
else:
if inputVal is not None:
outputRow.append(str(inputVal))
outputRow.append(str(outputVal))
metrics = modelResult.metrics
for metricName in self.__metricNames:
outputRow.append(metrics.get(metricName, 0.0))
#print "DEBUG: _BasicPredictionWriter: writing outputRow: %r" % (outputRow,)
self.__dataset.appendRecord(outputRow)
self.__dataset.flush()
return
def checkpoint(self, checkpointSink, maxRows):
""" [virtual method override] Save a checkpoint of the prediction output
stream. The checkpoint comprises up to maxRows of the most recent inference
records.
Parameters:
----------------------------------------------------------------------
checkpointSink: A File-like object where predictions checkpoint data, if
any, will be stored.
maxRows: Maximum number of most recent inference rows
to checkpoint.
"""
checkpointSink.truncate()
if self.__dataset is None:
if self.__checkpointCache is not None:
self.__checkpointCache.seek(0)
shutil.copyfileobj(self.__checkpointCache, checkpointSink)
checkpointSink.flush()
return
else:
# Nothing to checkpoint
return
self.__dataset.flush()
totalDataRows = self.__dataset.getDataRowCount()
if totalDataRows == 0:
# Nothing to checkpoint
return
# Open reader of prediction file (suppress missingValues conversion)
reader = FileRecordStream(self.__datasetPath, missingValues=[])
# Create CSV writer for writing checkpoint rows
writer = csv.writer(checkpointSink)
# Write the header row to checkpoint sink -- just field names
writer.writerow(reader.getFieldNames())
# Determine number of rows to checkpoint
numToWrite = min(maxRows, totalDataRows)
# Skip initial rows to get to the rows that we actually need to checkpoint
numRowsToSkip = totalDataRows - numToWrite
for i in xrange(numRowsToSkip):
reader.next()
# Write the data rows to checkpoint sink
numWritten = 0
while True:
row = reader.getNextRecord()
if row is None:
break
row = [str(element) for element in row]
#print "DEBUG: _BasicPredictionWriter: checkpointing row: %r" % (row,)
writer.writerow(row)
numWritten += 1
assert numWritten == numToWrite, \
"numWritten (%s) != numToWrite (%s)" % (numWritten, numToWrite)
checkpointSink.flush()
return
def _generateFile(filename, numRecords, categoryList, initProb,
firstOrderProb, secondOrderProb, seqLen, numNoise=0, resetsEvery=None):
""" Generate a set of records reflecting a set of probabilities.
Parameters:
----------------------------------------------------------------
filename: name of .csv file to generate
numRecords: number of records to generate
categoryList: list of category names
initProb: Initial probability for each category. This is a vector
of length len(categoryList).
firstOrderProb: A dictionary of the 1st order probabilities. The key
is the 1st element of the sequence, the value is
the probability of each 2nd element given the first.
secondOrderProb: A dictionary of the 2nd order probabilities. The key
is the first 2 elements of the sequence, the value is
the probability of each possible 3rd element given the
first two.
seqLen: Desired length of each sequence. The 1st element will
be generated using the initProb, the 2nd element by the
firstOrder table, and the 3rd and all successive
elements by the secondOrder table. None means infinite
length.
numNoise: Number of noise elements to place between each
sequence. The noise elements are evenly distributed from
all categories.
resetsEvery: If not None, generate a reset every N records
Here is an example of some parameters:
categoryList: ['cat1', 'cat2', 'cat3']
initProb: [0.7, 0.2, 0.1]
firstOrderProb: {'[0]': [0.3, 0.3, 0.4],
'[1]': [0.3, 0.3, 0.4],
'[2]': [0.3, 0.3, 0.4]}
secondOrderProb: {'[0,0]': [0.3, 0.3, 0.4],
'[0,1]': [0.3, 0.3, 0.4],
'[0,2]': [0.3, 0.3, 0.4],
'[1,0]': [0.3, 0.3, 0.4],
'[1,1]': [0.3, 0.3, 0.4],
'[1,2]': [0.3, 0.3, 0.4],
'[2,0]': [0.3, 0.3, 0.4],
'[2,1]': [0.3, 0.3, 0.4],
'[2,2]': [0.3, 0.3, 0.4]}
"""
# Create the file
print "Creating %s..." % (filename)
fields = [('reset', 'int', 'R'), ('name', 'string', '')]
outFile = FileRecordStream(filename, write=True, fields=fields)
# --------------------------------------------------------------------
# Convert the probabilitie tables into cumulative probabilities
initCumProb = initProb.cumsum()
firstOrderCumProb = dict()
for (key,value) in firstOrderProb.iteritems():
firstOrderCumProb[key] = value.cumsum()
secondOrderCumProb = dict()
for (key,value) in secondOrderProb.iteritems():
secondOrderCumProb[key] = value.cumsum()
# --------------------------------------------------------------------
# Write out the sequences
elementsInSeq = []
numElementsSinceReset = 0
maxCatIdx = len(categoryList) - 1
for i in xrange(numRecords):
# Generate a reset?
if numElementsSinceReset == 0:
reset = 1
else:
reset = 0
# Pick the next element, based on how are we are into the 2nd order
# sequence.
rand = numpy.random.rand()
if len(elementsInSeq) == 0:
catIdx = numpy.searchsorted(initCumProb, rand)
elif len(elementsInSeq) == 1:
catIdx = numpy.searchsorted(firstOrderCumProb[str(elementsInSeq)], rand)
elif (len(elementsInSeq) >=2) and \
(seqLen is None or len(elementsInSeq) < seqLen-numNoise):
catIdx = numpy.searchsorted(secondOrderCumProb[str(elementsInSeq[-2:])], rand)
else: # random "noise"
catIdx = numpy.random.randint(len(categoryList))
# Write out the record
catIdx = min(maxCatIdx, catIdx)
outFile.appendRecord([reset,categoryList[catIdx]])
#print categoryList[catIdx]
# ------------------------------------------------------------
# Increment counters
elementsInSeq.append(catIdx)
numElementsSinceReset += 1
# Generate another reset?
if resetsEvery is not None and numElementsSinceReset == resetsEvery:
numElementsSinceReset = 0
elementsInSeq = []
# Start another 2nd order sequence?
if seqLen is not None and (len(elementsInSeq) == seqLen+numNoise):
elementsInSeq = []
outFile.close()
def testMissingValues(self):
print "Beginning Missing Data test..."
filename = _getTempFileName()
# Some values missing of each type
# read dataset from disk, retrieve values
# string should return empty string, numeric types sentinelValue
print 'Creating tempfile:', filename
# write dataset to disk with float, int, and string fields
fields = [('timestamp', 'datetime', 'T'),
('name', 'string', ''),
('integer', 'int', ''),
('real', 'float', '')]
s = FileRecordStream(streamID=filename, write=True, fields=fields)
# Records
records = (
[datetime(day=1, month=3, year=2010), 'rec_1', 5, 6.5],
[datetime(day=2, month=3, year=2010), '', 8, 7.5],
[datetime(day=3, month=3, year=2010), 'rec_3', '', 8.5],
[datetime(day=4, month=3, year=2010), 'rec_4', 12, ''],
[datetime(day=5, month=3, year=2010), 'rec_5', -87657496599, 6.5],
[datetime(day=6, month=3, year=2010), 'rec_6', 12, -87657496599],
[datetime(day=6, month=3, year=2010), str(-87657496599), 12, 6.5])
for r in records:
s.appendRecord(list(r))
s.close()
# Read the standard file
s = FileRecordStream(streamID=filename, write=False)
fieldsRead = s.getFields()
self.assertTrue(fields == fieldsRead)
recordsRead = []
while True:
r = s.getNextRecord()
if r is None:
break
print 'Reading record ...'
print r
recordsRead.append(r)
# sort the records by date, so we know for sure which is which
sorted(recordsRead, key=lambda rec: rec[0])
# empty string
self.assertTrue(recordsRead[1][1] == SENTINEL_VALUE_FOR_MISSING_DATA)
# missing int
self.assertTrue(recordsRead[2][2] == SENTINEL_VALUE_FOR_MISSING_DATA)
# missing float
self.assertTrue(recordsRead[3][3] == SENTINEL_VALUE_FOR_MISSING_DATA)
# sentinel value in input handled correctly for int field
self.assertTrue(recordsRead[4][2] != SENTINEL_VALUE_FOR_MISSING_DATA)
# sentinel value in input handled correctly for float field
self.assertTrue(recordsRead[5][3] != SENTINEL_VALUE_FOR_MISSING_DATA)
# sentinel value in input handled correctly for string field
# this should leave the string as-is, since a missing string
# is encoded not with a sentinel value but with an empty string
self.assertTrue(recordsRead[6][1] != SENTINEL_VALUE_FOR_MISSING_DATA)
def _generateOverlapping(filename="overlap.csv", numSequences=2, elementsPerSeq=3,
numRepeats=10, hub=[0,1], hubOffset=1, resets=False):
""" Generate a temporal dataset containing sequences that overlap one or more
elements with other sequences.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
hub: sub-sequence to place within each other sequence
hubOffset: where, within each sequence, to place the hub
resets: if True, turn on reset at start of each sequence
"""
# Check for conflicts in arguments
assert (hubOffset + len(hub) <= elementsPerSeq)
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('reset', 'int', 'R'), ('category', 'int', 'C'),
('field1', 'string', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences with the hub in the middle
sequences = []
nextElemIdx = max(hub)+1
for _ in range(numSequences):
seq = []
for j in range(hubOffset):
seq.append(nextElemIdx)
nextElemIdx += 1
for j in hub:
seq.append(j)
j = hubOffset + len(hub)
while j < elementsPerSeq:
seq.append(nextElemIdx)
nextElemIdx += 1
j += 1
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for _ in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
reset = int(resets)
seq = sequences[seqIdx]
for (x) in seq:
outFile.appendRecord([reset, str(seqIdx), str(x)])
reset = 0
outFile.close()
class _BasicPredictionWriter(PredictionWriterIface):
""" This class defines the basic (file-based) implementation of
PredictionWriterIface, whose instances are returned by
BasicPredictionWriterFactory
"""
def __init__(self, experimentDir, label, inferenceType,
fields, metricNames=None, checkpointSource=None):
""" Constructor
experimentDir:
experiment directory path that contains description.py
label: A label string to incorporate into the filename.
inferenceElements:
inferenceType:
An constant from opfutils.InferenceType for the
requested prediction writer
fields: a non-empty sequence of nupic.data.fieldmeta.FieldMetaInfo
representing fields that will be emitted to this prediction
writer
metricNames: OPTIONAL - A list of metric names that well be emiited by this
prediction writer
checkpointSource:
If not None, a File-like object containing the
previously-checkpointed predictions for setting the initial
contents of this PredictionOutputStream. Will be copied
before returning, if needed.
"""
#assert len(fields) > 0
self.__experimentDir = experimentDir
# opfutils.InferenceType kind value
self.__inferenceType = inferenceType
# A tuple of nupic.data.fieldmeta.FieldMetaInfo
self.__inputFieldsMeta = tuple(copy.deepcopy(fields))
self.__numInputFields = len(self.__inputFieldsMeta)
self.__label = label
if metricNames is not None:
metricNames.sort()
self.__metricNames = metricNames
# Define our output field meta info
self.__outputFieldsMeta = []
# The list of inputs that we include in the prediction output
self._rawInputNames = []
# Output dataset
self.__datasetPath = None
self.__dataset = None
# Save checkpoint data until we're ready to create the output dataset
self.__checkpointCache = None
if checkpointSource is not None:
checkpointSource.seek(0)
self.__checkpointCache = StringIO.StringIO()
shutil.copyfileobj(checkpointSource, self.__checkpointCache)
return
def __openDatafile(self, modelResult):
"""Open the data file and write the header row"""
# Write reset bit
resetFieldMeta = FieldMetaInfo(
name="reset",
type=FieldMetaType.integer,
special = FieldMetaSpecial.reset)
self.__outputFieldsMeta.append(resetFieldMeta)
# -----------------------------------------------------------------------
# Write each of the raw inputs that go into the encoders
rawInput = modelResult.rawInput
rawFields = rawInput.keys()
rawFields.sort()
for field in rawFields:
if field.startswith('_') or field == 'reset':
continue
value = rawInput[field]
meta = FieldMetaInfo(name=field, type=FieldMetaType.string,
special=FieldMetaSpecial.none)
self.__outputFieldsMeta.append(meta)
self._rawInputNames.append(field)
# -----------------------------------------------------------------------
# Handle each of the inference elements
for inferenceElement, value in modelResult.inferences.iteritems():
inferenceLabel = InferenceElement.getLabel(inferenceElement)
# TODO: Right now we assume list inferences are associated with
# The input field metadata
if type(value) in (list, tuple):
# Append input and prediction field meta-info
self.__outputFieldsMeta.extend(self.__getListMetaInfo(inferenceElement))
elif isinstance(value, dict):
self.__outputFieldsMeta.extend(self.__getDictMetaInfo(inferenceElement,
value))
else:
if InferenceElement.getInputElement(inferenceElement):
self.__outputFieldsMeta.append(FieldMetaInfo(name=inferenceLabel+".actual",
type=FieldMetaType.string, special = ''))
self.__outputFieldsMeta.append(FieldMetaInfo(name=inferenceLabel,
type=FieldMetaType.string, special = ''))
if self.__metricNames:
for metricName in self.__metricNames:
metricField = FieldMetaInfo(
name = metricName,
type = FieldMetaType.float,
special = FieldMetaSpecial.none)
self.__outputFieldsMeta.append(metricField)
# Create the inference directory for our experiment
inferenceDir = _FileUtils.createExperimentInferenceDir(self.__experimentDir)
# Consctruct the prediction dataset file path
filename = (self.__label + "." +
opfutils.InferenceType.getLabel(self.__inferenceType) +
".predictionLog.csv")
self.__datasetPath = os.path.join(inferenceDir, filename)
# Create the output dataset
print "OPENING OUTPUT FOR PREDICTION WRITER AT: {0!r}".format(self.__datasetPath)
print "Prediction field-meta: {0!r}".format([tuple(i) for i in self.__outputFieldsMeta])
self.__dataset = FileRecordStream(streamID=self.__datasetPath, write=True,
fields=self.__outputFieldsMeta)
# Copy data from checkpoint cache
if self.__checkpointCache is not None:
self.__checkpointCache.seek(0)
reader = csv.reader(self.__checkpointCache, dialect='excel')
# Skip header row
try:
header = reader.next()
except StopIteration:
print "Empty record checkpoint initializer for {0!r}".format(self.__datasetPath)
else:
assert tuple(self.__dataset.getFieldNames()) == tuple(header), \
"dataset.getFieldNames(): {0!r}; predictionCheckpointFieldNames: {1!r}".format(
tuple(self.__dataset.getFieldNames()), tuple(header))
# Copy the rows from checkpoint
numRowsCopied = 0
while True:
try:
row = reader.next()
except StopIteration:
break
#print "DEBUG: restoring row from checkpoint: %r" % (row,)
self.__dataset.appendRecord(row)
numRowsCopied += 1
self.__dataset.flush()
print "Restored {0:d} rows from checkpoint for {1!r}".format(
numRowsCopied, self.__datasetPath)
# Dispose of our checkpoint cache
self.__checkpointCache.close()
self.__checkpointCache = None
return
def setLoggedMetrics(self, metricNames):
""" Tell the writer which metrics should be written
Parameters:
-----------------------------------------------------------------------
metricsNames: A list of metric lables to be written
"""
if metricNames is None:
self.__metricNames = set([])
else:
self.__metricNames = set(metricNames)
def close(self):
""" [virtual method override] Closes the writer (e.g., close the underlying
file)
"""
if self.__dataset:
self.__dataset.close()
self.__dataset = None
return
def __getListMetaInfo(self, inferenceElement):
""" Get field metadata information for inferences that are of list type
TODO: Right now we assume list inferences are associated with the input field
metadata
"""
fieldMetaInfo = []
inferenceLabel = InferenceElement.getLabel(inferenceElement)
for inputFieldMeta in self.__inputFieldsMeta:
if InferenceElement.getInputElement(inferenceElement):
outputFieldMeta = FieldMetaInfo(
name=inputFieldMeta.name + ".actual",
type=inputFieldMeta.type,
special=inputFieldMeta.special
)
predictionField = FieldMetaInfo(
name=inputFieldMeta.name + "." + inferenceLabel,
type=inputFieldMeta.type,
special=inputFieldMeta.special
)
fieldMetaInfo.append(outputFieldMeta)
fieldMetaInfo.append(predictionField)
return fieldMetaInfo
def __getDictMetaInfo(self, inferenceElement, inferenceDict):
"""Get field metadate information for inferences that are of dict type"""
fieldMetaInfo = []
inferenceLabel = InferenceElement.getLabel(inferenceElement)
if InferenceElement.getInputElement(inferenceElement):
fieldMetaInfo.append(FieldMetaInfo(name=inferenceLabel+".actual",
type=FieldMetaType.string,
special = ''))
keys = sorted(inferenceDict.keys())
for key in keys:
fieldMetaInfo.append(FieldMetaInfo(name=inferenceLabel+"."+str(key),
type=FieldMetaType.string,
special=''))
return fieldMetaInfo
def append(self, modelResult):
""" [virtual method override] Emits a single prediction as input versus
predicted.
modelResult: An opfutils.ModelResult object that contains the model input
and output for the current timestep.
"""
#print "DEBUG: _BasicPredictionWriter: writing modelResult: %r" % (modelResult,)
# If there are no inferences, don't write anything
inferences = modelResult.inferences
hasInferences = False
if inferences is not None:
for value in inferences.itervalues():
hasInferences = hasInferences or (value is not None)
if not hasInferences:
return
if self.__dataset is None:
self.__openDatafile(modelResult)
inputData = modelResult.sensorInput
sequenceReset = int(bool(inputData.sequenceReset))
outputRow = [sequenceReset]
# -----------------------------------------------------------------------
# Write out the raw inputs
rawInput = modelResult.rawInput
for field in self._rawInputNames:
outputRow.append(str(rawInput[field]))
# -----------------------------------------------------------------------
# Write out the inference element info
for inferenceElement, outputVal in inferences.iteritems():
inputElement = InferenceElement.getInputElement(inferenceElement)
if inputElement:
inputVal = getattr(inputData, inputElement)
else:
inputVal = None
if type(outputVal) in (list, tuple):
assert type(inputVal) in (list, tuple, None)
for iv, ov in zip(inputVal, outputVal):
# Write actual
outputRow.append(str(iv))
# Write inferred
outputRow.append(str(ov))
elif isinstance(outputVal, dict):
if inputVal is not None:
# If we have a predicted field, include only that in the actuals
if modelResult.predictedFieldName is not None:
outputRow.append(str(inputVal[modelResult.predictedFieldName]))
else:
outputRow.append(str(inputVal))
for key in sorted(outputVal.keys()):
outputRow.append(str(outputVal[key]))
else:
if inputVal is not None:
outputRow.append(str(inputVal))
outputRow.append(str(outputVal))
metrics = modelResult.metrics
for metricName in self.__metricNames:
outputRow.append(metrics.get(metricName, 0.0))
#print "DEBUG: _BasicPredictionWriter: writing outputRow: %r" % (outputRow,)
self.__dataset.appendRecord(outputRow)
self.__dataset.flush()
return
def checkpoint(self, checkpointSink, maxRows):
""" [virtual method override] Save a checkpoint of the prediction output
stream. The checkpoint comprises up to maxRows of the most recent inference
records.
Parameters:
----------------------------------------------------------------------
checkpointSink: A File-like object where predictions checkpoint data, if
any, will be stored.
maxRows: Maximum number of most recent inference rows
to checkpoint.
"""
checkpointSink.truncate()
if self.__dataset is None:
if self.__checkpointCache is not None:
self.__checkpointCache.seek(0)
shutil.copyfileobj(self.__checkpointCache, checkpointSink)
checkpointSink.flush()
return
else:
# Nothing to checkpoint
return
self.__dataset.flush()
totalDataRows = self.__dataset.getDataRowCount()
if totalDataRows == 0:
# Nothing to checkpoint
return
# Open reader of prediction file (suppress missingValues conversion)
reader = FileRecordStream(self.__datasetPath, missingValues=[])
# Create CSV writer for writing checkpoint rows
writer = csv.writer(checkpointSink)
# Write the header row to checkpoint sink -- just field names
writer.writerow(reader.getFieldNames())
# Determine number of rows to checkpoint
numToWrite = min(maxRows, totalDataRows)
# Skip initial rows to get to the rows that we actually need to checkpoint
numRowsToSkip = totalDataRows - numToWrite
for i in xrange(numRowsToSkip):
reader.next()
# Write the data rows to checkpoint sink
numWritten = 0
while True:
row = reader.getNextRecord()
if row is None:
break;
row = [str(element) for element in row]
#print "DEBUG: _BasicPredictionWriter: checkpointing row: %r" % (row,)
writer.writerow(row)
numWritten +=1
assert numWritten == numToWrite, \
"numWritten ({0!s}) != numToWrite ({1!s})".format(numWritten, numToWrite)
checkpointSink.flush()
return
def generateDataset(aggregationInfo, inputFilename, outputFilename=None):
"""Generate a dataset of aggregated values
Parameters:
----------------------------------------------------------------------------
aggregationInfo: a dictionary that contains the following entries
- fields: a list of pairs. Each pair is a field name and an
aggregation function (e.g. sum). The function will be used to aggregate
multiple values during the aggregation period.
aggregation period: 0 or more of unit=value fields; allowed units are:
[years months] |
[weeks days hours minutes seconds milliseconds microseconds]
NOTE: years and months are mutually-exclusive with the other units.
See getEndTime() and _aggregate() for more details.
Example1: years=1, months=6,
Example2: hours=1, minutes=30,
If none of the period fields are specified or if all that are specified
have values of 0, then aggregation will be suppressed, and the given
inputFile parameter value will be returned.
inputFilename: filename of the input dataset within examples/prediction/data
outputFilename: name for the output file. If not given, a name will be
generated based on the input filename and the aggregation params
retval: Name of the generated output file. This will be the same as the input
file name if no aggregation needed to be performed
If the input file contained a time field, sequence id field or reset field
that were not specified in aggregationInfo fields, those fields will be
added automatically with the following rules:
1. The order will be R, S, T, rest of the fields
2. The aggregation function for all will be to pick the first: lambda x: x[0]
Returns: the path of the aggregated data file if aggregation was performed
(in the same directory as the given input file); if aggregation did not
need to be performed, then the given inputFile argument value is returned.
"""
# Create the input stream
inputFullPath = resource_filename("nupic.datafiles", inputFilename)
inputObj = FileRecordStream(inputFullPath)
# Instantiate the aggregator
aggregator = Aggregator(aggregationInfo=aggregationInfo,
inputFields=inputObj.getFields())
# Is it a null aggregation? If so, just return the input file unmodified
if aggregator.isNullAggregation():
return inputFullPath
# ------------------------------------------------------------------------
# If we were not given an output filename, create one based on the
# aggregation settings
if outputFilename is None:
outputFilename = 'agg_%s' % \
os.path.splitext(os.path.basename(inputFullPath))[0]
timePeriods = 'years months weeks days '\
'hours minutes seconds milliseconds microseconds'
for k in timePeriods.split():
if aggregationInfo.get(k, 0) > 0:
outputFilename += '_%s_%d' % (k, aggregationInfo[k])
outputFilename += '.csv'
outputFilename = os.path.join(os.path.dirname(inputFullPath), outputFilename)
# ------------------------------------------------------------------------
# If some other process already started creating this file, simply
# wait for it to finish and return without doing anything
lockFilePath = outputFilename + '.please_wait'
if os.path.isfile(outputFilename) or \
os.path.isfile(lockFilePath):
while os.path.isfile(lockFilePath):
print('Waiting for %s to be fully written by another process' % \
lockFilePath)
time.sleep(1)
return outputFilename
# Create the lock file
lockFD = open(lockFilePath, 'w')
# -------------------------------------------------------------------------
# Create the output stream
outputObj = FileRecordStream(streamID=outputFilename, write=True,
fields=inputObj.getFields())
# -------------------------------------------------------------------------
# Write all aggregated records to the output
while True:
inRecord = inputObj.getNextRecord()
(aggRecord, aggBookmark) = aggregator.next(inRecord, None)
if aggRecord is None and inRecord is None:
break
if aggRecord is not None:
outputObj.appendRecord(aggRecord)
return outputFilename
def testSimpleMulticlassNetworkPY(self):
# Setup data record stream of fake data (with three categories)
filename = _getTempFileName()
fields = [("timestamp", "datetime", "T"),
("value", "float", ""),
("reset", "int", "R"),
("sid", "int", "S"),
("categories", "list", "C")]
records = (
[datetime(day=1, month=3, year=2010), 0.0, 1, 0, "0"],
[datetime(day=2, month=3, year=2010), 1.0, 0, 0, "1"],
[datetime(day=3, month=3, year=2010), 0.0, 0, 0, "0"],
[datetime(day=4, month=3, year=2010), 1.0, 0, 0, "1"],
[datetime(day=5, month=3, year=2010), 0.0, 0, 0, "0"],
[datetime(day=6, month=3, year=2010), 1.0, 0, 0, "1"],
[datetime(day=7, month=3, year=2010), 0.0, 0, 0, "0"],
[datetime(day=8, month=3, year=2010), 1.0, 0, 0, "1"])
dataSource = FileRecordStream(streamID=filename, write=True, fields=fields)
for r in records:
dataSource.appendRecord(list(r))
# Create the network and get region instances.
net = Network()
net.addRegion("sensor", "py.RecordSensor", "{'numCategories': 3}")
net.addRegion("classifier", "py.SDRClassifierRegion",
"{steps: '0', alpha: 0.001, implementation: 'py'}")
net.link("sensor", "classifier", "UniformLink", "",
srcOutput="dataOut", destInput="bottomUpIn")
net.link("sensor", "classifier", "UniformLink", "",
srcOutput="categoryOut", destInput="categoryIn")
sensor = net.regions["sensor"]
classifier = net.regions["classifier"]
# Setup sensor region encoder and data stream.
dataSource.close()
dataSource = FileRecordStream(filename)
sensorRegion = sensor.getSelf()
sensorRegion.encoder = MultiEncoder()
sensorRegion.encoder.addEncoder(
"value", ScalarEncoder(21, 0.0, 13.0, n=256, name="value"))
sensorRegion.dataSource = dataSource
# Get ready to run.
net.initialize()
# Train the network (by default learning is ON in the classifier, but assert
# anyway) and then turn off learning and turn on inference mode.
self.assertEqual(classifier.getParameter("learningMode"), 1)
net.run(8)
# Test the network on the same data as it trained on; should classify with
# 100% accuracy.
classifier.setParameter("inferenceMode", 1)
classifier.setParameter("learningMode", 0)
# Assert learning is OFF and that the classifier learned the dataset.
self.assertEqual(classifier.getParameter("learningMode"), 0,
"Learning mode is not turned off.")
self.assertEqual(classifier.getParameter("inferenceMode"), 1,
"Inference mode is not turned on.")
# make sure we can access all the parameters with getParameter
self.assertEqual(classifier.getParameter("maxCategoryCount"), 2000)
self.assertAlmostEqual(float(classifier.getParameter("alpha")), 0.001)
self.assertEqual(int(classifier.getParameter("steps")), 0)
self.assertTrue(classifier.getParameter("implementation") == "py")
self.assertEqual(classifier.getParameter("verbosity"), 0)
expectedCats = ([0.0], [1.0], [0.0], [1.0], [0.0], [1.0], [0.0], [1.0],)
dataSource.rewind()
for i in xrange(8):
net.run(1)
inferredCats = classifier.getOutputData("categoriesOut")
self.assertSequenceEqual(expectedCats[i], inferredCats.tolist(),
"Classififer did not infer expected category "
"for record number {}.".format(i))
# Close data stream, delete file.
dataSource.close()
os.remove(filename)
def testBasic(self):
"""Runs basic FileRecordStream tests."""
filename = _getTempFileName()
# Write a standard file
fields = [('name', 'string', ''),
('timestamp', 'datetime', 'T'),
('integer', 'int', ''),
('real', 'float', ''),
('reset', 'int', 'R'),
('sid', 'string', 'S'),
('categoryField', 'int', 'C'),]
fieldNames = ['name', 'timestamp', 'integer', 'real', 'reset', 'sid',
'categoryField']
print 'Creating temp file:', filename
s = FileRecordStream(streamID=filename, write=True, fields=fields)
self.assertTrue(s.getDataRowCount() == 0)
# Records
records = (
['rec_1', datetime(day=1, month=3, year=2010), 5, 6.5, 1, 'seq-1', 10],
['rec_2', datetime(day=2, month=3, year=2010), 8, 7.5, 0, 'seq-1', 11],
['rec_3', datetime(day=3, month=3, year=2010), 12, 8.5, 0, 'seq-1', 12])
self.assertTrue(s.getFields() == fields)
self.assertTrue(s.getNextRecordIdx() == 0)
print 'Writing records ...'
for r in records:
print list(r)
s.appendRecord(list(r))
self.assertTrue(s.getDataRowCount() == 3)
recordsBatch = (
['rec_4', datetime(day=4, month=3, year=2010), 2, 9.5, 1, 'seq-1', 13],
['rec_5', datetime(day=5, month=3, year=2010), 6, 10.5, 0, 'seq-1', 14],
['rec_6', datetime(day=6, month=3, year=2010), 11, 11.5, 0, 'seq-1', 15])
print 'Adding batch of records...'
for rec in recordsBatch:
print rec
s.appendRecords(recordsBatch)
self.assertTrue(s.getDataRowCount() == 6)
s.close()
# Read the standard file
s = FileRecordStream(filename)
self.assertTrue(s.getDataRowCount() == 6)
self.assertTrue(s.getFieldNames() == fieldNames)
# Note! this is the number of records read so far
self.assertTrue(s.getNextRecordIdx() == 0)
readStats = s.getStats()
print 'Got stats:', readStats
expectedStats = {
'max': [None, None, 12, 11.5, 1, None, 15],
'min': [None, None, 2, 6.5, 0, None, 10]
}
self.assertTrue(readStats == expectedStats)
readRecords = []
print 'Reading records ...'
while True:
r = s.getNextRecord()
print r
if r is None:
break
readRecords.append(r)
allRecords = records + recordsBatch
for r1, r2 in zip(allRecords, readRecords):
print 'Expected:', r1
print 'Read :', r2
self.assertTrue(r1 == r2)
s.close()
def testSimpleMulticlassNetwork(self):
# Setup data record stream of fake data (with three categories)
filename = _getTempFileName()
fields = [("timestamp", "datetime", "T"),
("value", "float", ""),
("reset", "int", "R"),
("sid", "int", "S"),
("categories", "list", "C")]
records = (
[datetime(day=1, month=3, year=2010), 0.0, 1, 0, ""],
[datetime(day=2, month=3, year=2010), 1.0, 0, 0, "1 2"],
[datetime(day=3, month=3, year=2010), 1.0, 0, 0, "1 2"],
[datetime(day=4, month=3, year=2010), 2.0, 0, 0, "0"],
[datetime(day=5, month=3, year=2010), 3.0, 0, 0, "1 2"],
[datetime(day=6, month=3, year=2010), 5.0, 0, 0, "1 2"],
[datetime(day=7, month=3, year=2010), 8.0, 0, 0, "0"],
[datetime(day=8, month=3, year=2010), 13.0, 0, 0, "1 2"])
dataSource = FileRecordStream(streamID=filename, write=True, fields=fields)
for r in records:
dataSource.appendRecord(list(r))
# Create the network and get region instances.
net = Network()
net.addRegion("sensor", "py.RecordSensor", "{'numCategories': 3}")
net.addRegion("classifier","py.KNNClassifierRegion",
"{'k': 2,'distThreshold': 0,'maxCategoryCount': 3}")
net.link("sensor", "classifier", "UniformLink", "",
srcOutput = "dataOut", destInput = "bottomUpIn")
net.link("sensor", "classifier", "UniformLink", "",
srcOutput = "categoryOut", destInput = "categoryIn")
sensor = net.regions["sensor"]
classifier = net.regions["classifier"]
# Setup sensor region encoder and data stream.
dataSource.close()
dataSource = FileRecordStream(filename)
sensorRegion = sensor.getSelf()
sensorRegion.encoder = MultiEncoder()
sensorRegion.encoder.addEncoder(
"value", ScalarEncoder(21, 0.0, 13.0, n=256, name="value"))
sensorRegion.dataSource = dataSource
# Get ready to run.
net.initialize()
# Train the network (by default learning is ON in the classifier, but assert
# anyway) and then turn off learning and turn on inference mode.
self.assertEqual(classifier.getParameter("learningMode"), 1)
net.run(8)
classifier.setParameter("inferenceMode", 1)
classifier.setParameter("learningMode", 0)
# Assert learning is OFF and that the classifier learned the dataset.
self.assertEqual(classifier.getParameter("learningMode"), 0,
"Learning mode is not turned off.")
self.assertEqual(classifier.getParameter("inferenceMode"), 1,
"Inference mode is not turned on.")
self.assertEqual(classifier.getParameter("categoryCount"), 3,
"The classifier should count three total categories.")
# classififer learns 12 patterns b/c there are 12 categories amongst the
# records:
self.assertEqual(classifier.getParameter("patternCount"), 12,
"The classifier should've learned 12 samples in total.")
# Test the network on the same data as it trained on; should classify with
# 100% accuracy.
expectedCats = ([0.0, 0.5, 0.5],
[0.0, 0.5, 0.5],
[0.0, 0.5, 0.5],
[0.5, 0.5, 0.0],
[0.0, 0.5, 0.5],
[0.0, 0.5, 0.5],
[0.5, 0.5, 0.0],
[0.0, 0.5, 0.5])
dataSource.rewind()
for i in xrange(8):
net.run(1)
inferredCats = classifier.getOutputData("categoriesOut")
self.assertSequenceEqual(expectedCats[i], inferredCats.tolist(),
"Classififer did not infer expected category probabilites for record "
"number {}.".format(i))
# Close data stream, delete file.
dataSource.close()
os.remove(filename)
class StreamReader(RecordStreamIface):
"""
Implements a stream reader. This is a high level class that owns one or more
underlying implementations of a
:class:`~nupic.data.record_stream.RecordStreamIface`. Each
:class:`~nupic.data.record_stream.RecordStreamIface` implements the raw
reading of records from the record store (which could be a file, hbase table
or something else).
In the future, we will support joining of two or more
:class:`~nupic.data.record_stream.RecordStreamIface`'s (which is why the
``streamDef`` accepts a list of 'stream' elements), but for now only 1 source
is supported.
The class also implements aggregation of the (in the future) joined records
from the sources.
This module parses the stream definition (as defined in
``/src/nupic/frameworks/opf/jsonschema/stream_def.json``), creates the
:class:`~nupic.data.record_stream.RecordStreamIface` for each source
('stream' element) defined in the stream def, performs aggregation, and
returns each record in the correct format according to the desired column
names specified in the streamDef.
This class implements the :class:`~nupic.data.record_stream.RecordStreamIface`
interface and thus can be used in place of a raw record stream.
This is an example streamDef:
.. code-block:: python
{
'version': 1
'info': 'test_hotgym',
'streams': [
{'columns': [u'*'],
'info': u'hotGym.csv',
'last_record': 4000,
'source': u'file://extra/hotgym/hotgym.csv'}.
],
'timeField': 'timestamp',
'aggregation': {
'hours': 1,
'fields': [
('timestamp', 'first'),
('gym', 'first'),
('consumption', 'sum')
],
}
}
:param streamDef: The stream definition, potentially containing multiple
sources (not supported yet). See
``src//nupic/frameworks/opf/jsonschema/stream_def.json`` for the
format of this dict
:param bookmark: Bookmark to start reading from. This overrides the
first_record field of the streamDef if provided.
:param saveOutput: If true, save the output to a csv file in a temp
directory. The path to the generated file can be found in the log
output.
:param isBlocking: should read operation block *forever* if the next row of
data is not available, but the stream is not marked as 'completed'
yet?
:param maxTimeout: if isBlocking is False, max seconds to wait for more data
before timing out; ignored when isBlocking is True.
:param eofOnTimeout: If True and we get a read timeout (isBlocking must be
False to get read timeouts), assume we've reached the end of the
input and produce the last aggregated record, if one can be
completed.
"""
def __init__(self, streamDef, bookmark=None, saveOutput=False,
isBlocking=True, maxTimeout=0, eofOnTimeout=False):
# Call superclass constructor
super(StreamReader, self).__init__()
loggerPrefix = 'com.numenta.nupic.data.StreamReader'
self._logger = logging.getLogger(loggerPrefix)
json_helpers.validate(streamDef,
schemaPath=pkg_resources.resource_filename(
jsonschema.__name__, "stream_def.json"))
assert len(streamDef['streams']) == 1, "Only 1 source stream is supported"
# Save constructor args
sourceDict = streamDef['streams'][0]
self._recordCount = 0
self._eofOnTimeout = eofOnTimeout
self._logger.debug('Reading stream with the def: %s', sourceDict)
# Dictionary to store record statistics (min and max of scalars for now)
self._stats = None
# ---------------------------------------------------------------------
# Get the stream definition params
# Limiting window of the stream. It would not return any records until
# 'first_record' ID is read (or very first with the ID above that). The
# stream will return EOS once it reads record with ID 'last_record' or
# above (NOTE: the name 'lastRecord' is misleading because it is NOT
# inclusive).
firstRecordIdx = sourceDict.get('first_record', None)
self._sourceLastRecordIdx = sourceDict.get('last_record', None)
# If a bookmark was given, then override first_record from the stream
# definition.
if bookmark is not None:
firstRecordIdx = None
# Column names must be provided in the streamdef json
# Special case is ['*'], meaning all available names from the record stream
self._streamFieldNames = sourceDict.get('columns', None)
if self._streamFieldNames != None and self._streamFieldNames[0] == '*':
self._needFieldsFiltering = False
else:
self._needFieldsFiltering = True
# Types must be specified in streamdef json, or in case of the
# file_recod_stream types could be implicit from the file
streamFieldTypes = sourceDict.get('types', None)
self._logger.debug('Types from the def: %s', streamFieldTypes)
# Validate that all types are valid
if streamFieldTypes is not None:
for dataType in streamFieldTypes:
assert FieldMetaType.isValid(dataType)
# Reset, sequence and time fields might be provided by streamdef json
streamResetFieldName = streamDef.get('resetField', None)
streamTimeFieldName = streamDef.get('timeField', None)
streamSequenceFieldName = streamDef.get('sequenceIdField', None)
self._logger.debug('r, t, s fields: %s, %s, %s', streamResetFieldName,
streamTimeFieldName,
streamSequenceFieldName)
# =======================================================================
# Open up the underlying record store
dataUrl = sourceDict.get('source', None)
assert dataUrl is not None
self._recordStore = self._openStream(dataUrl, isBlocking, maxTimeout,
bookmark, firstRecordIdx)
assert self._recordStore is not None
# =======================================================================
# Prepare the data structures we need for returning just the fields
# the caller wants from each record
recordStoreFields = self._recordStore.getFields()
self._recordStoreFieldNames = self._recordStore.getFieldNames()
if not self._needFieldsFiltering:
self._streamFieldNames = self._recordStoreFieldNames
# Build up the field definitions for each field. This is a list of tuples
# of (name, type, special)
self._streamFields = []
for dstIdx, name in enumerate(self._streamFieldNames):
if name not in self._recordStoreFieldNames:
raise RuntimeError("The column '%s' from the stream definition "
"is not present in the underlying stream which has the following "
"columns: %s" % (name, self._recordStoreFieldNames))
fieldIdx = self._recordStoreFieldNames.index(name)
fieldType = recordStoreFields[fieldIdx].type
fieldSpecial = recordStoreFields[fieldIdx].special
# If the types or specials were defined in the stream definition,
# then override what was found in the record store
if streamFieldTypes is not None:
fieldType = streamFieldTypes[dstIdx]
if streamResetFieldName is not None and streamResetFieldName == name:
fieldSpecial = FieldMetaSpecial.reset
if streamTimeFieldName is not None and streamTimeFieldName == name:
fieldSpecial = FieldMetaSpecial.timestamp
if (streamSequenceFieldName is not None and
streamSequenceFieldName == name):
fieldSpecial = FieldMetaSpecial.sequence
self._streamFields.append(FieldMetaInfo(name, fieldType, fieldSpecial))
# ========================================================================
# Create the aggregator which will handle aggregation of records before
# returning them.
self._aggregator = Aggregator(
aggregationInfo=streamDef.get('aggregation', None),
inputFields=recordStoreFields,
timeFieldName=streamDef.get('timeField', None),
sequenceIdFieldName=streamDef.get('sequenceIdField', None),
resetFieldName=streamDef.get('resetField', None))
# We rely on the aggregator to tell us the bookmark of the last raw input
# that contributed to the aggregated record
self._aggBookmark = None
# Compute the aggregation period in terms of months and seconds
if 'aggregation' in streamDef:
self._aggMonthsAndSeconds = nupic.support.aggregationToMonthsSeconds(
streamDef.get('aggregation'))
else:
self._aggMonthsAndSeconds = None
# ========================================================================
# Are we saving the generated output to a csv?
if saveOutput:
tmpDir = tempfile.mkdtemp()
outFilename = os.path.join(tmpDir, "generated_output.csv")
self._logger.info("StreamReader: Saving generated records to: '%s'" %
outFilename)
self._writer = FileRecordStream(streamID=outFilename,
write=True,
fields=self._streamFields)
else:
self._writer = None
@staticmethod
def _openStream(dataUrl,
isBlocking, # pylint: disable=W0613
maxTimeout, # pylint: disable=W0613
bookmark,
firstRecordIdx):
"""Open the underlying file stream
This only supports 'file://' prefixed paths.
:returns: record stream instance
:rtype: FileRecordStream
"""
filePath = dataUrl[len(FILE_PREF):]
if not os.path.isabs(filePath):
filePath = os.path.join(os.getcwd(), filePath)
return FileRecordStream(streamID=filePath,
write=False,
bookmark=bookmark,
firstRecord=firstRecordIdx)
def close(self):
""" Close the stream
"""
return self._recordStore.close()
def getNextRecord(self):
""" Returns combined data from all sources (values only).
:returns: None on EOF; empty sequence on timeout.
"""
# Keep reading from the raw input till we get enough for an aggregated
# record
while True:
# Reached EOF due to lastRow constraint?
if self._sourceLastRecordIdx is not None and \
self._recordStore.getNextRecordIdx() >= self._sourceLastRecordIdx:
preAggValues = None # indicates EOF
bookmark = self._recordStore.getBookmark()
else:
# Get the raw record and bookmark
preAggValues = self._recordStore.getNextRecord()
bookmark = self._recordStore.getBookmark()
if preAggValues == (): # means timeout error occurred
if self._eofOnTimeout:
preAggValues = None # act as if we got EOF
else:
return preAggValues # Timeout indicator
self._logger.debug('Read source record #%d: %r',
self._recordStore.getNextRecordIdx()-1, preAggValues)
# Perform aggregation
(fieldValues, aggBookmark) = self._aggregator.next(preAggValues, bookmark)
# Update the aggregated record bookmark if we got a real record back
if fieldValues is not None:
self._aggBookmark = aggBookmark
# Reached EOF?
if preAggValues is None and fieldValues is None:
return None
# Return it if we have a record
if fieldValues is not None:
break
# Do we need to re-order the fields in the record?
if self._needFieldsFiltering:
values = []
srcDict = dict(list(zip(self._recordStoreFieldNames, fieldValues)))
for name in self._streamFieldNames:
values.append(srcDict[name])
fieldValues = values
# Write to debug output?
if self._writer is not None:
self._writer.appendRecord(fieldValues)
self._recordCount += 1
self._logger.debug('Returning aggregated record #%d from getNextRecord(): '
'%r. Bookmark: %r',
self._recordCount-1, fieldValues, self._aggBookmark)
return fieldValues
def getDataRowCount(self):
"""
Iterates through stream to calculate total records after aggregation.
This will alter the bookmark state.
"""
inputRowCountAfterAggregation = 0
while True:
record = self.getNextRecord()
if record is None:
return inputRowCountAfterAggregation
inputRowCountAfterAggregation += 1
if inputRowCountAfterAggregation > 10000:
raise RuntimeError('No end of datastream found.')
def getNextRecordIdx(self):
"""
:returns: the index of the record that will be read next from
:meth:`getNextRecord`.
"""
return self._recordCount
def recordsExistAfter(self, bookmark):
"""
:returns: True if there are records left after the bookmark.
"""
return self._recordStore.recordsExistAfter(bookmark)
def getAggregationMonthsAndSeconds(self):
""" Returns the aggregation period of the record stream as a dict
containing 'months' and 'seconds'. The months is always an integer and
seconds is a floating point. Only one is allowed to be non-zero at a
time.
Will return the aggregation period from this call. This call is
used by the :meth:`nupic.data.record_stream.RecordStream.getNextRecordDict`
method to assign a record number to a record given its timestamp and the
aggregation interval.
:returns: aggregationPeriod (as a dict) where:
- ``months``: number of months in aggregation period
- ``seconds``: number of seconds in aggregation period
(as a float)
"""
return self._aggMonthsAndSeconds
def appendRecord(self, record):
raise RuntimeError("Not implemented in StreamReader")
def appendRecords(self, records, progressCB=None):
raise RuntimeError("Not implemented in StreamReader")
def seekFromEnd(self, numRecords):
raise RuntimeError("Not implemented in StreamReader")
def getFieldNames(self):
"""
Returns all fields in all inputs (list of plain names).
.. note:: currently, only one input is supported
"""
return [f.name for f in self._streamFields]
def getFields(self):
"""
:returns: a sequence of :class:`nupic.data.fieldmeta.FieldMetaInfo` for each
field in the stream.
"""
return self._streamFields
def getBookmark(self):
"""
:returns: a bookmark to the current position
"""
return self._aggBookmark
def clearStats(self):
""" Resets stats collected so far.
"""
self._recordStore.clearStats()
def getStats(self):
"""
TODO: This method needs to be enhanced to get the stats on the *aggregated*
records.
:returns: stats (like min and max values of the fields).
"""
# The record store returns a dict of stats, each value in this dict is
# a list with one item per field of the record store
# {
# 'min' : [f1_min, f2_min, f3_min],
# 'max' : [f1_max, f2_max, f3_max]
# }
recordStoreStats = self._recordStore.getStats()
# We need to convert each item to represent the fields of the *stream*
streamStats = dict()
for (key, values) in list(recordStoreStats.items()):
fieldStats = dict(list(zip(self._recordStoreFieldNames, values)))
streamValues = []
for name in self._streamFieldNames:
streamValues.append(fieldStats[name])
streamStats[key] = streamValues
return streamStats
def getError(self):
"""
:returns: errors saved in the stream.
"""
return self._recordStore.getError()
def setError(self, error):
""" Saves specified error in the stream.
:param error: to save
"""
self._recordStore.setError(error)
def isCompleted(self):
"""
:returns: True if all records have been read.
"""
return self._recordStore.isCompleted()
def setCompleted(self, completed=True):
"""
Marks the stream completed (True or False)
:param completed: (bool) is completed or not
"""
# CSV file is always considered completed, nothing to do
self._recordStore.setCompleted(completed)
def setTimeout(self, timeout):
"""Set the read timeout.
:param timeout: (float or int) timeout length
"""
self._recordStore.setTimeout(timeout)
def flush(self):
raise RuntimeError("Not implemented in StreamReader")
def testMissingValues(self):
print "Beginning Missing Data test..."
filename = _getTempFileName()
# Some values missing of each type
# read dataset from disk, retrieve values
# string should return empty string, numeric types sentinelValue
print 'Creating tempfile:', filename
# write dataset to disk with float, int, and string fields
fields = [
FieldMetaInfo('timestamp', FieldMetaType.datetime,
FieldMetaSpecial.timestamp),
FieldMetaInfo('name', FieldMetaType.string, FieldMetaSpecial.none),
FieldMetaInfo('integer', FieldMetaType.integer,
FieldMetaSpecial.none),
FieldMetaInfo('real', FieldMetaType.float, FieldMetaSpecial.none)
]
s = FileRecordStream(streamID=filename, write=True, fields=fields)
# Records
records = ([datetime(day=1, month=3, year=2010), 'rec_1', 5, 6.5], [
datetime(day=2, month=3, year=2010), '', 8, 7.5
], [datetime(day=3, month=3, year=2010), 'rec_3', '', 8.5], [
datetime(day=4, month=3, year=2010), 'rec_4', 12, ''
], [datetime(day=5, month=3, year=2010), 'rec_5', -87657496599, 6.5], [
datetime(day=6, month=3, year=2010), 'rec_6', 12, -87657496599
], [datetime(day=6, month=3, year=2010),
str(-87657496599), 12, 6.5])
for r in records:
s.appendRecord(list(r))
s.close()
# Read the standard file
s = FileRecordStream(streamID=filename, write=False)
fieldsRead = s.getFields()
self.assertEqual(fields, fieldsRead)
recordsRead = []
while True:
r = s.getNextRecord()
if r is None:
break
print 'Reading record ...'
print r
recordsRead.append(r)
# sort the records by date, so we know for sure which is which
sorted(recordsRead, key=lambda rec: rec[0])
# empty string
self.assertEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[1][1])
# missing int
self.assertEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[2][2])
# missing float
self.assertEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[3][3])
# sentinel value in input handled correctly for int field
self.assertNotEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[4][2])
# sentinel value in input handled correctly for float field
self.assertNotEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[5][3])
# sentinel value in input handled correctly for string field
# this should leave the string as-is, since a missing string
# is encoded not with a sentinel value but with an empty string
self.assertNotEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[6][1])
def testSimpleMulticlassNetwork(self):
# Setup data record stream of fake data (with three categories)
filename = _getTempFileName()
fields = [("timestamp", "datetime", "T"), ("value", "float", ""),
("reset", "int", "R"), ("sid", "int", "S"),
("categories", "list", "C")]
records = ([datetime(day=1, month=3, year=2010), 0.0, 1, 0, ""], [
datetime(day=2, month=3, year=2010), 1.0, 0, 0, "1 2"
], [datetime(day=3, month=3, year=2010), 1.0, 0, 0,
"1 2"], [datetime(day=4, month=3, year=2010), 2.0, 0, 0, "0"], [
datetime(day=5, month=3, year=2010), 3.0, 0, 0, "1 2"
], [datetime(day=6, month=3, year=2010), 5.0, 0, 0,
"1 2"], [datetime(day=7, month=3, year=2010), 8.0, 0, 0, "0"],
[datetime(day=8, month=3, year=2010), 13.0, 0, 0, "1 2"])
dataSource = FileRecordStream(streamID=filename,
write=True,
fields=fields)
for r in records:
dataSource.appendRecord(list(r))
# Create the network and get region instances.
net = Network()
net.addRegion("sensor", "py.RecordSensor", "{'numCategories': 3}")
net.addRegion("classifier", "py.KNNClassifierRegion",
"{'k': 2,'distThreshold': 0,'maxCategoryCount': 3}")
net.link("sensor",
"classifier",
"UniformLink",
"",
srcOutput="dataOut",
destInput="bottomUpIn")
net.link("sensor",
"classifier",
"UniformLink",
"",
srcOutput="categoryOut",
destInput="categoryIn")
sensor = net.regions["sensor"]
classifier = net.regions["classifier"]
# Setup sensor region encoder and data stream.
dataSource.close()
dataSource = FileRecordStream(filename)
sensorRegion = sensor.getSelf()
sensorRegion.encoder = MultiEncoder()
sensorRegion.encoder.addEncoder(
"value", ScalarEncoder(21, 0.0, 13.0, n=256, name="value"))
sensorRegion.dataSource = dataSource
# Get ready to run.
net.initialize()
# Train the network (by default learning is ON in the classifier, but assert
# anyway) and then turn off learning and turn on inference mode.
self.assertEqual(classifier.getParameter("learningMode"), 1)
net.run(8)
classifier.setParameter("inferenceMode", 1)
classifier.setParameter("learningMode", 0)
# Assert learning is OFF and that the classifier learned the dataset.
self.assertEqual(classifier.getParameter("learningMode"), 0,
"Learning mode is not turned off.")
self.assertEqual(classifier.getParameter("inferenceMode"), 1,
"Inference mode is not turned on.")
self.assertEqual(
classifier.getParameter("categoryCount"), 3,
"The classifier should count three total categories.")
# classififer learns 12 patterns b/c there are 12 categories amongst the
# records:
self.assertEqual(
classifier.getParameter("patternCount"), 12,
"The classifier should've learned 12 samples in total.")
# Test the network on the same data as it trained on; should classify with
# 100% accuracy.
expectedCats = ([0.0, 0.5, 0.5], [0.0, 0.5, 0.5], [0.0, 0.5, 0.5],
[0.5, 0.5, 0.0], [0.0, 0.5,
0.5], [0.0, 0.5,
0.5], [0.5, 0.5,
0.0], [0.0, 0.5, 0.5])
dataSource.rewind()
for i in xrange(8):
net.run(1)
inferredCats = classifier.getOutputData("categoriesOut")
self.assertSequenceEqual(
expectedCats[i], inferredCats.tolist(),
"Classififer did not infer expected category probabilites for record "
"number {}.".format(i))
# Close data stream, delete file.
dataSource.close()
os.remove(filename)
def _generateSimple(filename="simple.csv", numSequences=1, elementsPerSeq=3,
numRepeats=10):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences.
At the end of the dataset, we introduce missing records so that test
code can insure that the model didn't get confused by them.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('timestamp', 'datetime', 'T'),
('field1', 'string', ''),
('field2', 'float', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i*elementsPerSeq, (i+1)*elementsPerSeq)]
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
# Put 1 hour between each record
timestamp = datetime.datetime(year=2012, month=1, day=1, hour=0, minute=0,
second=0)
timeDelta = datetime.timedelta(hours=1)
# Write out the sequences without missing records
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
# Now, write some out with missing records
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for i,x in enumerate(seq):
if i != 1:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for i,x in enumerate(seq):
if i != 1:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
# Write out some more of the sequences *without* missing records
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
outFile.close()
def _generateOverlapping(filename="overlap.csv",
numSequences=2,
elementsPerSeq=3,
numRepeats=10,
hub=[0, 1],
hubOffset=1,
resets=False):
""" Generate a temporal dataset containing sequences that overlap one or more
elements with other sequences.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
hub: sub-sequence to place within each other sequence
hubOffset: where, within each sequence, to place the hub
resets: if True, turn on reset at start of each sequence
"""
# Check for conflicts in arguments
assert (hubOffset + len(hub) <= elementsPerSeq)
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('reset', 'int', 'R'), ('category', 'string', ''),
('field1', 'string', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences with the hub in the middle
sequences = []
nextElemIdx = max(hub) + 1
for _ in range(numSequences):
seq = []
for j in range(hubOffset):
seq.append(nextElemIdx)
nextElemIdx += 1
for j in hub:
seq.append(j)
j = hubOffset + len(hub)
while j < elementsPerSeq:
seq.append(nextElemIdx)
nextElemIdx += 1
j += 1
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for _ in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
for seqIdx in seqIdxs:
reset = int(resets)
seq = sequences[seqIdx]
for (x) in seq:
outFile.appendRecord([reset, str(seqIdx), str(x)])
reset = 0
outFile.close()
def testSimpleMulticlassNetworkPY(self):
# Setup data record stream of fake data (with three categories)
filename = _getTempFileName()
fields = [("timestamp", "datetime", "T"), ("value", "float", ""),
("reset", "int", "R"), ("sid", "int", "S"),
("categories", "list", "C")]
records = ([datetime(day=1, month=3, year=2010), 0.0, 1, 0, "0"], [
datetime(day=2, month=3, year=2010), 1.0, 0, 0, "1"
], [datetime(day=3, month=3, year=2010), 0.0, 0, 0,
"0"], [datetime(day=4, month=3, year=2010), 1.0, 0, 0,
"1"], [datetime(day=5, month=3, year=2010), 0.0, 0, 0, "0"],
[datetime(day=6, month=3, year=2010), 1.0, 0, 0, "1"
], [datetime(day=7, month=3, year=2010), 0.0, 0, 0, "0"],
[datetime(day=8, month=3, year=2010), 1.0, 0, 0, "1"])
dataSource = FileRecordStream(streamID=filename,
write=True,
fields=fields)
for r in records:
dataSource.appendRecord(list(r))
# Create the network and get region instances.
net = Network()
net.addRegion("sensor", "py.RecordSensor", "{'numCategories': 3}")
net.addRegion("classifier", "py.SDRClassifierRegion",
"{steps: '0', alpha: 0.001, implementation: 'py'}")
net.link("sensor",
"classifier",
"UniformLink",
"",
srcOutput="dataOut",
destInput="bottomUpIn")
net.link("sensor",
"classifier",
"UniformLink",
"",
srcOutput="categoryOut",
destInput="categoryIn")
sensor = net.regions["sensor"]
classifier = net.regions["classifier"]
# Setup sensor region encoder and data stream.
dataSource.close()
dataSource = FileRecordStream(filename)
sensorRegion = sensor.getSelf()
sensorRegion.encoder = MultiEncoder()
sensorRegion.encoder.addEncoder(
"value", ScalarEncoder(21, 0.0, 13.0, n=256, name="value"))
sensorRegion.dataSource = dataSource
# Get ready to run.
net.initialize()
# Train the network (by default learning is ON in the classifier, but assert
# anyway) and then turn off learning and turn on inference mode.
self.assertEqual(classifier.getParameter("learningMode"), 1)
net.run(8)
# Test the network on the same data as it trained on; should classify with
# 100% accuracy.
classifier.setParameter("inferenceMode", 1)
classifier.setParameter("learningMode", 0)
# Assert learning is OFF and that the classifier learned the dataset.
self.assertEqual(classifier.getParameter("learningMode"), 0,
"Learning mode is not turned off.")
self.assertEqual(classifier.getParameter("inferenceMode"), 1,
"Inference mode is not turned on.")
# make sure we can access all the parameters with getParameter
self.assertEqual(classifier.getParameter("maxCategoryCount"), 2000)
self.assertAlmostEqual(float(classifier.getParameter("alpha")), 0.001)
self.assertEqual(int(classifier.getParameter("steps")), 0)
self.assertTrue(classifier.getParameter("implementation") == "py")
self.assertEqual(classifier.getParameter("verbosity"), 0)
expectedCats = (
[0.0],
[1.0],
[0.0],
[1.0],
[0.0],
[1.0],
[0.0],
[1.0],
)
dataSource.rewind()
for i in xrange(8):
net.run(1)
inferredCats = classifier.getOutputData("categoriesOut")
self.assertSequenceEqual(
expectedCats[i], inferredCats.tolist(),
"Classififer did not infer expected category "
"for record number {}.".format(i))
# Close data stream, delete file.
dataSource.close()
os.remove(filename)
def generateDataset(aggregationInfo, inputFilename, outputFilename=None):
"""Generate a dataset of aggregated values
Parameters:
----------------------------------------------------------------------------
aggregationInfo: a dictionary that contains the following entries
- fields: a list of pairs. Each pair is a field name and an
aggregation function (e.g. sum). The function will be used to aggregate
multiple values during the aggregation period.
aggregation period: 0 or more of unit=value fields; allowed units are:
[years months] |
[weeks days hours minutes seconds milliseconds microseconds]
NOTE: years and months are mutually-exclusive with the other units.
See getEndTime() and _aggregate() for more details.
Example1: years=1, months=6,
Example2: hours=1, minutes=30,
If none of the period fields are specified or if all that are specified
have values of 0, then aggregation will be suppressed, and the given
inputFile parameter value will be returned.
inputFilename: filename (or relative path form NTA_DATA_PATH) of
the input dataset
outputFilename: name for the output file. If not given, a name will be
generated based on the input filename and the aggregation params
retval: Name of the generated output file. This will be the same as the input
file name if no aggregation needed to be performed
If the input file contained a time field, sequence id field or reset field
that were not specified in aggregationInfo fields, those fields will be
added automatically with the following rules:
1. The order will be R, S, T, rest of the fields
2. The aggregation function for all will be to pick the first: lambda x: x[0]
Returns: the path of the aggregated data file if aggregation was performed
(in the same directory as the given input file); if aggregation did not
need to be performed, then the given inputFile argument value is returned.
"""
# Create the input stream
inputFullPath = findDataset(inputFilename)
inputObj = FileRecordStream(inputFullPath)
# Instantiate the aggregator
aggregator = Aggregator(aggregationInfo=aggregationInfo,
inputFields=inputObj.getFields())
# Is it a null aggregation? If so, just return the input file unmodified
if aggregator.isNullAggregation():
return inputFullPath
# ------------------------------------------------------------------------
# If we were not given an output filename, create one based on the
# aggregation settings
if outputFilename is None:
outputFilename = 'agg_%s' % \
os.path.splitext(os.path.basename(inputFullPath))[0]
timePeriods = 'years months weeks days '\
'hours minutes seconds milliseconds microseconds'
for k in timePeriods.split():
if aggregationInfo.get(k, 0) > 0:
outputFilename += '_%s_%d' % (k, aggregationInfo[k])
outputFilename += '.csv'
outputFilename = os.path.join(os.path.dirname(inputFullPath), outputFilename)
# ------------------------------------------------------------------------
# If some other process already started creating this file, simply
# wait for it to finish and return without doing anything
lockFilePath = outputFilename + '.please_wait'
if os.path.isfile(outputFilename) or \
os.path.isfile(lockFilePath):
while os.path.isfile(lockFilePath):
print 'Waiting for %s to be fully written by another process' % \
lockFilePath
time.sleep(1)
return outputFilename
# Create the lock file
lockFD = open(lockFilePath, 'w')
# -------------------------------------------------------------------------
# Create the output stream
outputObj = FileRecordStream(streamID=outputFilename, write=True,
fields=inputObj.getFields())
# -------------------------------------------------------------------------
# Write all aggregated records to the output
while True:
inRecord = inputObj.getNextRecord()
(aggRecord, aggBookmark) = aggregator.next(inRecord, None)
if aggRecord is None and inRecord is None:
break
if aggRecord is not None:
outputObj.appendRecord(aggRecord)
return outputFilename
def _generateSimple(filename="simple.csv",
numSequences=1,
elementsPerSeq=3,
numRepeats=10):
""" Generate a simple dataset. This contains a bunch of non-overlapping
sequences.
At the end of the dataset, we introduce missing records so that test
code can insure that the model didn't get confused by them.
Parameters:
----------------------------------------------------
filename: name of the file to produce, including extension. It will
be created in a 'datasets' sub-directory within the
directory containing this script.
numSequences: how many sequences to generate
elementsPerSeq: length of each sequence
numRepeats: how many times to repeat each sequence in the output
"""
# Create the output file
scriptDir = os.path.dirname(__file__)
pathname = os.path.join(scriptDir, 'datasets', filename)
print "Creating %s..." % (pathname)
fields = [('timestamp', 'datetime', 'T'), ('field1', 'string', ''),
('field2', 'float', '')]
outFile = FileRecordStream(pathname, write=True, fields=fields)
# Create the sequences
sequences = []
for i in range(numSequences):
seq = [x for x in range(i * elementsPerSeq, (i + 1) * elementsPerSeq)]
sequences.append(seq)
# Write out the sequences in random order
seqIdxs = []
for i in range(numRepeats):
seqIdxs += range(numSequences)
random.shuffle(seqIdxs)
# Put 1 hour between each record
timestamp = datetime.datetime(year=2012,
month=1,
day=1,
hour=0,
minute=0,
second=0)
timeDelta = datetime.timedelta(hours=1)
# Write out the sequences without missing records
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
# Now, write some out with missing records
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for i, x in enumerate(seq):
if i != 1:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for i, x in enumerate(seq):
if i != 1:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
# Write out some more of the sequences *without* missing records
for seqIdx in seqIdxs:
seq = sequences[seqIdx]
for x in seq:
outFile.appendRecord([timestamp, str(x), x])
timestamp += timeDelta
outFile.close()
def testMissingValues(self):
print "Beginning Missing Data test..."
filename = _getTempFileName()
# Some values missing of each type
# read dataset from disk, retrieve values
# string should return empty string, numeric types sentinelValue
print "Creating tempfile:", filename
# write dataset to disk with float, int, and string fields
fields = [("timestamp", "datetime", "T"), ("name", "string", ""), ("integer", "int", ""), ("real", "float", "")]
s = FileRecordStream(streamID=filename, write=True, fields=fields)
# Records
records = (
[datetime(day=1, month=3, year=2010), "rec_1", 5, 6.5],
[datetime(day=2, month=3, year=2010), "", 8, 7.5],
[datetime(day=3, month=3, year=2010), "rec_3", "", 8.5],
[datetime(day=4, month=3, year=2010), "rec_4", 12, ""],
[datetime(day=5, month=3, year=2010), "rec_5", -87657496599, 6.5],
[datetime(day=6, month=3, year=2010), "rec_6", 12, -87657496599],
[datetime(day=6, month=3, year=2010), str(-87657496599), 12, 6.5],
)
for r in records:
s.appendRecord(list(r))
s.close()
# Read the standard file
s = FileRecordStream(streamID=filename, write=False)
fieldsRead = s.getFields()
self.assertEqual(fields, fieldsRead)
recordsRead = []
while True:
r = s.getNextRecord()
if r is None:
break
print "Reading record ..."
print r
recordsRead.append(r)
# sort the records by date, so we know for sure which is which
sorted(recordsRead, key=lambda rec: rec[0])
# empty string
self.assertEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[1][1])
# missing int
self.assertEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[2][2])
# missing float
self.assertEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[3][3])
# sentinel value in input handled correctly for int field
self.assertNotEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[4][2])
# sentinel value in input handled correctly for float field
self.assertNotEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[5][3])
# sentinel value in input handled correctly for string field
# this should leave the string as-is, since a missing string
# is encoded not with a sentinel value but with an empty string
self.assertNotEqual(SENTINEL_VALUE_FOR_MISSING_DATA, recordsRead[6][1])
