def convertGrokEnvToOPF(self):
# We need to create a task structure, most of which is taken verbatim
# from the Grok control dict
task = dict(self.__control)
task.pop('environment')
inferenceArgs = task.pop('inferenceArgs')
task['taskLabel'] = 'DefaultTask'
# Create the iterationCycle element that will be placed inside the
# taskControl.
iterationCount = task.get('iterationCount', -1)
iterationCountInferOnly = task.pop('iterationCountInferOnly', 0)
if iterationCountInferOnly == -1:
iterationCycle = [IterationPhaseSpecInferOnly(1000, inferenceArgs=inferenceArgs)]
elif iterationCountInferOnly > 0:
assert iterationCount > 0, "When iterationCountInferOnly is specified, "\
"iterationCount must also be specified and not be -1"
iterationCycle = [IterationPhaseSpecLearnAndInfer(iterationCount
-iterationCountInferOnly, inferenceArgs=inferenceArgs),
IterationPhaseSpecInferOnly(iterationCountInferOnly, inferenceArgs=inferenceArgs)]
else:
iterationCycle = [IterationPhaseSpecLearnAndInfer(1000, inferenceArgs=inferenceArgs)]
taskControl = dict(metrics = task.pop('metrics'),
loggedMetrics = task.pop('loggedMetrics'),
iterationCycle = iterationCycle)
task['taskControl'] = taskControl
# Create the new control
self.__control = dict(environment = OpfEnvironment.Grok,
tasks = [task])
# to one record from the (possibly aggregated) dataset. The task is
# terminated when either number of iterations reaches iterationCount or
# all records in the (possibly aggregated) database have been processed,
# whichever occurs first.
#
# iterationCount of -1 = iterate over the entire dataset
'iterationCount': -1,
# Task Control parameters for OPFTaskDriver (per opfTaskControlSchema.json)
'taskControl': {
# Iteration cycle list consisting of opftaskdriver.IterationPhaseSpecXXXXX
# instances.
'iterationCycle': [
#IterationPhaseSpecLearnOnly(1000),
IterationPhaseSpecLearnAndInfer(1000),
#IterationPhaseSpecInferOnly(10),
],
# Metrics: A list of MetricSpecs that instantiate the metrics that are
# computed for this experiment
'metrics': [
MetricSpec(metric='avg_err',
inferenceElement='classification',
params={'window': 200}),
MetricSpec(metric='neg_auc',
inferenceElement='classConfidences',
params={
'window': 200,
'computeEvery': 10
}),
# to one record from the (possibly aggregated) dataset. The task is
# terminated when either number of iterations reaches iterationCount or
# all records in the (possibly aggregated) database have been processed,
# whichever occurs first.
#
# iterationCount of -1 = iterate over the entire dataset
'iterationCount': -1,
# Task Control parameters for OPFTaskDriver (per opfTaskControlSchema.json)
'taskControl': {
# Iteration cycle list consisting of opftaskdriver.IterationPhaseSpecXXXXX
# instances.
'iterationCycle': [
#IterationPhaseSpecLearnOnly(1000),
IterationPhaseSpecLearnAndInfer(1000, inferenceArgs=None),
#IterationPhaseSpecInferOnly(10, inferenceArgs=None),
],
'metrics': [],
# Logged Metrics: A sequence of regular expressions that specify which of
# the metrics from the Inference Specifications section MUST be logged for
# every prediction. The regex's correspond to the automatically generated
# metric labels. This is similar to the way the optimization metric is
# specified in permutations.py.
'loggedMetrics': ['.*nupicScore.*'],
# Callbacks for experimentation/research (optional)
'callbacks': {
# Callbacks to be called at the beginning of a task, before model iterations.
# Signature: callback(<reference to OPFExperiment>); returns nothing
# to one record from the (possibly aggregated) dataset. The task is
# terminated when either number of iterations reaches iterationCount or
# all records in the (possibly aggregated) database have been processed,
# whichever occurs first.
#
# iterationCount of -1 = iterate over the entire dataset
'iterationCount': -1,
# Task Control parameters for OPFTaskDriver (per opfTaskControlSchema.json)
'taskControl': {
# Iteration cycle list consisting of opftaskdriver.IterationPhaseSpecXXXXX
# instances.
'iterationCycle': [
#IterationPhaseSpecLearnOnly(1000),
IterationPhaseSpecLearnAndInfer(
1000, dict(predictedField="consumption")),
#IterationPhaseSpecInferOnly(10),
],
'metrics': [
MetricSpec(metric='rmse',
field="consumption",
inferenceElement=InferenceElement.prediction),
],
# Callbacks for experimentation/research (optional)
'callbacks': {
# Callbacks to be called at the beginning of a task, before model iterations.
# Signature: callback(<reference to OPFExperiment>); returns nothing
'setup': [],
# Callbacks to be called after every learning/inference iteration