def runExperiment(experimentDirectory, experimentOptions):
o = experimentOptions
# Clean aggregation datasets if needed
#if o['clearAggregationDatasets']:
#helpers.cleanAggregationDatasets()
global experiment
experiment = Experiment(path=experimentDirectory,
runtimeOptions=experimentOptions)
# Create GUIs as needed
if not experimentOptions['postProcessOnly']:
if o['runGUI']:
# Only import these if the gui is requested, because
# they import further modules that require a console
# and exit if there is no console (as in the autobuild)
from nupic.frameworks.prediction.GUIs import TrainingGUI, InferenceGUI
if not (o['createNetworkOnly'] or o['runInferenceOnly']):
gui.append(TrainingGUI(experiment))
gui.append(InferenceGUI(experiment))
for i, g in enumerate(gui):
g.start()
else:
experiment.run()
if experimentOptions['postProcess']:
from nupic.frameworks.prediction.postprocess import postProcess
try:
postProcess(experiment)
experiment.writeResults()
if os.environ.has_key('NTA_AMAZON_SYNC_DATA'):
print "Deleting log files"
inferenceDir = os.path.join(experimentDirectory, "inference")
logFiles = glob.glob(os.path.join(inferenceDir, "*.txt"))
for f in logFiles:
try:
os.remove(f)
except:
print "Couldn't remove log file:", f
except Exception, e:
message = "Post processing has failed, %s" % str(e.args)
e.args = (message, ) + e.args[1:]
#traceback.print_exc(file=sys.stdout)
raise
def postProcess(experiment):
if not isinstance(experiment, Experiment):
experiment = Experiment(experiment)
inferenceDir = experiment.getInferenceDirectory()
# Use level1 rather than level1_sp because the latter
# is still in learning mode
# use "all" rather than level1 because level1_sp is not created by default
experiment.loadNetwork("all")
net = experiment.network
# Analyze the trained network and extract the auxiliary data structures
netInfo = _analyzeTrainedNet(net=net, options=dict())
# ------------------------------------------------------------------------
# Get the list of tests, and the list of baselines required
testNames = [test['name'] for test in experiment.description['infer']]
testOptions = [
test.get('ppOptions', dict())
for test in experiment.description['infer']
]
standaloneTests = list(testNames)
options = dict(zip(testNames, testOptions))
baselineTestNames = set()
for testName in testNames:
(datasetName, testVariant) = testName.split('_')
if testVariant == 'baseline':
baselineTestNames.add(testName)
# ------------------------------------------------------------------------
# For each baseline, train the classifier, and then run each variant that uses
# that baseline
for baselineTestName in baselineTestNames:
(baseDatasetName, baseVariantName) = baselineTestName.split('_')
assert (baseVariantName == 'baseline')
standaloneTests.remove(baselineTestName)
baseline = _analyzeTestRun(experiment,
netInfo=netInfo,
options=options[baselineTestName],
datasetName=baseDatasetName,
testVariantName=baseVariantName,
baseline=None)
# Now, run each variant that uses this baseline
for testName in testNames:
(datasetName, variantName) = testName.split('_')
if datasetName != baseDatasetName or variantName == 'baseline':
continue
standaloneTests.remove(testName)
_analyzeTestRun(experiment,
netInfo=netInfo,
options=options[testName],
datasetName=datasetName,
testVariantName=variantName,
baseline=baseline)
# Run each standalone test
for testName in standaloneTests:
(datasetName, variantName) = testName.split('_')
_analyzeTestRun(experiment,
netInfo=netInfo,
options=options[testName],
datasetName=datasetName,
testVariantName=variantName,
baseline=None)
