def distillKnowledge(deepNet,
filepath,
batchSize=50,
holdoutPercentage=0.5,
log=None):
import os
import numpy as np
from dataset.ingest.labeled import ingestImagery
from dataset.pickle import writePickleZip
from distill.net import DistilleryClassifier
if not isinstance(deepNet, DistilleryClassifier):
raise ValueError('The network must be setup as a DistilleryNetwork.')
# build a new pickle with this information
# TODO: NETWORKS NEED UNIQUE LABEL IDENTIFIERS WHICH CAN BE ADDED HERE
rootpath = os.path.abspath(filepath)
outputFile = os.path.join(
rootpath,
os.path.basename(rootpath) + '_darkLabels' + '_holdout_' +
str(holdoutPercentage) + '_batch_' + str(batchSize) + '.pkl.gz')
if os.path.exists(outputFile):
if log is not None:
log.info('Pickle exists for this dataset [' + outputFile +
']. Using this instead.')
return outputFile
# NOTE: The pickleDataset will silently use previously created pickles if
# one exists (for efficiency). So watch out for stale pickles!
train, test, labels = ingestImagery(filepath=filepath,
shared=False,
batchSize=batchSize,
holdoutPercentage=holdoutPercentage,
log=log)
if log is not None:
log.info('Distilling knowledge from deep network')
# distill knowledge into a pickle which can be used to train other networks
batchSize = train[0].shape[0]
darkLabels = [deepNet.softTarget(dataset) for dataset in train[0]]
labelDims = [len(darkLabels)] + list(darkLabels[0].shape)
darkLabels = np.reshape(np.concatenate(darkLabels), labelDims)
train = train[0], train[1], darkLabels
# pickle the dataset
writePickleZip(outputFile, (train, test, labels), log)
# return the output filename
return outputFile
def distillKnowledge(deepNet, filepath, batchSize=50,
holdoutPercentage=0.5, log=None) :
import os
import numpy as np
from dataset.ingest.labeled import ingestImagery
from dataset.pickle import writePickleZip
from distill.net import DistilleryClassifier
if not isinstance(deepNet, DistilleryClassifier) :
raise ValueError('The network must be setup as a DistilleryNetwork.')
# build a new pickle with this information
# TODO: NETWORKS NEED UNIQUE LABEL IDENTIFIERS WHICH CAN BE ADDED HERE
rootpath = os.path.abspath(filepath)
outputFile = os.path.join(rootpath, os.path.basename(rootpath) +
'_darkLabels' +
'_holdout_' + str(holdoutPercentage) +
'_batch_' + str(batchSize) +
'.pkl.gz')
if os.path.exists(outputFile) :
if log is not None :
log.info('Pickle exists for this dataset [' + outputFile +
']. Using this instead.')
return outputFile
# NOTE: The pickleDataset will silently use previously created pickles if
# one exists (for efficiency). So watch out for stale pickles!
train, test, labels = ingestImagery(filepath=filepath, shared=False,
batchSize=batchSize,
holdoutPercentage=holdoutPercentage,
log=log)
if log is not None :
log.info('Distilling knowledge from deep network')
# distill knowledge into a pickle which can be used to train other networks
batchSize = train[0].shape[0]
darkLabels = [deepNet.softTarget(dataset) for dataset in train[0]]
labelDims = [len(darkLabels)] + list(darkLabels[0].shape)
darkLabels = np.reshape(np.concatenate(darkLabels), labelDims)
train = train[0], train[1], darkLabels
# pickle the dataset
writePickleZip(outputFile, (train, test, labels), log)
# return the output filename
return outputFile
prof = Profiler(log=log, name=logName, profFile=options.profile)
# create a random number generator for efficiency
from numpy.random import RandomState
from operator import mul
from time import time
rng = RandomState(int(time()))
#rng = RandomState(4567) # always initialize the same
# NOTE: The pickleDataset will silently use previously created pickles if
# one exists (for efficiency). So watch out for stale pickles!
# NOTE: User may pass a dark pickle into here, and the logic will react
# appropriately to the situation.
train, test, labels = ingestImagery(filepath=options.data,
shared=True,
batchSize=options.batchSize,
holdoutPercentage=options.holdout,
log=log)
inputSize = train[0].shape.eval()
# create a file with pre-initialized weights so both networks use the same
# baseline for testing.
if options.synapse is None:
networkFile = createNetwork(inputSize=inputSize[1:],
numKernels=options.kernel,
numNeurons=options.neuron,
numLabels=labels.shape[0])
else:
networkFile = options.synapse
regScale = 1. / (2 * options.kernel * 5 * 5 + options.neuron +
parser = argparse.ArgumentParser()
addLoggingParams(parser)
addUnsupDataParams(parser, 'saeClass', multiLoad=True)
addDebuggingParams(parser)
parser.add_argument('--percent', dest='percentReturned', type=float,
default=100.,
help='Return some percentage of the highest related ' +
'examples. All others will not be returned to ' +
'the user.')
parser.add_argument('--csv', dest='csvFile',type=str, default='output.csv',
help='Name of the CSV file output.')
options = parser.parse_args()
# setup the logger
log, prof = setupLogging(options, 'SAE-Classification Benchmark')
# NOTE: The pickleDataset will silently use previously created pickles if
# one exists (for efficiency). So watch out for stale pickles!
train, test, labels = ingestImagery(filepath=options.data, shared=False,
batchSize=options.batchSize,
holdoutPercentage=options.holdout,
log=log)
# load all networks initialized to the target imagery
nets = createNetworks(options.maxTarget, options.batchSize,
options.synapse, prof, options.debug)
# test the training data for similarity to the target
buildCSV(options.csvFile, nets, options.targetDir, test,
False, options.percentReturned, options.debug)
default=None,
help='Load from a previously saved network.')
parser.add_argument('data',
help='Directory or pkl.gz file for the ' +
'training and test sets')
options = parser.parse_args()
# setup the logger
logName = 'SAE-Classification Benchmark: ' + options.data
log = setupLogging(logName, options.level, options.logfile)
prof = Profiler(log=log, name=logName, profFile=options.profile)
# NOTE: The pickleDataset will silently use previously created pickles if
# one exists (for efficiency). So watch out for stale pickles!
train, test, labels = ingestImagery(filepath=options.data,
shared=True,
batchSize=options.batchSize,
log=log)
# load example imagery --
# these are confirmed objects we are attempting to identify
target = readTargetData(options.targetDir)
if options.synapse is None:
regType = 'L2'
regValue = .001
trainer = buildTrainerSAENetwork(train,
regType,
regValue,
target,
prof=prof,
kernelConv=options.kernel,