def run(
nb_filters=nb_filters,
filterPolicy=filterPolicy,
poolings=poolings,
extractor=extractor,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
fixedSize=fixedSize,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
random=random,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
nbTrees=nbTrees,
maxFeatures=maxFeatures,
maxDepth=maxDepth,
minSamplesSplit=minSamplesSplit,
minSamplesLeaf=minSamplesLeaf,
bootstrap=bootstrap,
randomClassif=randomClassif,
nbJobsEstimator=nbJobsEstimator,
verbose=verbose,
learningUse=learningUse,
testingUse=testingUse,
saveFile=saveFile,
shouldSave=shouldSave,
):
randomState = None
if not randomClassif:
randomState = 100
lsSize = learningUse
if learningUse > maxLearningSize:
lsSize = maxLearningSize
tsSize = testingUse
if testingUse > maxTestingSize:
tsSize = maxTestingSize
# ======INSTANTIATING========#
# --RandConv--
randConvCoord = coordinatorRandConvFactory(
nbFilters=nb_filters,
filterPolicy=filterPolicy,
poolings=poolings,
extractor=extractor,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
random=random,
)
randConvCoord = LoadCoordinator(randConvCoord, learnFile, testFile)
# --Extra-tree--
baseClassif = ExtraTreesClassifier(
nbTrees,
max_features=maxFeatures,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
bootstrap=bootstrap,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose,
)
# --Classifier
classifier = Classifier(randConvCoord, baseClassif)
# --Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
# =====COMPUTATION=====#
# --Learning--#
print "Starting learning"
fitStart = time()
classifier.fit(learningSet)
fitEnd = time()
print "Learning done", formatDuration(fitEnd - fitStart)
sys.stdout.flush()
# --Testing--#
y_truth = testingSet.getLabels()
predStart = time()
y_prob, y_pred = classifier.predict_predict_proba(testingSet)
predEnd = time()
accuracy = classifier.accuracy(y_pred, y_truth)
confMat = classifier.confusionMatrix(y_pred, y_truth)
# ====ANALYSIS=====#
importance, order = randConvCoord.importancePerFeatureGrp(baseClassif)
print "==================RandConv================"
print "-----------Filtering--------------"
print "nb_filters", nb_filters
print "filterPolicy", filterPolicy
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowMinSizeRatio", subwindowMinSizeRatio
print "subwindowMaxSizeRatio", subwindowMaxSizeRatio
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "fixedSize", fixedSize
print "------------Misc-----------------"
print "includeOriginalImage", includeOriginalImage
print "random", random
print "tempFolder", tempFolder
print "verbosity", verbosity
print "nbJobs", nbJobs
print "--------ExtraTrees----------"
print "nbTrees", nbTrees
print "maxFeatures", maxFeatures
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "bootstrap", bootstrap
print "nbJobsEstimator", nbJobsEstimator
print "verbose", verbose
print "randomState", randomState
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
if shouldSave:
print "saveFile", saveFile
print "Fit time", formatDuration(fitEnd - fitStart)
print "Classifcation time", formatDuration(predEnd - predStart)
print "Accuracy", accuracy
if shouldSave:
np.save(saveFile, y_prob)
return accuracy, confMat, importance, order
def run(lsFile, tsFile, **kwargs):
randomState = None
if random:
randomState = 100
#======INSTANTIATING========#
os.environ["JOBLIB_TEMP_FOLDER"] = "/home/jmbegon/jmbegon/code/work/tmp/"
#--Pixit--
randConvCoord = coordinatorRandConvFactory(
nbFilters=nb_filters,
filterMinVal=filter_min_val,
filterMaxVal=filter_max_val,
filterMinSize=filterMinSize,
filterMaxSize=filterMaxSize,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
poolings=poolings,
filterNormalisation=filterNormalisation,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
random=random)
#--Extra-tree--
baseClassif = ExtraTreesClassifier(nbTrees,
max_features=maxFeatures,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
bootstrap=bootstrap,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose)
#--Classifier
classifier = Classifier(randConvCoord, baseClassif)
#--Data--
with open(lsFile, "wb") as f:
lsSize, Xls, yls = pickle.load(f, protocol=2)
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
with open(tsFile, "wb") as f:
tsSize, Xts, yts = pickle.load(f, protocol=2)
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
#=====COMPUTATION=====#
#--Learning--#
print "Starting learning"
fitStart = time()
baseClassif.fit(Xls, yls)
fitEnd = time()
print "Learning done", (fitEnd-fitStart), "seconds (no extraction)"
sys.stdout.flush()
#--Testing--#
y_truth = testingSet.getLabels()
predStart = time()
y_pred = classifier._predict(Xts, lsSize)
predEnd = time()
#====ANALYSIS=====#
accuracy = classifier.accuracy(y_pred, y_truth)
confMat = classifier.confusionMatrix(y_pred, y_truth)
importance, order = randConvCoord.importancePerFeatureGrp(baseClassif)
print "========================================="
print "--------ExtraTrees----------"
print "nbTrees", nbTrees
print "maxFeatures", maxFeatures
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "bootstrap", bootstrap
print "nbJobsEstimator", nbJobsEstimator
print "verbose", verbose
print "randomState", randomState
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
print "Fit time (no extraction)", (fitEnd-fitStart), "seconds"
print "Classifcation time (no extraction)", (predEnd-predStart), "seconds"
print "Accuracy", accuracy
return accuracy, confMat, importance, order
from sklearn.metrics import confusion_matrix
from CifarLoader import CifarFromNumpies
from ImageBuffer import FileImageBuffer, NumpyImageLoader
testingSetDir = "test/"
testingIndexFile = "0index"
if __name__ == "__main__":
Y = np.load(sys.argv[0])
for filename in sys.argv[1:]:
Y += np.load(filename)
y_pred = np.argmax(Y, axis=1)
tsSize = len(y_pred)
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
y_truth = testingSet.getLabels()
accuracy = sum(map((lambda x, y: x == y), y_pred, y_truth))/float(len(y_truth))
confMat = confusion_matrix(y_truth, y_pred)
print "==================Classif aggregation================"
print "Files :\n", sys.argv
print "Accuracy", accuracy
print "Confusion matrix :\n", confMat
def run(nb_filters=nb_filters,
filterPolicy=filterPolicy,
poolings=poolings,
extractor=extractor,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
fixedSize=fixedSize,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
random=random,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
nbTrees=nbTrees,
maxFeatures=maxFeatures,
maxDepth=maxDepth,
minSamplesSplit=minSamplesSplit,
minSamplesLeaf=minSamplesLeaf,
bootstrap=bootstrap,
randomClassif=randomClassif,
nbJobsEstimator=nbJobsEstimator,
verbose=verbose,
learningUse=learningUse,
testingUse=testingUse,
saveFile=saveFile,
shouldSave=shouldSave):
randomState = None
if not randomClassif:
randomState = 100
lsSize = learningUse
if learningUse > maxLearningSize:
lsSize = maxLearningSize
tsSize = testingUse
if testingUse > maxTestingSize:
tsSize = maxTestingSize
#======INSTANTIATING========#
#--RandConv--
randConvCoord = coordinatorRandConvFactory(
nbFilters=nb_filters,
filterPolicy=filterPolicy,
poolings=poolings,
extractor=extractor,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
random=random)
randConvCoord = LoadCoordinator(randConvCoord, learnFile, testFile)
#--Extra-tree--
baseClassif = ExtraTreesClassifier(nbTrees,
max_features=maxFeatures,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
bootstrap=bootstrap,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose)
#--Classifier
classifier = Classifier(randConvCoord, baseClassif)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
#=====COMPUTATION=====#
#--Learning--#
print "Starting learning"
fitStart = time()
classifier.fit(learningSet)
fitEnd = time()
print "Learning done", formatDuration(fitEnd - fitStart)
sys.stdout.flush()
#--Testing--#
y_truth = testingSet.getLabels()
predStart = time()
y_prob, y_pred = classifier.predict_predict_proba(testingSet)
predEnd = time()
accuracy = classifier.accuracy(y_pred, y_truth)
confMat = classifier.confusionMatrix(y_pred, y_truth)
#====ANALYSIS=====#
importance, order = randConvCoord.importancePerFeatureGrp(baseClassif)
print "==================RandConv================"
print "-----------Filtering--------------"
print "nb_filters", nb_filters
print "filterPolicy", filterPolicy
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowMinSizeRatio", subwindowMinSizeRatio
print "subwindowMaxSizeRatio", subwindowMaxSizeRatio
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "fixedSize", fixedSize
print "------------Misc-----------------"
print "includeOriginalImage", includeOriginalImage
print "random", random
print "tempFolder", tempFolder
print "verbosity", verbosity
print "nbJobs", nbJobs
print "--------ExtraTrees----------"
print "nbTrees", nbTrees
print "maxFeatures", maxFeatures
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "bootstrap", bootstrap
print "nbJobsEstimator", nbJobsEstimator
print "verbose", verbose
print "randomState", randomState
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
if shouldSave:
print "saveFile", saveFile
print "Fit time", formatDuration(fitEnd - fitStart)
print "Classifcation time", formatDuration(predEnd - predStart)
print "Accuracy", accuracy
if shouldSave:
np.save(saveFile, y_prob)
return accuracy, confMat, importance, order
def run(poolings=poolings,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
fixedSize=fixedSize,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
random=random,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
nbTrees=nbTrees,
maxFeatures=maxFeatures,
maxDepth=maxDepth,
minSamplesSplit=minSamplesSplit,
minSamplesLeaf=minSamplesLeaf,
bootstrap=bootstrap,
nbJobsEstimator=nbJobsEstimator,
verbose=verbose,
learningUse=learningUse,
testingUse=testingUse):
randomState = None
if random:
randomState = 100
lsSize = learningUse
if learningUse > maxLearningSize:
lsSize = maxLearningSize
tsSize = testingUse
if testingUse > maxTestingSize:
tsSize = maxTestingSize
#======INSTANTIATING========#
os.environ["JOBLIB_TEMP_FOLDER"] = "/home/jmbegon/jmbegon/code/work/tmp/"
#--customRandConv--
randConvCoord = customRandConvFactory(
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
poolings=poolings,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
random=random)
#--Extra-tree--
baseClassif = ExtraTreesClassifier(nbTrees,
max_features=maxFeatures,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
bootstrap=bootstrap,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose)
#--Classifier
classifier = Classifier(randConvCoord, baseClassif)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
#=====COMPUTATION=====#
#--Learning--#
print "Starting learning"
fitStart = time()
classifier.fit(learningSet)
fitEnd = time()
print "Learning done", (fitEnd-fitStart), "seconds"
sys.stdout.flush()
#--Testing--#
y_truth = testingSet.getLabels()
predStart = time()
y_pred = classifier.predict(testingSet)
predEnd = time()
accuracy = classifier.accuracy(y_pred, y_truth)
confMat = classifier.confusionMatrix(y_pred, y_truth)
#====ANALYSIS=====#
importance, order = randConvCoord.importancePerFeatureGrp(baseClassif)
print "==================CUSTOM======================="
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowMinSizeRatio", subwindowMinSizeRatio
print "subwindowMaxSizeRatio", subwindowMaxSizeRatio
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "fixedSize", fixedSize
print "------------Misc-----------------"
print "includeOriginalImage", includeOriginalImage
print "random", random
print "tempFolder", tempFolder
print "verbosity", verbosity
print "nbJobs", nbJobs
print "--------ExtraTrees----------"
print "nbTrees", nbTrees
print "maxFeatures", maxFeatures
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "bootstrap", bootstrap
print "nbJobsEstimator", nbJobsEstimator
print "verbose", verbose
print "randomState", randomState
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
print "Fit time", (fitEnd-fitStart), "seconds"
print "Classifcation time", (predEnd-predStart), "seconds"
print "Accuracy", accuracy
return accuracy, confMat, importance, order
def run(nb_filters=nb_filters,
poolings=poolings,
nbSubwindows=nbSubwindows,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
nbTrees=nbTrees,
maxFeatures=maxFeatures,
maxDepth=maxDepth,
minSamplesSplit=minSamplesSplit,
minSamplesLeaf=minSamplesLeaf,
bootstrap=bootstrap,
nbJobsEstimator=nbJobsEstimator,
verbose=verbose,
learningUse=learningUse,
testingUse=testingUse):
lsSize = learningUse
if learningUse > maxLearningSize:
lsSize = maxLearningSize
tsSize = testingUse
if testingUse > maxTestingSize:
tsSize = maxTestingSize
totalNbFeatures = nb_filters*len(poolings)*subwindowTargetWidth*subwindowTargetHeight*3
totalNbObj = lsSize*nbSubwindows
nbFeatures = totalNbFeatures/nbJobs
floatSize = np.zeros().itemsize
singleArraySize = nbFeatures*totalNbObj*floatSize
totalArraySize = totalNbFeatures*totalNbObj*floatSize
#======INSTANTIATING========#
os.environ["JOBLIB_TEMP_FOLDER"] = "/home/jmbegon/jmbegon/code/work/tmp/"
#--Pixit--
memCoord = MemroyTestCoordinator(nbFeatures, totalNbObj)
if nbJobs != 1:
memCoord.parallelize(nbJobs, tempFolder)
#--Extra-tree--
baseClassif = ExtraTreesClassifier(nbTrees,
max_features=maxFeatures,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
bootstrap=bootstrap,
n_jobs=nbJobsEstimator,
verbose=verbose)
#--Classifier
classifier = Classifier(memCoord, baseClassif)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
#=====COMPUTATION=====#
#--Learning--#
classifier.fit(learningSet)
print "========================================="
print "-----------Filtering--------------"
print "nb_filters", nb_filters
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "------------Misc-----------------"
print "tempFolder", tempFolder
print "verbosity", verbosity
print "nbJobs", nbJobs
print "--------ExtraTrees----------"
print "nbTrees", nbTrees
print "maxFeatures", maxFeatures
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "bootstrap", bootstrap
print "nbJobsEstimator", nbJobsEstimator
print "verbose", verbose
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
print "totalNbFeatures", totalNbFeatures
print "totalNbObj", totalNbObj
print "singleArraySize", singleArraySize
print "totalArraySize", totalArraySize
def run():
#======HYPER PARAMETERS======#
#----RandConv param
#Filtering
nb_filters = 100
filter_min_val = -1
filter_max_val = 1
filterMinSize = 2
filterMaxSize = 32
filterNormalisation = FilterGenerator.NORMALISATION_MEANVAR
#Aggregation
poolings = [(2, 2, Const.POOLING_AGGREG_AVG)]
#Subwindow
nbSubwindows = 10
subwindowMinSizeRatio = 0.75
subwindowMaxSizeRatio = 1.
subwindowTargetWidth = 16
subwindowTargetHeight = 16
fixedSize = False
subwindowInterpolation = SubWindowExtractor.INTERPOLATION_BILINEAR
#Misc.
includeOriginalImage = True
#Compressor
compressorType = "Sampling"
nbCompressedFeatures = 20
compressOriginalImage = True
#Parrallelization & Logger
nbJobs = -1
verbosity = 8
tempFolder = "tmp/"
#-----Extratree param
nbTrees = 30
maxFeatures = "auto"
maxDepth = None
minSamplesSplit = 2
minSamplesLeaf = 1
bootstrap = False
nbJobsEstimator = -1
randomState = None
verbose = 8
#=====DATA=====#
# maxLearningSize = 50000
# maxTestingSize = 10000
learningUse = 500
learningSetDir = "learn/"
learningIndexFile = "0index"
testingUse = 500
testingSetDir = "test/"
testingIndexFile = "0index"
#======INSTANTIATING========#
os.environ["JOBLIB_TEMP_FOLDER"] = "/home/jmbegon/jmbegon/code/work/tmp/"
#--Pixit--
randConvCoord = coordinatorCompressRandConvFactory(
nbFilters=nb_filters,
filterMinVal=filter_min_val,
filterMaxVal=filter_max_val,
filterMinSize=filterMinSize,
filterMaxSize=filterMaxSize,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
poolings=poolings,
filterNormalisation=filterNormalisation,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
compressorType=compressorType,
nbCompressedFeatures=nbCompressedFeatures,
compressOriginalImage=compressOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder)
#--Extra-tree--
baseClassif = ExtraTreesClassifier(nbTrees,
max_features=maxFeatures,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
bootstrap=bootstrap,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose)
#--Classifier
classifier = Classifier(randConvCoord, baseClassif)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:learningUse]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:testingUse]
#=====COMPUTATION=====#
#--Learning--#
print "Starting learning"
fitStart = time()
classifier.fit(learningSet)
fitEnd = time()
print "Learning done", (fitEnd-fitStart), "seconds"
sys.stdout.flush()
#--Testing--#
y_truth = testingSet.getLabels()
predStart = time()
y_pred = classifier.predict(testingSet)
predEnd = time()
accuracy = classifier.accuracy(y_pred, y_truth)
confMat = classifier.confusionMatrix(y_pred, y_truth)
#====ANALYSIS=====#
importance, order = randConvCoord.importancePerFeatureGrp(baseClassif)
print "========================================="
print "-----------Filtering--------------"
print "nb_filters", nb_filters
print "filter_min_val", filter_min_val
print "filter_max_val", filter_max_val
print "filterMinSize", filterMinSize
print "filterMaxSize", filterMaxSize
print "filterNormalisation", filterNormalisation
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowMinSizeRatio", subwindowMinSizeRatio
print "subwindowMaxSizeRatio", subwindowMaxSizeRatio
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "fixedSize", fixedSize
print "------------Misc-----------------"
print "includeOriginalImage", includeOriginalImage
print "compressorType", compressorType
print "nbCompressedFeatures", nbCompressedFeatures
print "compressOriginalImage", compressOriginalImage
print "nbJobs", nbJobs
print "--------ExtraTrees----------"
print "nbTrees", nbTrees
print "maxFeatures", maxFeatures
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "bootstrap", bootstrap
print "nbJobsEstimator", nbJobsEstimator
print "randomState", randomState
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
print "Fit time", (fitEnd-fitStart), "seconds"
print "Classifcation time", (predEnd-predStart), "seconds"
print "Accuracy", accuracy
print "Confusion matrix :\n", confMat
return accuracy, confMat, importance, order
def run(lsName, tsName, **kwargs):
lsSize = learningUse
if learningUse > maxLearningSize:
lsSize = maxLearningSize
tsSize = testingUse
if testingUse > maxTestingSize:
tsSize = maxTestingSize
#======INSTANTIATING========#
os.environ["JOBLIB_TEMP_FOLDER"] = "/home/jmbegon/jmbegon/code/work/tmp/"
#--Coordinator--
randConvCoord = coordinatorRandConvFactory(
nbFilters=nb_filters,
filterMinVal=filter_min_val,
filterMaxVal=filter_max_val,
filterMinSize=filterMinSize,
filterMaxSize=filterMaxSize,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
poolings=poolings,
filterNormalisation=filterNormalisation,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
random=random)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
#=====COMPUTATION=====#
#--Learning--#
print "Starting learning"
fitStart = time()
X, y = randConvCoord.process(learningSet, True)
with open(lsName, "wb") as f:
pickle.dump((lsSize, X, y), f, protocol=2)
fitEnd = time()
print "Learning done", (fitEnd-fitStart), "seconds"
sys.stdout.flush()
#--Testing--#
predStart = time()
X, y = randConvCoord.process(testingSet, False)
with open(tsName, "wb") as f:
pickle.dump((tsSize, X, y), f, protocol=2)
predEnd = time()
print "========================================="
print "-----------Filtering--------------"
print "nb_filters", nb_filters
print "filter_min_val", filter_min_val
print "filter_max_val", filter_max_val
print "filterMinSize", filterMinSize
print "filterMaxSize", filterMaxSize
print "filterNormalisation", filterNormalisation
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowMinSizeRatio", subwindowMinSizeRatio
print "subwindowMaxSizeRatio", subwindowMaxSizeRatio
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "fixedSize", fixedSize
print "------------Misc-----------------"
print "includeOriginalImage", includeOriginalImage
print "random", random
print "tempFolder", tempFolder
print "verbosity", verbosity
print "nbJobs", nbJobs
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
print "Fit time", (fitEnd-fitStart), "seconds"
print "Classifcation time", (predEnd-predStart), "seconds"
baseClassif = ExtraTreesClassifier(nbTrees,
max_features=maxFeatures,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
bootstrap=bootstrap,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose)
#--Classifier
classifier = Classifier(pixitCoord, baseClassif)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:learningUse]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:testingUse]
#=====COMPUTATION=====#
#--Learning--#
fitStart = time()
classifier.fit(learningSet)
fitEnd = time()
#--Testing--#
y_truth = testingSet.getLabels()
def run(nb_filters=nb_filters,
filterPolicy=filterPolicy,
poolings=poolings,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
fixedSize=fixedSize,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
random=random,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
nbTrees=nbTrees,
maxDepth=maxDepth,
minSamplesSplit=minSamplesSplit,
minSamplesLeaf=minSamplesLeaf,
randomClassif=randomClassif,
nbJobsEstimator=nbJobsEstimator,
verbose=verbose,
learningUse=learningUse,
testingUse=testingUse,
saveFile=saveFile,
shouldSave=shouldSave):
randomState = None
if not randomClassif:
randomState = 100
lsSize = learningUse
if learningUse > maxLearningSize:
lsSize = maxLearningSize
tsSize = testingUse
if testingUse > maxTestingSize:
tsSize = maxTestingSize
#======INSTANTIATING========#
#--randconv--
randConvCoord = coordinatorRandConvFactory(
nbFilters=nb_filters,
filterPolicy=filterPolicy,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
poolings=poolings,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
random=random)
nb_filters = len(randConvCoord.getFilters())
#--SVM--
baseClassif = LinearSVC(verbose=verbose, random_state=randomState)
#--Classifier
classifier = uClassifier(coordinator=randConvCoord,
base_classifier=baseClassif,
n_estimators=nbTrees,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
#=====COMPUTATION=====#
#--Learning--#
print "Starting learning"
fitStart = time()
hist = classifier._preprocess(learningSet, learningPhase=True)
y = learningSet.getLabels()
if shouldSave:
np.savez(saveFile,
data=hist.data,
indices=hist.indices,
indptr=hist.indptr,
shape=hist.shape)
classifier.fit_histogram(hist, y)
fitEnd = time()
print "Learning done", formatDuration(fitEnd - fitStart)
sys.stdout.flush()
#--Testing--#
y_truth = testingSet.getLabels()
predStart = time()
y_pred = classifier.predict(testingSet)
predEnd = time()
accuracy = classifier.accuracy(y_pred, y_truth)
confMat = classifier.confusionMatrix(y_pred, y_truth)
#====ANALYSIS=====#
importance, order = randConvCoord.importancePerFeatureGrp(
classifier._visualBagger)
print "==================Bag of Visual Words======================="
print "-----------Filtering--------------"
print "nb_filters", nb_filters
print "filterPolicy", filterPolicy
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowMinSizeRatio", subwindowMinSizeRatio
print "subwindowMaxSizeRatio", subwindowMaxSizeRatio
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "fixedSize", fixedSize
print "------------Misc-----------------"
print "includeOriginalImage", includeOriginalImage
print "random", random
print "tempFolder", tempFolder
print "verbosity", verbosity
print "nbJobs", nbJobs
print "--------Bag of words params + SVC----------"
print "nbTrees", nbTrees
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "nbJobsEstimator", nbJobsEstimator
print "verbose", verbose
print "randomState", randomState
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
if shouldSave:
print "saveFile", saveFile
print "Fit time", formatDuration(fitEnd - fitStart)
print "Classifcation time", formatDuration(predEnd - predStart)
print "Accuracy", accuracy
print "Leafs", formatBigNumber(classifier.histoSize)
return accuracy, confMat, importance, order
def run(nb_filters=nb_filters,
filterPolicy=filterPolicy,
poolings=poolings,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
fixedSize=fixedSize,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
random=random,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
nbTrees=nbTrees,
maxDepth=maxDepth,
minSamplesSplit=minSamplesSplit,
minSamplesLeaf=minSamplesLeaf,
randomClassif=randomClassif,
nbJobsEstimator=nbJobsEstimator,
verbose=verbose,
learningUse=learningUse,
testingUse=testingUse,
saveFile=saveFile,
shouldSave=shouldSave):
randomState = None
if not randomClassif:
randomState = 100
lsSize = learningUse
if learningUse > maxLearningSize:
lsSize = maxLearningSize
tsSize = testingUse
if testingUse > maxTestingSize:
tsSize = maxTestingSize
#======INSTANTIATING========#
#--randconv--
randConvCoord = coordinatorRandConvFactory(
nbFilters=nb_filters,
filterPolicy=filterPolicy,
nbSubwindows=nbSubwindows,
subwindowMinSizeRatio=subwindowMinSizeRatio,
subwindowMaxSizeRatio=subwindowMaxSizeRatio,
subwindowTargetWidth=subwindowTargetWidth,
subwindowTargetHeight=subwindowTargetHeight,
poolings=poolings,
subwindowInterpolation=subwindowInterpolation,
includeOriginalImage=includeOriginalImage,
nbJobs=nbJobs,
verbosity=verbosity,
tempFolder=tempFolder,
random=random)
nb_filters = len(randConvCoord.getFilters())
#--SVM--
baseClassif = LinearSVC(verbose=verbose, random_state=randomState)
#--Classifier
classifier = uClassifier(coordinator=randConvCoord,
base_classifier=baseClassif,
n_estimators=nbTrees,
max_depth=maxDepth,
min_samples_split=minSamplesSplit,
min_samples_leaf=minSamplesLeaf,
n_jobs=nbJobsEstimator,
random_state=randomState,
verbose=verbose)
#--Data--
loader = CifarFromNumpies(learningSetDir, learningIndexFile)
learningSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
learningSet = learningSet[0:lsSize]
loader = CifarFromNumpies(testingSetDir, testingIndexFile)
testingSet = FileImageBuffer(loader.getFiles(), NumpyImageLoader())
testingSet = testingSet[0:tsSize]
#=====COMPUTATION=====#
#--Learning--#
print "Starting learning"
fitStart = time()
hist = classifier._preprocess(learningSet, learningPhase=True)
y = learningSet.getLabels()
if shouldSave:
np.savez(saveFile, data=hist.data, indices=hist.indices,
indptr=hist.indptr, shape=hist.shape)
classifier.fit_histogram(hist, y)
fitEnd = time()
print "Learning done", formatDuration(fitEnd-fitStart)
sys.stdout.flush()
#--Testing--#
y_truth = testingSet.getLabels()
predStart = time()
y_pred = classifier.predict(testingSet)
predEnd = time()
accuracy = classifier.accuracy(y_pred, y_truth)
confMat = classifier.confusionMatrix(y_pred, y_truth)
#====ANALYSIS=====#
importance, order = randConvCoord.importancePerFeatureGrp(classifier._visualBagger)
print "==================Bag of Visual Words======================="
print "-----------Filtering--------------"
print "nb_filters", nb_filters
print "filterPolicy", filterPolicy
print "----------Pooling--------------"
print "poolings", poolings
print "--------SW extractor----------"
print "#Subwindows", nbSubwindows
print "subwindowMinSizeRatio", subwindowMinSizeRatio
print "subwindowMaxSizeRatio", subwindowMaxSizeRatio
print "subwindowTargetWidth", subwindowTargetWidth
print "subwindowTargetHeight", subwindowTargetHeight
print "fixedSize", fixedSize
print "------------Misc-----------------"
print "includeOriginalImage", includeOriginalImage
print "random", random
print "tempFolder", tempFolder
print "verbosity", verbosity
print "nbJobs", nbJobs
print "--------Bag of words params + SVC----------"
print "nbTrees", nbTrees
print "maxDepth", maxDepth
print "minSamplesSplit", minSamplesSplit
print "minSamplesLeaf", minSamplesLeaf
print "nbJobsEstimator", nbJobsEstimator
print "verbose", verbose
print "randomState", randomState
print "------------Data---------------"
print "LearningSet size", len(learningSet)
print "TestingSet size", len(testingSet)
print "-------------------------------"
if shouldSave:
print "saveFile", saveFile
print "Fit time", formatDuration(fitEnd-fitStart)
print "Classifcation time", formatDuration(predEnd-predStart)
print "Accuracy", accuracy
print "Leafs", formatBigNumber(classifier.histoSize)
return accuracy, confMat, importance, order