def runSVM(trainF, testF):
"""
[6/27/2011]: This script serves little purpose as it takes to long
to interate over 0-10,000 to optimize C automatically.
It is more effiecinetly done by hand (for now at least).
Go!
"""
import PyML as ml
import numpy as np
znorm = ml.Standardizer()
# Init the SVM
s = SVM()
print(s)
# Reformat the data to csv (aka Vector) so feature based
# normalization can occur, then normalize that train and
# test data
test = ml.SparseDataSet(testF)
vectestF = 'vec_' + testF
test.save(vectestF, format='csv')
vecTest = ml.VectorDataSet(vectestF, labelsColumn=1, idColumn=0)
znorm.train(vecTest)
trainedSVM = []
# Returns a SVM trained class
if isinstance(trainF, str):
train = ml.SparseDataSet(trainF)
vectrainF = 'vec_' + trainF
train.save(vectrainF, format='csv')
vecTrain = ml.VectorDataSet(vectrainF, labelsColumn=1, idColumn=0)
znorm.train(vecTrain)
# Optimize C
param = ml.modelSelection.Param(s, 'C', list(np.arange(0, 10000, .5)))
m = ml.modelSelection.ModelSelector(param)
trainedSVM = m.train(vecTrain) # Optimize C
trainedSVM.save('svm_' + testF)
else:
trainedSVM = trainF
cross = trainedSVM.stratifiedCV(trainedSVM, 10)
results = trainedSVM(vecTest)
return results
def zSparse(fname):
"""
Converts a sparse formated SVMLIB data file to Vector/CSV format
and then znomralizes on a feature basis and writes out that file
as fname.
"""
znorm = ml.Standardizer()
sparse = ml.SparseDataSet(fname)
sparse.save('temp',format='csv')
vec = ml.VectorDataSet('temp',labelsColumn=1,idColumn=0)
znorm.train(vec)
vecName = 'vec_' + fname
# Verbal overwrite of priors
if os.path.exists(vecName):
print('Overwriting {0}.'.format(vecName))
os.remove(vecName)
vec.save(vecName)
