MKL hangs with Chi2 kernels

[iarroyof]

Hi all. I need your help. I'm trying to setup an evaluation framework for MKL by using different types of kernels. However while I try to train MKL weights for multiple Chi2 basis, the program hung at the weight learning stage [1] (See at the bottom). This is my code:

from modshogun import *
from tools.load import LoadMatrix

# Loading toy data from files
def load_Toy(dataRoute, fileTrain, fileLabels):
    lm = LoadMatrix()
    dataSet = lm.load_numbers(dataRoute + fileTrain)
    labels = lm.load_labels(dataRoute + fileLabels)

return (dataSet.T[0:3*len(dataSet.T)/4].T,  # Return the training set, 3/4 * dataSet
           dataSet.T[(3*len(dataSet.T)/4):].T,  # Return the test set, 1/4 * dataSet
           labels[0:3*len(labels)/4],               # Return corresponding train and test labels
           labels[(3*len(labels)/4):])

# Data importation
[traindata, testdata, trainlab, testlab] = load_Toy('../shogun-data/toy/',  # Data rute
               'fm_train_multiclass_digits500.dat',     # Multiclass dataSet examples file name
               'label_train_multiclass_digits500.dat')          # Multiclass Labels file name
feats_train = RealFeatures(traindata)   # train examples
labelsTr = MulticlassLabels(trainlab)   # train multiclass labels
feats_test = RealFeatures(testdata)     # test examples
labelsTs = MulticlassLabels(testlab)    # test multiclass labels

# Parameters
weightNorm = 2
regParam = 2
epsilon = 1e-5
threads = 2
mkl_epsilon = 0.001

# Setting up the MKL machine
mkl = MKLMulticlass()                   # MKL object 
mkl.set_C(regParam)                     # Setting multiclass regularization parameter
mkl.set_mkl_norm(weightNorm)        # Setting the weight vector norm
mkl.set_epsilon(epsilon)                # setting the transducer epsilon
mkl.set_mkl_epsilon(mkl_epsilon)
kernels = []
kernels.append(Chi2Kernel(l = feats_train, r = feats_train, width = 5))
kernels.append(Chi2Kernel(l = feats_train, r = feats_train, width = 25))
kernels.append(Chi2Kernel(l = feats_train, r = feats_train, width = 105))

combKer = CombinedKernel()
for k in kernels:
    combKer.append_kernel(k)

combKer.init(feats_train, feats_train)
print '[0] Kernel fitted...'
mkl.set_kernel(combKer)
mkl.set_labels(labelsTr)
print '[1] Going to train the mkl machine...'
# Train the weights to return the learnt kernel
mkl.train()
print '[2] Kernel trained... Weights: ', combKer.get_subkernel_weights()
# Now with test samples. 
combKer.init(feats_train, feats_test)       # The inner product between training
mkl.set_kernel(combKer)                 # and test examples generates the corresponding Gramm Matrix.
out = mkl.apply()                       # Applying the obtained Gramm Matrix
# Evaluating the test performance
evalua = MulticlassAccuracy()
testerr = evalua.evaluate(out, labelsTs)

print '[3] Kernel evaluation ready. The precision was: ', testerr*100, '%'

Any help would be very appreciated. Thank you in advance.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.