def getOptCG(labels, values):
# Format data for subsequent methods
# Set up cross-validation settings
param = Param()
param.cset = range(-5, 15, 2)
param.gset = range(3, -15, -2)
param.nfold = 10
prob = svm_problem(labels, values)
rVal = [[0 for col in range(len(param.cset))] for row in range(len(param.gset))];
# Cross-validation to get optimal parameters
for i in range(len(param.gset)):
param.g = 2 ** param.gset[i]
for j in range(len(param.cset)):
param.c = 2 ** param.cset[j]
testParam = svm_parameter(param.libsvm)
# Train on learning data with x-validation and store result
rVal[i][j] = svm_train(prob, param.libsvm + " -v " + str(param.nfold))
# Select the parameters with highest accuracy
min_val, loc = getMax(rVal)
g = 2 ** param.gset[loc[0]]
c = 2 ** param.cset[loc[1]]
return c, g
def svmtrain(labels, values=None, c=None, g=None):
# If Dictionary
if isinstance(labels, dict):
values = [j for i in labels.itervalues() for j in i.itervalues()]
labels = [i + 1 for i in range(len(labels.values())) for j in range(len(labels[labels.keys()[i]]))]
if values != None:
optParam = Param()
optParam.c = c
optParam.g = g
if c == None or g == None:
# Retrieve optimal c and g
optParam.c, optParam.g = getOptCG(labels, values)
# Train model with optimal c and g
prob = svm_problem(labels, values)
m = svm_train(prob, optParam.libsvm)
# Return model
return m
else:
raise TypeError("Values not provided for the arguments")
def xTrain(labels, values=None, k=10, rand=True):
# Split data into k partitions
partitions = split(labels, values, k, rand)
best_model = None
best_acc = -1
avg_acc = 0
count = 0
if isinstance(labels, dict):
values = [j for i in labels.itervalues() for j in i.itervalues()]
labels = [i + 1 for i in range(len(labels.values())) for j in range(len(labels[labels.keys()[i]]))]
if values != None:
optParam = Param()
print "Searching for optimal parameters..."
optParam.c, optParam.g = getOptCG(labels, values)
print "c: " + str(optParam.c) + " g: " + str(optParam.g)
print " "
# For each partition, train a model and check the accuracy of the partition's test data against
# the model. The highest accuracy model will be the model returned. The accuracy returned is the
# average accuracy of all the partitions
for i in partitions.iter:
print "Training iteration " + str(count + 1) + ".."
# Train the model using the partition training data
model = svmtrain(i.train.labels, i.train.values, optParam.c, optParam.g)
# Get a list of predictions for the testing data
pred = svmtest(i.test.values, model)
# Find the accuracy of the test data predicted by the model
acc, x1, x2 = evaluations(i.test.labels, pred)
# Store the model with the best accuracy
if acc > best_acc:
best_acc = acc
best_model = model
print "Iteration " + str(count + 1) + " accuracy: " + str(acc)
avg_acc += acc
count += 1
# Get the avg accuracy
avg_acc /= count
print " "
print "xTrain completed."
return model, avg_acc