def train(self, X, Y):
# Get relevant data
idx = Y==self.i
jdx = Y==self.j
ijdx = np.logical_or(idx, jdx)
x = X[ijdx]
y = Y[ijdx]
# Project y onto {-1, +1}
y[y==self.i] = -1.0
y[y==self.j] = 1.0
self.Y = y
self.G = self.G[ijdx]
G = self.G.T[ijdx].T
# Train SMO
self.a, self.b = train(x, y, G, C=1e-5, tol=1e-2, eps=1e-3)
def proc((c,code)):
print c,code
coded_labels = train_labels.copy()
# Create the code-mapped label vector
for i, label in enumerate(train_labels):
coded_labels[i] = code[int(label)]
# SVM for this code
w, b = train(train_data, coded_labels, C=0.00001, tol=0.00001, eps=1e-2)
# Hypothesis
h_train = hypothesis(train_data, w, b)
h_test = hypothesis(test_data, w, b)
return c, h_train, h_test
data_file = "../data/spambase/spambase.data"
dmat = []
f = open(data_file, "r")
for line in f:
x = line.split(',')
x = [float(e) for e in x]
dmat.append(x)
data = np.array(dmat)
# k-folds xvalidation
k = 10
kfolder = KFolder(data, k, standard=True, shuffle=True)
for i in range(k-1):
print "Fold:", i+1
# Get data and labels at fold k
X,Y = kfolder.testing(i+1)
# Get the testing data
Xi,Yi = kfolder.testing(i)
Yi[Yi==0] = -1.0
# Train
Y[Y==0] = -1.0
G, Gi = gram(X), tgram(X, Xi)
a, b = train(X, Y.ravel(), G, C=1e-4, tol=1e-4, eps=1e-3)
# Test
print "Training accuracy:", test(Y, Y, G, a, b)
print "Testing accuracy:", test(Y, Yi, Gi, a, b)