def run(input_dir, num_samples_per_site, onevone = True):
input_dir = sys.argv[1]
num_samples_per_site = int(sys.argv[2])
target_sites = get_target_sites(input_dir)
labels = dict(map(lambda (i, s): (s, i), enumerate(target_sites)))
X, Y = load_feature_vectors(input_dir, num_samples_per_site, labels)
#X, Y, testX, testY = select_test_set(X, Y, (num_samples_per_site / 2) * len(target_sites))
#X = [[100.0, 100.0], [101.0, 101.0], [200.0, 200.0], [201.0, 201.0], [300.0, 301.0], [300.0, 301.0]]
#Y = [0.0, 0.0, 1.0, 1.0, 2.0, 2.0]
iters = select_cross_validation_subsets(X, Y, 6)
#crange = map(lambda c: c/10.0, range(1, 11))
#crange = map(lambda c: c/1000.0, range(1, 11))
#crange.extend(map(lambda c: c/100.0, range(2, 11)))
#crange = [0.00001, 0.01, 1.0, 100.0, 10000.0]
crange = [1.0, 100.0]
for c in crange:
print "C=%f"%c
print "Linear kernel:"
for d in iters:
trainX, testX = scale(d["train"][0], d["test"][0])
classify(trainX, d["train"][1], testX, d["test"][1], onevone, c)
print "RBF kernel:"
for d in iters:
trainX, testX = scale(d["train"][0], d["test"][0])
classify(trainX, d["train"][1], testX, d["test"][1], onevone, c, 'rbf')
def anomaly_detection(labels, Xnew, Y, testXnew, testY):
print "Anomaly detection"
for test_class in range(len(labels)):
anomX = []
for (i, x) in enumerate(Xnew):
if Y[i] == test_class:
anomX.append(x)
minD = []
# Translate so all coordinates are positive
for d in range(len(anomX[0])):
minD.append(min(map(lambda x: x[d], anomX)))
anomX = translate(minD, anomX)
anomX = scale(anomX)
rho, alphas = AnomDet_fit(anomX, 0.1, rbf)
test = testXnew
test = translate(minD, test)
test = scale(test)
num_correct = [0, 0]
predictions = [0, 0]
for (i, x) in enumerate(test):
classify = AnomDet_classify(x, alphas, rho, anomX, rbf)
#print "Label: %d, classification: %d"%(testY[i], classify)
if testY[i] == test_class:
predictions[0] = predictions[0] + 1
if classify == 0.0:
num_correct[0] = num_correct[0] + 1
else:
predictions[1] = predictions[1] + 1
if classify != 0.0:
num_correct[1] = num_correct[1] + 1
print "Test class %d. Normal correct: %d/%d, anomaly correct: %d/%d"%(
test_class, num_correct[0], predictions[0], num_correct[1],
predictions[1])
def multiclass_svm(X, Y, testX, testY, labels, c=1.0):
X, testX = scale(X, testX)
try:
shutil.rmtree("tmp_x")
except:
print "No tmp directory to delete"
os.mkdir("tmp_x")
out = open("tmp_x/x.dat", mode="w+")
for x in X:
out.write(json.dumps(x.tolist()))
out.write("\n")
out.close()
thetas, bs, slacks = SVM_fit("tmp_x/x.dat", Y, len(labels), len(X[0]), c)
num_correct = 0
for (i, x) in enumerate(testX):
result = SVM_classify(x, thetas, bs)
if result == testY[i]:
num_correct = num_correct + 1
print "Num correct: %d/%d"%(num_correct, len(testY))
def multiclass_svm(Xnew, testXnew, Y, testY, labels):
print "Classifying with a multiclass SVM"
Xnew, testXnew = scale(Xnew, testXnew)
try:
shutil.rmtree("tmp_x")
except:
print "No tmp directory to delete"
os.mkdir("tmp_x")
out = open("tmp_x/x.dat", mode="w+")
for x in Xnew:
out.write(json.dumps(x.tolist()))
out.write("\n")
out.close()
out = open("tmp_x/xtest.dat", mode="w+")
for x in testXnew:
out.write(json.dumps(x.tolist()))
out.write("\n")
out.close()
d = len(Xnew[0])
del Xnew
del testXnew
thetas, bs, slacks = SVM_fit("tmp_x/x.dat", Y, len(labels), d, 0.05)
num_correct = 0
tests = open("tmp_x/xtest.dat")
i = 0
for l in tests:
x = json.loads(l)
if (SVM_classify(x, thetas, bs) == testY[i]):
num_correct = num_correct + 1
i = i + 1
print "Num correct: %d/%d"%(num_correct, len(testY))
