def runFM(train_X, train_y, test_X, test_y, test_X2, label, dev_index,
val_index):
class_weights = {
'toxic': 1.0,
'severe_toxic': 0.2,
'obscene': 1.0,
'threat': 0.1,
'insult': 0.8,
'identity_hate': 0.2
}
model = FM_FTRL(alpha=0.02,
beta=0.01,
L1=0.00001,
L2=30.0,
D=train_X.shape[1],
alpha_fm=0.1,
L2_fm=0.5,
init_fm=0.01,
weight_fm=50.0,
D_fm=200,
e_noise=0.0,
iters=3,
inv_link="identity",
e_clip=1.0,
threads=4,
use_avx=1,
verbose=1)
train_weight = np.array(
[1.0 if x == 1 else class_weights[label] for x in train_y])
model.fit(train_X, train_y, train_weight, reset=False)
pred_test_y = sigmoid(model.predict(test_X))
pred_test_y2 = sigmoid(model.predict(test_X2))
return pred_test_y, pred_test_y2
def testSigmoid(self):
self.assertAlmostEqual(0.5, preprocess.sigmoid(0))
self.assertAlmostEqual(0.7310585786300049, preprocess.sigmoid(1))
self.assertAlmostEqual(0.2689414213699951, preprocess.sigmoid(-1))
t = numpy.array([-2, -1, 0, 1, 2])
expected = numpy.array([0.11920292202211755, 0.2689414213699951, 0.5, 0.7310585786300049, 0.8807970779778823])
actual = preprocess.sigmoid(t)
self.assertAlmostEqual(0.0, linalg.norm(expected - actual))
def test(x):
mu, sigma = preprocess.muSigma(x)
self.assertAlmostEqual(1.23902738264240, x[1][2])
self.assertEqual(5, len(mu))
self.assertEqual(5, len(sigma))
self.assertAlmostEqual(2.87969736221038, mu[0])
self.assertAlmostEqual(2.04868506865762, sigma[0])
self.assertAlmostEqual(-0.99025024303433, (x[0][0] - mu[0]) / sigma[0])
self.assertAlmostEqual(1.97861578296198, mu[2])
self.assertAlmostEqual(2.33076030134340, sigma[2])
self.assertAlmostEqual(-0.31731637092553, (x[1][2] - mu[2]) / sigma[2])
y = preprocess.normalize(x, mu, sigma)
m, n = y.shape
self.assertEqual(4, m)
self.assertEqual(5, n)
self.assertAlmostEqual(-0.99025024303433, y[0][0])
self.assertAlmostEqual(-0.31731637092553, y[1][2])
u = preprocess.sigmoid(y)
self.assertAlmostEqual(0.27086265279957, u[0][0])
self.assertAlmostEqual(0.42132990768430, u[1][2])
def normalize(self, data):
x, y = data
z = preprocess.sigmoid(preprocess.normalize(x, self.mu, self.sigma))
return numpy.array(z, dtype = numpy.float32), y
def runChainedFM(train_X, train_y, test_X, test_y, test_X2, label, dev_index,
val_index):
print_step('Loading Lvl1')
lvl1_train, lvl1_test = load_cache('lvl1_fm')
[
lvl1_train[c].apply(lambda x: 0 if x < 0.5 else 1)
for c in lvl1_train.columns if 'fm_' in c and c != label
]
lvl1_train = csr_matrix(
pd.concat([
lvl1_train[c].apply(lambda x: 0 if x < 0.5 else 1)
for c in lvl1_train.columns if 'fm_' in c and c != label
],
axis=1).values)
lvl1_test = csr_matrix(
pd.concat([
lvl1_test[c].apply(lambda x: 0 if x < 0.5 else 1)
for c in lvl1_test.columns if 'fm_' in c and c != label
],
axis=1).values)
print_step('Merging 1/3')
lvl1_valid = lvl1_train[val_index]
lvl1_train = lvl1_train[dev_index]
train_X = csr_matrix(hstack([train_X, lvl1_train]))
print_step('Merging 2/3')
test_X = csr_matrix(hstack([test_X, lvl1_valid]))
print_step('Merging 3/3')
test_X2 = csr_matrix(hstack([test_X2, lvl1_test]))
print_step('Modeling')
class_weights = {
'toxic': 1.0,
'severe_toxic': 0.2,
'obscene': 1.0,
'threat': 0.1,
'insult': 0.8,
'identity_hate': 0.2
}
model = FM_FTRL(alpha=0.02,
beta=0.01,
L1=0.00001,
L2=30.0,
D=train_X.shape[1],
alpha_fm=0.1,
L2_fm=0.5,
init_fm=0.01,
weight_fm=50.0,
D_fm=200,
e_noise=0.0,
iters=3,
inv_link="identity",
e_clip=1.0,
threads=4,
use_avx=1,
verbose=1)
train_weight = np.array(
[1.0 if x == 1 else class_weights[label] for x in train_y])
model.fit(train_X, train_y, train_weight, reset=False)
pred_test_y = sigmoid(model.predict(test_X))
pred_test_y2 = sigmoid(model.predict(test_X2))
return pred_test_y, pred_test_y2
