def test(self):
x_train = np.linspace(-5, 5, 20)
x_test = np.linspace(-6, 6, 1000)
def f(x):
return 0.01 * x**3 + np.log(np.abs(x))
y_train = f(x_train)
y_test = f(x_test)
x_train = x_train.reshape(-1, 1)
x_test = x_test.reshape(-1, 1)
mgcv = MGCV().fit(x_train, y_train)
self.assertLess(np.abs(mgcv.predict(x_test) - y_test).mean(), 0.15)
self.assertIsInstance(mgcv.get_details(), dict)
def test_splines(self):
mgcv = MGCV().fit(self.x_train, self.y_train)
details = mgcv.get_details()
self.assertEqual(
details['formula'], 'y ~ s(x1) + s(x2) + s(x3) + s(x4) + s(x5) '
'+ s(x6) + s(x7) + s(x8) + s(x9) + s(x10)')
for formula_ in [
'y ~ x1 + s(x1)', 'y ~ ti(x1) + ti(x2)', 'y ~ t2(x1) + s(x2)',
'y ~ s(x1, x2, x3)', 'y ~ s(x1, x2) + ti(x3)'
]:
mgcv = MGCV(formula=formula_).fit(self.x_train, self.y_train)
details = mgcv.get_details()
self.assertEqual(details['formula'], formula_)
formula_ = details['formula'].split('~')[1].replace(' ', '')
formula_parts = formula_.split('+')
for fp in formula_parts:
if 'Intercept' in fp:
continue
if fp.startswith('s') or fp.startswith('ti') \
or fp.startswith('t2'):
assert fp in details['smooth_terms']['edf']
else:
assert fp in details['parametric_terms']['Estimate']
def test_consistency(self):
mgcv1 = MGCV(formula='y ~ s(x1)').fit(self.x_train, self.y_train)
mgcv2 = MGCV(formula='y ~ s(x1)').fit(self.x_train, self.y_train)
self.assertEqual(
np.linalg.norm(
mgcv1.predict(self.x_test) - mgcv1.predict(self.x_test)), 0.0)
self.assertEqual(
np.linalg.norm(
mgcv1.predict(self.x_test) - mgcv2.predict(self.x_test)), 0.0)
mgcv3 = MGCV(formula='y ~ s(x1)',
gam_type='bam').fit(self.x_train, self.y_train)
self.assertLess(
np.linalg.norm(
mgcv1.predict(self.x_test) - mgcv3.predict(self.x_test)), 0.02)
def test_splies_similarity(self):
mgcv1 = MGCV(formula='y ~ s(x1, bs="tp")').fit(self.x_train,
self.y_train)
mgcv2 = MGCV(formula='y ~ s(x1)').fit(self.x_train, self.y_train)
self.assertEqual(
np.linalg.norm(
mgcv1.predict(self.x_test) - mgcv1.predict(self.x_test)), 0.0)
self.assertEqual(
np.linalg.norm(
mgcv1.predict(self.x_test) - mgcv2.predict(self.x_test)), 0.0)
mgcv3 = MGCV(formula='y ~ s(x1, bs="ps")').fit(self.x_train,
self.y_train)
self.assertLess(
np.abs(mgcv1.predict(self.x_test) -
mgcv3.predict(self.x_test)).mean(), 0.06)
mgcv4 = MGCV(formula='y ~ s(x1, bs="ad")').fit(self.x_train,
self.y_train)
self.assertLess(
np.abs(mgcv1.predict(self.x_test) -
mgcv4.predict(self.x_test)).mean(), 0.06)