def test_linear_weight_init_in_sofm(self):
sofm = algorithms.SOFM(
n_inputs=4,
features_grid=(3, 3),
weight='init_pca',
)
input_data = np.random.random((100, 4))
self.assertTrue(callable(sofm.weight))
sofm.train(input_data, epochs=0)
self.assertFalse(callable(sofm.weight))
self.assertEqual(sofm.weight.shape, (4, 9))
for row in (0, 3, 6):
left = sofm.weight[:, row]
center = sofm.weight[:, row + 1]
right = sofm.weight[:, row + 2]
self.assertLess(np.linalg.norm((left - center)**2),
np.linalg.norm((left - right)**2))
for i in range(3):
top = sofm.weight[:, i]
center = sofm.weight[:, i + 3]
bottom = sofm.weight[:, i + 6]
self.assertLess(np.linalg.norm((top - center)**2),
np.linalg.norm((top - bottom)**2))
def test_sofm_hexagon_grid(self):
data = make_circle(max_samples=100)
sofm = algorithms.SOFM(
n_inputs=2,
n_outputs=9,
learning_radius=1,
reduce_radius_after=4,
features_grid=(3, 3),
verbose=False,
grid_type='hexagon',
)
sofm.train(data, epochs=10)
grid = sofm.weight.reshape((2, 3, 3))
center = grid[:, 1, 1]
top_left = grid[:, 0, 0]
top_right = grid[:, 0, 2]
distance_top_left = np.linalg.norm(center - top_left)
distance_top_right = np.linalg.norm(center - top_right)
self.assertLess(distance_top_right, distance_top_left)
bottom_left = grid[:, 2, 0]
bottom_right = grid[:, 2, 2]
distance_bottom_left = np.linalg.norm(center - bottom_left)
distance_bottom_right = np.linalg.norm(center - bottom_right)
self.assertLess(distance_bottom_right, distance_bottom_left)
def test_invalid_attrs(self):
with self.assertRaisesRegexp(ValueError, "Feature grid"):
# Invalid feature grid shape
algorithms.SOFM(n_inputs=2, n_outputs=4, features_grid=(2, 3))
with self.assertRaisesRegexp(ValueError, "n_outputs, features_grid"):
algorithms.SOFM(n_inputs=2)
with self.assertRaisesRegexp(ValueError, "more than 2 dimensions"):
sofm = algorithms.SOFM(n_inputs=2, n_outputs=3, weight=self.weight)
sofm.train(np.zeros((10, 2, 1)))
with self.assertRaisesRegexp(ValueError, "more than 2 dimensions"):
sofm = algorithms.SOFM(n_inputs=2, n_outputs=3, weight=self.weight)
sofm.predict(np.zeros((10, 2, 1)))
with self.assertRaisesRegexp(ValueError, "Input data expected"):
sofm = algorithms.SOFM(n_inputs=2, n_outputs=3, weight=self.weight)
sofm.train(np.zeros((10, 10)))
with self.assertRaisesRegexp(ValueError, "Input data expected"):
sofm = algorithms.SOFM(n_inputs=2, n_outputs=3, weight=self.weight)
sofm.predict(np.zeros((10, 10)))
with self.assertRaisesRegexp(ValueError, "one or two dimensional"):
algorithms.SOFM(n_inputs=2,
features_grid=(3, 1, 1),
grid_type='hexagon')
def test_sofm_init_during_the_training(self):
sofm = algorithms.SOFM(
n_inputs=4,
n_outputs=7,
weight='sample_from_data',
)
X = np.random.random((10, 4))
self.assertTrue(callable(sofm.weight))
sofm.train(X, epochs=1)
self.assertFalse(callable(sofm.weight))
def test_sample_data_weight_init_in_sofm(self):
sofm = algorithms.SOFM(
n_inputs=4,
n_outputs=7,
weight='sample_from_data',
)
input_data = np.random.random((10, 4))
self.assertTrue(callable(sofm.weight))
sofm.train(input_data, epochs=0)
self.assertFalse(callable(sofm.weight))
self.assertEqual(sofm.weight.shape, (4, 7))
def test_sofm_storage(self):
input_data = np.random.random((100, 10))
sofm = algorithms.SOFM(n_inputs=10,
features_grid=(10, 10),
std=1,
step=0.5,
learning_radius=2,
weight=np.random.random((10, 100)))
sofm.train(input_data, epochs=10)
self.assertPickledNetwork(sofm, input_data)
parameters = sofm.get_params()
self.assertIn('weight', parameters)
self.assertIsInstance(parameters['weight'], np.ndarray)
def test_sofm_training_with_4d_grid(self):
sofm = algorithms.SOFM(
n_inputs=4,
n_outputs=8,
features_grid=(2, 2, 2),
verbose=False,
)
data = np.concatenate([input_data, input_data], axis=1)
sofm.train(data, epochs=1)
error_after_first_epoch = sofm.errors.last()
sofm.train(data, epochs=9)
self.assertLess(sofm.errors.last(), error_after_first_epoch)
def test_sofm_weight_norm_after_training_with_custom_weights(self):
sofm = algorithms.SOFM(n_inputs=2,
n_outputs=3,
distance='cos',
learning_radius=1,
features_grid=(3, 1),
weight=X[(0, 2, 4), :].T,
verbose=False)
actual_norms = np.linalg.norm(sofm.weight, axis=0)
expected_norms = np.array([1, 1, 1])
np.testing.assert_array_almost_equal(expected_norms, actual_norms)
sofm.train(X, epochs=6)
actual_norms = np.linalg.norm(sofm.weight, axis=0)
expected_norms = np.array([1, 1, 1])
np.testing.assert_array_almost_equal(expected_norms, actual_norms)
