def normalization_test():
X = np.random.rand(100, 100) - 100
estimator = DictionaryLearning(k=10, dict_normalization=1)
estimator.fit(X)
C, D = estimator.decomposition()
for i in range(D.shape[0]):
np.testing.assert_approx_equal(np.linalg.norm(D[0, :]), 1)
def right_number_of_atoms_test():
X = np.random.rand(100, 100)
estimator = DictionaryLearning(k=10)
estimator.fit(X)
C, D = estimator.decomposition()
assert C.shape[1] == 10
assert D.shape[0] == 10
def non_negativity_decomposition_on_coefficients_test():
X = np.random.rand(100, 100) - 100
estimator = DictionaryLearning(k=5, non_negativity="coeff")
estimator.fit(X)
C, D = estimator.decomposition()
assert np.min(C) >= 0
assert np.min(D) < 0
def non_negativity_decomposition_on_both_test():
X = np.abs(np.random.rand(100, 100))
estimator = DictionaryLearning(k=5, non_negativity="both")
estimator.fit(X)
C, D = estimator.decomposition()
assert np.min(C) >= 0
assert np.min(D) >= 0
def wrong_range2_test():
X, _, _ = synthetic_data_negative()
estimator = DictionaryLearning(k=5,
coeff_penalty=L1Penalty(1),
dict_penalty=(L1Penalty(2)),
non_negativity="coeff")
tune_parameters_DL(X,
estimator,
analysis=1,
distributed=0,
dict_penalty_range=(-1, 1, 5),
coeff_penalty_range=(-1, 1, 5))
import matplotlib.pyplot as plt
from dalila.dictionary_learning import DictionaryLearning
from dalila.penalty import GroupLassoPenalty
from dalila.dataset_generator import group_lasso_dataset_generator
logging.basicConfig(level=logging.INFO)
signals, coefficients, atoms = group_lasso_dataset_generator()
plt.pcolor(coefficients, cmap=plt.cm.Reds)
plt.show()
decompositions = []
for i in np.linspace(0.1, 10, 15):
estimator = DictionaryLearning(k=6,
dict_penalty=None,
coeff_penalty=GroupLassoPenalty(groups, i),
non_negativity="coeff")
estimator.fit(signals, n_iter=20000)
decompositions.append(estimator.decomposition())
print("finito", i)
print("reconstruction error", estimator.reconstruction_error())
print("\n\n")
for (i, d) in enumerate(decompositions):
# it is necessary to reorder the atoms as they are declared in atoms to
# see if the algorithm works
C = d[0]
D = d[1]
new_D = np.empty_like(D)
new_C = np.empty_like(C)
def wrong_penalty_type_coeff_test():
X = np.random.rand(100, 100)
estimator = DictionaryLearning(k=5, coeff_penalty="")
estimator.fit(X)
def non_negative_parameter_but_negative_matrix_test():
X = np.random.rand(100, 100) - 50
estimator = DictionaryLearning(k=5, non_negativity="both")
estimator.fit(X)
def non_negative_parameter_test():
X = np.random.rand(100, 100)
estimator = DictionaryLearning(k=5, non_negativity="not_considered")
estimator.fit(X)
def negative_atoms_test():
X = np.random.rand(100, 100)
estimator = DictionaryLearning(k=-1, dict_penalty=None)
estimator.fit(X)