def test_diagonal_empty_codes(self):
importance_matrix = np.array([[
1.,
0.,
], [0., 1.], [0., 0.]])
result = modularity_explicitness.modularity(importance_matrix)
np.testing.assert_allclose(result, 2. / 3.)
def aggregation_modularity(matrix, ys):
"""Aggregation function of the modularity score."""
del ys
score = {}
score["modularity"] = modularity_explicitness.modularity(matrix)
modularity_per_code = compute_modularity_per_code(matrix)
assert len(modularity_per_code) == matrix.shape[0], "Wrong length."
for i in range(len(modularity_per_code)):
score["modularity.code_{}".format(i)] = modularity_per_code[i]
return score
def compute_local_modularity(ground_truth_data,
representation_function,
random_state,
artifact_dir=None,
num_train=gin.REQUIRED,
num_local_clusters=gin.REQUIRED,
batch_size=16):
"""Computes the modularity metric according to Sec 3.
Args:
ground_truth_data: GroundTruthData to be sampled from.
representation_function: Function that takes observations as input and
outputs a dim_representation sized representation for each observation.
random_state: Numpy random state used for randomness.
artifact_dir: Optional path to directory where artifacts can be saved.
num_train: Number of points used for training.
num_local_clusters: how many times to run the local mig calculation.
batch_size: Batch size for sampling.
Returns:
Dictionary with average modularity score and average explicitness
(train and test).
"""
del artifact_dir
mod_results = []
for modrun in range(num_local_clusters):
#print("Generating training set %d." % modrun)
mus_train, ys_train = utils.generate_local_batch_factor_code(
ground_truth_data, representation_function, num_train,
random_state, batch_size)
discretized_mus = utils.make_discretizer(mus_train)
#print(mus_train.shape, ys_train.shape)
mutual_information = utils.discrete_mutual_info(
discretized_mus, ys_train)
# Mutual information should have shape [num_codes, num_factors].
assert mutual_information.shape[0] == mus_train.shape[0]
assert mutual_information.shape[1] == ys_train.shape[0]
mod_results.append(
modularity_explicitness.modularity(mutual_information))
mod_results = np.array(mod_results)
scores = {}
scores["modularity_score"] = np.mean(mod_results)
scores["local_modularity_scores_samples"] = mod_results.tolist()
return scores
def test_one_code_two_factors(self):
importance_matrix = np.diag(5. * np.ones(5))
importance_matrix = np.hstack([importance_matrix, importance_matrix])
result = modularity_explicitness.modularity(importance_matrix)
np.testing.assert_allclose(result, 1. - 1. / 9)
def test_missed_factors(self):
importance_matrix = np.diag(5. * np.ones(5))
result = modularity_explicitness.modularity(importance_matrix[:2, :])
np.testing.assert_allclose(result, 1.0)
def test_redundant_codes(self):
importance_matrix = np.diag(5. * np.ones(5))
importance_matrix = np.vstack([importance_matrix, importance_matrix])
result = modularity_explicitness.modularity(importance_matrix)
np.testing.assert_allclose(result, 1.)
def test_zero(self):
importance_matrix = np.zeros(shape=[10, 10], dtype=np.float64)
result = modularity_explicitness.modularity(importance_matrix)
np.testing.assert_allclose(result, .0)
def test_diagonal(self):
importance_matrix = np.diag(5. * np.ones(5))
result = modularity_explicitness.modularity(importance_matrix)
np.testing.assert_allclose(result, 1.0)
