def _equalize(self, equality_sets):
# pylint: disable=R0914
self._is_direction_identified()
if self._verbose:
words_data = []
for equality_set_index, equality_set_words in enumerate(equality_sets):
equality_set_vectors = [
normalize(self[word]) for word in equality_set_words
]
center = np.mean(equality_set_vectors, axis=0)
(projected_center,
rejected_center) = project_reject_vector(center, self.direction)
scaling = np.sqrt(1 - np.linalg.norm(rejected_center)**2)
for word, vector in zip(equality_set_words, equality_set_vectors):
projected_vector = project_vector(vector, self.direction)
projected_part = normalize(projected_vector - projected_center)
# In the code it is different of Bolukbasi
# It behaves the same only for equality_sets
# with size of 2 (pairs) - not sure!
# However, my code is the same as the article
# equalized_vector = rejected_center + scaling * self.direction
# https://github.com/tolga-b/debiaswe/blob/10277b23e187ee4bd2b6872b507163ef4198686b/debiaswe/debias.py#L36-L37
# For pairs, projected_part_vector1 == -projected_part_vector2,
# and this is the same as
# projected_part_vector1 == self.direction
equalized_vector = rejected_center + scaling * projected_part
update_word_vector(self.model, word, equalized_vector)
if self._verbose:
words_data.append({
'equality_set_index':
equality_set_index,
'word':
word,
'scaling':
scaling,
'projected_scalar':
vector @ self.direction,
'equalized_projected_scalar':
(equalized_vector @ self.direction),
})
if self._verbose:
print('Equalize Words Data '
'(all equal for 1-dim bias space (direction):')
words_data_df = (pd.DataFrame(words_data).set_index(
['equality_set_index', 'word']))
print(tabulate(words_data_df, headers='keys'))
self.model.init_sims(replace=True)
def test_project_params():
v = np.array([1, 2, 3])
u = np.array([-4, 5, -6])
(_, projected_vector_v1, rejected_vector_v1) = project_params(v, u)
projected_vector_v2, rejected_vector_v2 = project_reject_vector(u, v)
np.testing.assert_allclose(projected_vector_v1, projected_vector_v2)
np.testing.assert_allclose(rejected_vector_v1, rejected_vector_v2)
def test_project_params():
# pylint: disable=arguments-out-of-order
v = np.array([1, 2, 3])
u = np.array([-4, 5, -6])
(_, projected_vector_v1, rejected_vector_v1) = project_params(v, u)
projected_vector_v2, rejected_vector_v2 = project_reject_vector(u, v)
np.testing.assert_allclose(projected_vector_v1, projected_vector_v2)
np.testing.assert_allclose(rejected_vector_v1, rejected_vector_v2)
def test_equalize(gender_biased_w2v_small, is_preforming=True):
"""Test _equalize method in GenderBiasWE."""
# pylint: disable=line-too-long
equality_sets = {
tuple(w)
for w in gender_biased_w2v_small._data['equalize_pairs']
}
equality_sets |= {
tuple(w)
for w in gender_biased_w2v_small._data['definitional_pairs']
}
equality_sets = gender_biased_w2v_small._generate_pair_candidates(
equality_sets)
if is_preforming:
gender_biased_w2v_small._equalize(equality_sets)
for equality_set in equality_sets:
projection_vectors = []
rejection_vectors = []
for equality_word in equality_set:
vector = gender_biased_w2v_small[equality_word]
np.testing.assert_allclose(np.linalg.norm(vector), 1, atol=ATOL)
(projection_vector, rejection_vector) = project_reject_vector(
vector, gender_biased_w2v_small.direction)
projection_vectors.append(projection_vector)
rejection_vectors.append(rejection_vector)
# <e1, d> == -<e2, d>
# assuming equality sets of size 2
assert len(projection_vectors) == 2
np.testing.assert_allclose(
projection_vectors[0] @ gender_biased_w2v_small.direction,
-projection_vectors[1] @ gender_biased_w2v_small.direction,
atol=ATOL)
# all rejection part is equal for all the vectors
for rejection_vector in rejection_vectors[1:]:
np.testing.assert_allclose(rejection_vectors[0],
rejection_vector,
atol=ATOL)
check_all_vectors_unit_length(gender_biased_w2v_small)