def test_entropy(distributions, reduction):
np.random.seed(1337)
entropy = Entropy(reduction=reduction)
marg = entropy(distributions)
assert np.all(marg == [1, 2, 0]), "Entropy is not right {}".format(marg)
entropy = Entropy(0.9, reduction=reduction)
marg = entropy(distributions)
assert np.any(marg != [1, 2, 0])
def test_heuristics_reorder_list():
# we are just testing if given calculated uncertainty measures for chunks of data
# the `reorder_indices` would make correct decision. Here index 0 has the
# highest uncertainty chosen but both methods (uncertainties1 and uncertainties2)
streaming_prediction = [np.array([0.98]), np.array([0.87, 0.68]),
np.array([0.96, 0.54])]
heuristic = BALD()
ranks = heuristic.reorder_indices(streaming_prediction)
assert np.all(ranks == [0, 3, 1, 2, 4]), "reorder list for BALD is not right {}".format(ranks)
heuristic = Variance()
ranks = heuristic.reorder_indices(streaming_prediction)
assert np.all(ranks == [0, 3, 1, 2, 4]), "reorder list for Variance is not right {}".format(
ranks)
heuristic = Entropy()
ranks = heuristic.reorder_indices(streaming_prediction)
assert np.all(ranks == [0, 3, 1, 2, 4]), "reorder list for Entropy is not right {}".format(
ranks)
heuristic = Margin()
ranks = heuristic.reorder_indices(streaming_prediction)
assert np.all(ranks == [4, 2, 1, 3, 0]), "reorder list for Margin is not right {}".format(ranks)
heuristic = Certainty()
ranks = heuristic.reorder_indices(streaming_prediction)
assert np.all(ranks == [4, 2, 1, 3, 0]), "reorder list for Certainty is not right {}".format(
ranks)
heuristic = Random()
ranks = heuristic.reorder_indices(streaming_prediction)
assert ranks.size == 5, "reorder list for Random is not right {}".format(
ranks)
return logits
probability_distribution = wrapped(None, logits)
assert np.alltrue((probability_distribution >= 0)
& (probability_distribution <= 1)).all()
def test_that_precomputed_passes_back_predictions():
precomputed = Precomputed()
ranks = np.arange(10)
assert (precomputed(ranks) == ranks).all()
@pytest.mark.parametrize('heuristic1, heuristic2, weights',
[(BALD(), Variance(), [0.7, 0.3]),
(BALD(), Entropy(reduction='mean'), [0.9, 0.8]),
(Entropy(), Variance(), [4, 8]),
(Certainty(), Variance(), [9, 2]),
(Certainty(), Certainty(reduction='mean'), [1, 3])])
def test_combine_heuristics(heuristic1, heuristic2, weights):
np.random.seed(1337)
predictions = [distributions_3d, distributions_5d]
if isinstance(heuristic1,
Certainty) and not isinstance(heuristic2, Certainty):
with pytest.raises(Exception) as e_info:
heuristics = CombineHeuristics([heuristic1, heuristic2],
weights=weights,
reduction='mean')
assert 'heuristics should have the same value for `revesed` parameter' in str(
e_info.value)