def test_cascade_first_stage_has_no_score_mask():
X, _, qid = fixtures.train_data()
cascade = factories.cascade(num_stages=1, cutoffs=[5])
ranker = cascade.rankers[0]
cascade.predict(X, qid)
assert Cascade.SCORE_MASK not in ranker.predict
def cascade(num_stages, cutoffs, score_type='independent'):
X, y, qid = fixtures.train_data()
X_val, y_val, qid_val = fixtures.valid_data()
config = fixtures.cascade_config(num_stages=num_stages, cutoffs=cutoffs)
config['score_type'] = score_type
cost = fixtures.feature_costs()
cascade = Cascade(config, cost)
cascade.create_boosters(X, y, qid, X_val, y_val, qid_val)
return cascade
def test_cascade_first_stage_applies_cutoff():
X, _, qid = fixtures.train_data()
cascade = factories.cascade(num_stages=1, cutoffs=[2])
ranker = cascade.rankers[0]
ranker.booster.update()
offsets = group_offsets(qid)
a, b = next(offsets)
cascade.predict(X, qid)
expected = (b - a) * [0.01948363]
np.testing.assert_almost_equal(ranker.kappa[a:b], expected)
def test_cascade_score_mask_does_not_appear_in_first_stage():
X, _, qid = fixtures.train_data()
cascade = factories.cascade(num_stages=2, cutoffs=[4, 2])
cascade.update()
ranker_one = cascade.rankers[0]
ranker_two = cascade.rankers[1]
offsets = group_offsets(qid)
a, b = next(offsets)
cascade.predict(X, qid, is_train=True)
assert Cascade.SCORE_MASK not in ranker_one.predict
def test_cascade_second_stage_applies_cutoff():
X, _, qid = fixtures.train_data()
cascade = factories.cascade(num_stages=2, cutoffs=[4, 2])
cascade.update()
ranker_two = cascade.rankers[1]
offsets = group_offsets(qid)
a, b = next(offsets)
cascade.predict(X, qid)
topk = sorted(ranker_two.predict[a:b], reverse=True)
expected = (b - a) * [topk[1]]
np.testing.assert_almost_equal(ranker_two.kappa[a:b], expected)
def test_cascade_second_stage_applies_mask():
X, _, qid = fixtures.train_data()
cascade = factories.cascade(num_stages=2, cutoffs=[4, 2])
cascade.update()
ranker_one = cascade.rankers[0]
ranker_two = cascade.rankers[1]
offsets = group_offsets(qid)
a, b = next(offsets)
cascade.predict(X, qid)
expected = [1, 0, 1, 1, 1]
np.testing.assert_almost_equal(ranker_one.mask[a:b], expected)
expected = [0, 0, 1, 1, 0]
np.testing.assert_almost_equal(ranker_two.mask[a:b], expected)
def test_cascade_uses_score_mask():
"""As per previous implementation, always use the SCORE_MASK during predict
regardless of whether we are doing training or inference.
"""
X, _, qid = fixtures.train_data()
cascade = factories.cascade(num_stages=2, cutoffs=[4, 2])
cascade.update()
ranker_one = cascade.rankers[0]
ranker_two = cascade.rankers[1]
offsets = group_offsets(qid)
a, b = next(offsets)
for is_train in [True, False]:
cascade.predict(X, qid, is_train=is_train)
assert Cascade.SCORE_MASK in ranker_two.predict