def test_basic_successive_halving(c, s, a, b):
# Most of the basics are tested through Hyperband (which relies on
# successive halving)
model = SGDClassifier(tol=1e-3)
params = {"alpha": np.logspace(-3, 0, num=1000)}
n, r = 10, 5
search = SuccessiveHalvingSearchCV(model, params, n, r)
X, y = make_classification()
yield search.fit(X, y, classes=np.unique(y))
assert search.best_score_ > 0
assert isinstance(search.best_estimator_, SGDClassifier)
def test_successive_halving_params(c, s, a, b):
# Makes sure when SHAs are fit with values from the "SuccessiveHalvingSearchCV
# params" key, the number of models/calls stay the same as Hyperband.
# This sanity check again makes sure parameters passed correctly
# (similar to `test_params_passed`)
X, y = make_classification(n_samples=10, n_features=4, chunks=10)
model = ConstantFunction()
params = {"value": scipy.stats.uniform(0, 1)}
alg = HyperbandSearchCV(model, params, max_iter=27, random_state=42)
kwargs = [
v["SuccessiveHalvingSearchCV params"] for v in alg.metadata["brackets"]
]
SHAs = [SuccessiveHalvingSearchCV(model, params, **v) for v in kwargs]
metadata = alg.metadata["brackets"]
for k, (true_meta, SHA) in enumerate(zip(metadata, SHAs)):
yield SHA.fit(X, y)
n_models = len(SHA.model_history_)
pf_calls = [
v[-1]["partial_fit_calls"] for v in SHA.model_history_.values()
]
assert true_meta["n_models"] == n_models
assert true_meta["partial_fit_calls"] == sum(pf_calls)
def test_search_patience_infeasible_tol(c, s, a, b):
X, y = make_classification(n_samples=100, n_features=5)
params = {"value": np.random.RandomState(42).rand(1000)}
model = ConstantFunction()
search = SuccessiveHalvingSearchCV(
model,
params,
patience=2,
tol=np.nan,
n_initial_parameters=20,
n_initial_iter=4,
max_iter=1000,
)
yield search.fit(X, y, classes=[0, 1])
assert search.metadata_["partial_fit_calls"] == search.metadata[
"partial_fit_calls"]
assert search.metadata_ == search.metadata
def _test_sha_max_iter(c, s, a, b):
model = SGDClassifier(tol=1e-3)
params = {"alpha": np.logspace(-3, 0, num=1000)}
search = SuccessiveHalvingSearchCV(model,
params,
n_initial_parameters=n,
n_initial_iter=r)
X, y = make_classification()
yield search.fit(X, y, classes=np.unique(y))
calls = set(search.cv_results_["partial_fit_calls"]) - {1}
assert min(calls) == r
# One model trained to completion
assert (search.cv_results_["partial_fit_calls"]
== max(calls)).sum() < search.aggressiveness
assert search.metadata == search.metadata_
assert set(search.metadata.keys()) == {
"partial_fit_calls",
"n_models",
"max_iter",
}
