def test_lars_copy_X(datatype):
X, y = load_boston(return_X_y=True)
X = cp.asarray(X, dtype=datatype, order='F')
y = cp.asarray(y, dtype=datatype, order='F')
X0 = cp.copy(X)
culars1 = cuLars(precompute=False, copy_X=True)
culars1.fit(X, y)
# Test that array was not changed
assert cp.all(X0 == X)
def test_lars_collinear(datatype, nrows, column_info, precompute):
ncols, n_info = column_info
X_train, X_test, y_train, y_test = make_regression_dataset(
datatype, nrows, ncols, n_info)
n_duplicate = min(ncols, 100)
X_train = np.concatenate((X_train, X_train[:, :n_duplicate]), axis=1)
X_test = np.concatenate((X_test, X_test[:, :n_duplicate]), axis=1)
params = {"precompute": precompute, "n_nonzero_coefs": ncols + n_duplicate}
culars = cuLars(**params)
culars.fit(X_train, y_train)
assert culars.score(X_train, y_train) > 0.85
assert culars.score(X_test, y_test) > 0.85
def test_lars_model(datatype, nrows, column_info, precompute, normalize):
ncols, n_info = column_info
X_train, X_test, y_train, y_test = make_regression_dataset(
datatype, nrows, ncols, n_info)
if precompute == 'precompute' or not normalize:
# Apply normalization manually, because the solver expects normalized
# input data
X_train, y_train, x_mean, x_scale, y_mean = \
normalize_data(X_train, y_train)
y_test = y_test - y_mean
X_test = (X_test - x_mean) / x_scale
if precompute == 'precompute':
precompute = np.dot(X_train.T, X_train)
params = {'precompute': precompute, 'normalize': normalize}
# Initialization of cuML's LARS
culars = cuLars(**params)
# fit and predict cuml LARS
culars.fit(X_train, y_train)
cu_score_train = culars.score(X_train, y_train)
cu_score_test = culars.score(X_test, y_test)
if nrows < 500000:
# sklearn model initialization, fit and predict
sklars = skLars(**params)
sklars.fit(X_train, y_train)
# Set tolerance to include the 95% confidence interval around
# scikit-learn accuracy.
accuracy_target = sklars.score(X_test, y_test)
tol = 1.96 * np.sqrt(accuracy_target * (1.0 - accuracy_target) / 100.0)
if tol < 0.001:
tol = 0.001 # We allow at least 0.1% tolerance
print(cu_score_train, cu_score_test, accuracy_target, tol)
assert cu_score_train >= sklars.score(X_train, y_train) - tol
assert cu_score_test >= accuracy_target - tol
else:
assert cu_score_test > 0.95
def test_lars_attributes(datatype, params):
X, y = load_boston(return_X_y=True)
X = X.astype(datatype)
y = y.astype(datatype)
culars = cuLars(**params)
culars.fit(X, y)
sklars = skLars(**params)
sklars.fit(X, y)
assert culars.score(X, y) >= sklars.score(X, y) - 0.01
limit_max_iter = "n_nonzero_coefs" in params
if limit_max_iter:
n_iter_tol = 0
else:
n_iter_tol = 2
assert abs(culars.n_iter_ - sklars.n_iter_) <= n_iter_tol
tol = 1e-4 if params.pop("fit_intercept", True) else 1e-1
n = min(culars.n_iter_, sklars.n_iter_)
assert array_equal(culars.alphas_[:n],
sklars.alphas_[:n],
unit_tol=tol,
total_tol=1e-4)
assert array_equal(culars.active_[:n], sklars.active_[:n])
if limit_max_iter:
assert array_equal(culars.coef_, sklars.coef_)
if hasattr(sklars, 'coef_path_'):
assert array_equal(culars.coef_path_,
sklars.coef_path_[sklars.active_],
unit_tol=1e-3)
intercept_diff = abs(culars.intercept_ - sklars.intercept_)
if abs(sklars.intercept_) > 1e-6:
intercept_diff /= sklars.intercept_
assert intercept_diff <= 1e-3
def test_lars_collinear(datatype, nrows, column_info, precompute):
ncols, n_info = column_info
if nrows == 500000 and ncols == 1000 and pytest.max_gpu_memory < 32:
if pytest.adapt_stress_test:
nrows = nrows * pytest.max_gpu_memory // 32
else:
pytest.skip("Insufficient GPU memory for this test."
"Re-run with 'CUML_ADAPT_STRESS_TESTS=True'")
X_train, X_test, y_train, y_test = make_regression_dataset(
datatype, nrows, ncols, n_info)
n_duplicate = min(ncols, 100)
X_train = np.concatenate((X_train, X_train[:, :n_duplicate]), axis=1)
X_test = np.concatenate((X_test, X_test[:, :n_duplicate]), axis=1)
params = {"precompute": precompute, "n_nonzero_coefs": ncols + n_duplicate}
culars = cuLars(**params)
culars.fit(X_train, y_train)
assert culars.score(X_train, y_train) > 0.85
assert culars.score(X_test, y_test) > 0.85