def test_logistic_regression(strategy,
use_wrapper,
nrows,
num_classes,
column_info,
dtype=np.float32):
ncols, n_info = column_info
X_train, X_test, y_train, y_test = make_classification_dataset(
datatype=dtype,
nrows=nrows,
ncols=ncols,
n_info=n_info,
num_classes=num_classes)
y_train = y_train.astype(dtype)
y_test = y_test.astype(dtype)
culog = cuLog()
if use_wrapper:
cls = cu_multiclass.MulticlassClassifier(culog, strategy=strategy)
else:
if (strategy == 'ovo'):
cls = cu_multiclass.OneVsOneClassifier(culog)
else:
cls = cu_multiclass.OneVsRestClassifier(culog)
cls.fit(X_train, y_train)
test_score = cls.score(X_test, y_test)
assert test_score > 0.7
def test_logistic_regression_decision_function(
dtype, nrows, column_info, num_classes, fit_intercept
):
ncols, n_info = column_info
X_train, X_test, y_train, y_test = make_classification_dataset(
datatype=dtype, nrows=nrows, ncols=ncols,
n_info=n_info, num_classes=num_classes
)
y_train = y_train.astype(dtype)
y_test = y_test.astype(dtype)
culog = cuLog(fit_intercept=fit_intercept, output_type="numpy")
culog.fit(X_train, y_train)
sklog = skLog(fit_intercept=fit_intercept)
sklog.coef_ = culog.coef_.T
if fit_intercept:
sklog.intercept_ = culog.intercept_
else:
skLog.intercept_ = 0
sklog.classes_ = np.arange(num_classes)
cu_dec_func = culog.decision_function(X_test)
if num_classes > 2:
cu_dec_func = cu_dec_func.T
sk_dec_func = sklog.decision_function(X_test)
assert array_equal(cu_dec_func, sk_dec_func)
def test_logistic_regression_predict_proba(dtype, nrows, column_info,
num_classes, fit_intercept):
ncols, n_info = column_info
X_train, X_test, y_train, y_test = \
make_classification_dataset(datatype=dtype, nrows=nrows,
ncols=ncols, n_info=n_info,
num_classes=num_classes)
y_train = y_train.astype(dtype)
y_test = y_test.astype(dtype)
culog = cuLog(fit_intercept=fit_intercept)
culog.fit(X_train, y_train)
if num_classes > 2:
sklog = skLog(fit_intercept=fit_intercept, solver="lbfgs",
multi_class="multinomial")
else:
sklog = skLog(fit_intercept=fit_intercept)
sklog.coef_ = culog.coef_.copy_to_host().T
if fit_intercept:
sklog.intercept_ = culog.intercept_.copy_to_host()
else:
skLog.intercept_ = 0
sklog.classes_ = np.arange(num_classes)
cu_proba = culog.predict_proba(X_test).get()
sk_proba = sklog.predict_proba(X_test)
cu_log_proba = culog.predict_log_proba(X_test).get()
sk_log_proba = sklog.predict_log_proba(X_test)
assert array_equal(cu_proba, sk_proba)
assert array_equal(cu_log_proba, sk_log_proba)
def test_logistic_predict_convert_dtype(train_dtype, test_dtype):
X, y = make_classification(n_samples=50, n_features=10, random_state=0)
X = X.astype(train_dtype)
X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.8,
random_state=0)
clf = cuLog()
clf.fit(X_train, y_train)
clf.predict(X_test.astype(test_dtype))
def test_logistic_regression_attributes():
X, y = make_blobs()
clf = cuLog().fit(X, y, convert_dtype=True)
attrs = ["dtype", "solver_model", "coef_", "intercept_",
"l1_ratio", "n_cols", "C", "penalty",
"fit_intercept", "solver"]
for attr in attrs:
assert hasattr(clf, attr)
def test_logistic_regression(num_classes, dtype, penalty, l1_ratio,
fit_intercept, nrows, column_info, C, tol):
if penalty in ['l1', 'elasticnet']:
pytest.xfail("OWL numerical stability is being improved")
ncols, n_info = column_info
# Checking sklearn >= 0.21 for testing elasticnet
sk_check = LooseVersion(str(sklearn.__version__)) >= LooseVersion("0.21.0")
if not sk_check and penalty == 'elasticnet':
pytest.skip("Need sklearn > 0.21 for testing logistic with"
"elastic net.")
X_train, X_test, y_train, y_test = \
make_classification_dataset(datatype=dtype, nrows=nrows,
ncols=ncols, n_info=n_info,
num_classes=num_classes)
y_train = y_train.astype(dtype)
y_test = y_test.astype(dtype)
culog = cuLog(penalty=penalty,
l1_ratio=l1_ratio,
C=C,
fit_intercept=fit_intercept,
tol=tol,
verbose=0)
culog.fit(X_train, y_train)
# Only solver=saga supports elasticnet in scikit
if penalty in ['elasticnet', 'l1']:
if sk_check:
sklog = skLog(penalty=penalty,
l1_ratio=l1_ratio,
solver='saga',
C=C,
fit_intercept=fit_intercept,
multi_class='auto')
else:
sklog = skLog(penalty=penalty,
solver='saga',
C=C,
fit_intercept=fit_intercept,
multi_class='auto')
else:
sklog = skLog(penalty=penalty,
solver='lbfgs',
C=C,
fit_intercept=fit_intercept,
multi_class='auto')
sklog.fit(X_train, y_train)
# Setting tolerance to lowest possible per loss to detect regressions
# as much as possible
assert culog.score(X_test, y_test) >= sklog.score(X_test, y_test) - 0.06
def test_logistic_regression_model_default(dtype):
X_train, X_test, y_train, y_test = small_classification_dataset(dtype)
y_train = y_train.astype(dtype)
y_test = y_test.astype(dtype)
culog = cuLog()
culog.fit(X_train, y_train)
sklog = skLog(multi_class="auto")
sklog.fit(X_train, y_train)
assert culog.score(X_test, y_test) >= sklog.score(X_test, y_test) - 0.022
def test_logistic_regression_input_type_consistency(constructor, dtype):
from cudf.core.frame import Frame
X = constructor([[5, 10], [3, 1], [7, 8]]).astype(dtype)
y = constructor([0, 1, 1]).astype(dtype)
clf = cuLog().fit(X, y, convert_dtype=True)
original_type = type(X)
if constructor == cudf.DataFrame:
original_type = Frame
assert isinstance(clf.predict_proba(X), original_type)
assert isinstance(clf.predict(X), original_type)
def test_logistic_regression_weighting(regression_dataset, option,
test_status):
regression_type, data, coef, output = regression_dataset[test_status]
class_weight = None
sample_weight = None
if option == 'sample_weight':
n_samples = data.shape[0]
sample_weight = np.abs(np.random.rand(n_samples))
elif option == 'class_weight':
class_weight = np.random.rand(2)
class_weight = {0: class_weight[0], 1: class_weight[1]}
elif option == 'balanced':
class_weight = 'balanced'
culog = cuLog(fit_intercept=False, class_weight=class_weight)
culog.fit(data, output, sample_weight=sample_weight)
sklog = skLog(fit_intercept=False, class_weight=class_weight)
sklog.fit(data, output, sample_weight=sample_weight)
skcoef = np.squeeze(sklog.coef_)
cucoef = np.squeeze(culog.coef_)
if regression_type == 'binary':
skcoef /= np.linalg.norm(skcoef)
cucoef /= np.linalg.norm(cucoef)
unit_tol = 0.04
total_tol = 0.08
elif regression_type.startswith('multiclass'):
skcoef = skcoef.T
skcoef /= np.linalg.norm(skcoef, axis=1)[:, None]
cucoef /= np.linalg.norm(cucoef, axis=1)[:, None]
unit_tol = 0.2
total_tol = 0.3
equality = array_equal(skcoef,
cucoef,
unit_tol=unit_tol,
total_tol=total_tol)
if not equality:
print('\ncoef.shape: ', coef.shape)
print('coef:\n', coef)
print('cucoef.shape: ', cucoef.shape)
print('cucoef:\n', cucoef)
assert equality
cuOut = culog.predict(data)
skOut = sklog.predict(data)
assert array_equal(skOut, cuOut, unit_tol=unit_tol, total_tol=total_tol)
def test_logistic_regression_unscaled(dtype, penalty, l1_ratio):
# Test logistic regression on the breast cancer dataset. We do not scale
# the dataset which could lead to numerical problems (fixed in PR #2543).
X, y = load_breast_cancer(return_X_y=True)
X = X.astype(dtype)
y = y.astype(dtype)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
params = {"penalty": penalty, "C": 1, "tol": 1e-4, "fit_intercept": True,
'max_iter': 5000, "l1_ratio": l1_ratio}
culog = cuLog(**params)
culog.fit(X_train, y_train)
score_test = 0.94
score_train = 0.94
assert culog.score(X_train, y_train) >= score_train
assert culog.score(X_test, y_test) >= score_test
def test_logistic_regression_sparse_only(dtype, nlp_20news):
# sklearn score with max_iter = 10000
sklearn_score = 0.878
acceptable_score = sklearn_score - 0.01
X, y = nlp_20news
X = csr_matrix(X.astype(dtype))
y = y.get().astype(dtype)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
culog = cuLog()
culog.fit(X_train, y_train)
score = culog.score(X_test, y_test)
assert score >= acceptable_score
def test_logistic_regression_predict_proba(
dtype, nrows, column_info, num_classes, fit_intercept, sparse_input
):
ncols, n_info = column_info
X_train, X_test, y_train, y_test = make_classification_dataset(
datatype=dtype, nrows=nrows, ncols=ncols,
n_info=n_info, num_classes=num_classes
)
X_train = csr_matrix(X_train) if sparse_input else X_train
X_test = csr_matrix(X_test) if sparse_input else X_test
y_train = y_train.astype(dtype)
y_test = y_test.astype(dtype)
culog = cuLog(fit_intercept=fit_intercept, output_type="numpy")
culog.fit(X_train, y_train)
if num_classes > 2:
sklog = skLog(
fit_intercept=fit_intercept,
solver="lbfgs",
multi_class="multinomial"
)
else:
sklog = skLog(fit_intercept=fit_intercept)
sklog.coef_ = culog.coef_.T
if fit_intercept:
sklog.intercept_ = culog.intercept_
else:
skLog.intercept_ = 0
sklog.classes_ = np.arange(num_classes)
cu_proba = culog.predict_proba(X_test)
sk_proba = sklog.predict_proba(X_test)
cu_log_proba = culog.predict_log_proba(X_test)
sk_log_proba = sklog.predict_log_proba(X_test)
assert array_equal(cu_proba, sk_proba)
assert array_equal(cu_log_proba, sk_log_proba)
def test_logistic_regression_model_digits(
dtype, order, sparse_input, fit_intercept, penalty):
# smallest sklearn score with max_iter = 10000
# put it as a constant here, because sklearn 0.23.1 needs a lot of iters
# to converge and has a bug returning an unrelated error if not converged.
acceptable_score = 0.95
digits = load_digits()
X_dense = digits.data.astype(dtype)
X_dense.reshape(X_dense.shape, order=order)
X = csr_matrix(X_dense) if sparse_input else X_dense
y = digits.target.astype(dtype)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
culog = cuLog(fit_intercept=fit_intercept, penalty=penalty)
culog.fit(X_train, y_train)
score = culog.score(X_test, y_test)
assert score >= acceptable_score
def test_logistic_regression(
num_classes, dtype, penalty, l1_ratio,
fit_intercept, nrows, column_info, C, tol
):
ncols, n_info = column_info
# Checking sklearn >= 0.21 for testing elasticnet
sk_check = LooseVersion(str(sklearn.__version__)) >= LooseVersion("0.21.0")
if not sk_check and penalty == "elasticnet":
pytest.skip(
"Need sklearn > 0.21 for testing logistic with" "elastic net."
)
X_train, X_test, y_train, y_test = make_classification_dataset(
datatype=dtype, nrows=nrows, ncols=ncols,
n_info=n_info, num_classes=num_classes
)
y_train = y_train.astype(dtype)
y_test = y_test.astype(dtype)
culog = cuLog(
penalty=penalty, l1_ratio=l1_ratio, C=C,
fit_intercept=fit_intercept, tol=tol
)
culog.fit(X_train, y_train)
# Only solver=saga supports elasticnet in scikit
if penalty in ["elasticnet", "l1"]:
if sk_check:
sklog = skLog(
penalty=penalty,
l1_ratio=l1_ratio,
solver="saga",
C=C,
fit_intercept=fit_intercept,
multi_class="auto",
)
else:
sklog = skLog(
penalty=penalty,
solver="saga",
C=C,
fit_intercept=fit_intercept,
multi_class="auto",
)
else:
sklog = skLog(
penalty=penalty,
solver="lbfgs",
C=C,
fit_intercept=fit_intercept,
multi_class="auto",
)
sklog.fit(X_train, y_train)
# Setting tolerance to lowest possible per loss to detect regressions
# as much as possible
cu_preds = culog.predict(X_test)
tol_test = 0.012
tol_train = 0.006
if num_classes == 10 and penalty in ["elasticnet", "l1"]:
tol_test *= 10
tol_train *= 10
assert culog.score(X_train, y_train) >= sklog.score(X_train, y_train) - \
tol_train
assert culog.score(X_test, y_test) >= sklog.score(X_test, y_test) - \
tol_test
assert len(np.unique(cu_preds)) == len(np.unique(y_test))
def test_logistic_regression(num_classes, dtype, penalty, l1_ratio,
fit_intercept):
# Checking sklearn >= 0.21 for testing elasticnet
sk_check = LooseVersion(str(sklearn.__version__)) >= LooseVersion("0.21.0")
if not sk_check and penalty == 'elasticnet':
pytest.skip("Need sklearn > 0.21 for testing logistic with"
"elastic net.")
nrows = 100000
train_rows = np.int32(nrows * 0.8)
X, y = make_classification(n_samples=nrows,
n_features=num_classes,
n_redundant=0,
n_informative=2)
X_test = np.asarray(X[train_rows:, 0:]).astype(dtype)
X_train = np.asarray(X[0:train_rows, :]).astype(dtype)
y_train = np.asarray(y[0:train_rows, ]).astype(dtype)
culog = cuLog(penalty=penalty,
l1_ratio=l1_ratio,
C=5.0,
fit_intercept=fit_intercept,
tol=1e-8)
culog.fit(X_train, y_train)
# Only solver=saga supports elasticnet in scikit
if penalty in ['elasticnet', 'l1']:
if sk_check:
sklog = skLog(penalty=penalty,
l1_ratio=l1_ratio,
solver='saga',
C=5.0,
fit_intercept=fit_intercept)
else:
sklog = skLog(penalty=penalty,
solver='saga',
C=5.0,
fit_intercept=fit_intercept)
elif penalty == 'l2':
sklog = skLog(penalty=penalty,
solver='lbfgs',
C=5.0,
fit_intercept=fit_intercept)
else:
if sk_check:
sklog = skLog(penalty=penalty,
solver='lbfgs',
C=5.0,
fit_intercept=fit_intercept)
else:
sklog = skLog(penalty='l2',
solver='lbfgs',
C=1e9,
fit_intercept=fit_intercept)
sklog.fit(X_train, y_train)
preds = culog.predict(X_test)
skpreds = sklog.predict(X_test)
# Setting tolerance to lowest possible per loss to detect regressions
# as much as possible
if penalty in ['elasticnet', 'l1', 'l2']:
assert np.sum(preds.to_array() != skpreds) / 20000 < 1e-1
else:
# This is the only case where cuml and sklearn actually do a similar
# lbfgs, other cases cuml does owl or sklearn does saga
assert np.sum(preds.to_array() != skpreds) / 20000 < 1e-3