def test_MaxentModel_feature_types():
all = models.MaxentModel(feature_types="lqpht", beta_multiplier=1.5)
auto = models.MaxentModel(feature_types=["auto"])
linear = models.MaxentModel(feature_types=["linear"])
qp = models.MaxentModel(feature_types=["quadratic", "product"],
convergence_tolerance=2e-7,
beta_lqp=1.5)
ht = models.MaxentModel(
feature_types=["hinge", "threshold"],
beta_threshold=1.5,
beta_hinge=0.75,
n_threshold_features=10,
n_hinge_features=10,
)
all.fit(x, y)
auto.fit(x, y)
linear.fit(x, y)
qp.fit(x, y)
ht.fit(x, y)
for model in [all, linear, qp, ht, auto]:
ypred = model.predict(x)
auc_score = metrics.roc_auc_score(y, ypred)
assert 0.5 <= auc_score <= 1.0
def test_tau_scaler():
model = models.MaxentModel(tau=0.5)
model.fit(x, y)
ypred = model.predict(x, transform="logistic")
fit_mean_pt5 = ypred[y == 1].mean()
assert 0.48 < ypred[y == 1].mean() < 0.75
model = models.MaxentModel(tau=0.25)
model.fit(x, y)
ypred = model.predict(x, transform="logistic")
fit_mean_pt25 = ypred[y == 1].mean()
assert fit_mean_pt25 < fit_mean_pt5
def test_MaxentModel_flow():
model = models.MaxentModel()
model.fit(x, y)
ypred = model.predict(x)
assert len(ypred) == len(y)
assert ypred.max() <= 1.0
assert ypred.min() >= 0.0
def test_MaxentModel_best_lambdas():
model = models.MaxentModel(use_lambdas="best")
model.fit(x, y)
ypred = model.predict(x, transform="logistic")
auc_score = metrics.roc_auc_score(y, ypred)
assert 0.5 <= auc_score <= 1.0
assert 0.48 < ypred[y == 1].mean() < 0.75
def test_MaxentModel_performance():
model = models.MaxentModel(use_lambdas="last", scorer="roc_auc")
model.fit(x, y)
ypred = model.predict(x)
auc_score = metrics.roc_auc_score(y, ypred)
# check that model is not nonsense
assert 0.5 <= auc_score <= 1.0
# check that full model close to cross-val performance (+- stderr)
assert abs(auc_score - model.estimator.cv_mean_score_[-1]
) < 0.1 + model.estimator.cv_standard_error_[-1]