def infer_distribution_test():
from h2o.utils import CustomDistributionGeneric, CustomDistributionGaussian
class CustomDistributionGaussian2(CustomDistributionGeneric):
def link(self):
return "identity"
def init(self, w, o, y):
return [w * (y - o), w]
def gradient(self, y, f):
return y - f
def gamma(self, w, y, z, f):
return [w * z, w]
custom_dist1 = h2o.upload_custom_distribution(CustomDistributionGaussian)
custom_dist2 = h2o.upload_custom_distribution(CustomDistributionGaussian2)
for dist in [
"poisson", "laplace", "tweedie", "gaussian", "huber", "gamma",
"quantile", "bernoulli", "quasibinomial", "multinomial"
]:
infer_distribution_helper(dist, dist)
# custom distribution
infer_distribution_helper("custom", "custom",
dict(custom_distribution_func=custom_dist1),
dict(custom_distribution_func=custom_dist1))
# revert to default
infer_distribution_helper("tweedie", "gaussian", dict(tweedie_power=1.2))
infer_distribution_helper("huber", "gaussian", dict(huber_alpha=0.2))
infer_distribution_helper("quantile", "gaussian", dict(quantile_alpha=0.2))
infer_distribution_helper("custom", "gaussian",
dict(custom_distribution_func=custom_dist1),
dict(custom_distribution_func=custom_dist2))
# unaffected by param for different distribution
infer_distribution_helper("quantile", "quantile", dict(tweedie_power=1.2))
infer_distribution_helper("tweedie", "tweedie", dict(huber_alpha=0.2))
infer_distribution_helper("huber", "huber", dict(quantile_alpha=0.2))
infer_distribution_helper(
"custom", "custom", dict(custom_distribution_func=custom_dist1),
dict(custom_distribution_func=custom_dist1, tweedie_power=1.2))
def test_scenario():
expected_dist = distribution_name
df = make_data(scenario.get("nrows", 264))
# Hack so we don't remove the custom distribution function
if expected_dist == "custom":
from h2o.utils.distributions import CustomDistributionGaussian
custom_dist = h2o.upload_custom_distribution(
CustomDistributionGaussian)
scenario["distribution"][
"custom_distribution_func"] = custom_dist
aml = H2OAutoML(max_models=scenario.get("max_models", 12),
distribution=scenario["distribution"],
seed=seed,
max_runtime_secs_per_model=1)
try:
aml.train(y=scenario["response"], training_frame=df)
except Exception:
assert scenario.get(
'fail',
False), "This distribution should not have failed."
return
assert not scenario.get(
'fail', False), "This distribution should have failed."
if aml.leaderboard.nrow == 0:
algos = []
else:
algos = list(
set(
get_leaderboard(
aml, "algo").as_data_frame()["algo"].unique()))
for expected in [
'DeepLearning', "DRF", 'GBM', 'GLM', 'StackedEnsemble',
'XGBoost'
]:
assert expected in algos, "Expected {expected} but no found.".format(
expected=expected)
for model_id in aml.leaderboard.as_data_frame()["model_id"]:
distribution = get_distribution(model_id)
assert distribution == expected_dist or \
h2o.get_model(model_id).algo not in [a.lower() for a in scenario["algos"]], (
"{model}: Expected distribution {s_dist} but {distribution} found!".format(
model=model_id,
s_dist=expected_dist,
distribution=distribution
))
def test_custom_distribution_reuse():
from h2o.utils.distributions import CustomDistributionGaussian
train = h2o.import_file(
pyunit_utils.locate("smalldata/iris/iris_train.csv"))
y = "petal_wid"
x = train.columns
x.remove(y)
nfolds = 2
for i in range(3):
test_wrong_multinomial()
custom_dist1 = h2o.upload_custom_distribution(
CustomDistributionGaussian)
gbm = H2OGradientBoostingEstimator(
nfolds=nfolds,
fold_assignment="Modulo",
keep_cross_validation_predictions=True,
distribution="custom",
custom_distribution_func=custom_dist1)
gbm.train(x=x, y=y, training_frame=train)
def custom_distribution_bernoulli():
return h2o.upload_custom_distribution(CustomDistributionBernoulli,
func_name="custom_bernoulli",
func_file="custom_bernoulli.py")
def upload_distribution(distribution, name):
return h2o.upload_custom_distribution(distribution,
func_name="custom_" + name,
func_file="custom_" + name + ".py")
