H2O Cheat Sheet

1Install & init3.46

pip install h2o import h2o★
Requires a Java runtime for the backend. The pip package bundles the H2O jar.
h2o.init()★
Starts (or connects to) the H2O cluster. h2o.init(nthreads=-1, max_mem_size="8G") to size it; connect to a remote one with ip=/port=.
h2o.cluster().show_status() · h2o.no_progress()
Inspect the cluster; silence progress bars in scripts.

2H2OFramedata in the cluster

df = h2o.import_file("data.csv")★
Loads on the cluster (fast, out-of-core), not into Python memory. Also from S3/HDFS URLs.
df = h2o.H2OFrame(pandas_df) · df.as_data_frame() # back to pandas★
Convert to/from pandas. as_data_frame() pulls the (whole) frame into local memory — fine for results, not huge data.
df.shape · df.columns · df.head() · df.describe() · df.types★
pandas-like introspection. Indexing: df["col"], df[df["age"] > 30].

3Data prepbefore training

df["target"] = df["target"].asfactor()★
Critical: for classification, cast the target to a factor (categorical). A numeric target trains regression — a silent, common mistake.
train, valid, test = df.split_frame([0.7, 0.15], seed=1)★
Split ratios (the remainder is the last frame). Splits happen in the cluster.
y = "target"; x = [c for c in df.columns if c != y]★
H2O trains by column names: y = target name, x = list of predictor names.
df.impute(...) · df["c"].asnumeric() · df.na_omit()
In-cluster transforms. Note: tree models handle NAs natively — imputation often unnecessary.

4AutoMLthe easy button

from h2o.automl import H2OAutoML aml = H2OAutoML(max_models=20, max_runtime_secs=300, seed=1)★
Trains a diverse set (GBMs, RF, GLM, XGBoost, Deep Learning) plus stacked ensembles. Bound by model count or wall-clock.
aml.train(x=x, y=y, training_frame=train)★
One call does the whole search & ensembling. Add leaderboard_frame= for held-out ranking.
aml.leaderboard # ranked models · aml.leader★
The leaderboard ranks by a task-appropriate metric; aml.leader is the best model (often a stacked ensemble).
preds = aml.leader.predict(test)★
Predict returns an H2OFrame (predict column + per-class probabilities for classification).

5Estimatorsspecific models

from h2o.estimators import H2OGradientBoostingEstimator gbm = H2OGradientBoostingEstimator(ntrees=200, max_depth=5, learn_rate=0.1) gbm.train(x, y, training_frame=train, validation_frame=valid)★
Same train(x, y, training_frame) API for every estimator.
H2ORandomForestEstimator · H2OGeneralizedLinearEstimator · H2OXGBoostEstimator★
The workhorses (RF, GLM, XGBoost). Also H2ODeepLearningEstimator, H2OIsolationForestEstimator (anomaly), H2OKMeansEstimator.
H2OGradientBoostingEstimator(nfolds=5) # built-in cross-validation
nfolds does k-fold CV internally — no manual loop. H2OGridSearch for hyperparameter search.

6Performanceevaluate

perf = model.model_performance(test) perf.auc() · perf.rmse() · perf.confusion_matrix()★
Metrics on a test frame — AUC/logloss (classification), RMSE/MAE (regression), plus confusion matrix & gains/lift.
model.varimp(use_pandas=True) · model.varimp_plot()★
Feature importances (tree/GLM models).
model.cross_validation_metrics_summary()
CV metrics when trained with nfolds.

7Explainabilityunderstand models

aml.explain(test) · model.explain(test)★
One call produces a full report: variable importance, SHAP summary, partial dependence, model correlation — great in a notebook.
model.shap_summary_plot(test) · model.pd_plot(test, column="age")
SHAP & partial-dependence plots for individual models.
model.explain_row(test, row_index=0) # per-prediction explanation
Local explanation for a single prediction (SHAP contributions).

8Save & MOJOproduction

path = h2o.save_model(model, path="models/") model = h2o.load_model(path)★
Binary save/load — but tied to the H2O version. For long-term/production use MOJO instead.
model.download_mojo("model.zip")★
MOJO = a portable, version-stable scoring artifact that runs with a small Java runtime without an H2O cluster — the production deployment path.
h2o.cluster().shutdown()
Tear down the cluster when done (frees the JVM & memory).

!Common gotchasread before shipping

asfactor() for classification
The #1 gotcha: a numeric target trains a regression. Cast the target to a factor for classification.
data lives in the cluster
An H2OFrame is a handle to data in the JVM, not a local array. as_data_frame() materializes it locally — don't do that for huge frames.
train by column names
x/y are column-name strings/lists, not arrays. Exclude the target (and any id/leaky columns) from x.
MOJO for production, not save_model
Binary models break across H2O versions. Export a MOJO for stable, cluster-free scoring.
Java + memory
Needs a JVM; set max_mem_size for big data or the cluster OOMs. Remember to shutdown().

H2O cheat sheet