Other functions

The verbs that don't fit neatly into Initialize / Train / Optimize / Analyze / Deploy.

pull()#

Returns the most recent metrics DataFrame. Every metrics-emitting verb (create_model, tune_model, compare_models, ensemble_model, blend_models, stack_models, calibrate_model, finalize_model, predict_model) updates this slot before returning.

exp.create_model("rf")
metrics = exp.pull()  # DataFrame with Fold/Mean/Std rows

exp.compare_models()
leaderboard = exp.pull()  # the leaderboard

In 3.x, pull() was the only way to get the metrics. In 4.0, every result object already carries them — so you rarely need pull() unless you want the most recent one without hanging onto the result.

models(internal=False)#

Returns a DataFrame describing every model in the registry.

exp.models()
#               Name             Reference            Turbo
# ID
# lr            Logistic Regr.   sklearn.linear...    True
# knn           K Neighbors      sklearn.neighbo...   True
# nb            Naive Bayes      sklearn.naive_b...   True
# ...

Pass internal=True for the full container shape (Class, Equality, Args, …) — useful when you're inspecting the registry programmatically.

For time-series, the list is filtered to entries whose model_type is in TSModelTypes — so ensemble_forecaster (which can't be constructed in isolation) is hidden.

get_metrics()#

Returns a DataFrame describing every metric in the registry.

exp.get_metrics()
#         Name                          Display Name   Score Function   ...
# ID
# acc     Accuracy                      Accuracy       sklearn.metr...  ...
# auc     AUC                           AUC            sklearn.metr...  ...
# recall  Recall                        Recall         sklearn.metr...  ...
# ...

Custom metrics added via add_metric show up here too.

add_metric(id, name, score_func, ...)#

Registers a custom metric. The metric runs in CV (across every create_model call) and the leaderboard.

from sklearn.metrics import balanced_accuracy_score

container = exp.add_metric(
    "balanced",
    "Balanced Accuracy",
    balanced_accuracy_score,
    greater_is_better=True,
)

# The next compare_models / create_model includes "Balanced Accuracy"
# in the leaderboard / fold metrics.
exp.compare_models()

For metrics that need extra args (e.g. sample_weight, pos_label), pass them via args={...} and they're forwarded to every call.

For probabilistic metrics, set target="pred_proba" so the metric receives predict_proba output instead of predict.

Currently supported on classification + regression. Clustering / anomaly / time-series will gain it in a future session — for now, pre-compute and add by hand if needed.

remove_metric(name_or_id)#

Drops a metric from the registry. Subsequent verbs won't compute it.

exp.remove_metric("kappa")
exp.compare_models()  # leaderboard no longer has Kappa

get_config(key) / set_config(key, value)#

Read or mutate a slot on exp._fit_state. Tightly scoped: only a small allowlist is settable (session_id, n_jobs, verbose, fold, log_experiment).

exp.get_config("X_train")    # the train DataFrame
exp.get_config("session_id") # the random seed
exp.set_config("n_jobs", 4)  # adjust parallelism for subsequent verbs

The 3.x set_config accepted any internal slot. In 4.0 the surface is intentionally narrow — most "config" lives on the constructor where it's checked at fit time.

What's removed#