Development version

You are reading the latest (development) docs. For stable documentation, see stable.

Resume a Stopped Search from a Checkpoint

Time: 8 min | Difficulty: Intermediate

What This Solves

Long searches can be interrupted (OOM, timeout, instance preemption). Checkpointing saves the search state after each generation so you can resume without re-evaluating already-visited candidates.

Recipe: Save a Checkpoint

from sklearn.datasets import load_breast_cancer
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import StratifiedKFold, train_test_split

from sklearn_genetic import GASearchCV, EvolutionConfig, RuntimeConfig
from sklearn_genetic.callbacks import ConsecutiveStopping
from sklearn_genetic.space import Continuous, Integer

X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42
)

param_grid = {
    "n_estimators": Integer(50, 300),
    "max_depth":    Integer(3, 20),
    "max_features": Continuous(0.1, 1.0),
}

cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)

ga = GASearchCV(
    estimator=RandomForestClassifier(random_state=42, n_jobs=-1),
    param_grid=param_grid,
    scoring="roc_auc",
    cv=cv,
    evolution_config=EvolutionConfig(population_size=20, generations=30, elitism=True),
    runtime_config=RuntimeConfig(n_jobs=-1, verbose=True),
    random_state=42,
)

# checkpoint_path: file to save state after each generation
ga.fit(
    X_train, y_train,
    checkpoint_path="./ga_search_checkpoint.pkl",
    callbacks=[ConsecutiveStopping(generations=5, metric="fitness_best")],
)

print(f"Best CV ROC AUC: {ga.best_score_:.4f}")

Recipe: Resume from a Checkpoint

import joblib
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split

X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42
)

# Load the saved search
ga = joblib.load("./ga_search_checkpoint.pkl")

print(f"Resumed from generation {ga.history[-1]['gen']}")
print(f"Best CV ROC AUC so far: {ga.best_score_:.4f}")

# Continue from where it stopped — provide the same data
ga.fit(
    X_train, y_train,
    checkpoint_path="./ga_search_checkpoint.pkl",
    # Increase generations or remove early stopping to run more
)

print(f"Final best CV ROC AUC: {ga.best_score_:.4f}")

Key Points

• checkpoint_path: Saves the full GASearchCV object (including population, history, best params) after each generation using joblib.dump.
• Same data required: Resume with the same X_train, y_train and cv splits. Different data produces incorrect results.
• random_state preserved: The checkpoint restores the PRNG state, so resumed generations are deterministic.
• Cache hits: Evaluated candidates from the checkpoint are cached — they won't be re-evaluated on resume.
• File size: Checkpoints include the full population. For large estimators, this can be 10–100 MB.

See Also

• Reproducibility and Checkpointing — full guide
• Checkpointing & Resume Example — annotated example
• Stop After a Time Budget — stop cleanly at a time limit, then resume