Custom Callbacks

You can extend the built-in callbacks or write your own from scratch by subclassing BaseCallback.

Prerequisites

• Read Callbacks first

The BaseCallback Interface

from sklearn_genetic.callbacks.base import BaseCallback

class MyCallback(BaseCallback):
    def on_start(self, estimator=None):
        """Called once before generation 0. Return None."""
        pass

    def on_step(self, record=None, logbook=None, estimator=None):
        """Called after each generation.
        Return True to STOP the search, False to continue."""
        return False

    def on_end(self, logbook=None, estimator=None):
        """Called after the final generation. Return None."""
        pass

Key rule: on_step returning True stops the search. False (or None) continues.

Parameters available in on_step

Parameter	Type	Description
record	dict	Current generation stats: gen, nevals, fitness, fitness_std, fitness_best, fitness_max, fitness_min, genotype_diversity, stagnation_generations, etc.
logbook	DEAP Logbook	All previous generation records. Use logbook.select("fitness") to get a list of values
estimator	GASearchCV	The search estimator with all its attributes

Example: Stop After N Generations Below a Threshold

This callback stops the search if more than N fitness values in the history fall below a threshold:

from sklearn_genetic.callbacks.base import BaseCallback


class DummyThreshold(BaseCallback):
    def __init__(self, threshold, N, metric="fitness"):
        self.threshold = threshold
        self.N = N
        self.metric = metric

    def on_start(self, estimator=None):
        print("Search starting!")

    def on_step(self, record=None, logbook=None, estimator=None):
        if len(logbook) <= self.N:
            return False  # not enough history yet

        # Get the last N+1 metric values
        stats = logbook.select(self.metric)[-(self.N + 1):]
        n_below = sum(1 for x in stats if x < self.threshold)

        if n_below > self.N:
            return True  # stop

        return False

    def on_end(self, logbook=None, estimator=None):
        print(f"Search ended after {len(logbook)} generations.")

Use it like any other callback:

callback = DummyThreshold(threshold=0.85, N=4, metric="fitness")
evolved_estimator.fit(X_train, y_train, callbacks=callback)

Example: Log to a File

import json
from sklearn_genetic.callbacks.base import BaseCallback


class JsonLogger(BaseCallback):
    def __init__(self, path):
        self.path = path
        self._file = None

    def on_start(self, estimator=None):
        self._file = open(self.path, "w")

    def on_step(self, record=None, logbook=None, estimator=None):
        if record is not None:
            # record contains only JSON-serializable types
            self._file.write(json.dumps({
                "gen": record.get("gen"),
                "fitness_best": record.get("fitness_best"),
                "genotype_diversity": record.get("genotype_diversity"),
            }) + "\n")
            self._file.flush()
        return False

    def on_end(self, logbook=None, estimator=None):
        if self._file:
            self._file.close()

Example: Alert When Score Plateaus

from sklearn_genetic.callbacks.base import BaseCallback


class PlateauAlert(BaseCallback):
    """Send an alert when the best score hasn't improved for N generations."""

    def __init__(self, patience=10):
        self.patience = patience

    def on_step(self, record=None, logbook=None, estimator=None):
        stagnation = record.get("stagnation_generations", 0) if record else 0
        if stagnation >= self.patience:
            print(
                f"[PlateauAlert] Best score has not improved for "
                f"{stagnation} generations. Consider stopping early."
            )
        return False

Tips & Gotchas

• Always return False from on_step unless you explicitly want to stop the search.
• Call print(record.keys()) inside on_step during development to see all available fields.
• The logbook is a DEAP Logbook object. Use logbook.select("metric_name") to get a Python list of values across all generations.
• You can combine custom callbacks with built-in ones in the same list.
• Heavy computation in on_step (e.g., writing to a remote server) can slow down the search. For non-critical logging, consider buffering and flushing only in on_end.

Next Steps

• Callbacks — all built-in callbacks and how to combine them
• MLflow Integration — log experiments to MLflow