Checkpointing

In this tutorial, we will present how persistence can be achieved in your evolutions. The only required tools are a simple dict and a serialization method. Important data will be inserted in the dictionary and serialized to a file so that if something goes wrong, the evolution can be restored from the last saved checkpoint. It can also serve to continue an evolution beyond the pre-fixed termination criterion.

Checkpointing is not offered in standard algorithms such as eaSimple, eaMuPlus/CommaLambda and eaGenerateUpdate. You must create your own algorithm (or copy an existing one) and introduce this feature yourself.

Starting with a very basic example, we will cover the necessary stuff to checkpoint everything needed to restore an evolution. We skip the class definition and registration of tools in the toolbox to go directly to the algorithm and the main function. Our main function receives an optional string argument containing the path of the checkpoint file to restore.

import pickle

def main(checkpoint=None):
    if checkpoint:
        # A file name has been given, then load the data from the file
        cp = pickle.load(open(checkpoint, "r"))
        population = cp["population"]
        start_gen = cp["generation"]
        halloffame = cp["halloffame"]
        logbook = cp["logbook"]
        random.setstate(cp["rndstate"])
    else:
        # Start a new evolution
        population = toolbox.population(n=300)
        start_gen = 0
        halloffame = tools.HallOfFame(maxsize=1)
        logbook = tools.Logbook()

    stats = tools.Statistics(lambda ind: ind.fitness.values)
    stats.register("avg", numpy.mean)
    stats.register("max", numpy.max)

    for gen in range(start_gen, NGEN):
        population = algorithms.varAnd(population, toolbox, cxpb=CXPB, mutpb=MUTPB)

        # Evaluate the individuals with an invalid fitness
        invalid_ind = [ind for ind in population if not ind.fitness.valid]
        fitnesses = toolbox.map(toolbox.evaluate, invalid_ind)
        for ind, fit in zip(invalid_ind, fitnesses):
            ind.fitness.values = fit

        halloffame.update(population)
        record = stats.compile(population)
        logbook.record(gen=gen, evals=len(invalid_ind), **record)

        population = toolbox.select(population, k=len(population))

        if gen % FREQ == 0:
            # Fill the dictionary using the dict(key=value[, ...]) constructor
            cp = dict(population=population, generation=gen, halloffame=halloffame,
                      logbook=logbook, rndstate=random.getstate())
            pickle.dump(cp, open("checkpoint_name.pkl", "w"))

Now, the whole data will be written in a pickled dictionary every FREQ generations. Loading the checkpoint is done if the main function is given a path to a checkpoint file. In that case, the evolution continues from where it was in the last checkpoint. It will produce the exact same results as if it was not stopped and reloaded because we also restored the random module state. If you use numpy’s random numbers, don’t forget to save and reload their state too.