User Priors over the Optimum

Example for optimizing a Multi-Layer Perceptron (MLP) setting priors over the optimum on the hyperparameters. These priors are derived from user knowledge (from previous runs on similar tasks, common knowledge or intuition gained from manual tuning). To create the priors, we make use of the Normal and Beta Hyperparameters, as well as the “weights” property of the CategoricalHyperparameter. This can be integrated into the optimiztion for any SMAC facade, but we stick with the hyperparameter optimization facade here. To incorporate user priors into the optimization, you have to change the acquisition function to PriorAcquisitionFunction.

[INFO][abstract_initial_design.py:147] Using 10 initial design configurations and 1 additional configurations.
[WARNING][prior_acqusition_function.py:107] Discretizing the prior for random forest models.
[INFO][abstract_intensifier.py:516] Added config 145e6c as new incumbent because there are no incumbents yet.
[INFO][abstract_intensifier.py:595] Added config 2aab89 and rejected config 145e6c as incumbent because it is not better than the incumbents on 1 instances:
[INFO][abstract_intensifier.py:595] Added config 1dd478 and rejected config 2aab89 as incumbent because it is not better than the incumbents on 1 instances:
[INFO][abstract_intensifier.py:595] Added config 8111d4 and rejected config 1dd478 as incumbent because it is not better than the incumbents on 1 instances:
[INFO][abstract_intensifier.py:595] Added config 4fcbbb and rejected config 8111d4 as incumbent because it is not better than the incumbents on 1 instances:
[INFO][abstract_intensifier.py:595] Added config a18544 and rejected config 4fcbbb as incumbent because it is not better than the incumbents on 1 instances:
[INFO][abstract_intensifier.py:595] Added config ece60c and rejected config a18544 as incumbent because it is not better than the incumbents on 1 instances:
[INFO][abstract_intensifier.py:595] Added config fee5c4 and rejected config ece60c as incumbent because it is not better than the incumbents on 1 instances:
[INFO][abstract_intensifier.py:595] Added config 649016 and rejected config fee5c4 as incumbent because it is not better than the incumbents on 1 instances:
[INFO][smbo.py:328] Configuration budget is exhausted:
[INFO][smbo.py:329] --- Remaining wallclock time: inf
[INFO][smbo.py:330] --- Remaining cpu time: inf
[INFO][smbo.py:331] --- Remaining trials: 0
Default cost: 0.038952336737852145
Default cost: 0.02838904363973993

import warnings

import numpy as np
from ConfigSpace import (
    BetaIntegerHyperparameter,
    CategoricalHyperparameter,
    Configuration,
    ConfigurationSpace,
    NormalFloatHyperparameter,
    UniformIntegerHyperparameter,
)
from sklearn.datasets import load_digits
from sklearn.exceptions import ConvergenceWarning
from sklearn.model_selection import StratifiedKFold, cross_val_score
from sklearn.neural_network import MLPClassifier

from smac import HyperparameterOptimizationFacade, Scenario
from smac.acquisition.function import PriorAcquisitionFunction

__copyright__ = "Copyright 2021, AutoML.org Freiburg-Hannover"
__license__ = "3-clause BSD"


digits = load_digits()


class MLP:
    @property
    def configspace(self) -> ConfigurationSpace:
        # Build Configuration Space which defines all parameters and their ranges.
        # To illustrate different parameter types,
        # we use continuous, integer and categorical parameters.
        cs = ConfigurationSpace()

        # We do not have an educated belief on the number of layers beforehand
        # As such, the prior on the HP is uniform
        n_layer = UniformIntegerHyperparameter(
            "n_layer",
            lower=1,
            upper=5,
        )

        # We believe the optimal network is likely going to be relatively wide,
        # And place a Beta Prior skewed towards wider networks in log space
        n_neurons = BetaIntegerHyperparameter(
            "n_neurons",
            lower=8,
            upper=256,
            alpha=4,
            beta=2,
            log=True,
        )

        # We believe that ReLU is likely going to be the optimal activation function about
        # 60% of the time, and thus place weight on that accordingly
        activation = CategoricalHyperparameter(
            "activation",
            ["logistic", "tanh", "relu"],
            weights=[1, 1, 3],
            default_value="relu",
        )

        # Moreover, we believe ADAM is the most likely optimizer
        optimizer = CategoricalHyperparameter(
            "optimizer",
            ["sgd", "adam"],
            weights=[1, 2],
            default_value="adam",
        )

        # We do not have an educated opinion on the batch size, and thus leave it as-is
        batch_size = UniformIntegerHyperparameter(
            "batch_size",
            16,
            512,
            default_value=128,
        )

        # We place a log-normal prior on the learning rate, so that it is centered on 10^-3,
        # with one unit of standard deviation per multiple of 10 (in log space)
        learning_rate_init = NormalFloatHyperparameter(
            "learning_rate_init",
            lower=1e-5,
            upper=1.0,
            mu=1e-3,  # will be transformed to log space later
            sigma=10,  # will be transformed to log space later
            log=True,
        )

        # Add all hyperparameters at once:
        cs.add([n_layer, n_neurons, activation, optimizer, batch_size, learning_rate_init])

        return cs

    def train(self, config: Configuration, seed: int = 0) -> float:
        with warnings.catch_warnings():
            warnings.filterwarnings("ignore", category=ConvergenceWarning)

            classifier = MLPClassifier(
                hidden_layer_sizes=[config["n_neurons"]] * config["n_layer"],
                solver=config["optimizer"],
                batch_size=config["batch_size"],
                activation=config["activation"],
                learning_rate_init=config["learning_rate_init"],
                random_state=seed,
                max_iter=5,
            )

            # Returns the 5-fold cross validation accuracy
            cv = StratifiedKFold(n_splits=5, random_state=seed, shuffle=True)  # to make CV splits consistent
            score = cross_val_score(classifier, digits.data, digits.target, cv=cv, error_score="raise")

        return 1 - np.mean(score)


if __name__ == "__main__":
    mlp = MLP()
    default_config = mlp.configspace.get_default_configuration()

    # Define our environment variables
    scenario = Scenario(mlp.configspace, n_trials=40)

    # We also want to include our default configuration in the initial design
    initial_design = HyperparameterOptimizationFacade.get_initial_design(
        scenario,
        additional_configs=[default_config],
    )

    # We define the prior acquisition function, which conduct the optimization using priors over the optimum
    acquisition_function = PriorAcquisitionFunction(
        acquisition_function=HyperparameterOptimizationFacade.get_acquisition_function(scenario),
        decay_beta=scenario.n_trials / 10,  # Proven solid value
    )

    # We only want one config call (use only one seed in this example)
    intensifier = HyperparameterOptimizationFacade.get_intensifier(
        scenario,
        max_config_calls=1,
    )

    # Create our SMAC object and pass the scenario and the train method
    smac = HyperparameterOptimizationFacade(
        scenario,
        mlp.train,
        initial_design=initial_design,
        acquisition_function=acquisition_function,
        intensifier=intensifier,
        overwrite=True,
    )

    incumbent = smac.optimize()

    # Get cost of default configuration
    default_cost = smac.validate(default_config)
    print(f"Default cost: {default_cost}")

    # Let's calculate the cost of the incumbent
    incumbent_cost = smac.validate(incumbent)
    print(f"Default cost: {incumbent_cost}")

previous

Continue an Optimization

next

Multi-Fidelity and Multi-Instances