BOHB-HPO 0.3.1

Bayesian Optimization Hyperband Hyperparameter Optimization

Bayesian Optimization Hyperband Hyperparameter Optimization

Implementation for BOHB

Requirements

- numpy
- scipy
- statsmodels
- torch (example)

Installation

pip3 install bohb-hpo

Usage

from bohb import BOHB
import bohb.configspace as cs


def objective(step, alpha, beta):
    return 1 / (alpha * step + 0.1) + beta


def evaluate(params, n_iterations):
    loss = 0.0
    for i in range(int(n_iterations)):
        loss += objective(**params, step=i)
    return loss/n_iterations


alpha = cs.CategoricalHyperparameter('alpha', [0.001, 0.01, 0.1])
beta = cs.CategoricalHyperparameter('beta', [1, 2, 3])
configspace = cs.ConfigurationSpace([alpha, beta], seed=123)

opt = BOHB(configspace, evaluate, max_budget=10, min_budget=1)
logs = opt.optimize()

See examples

Configspace Examples

• Basic

import dehb.configspace as cs
lr = cs.UniformHyperparameter('lr', 1e-4, 1e-1, log=True)
batch_size = cs.CategoricalHyperparameter('batch_size', [8, 16, 32])
configspace = cs.ConfigurationSpace([lr, batch_size], seed=123)

• Conditional Parameters

import bohb.configspace as cs
a = cs.IntegerUniformHyperparameter('a', 0, 4)
b = cs.CategoricalHyperparameter('b', ['a', 'b', 'c'], a == 0)
b_default = cs.CategoricalHyperparameter('b', ['d'], ~b.cond)
configspace = cs.ConfigurationSpace([a, b, b_default], seed=123)

• Complex Conditional Parameters

import bohb.configspace as cs
a = cs.IntegerUniformHyperparameter('a', 0, 4)
b1 = cs.UniformHyperparameter('b', 0, 0.5, a <= 1)
b2 = cs.UniformHyperparameter('b', 0.5, 1, ~b1.cond)
c1 = cs.CategoricalHyperparameter('c', ['a', 'b', 'c'], b1 < 0.25)
c2 = cs.CategoricalHyperparameter('c', ['c', 'd', 'e'], ~c1.cond)
d1 = cs.UniformHyperparameter('d', 0, 1, (b1 < 0.125) & (c1 == 'b'))
d2 = cs.UniformHyperparameter('d', 0, 0, ~d1.cond)
configspace = cs.ConfigurationSpace([a, b1, b2, c1, c2, d1, d2], seed=123)

TODO

- Parallel Optimization (Implemented but not working properly)
- More Hyperparameters

License

bohb-hpo is licensed under the MIT License.