Skip to content
/ lipo Public

Global, derivative-free optimization for hyperparameter tuning

License

Notifications You must be signed in to change notification settings

jdb78/lipo

Repository files navigation

LIPO is a package for derivative- and parameter-free global optimization, e.g. for hyperparameter tuning. Is based on the dlib package and provides wrappers around its optimization routine.

The algorithm outperforms random search - sometimes by margins as large as 10000x. It is often preferable to Bayesian optimization which requires "tuning of the tuner". Performance is on par with moderately to well tuned Bayesian optimization.

The provided implementation has the option to automatically enlarge the search space if bounds are found to be too restrictive (i.e. the optimum being to close to one of them).

See the LIPO algorithm implementation for details.

A great blog post by the author of dlib exists, describing how it works.

Installation

Execute

pip install lipo

Usage

from lipo import GlobalOptimizer

def function(x, y, z):
    zdict = {"a": 1, "b": 2}
    return -((x - 1.23) ** 6) + -((y - 0.3) ** 4) * zdict[z]

pre_eval_x = dict(x=2.3, y=13, z="b")
evaluations = [(pre_eval_x, function(**pre_eval_x))]

search = GlobalOptimizer(
    function,
    lower_bounds={"x": -10.0, "y": -10},
    upper_bounds={"x": 10.0, "y": -3},
    categories={"z": ["a", "b"]},
    evaluations=evaluations,
    maximize=True,
)

num_function_calls = 1000
search.run(num_function_calls)

The optimizer will automatically extend the search bounds if necessary.

Further, the package provides an implementation of the scikit-learn interface for hyperparamter search.

from lipo import LIPOSearchCV

search = LIPOSearchCV(
    estimator,
    param_space={"param_1": [0.1, 100], "param_2": ["category_1", "category_2"]},
    n_iter=100
)
search.fit(X, y)
print(search.best_params_)

Comparison to other frameworks

For benchmarks, see the notebook in the benchmark directory.

This is a Bayesian framework.

+ A well-chosen prior can lead to very good results slightly faster

- If the wrong prior is chosen, tuning can take long

- It is not parameter-free - one can get stuck in a local minimum which means tuning of the tuner can be required

- LIPO can converge faster when it is close to the minimum using a quadratic approximation

- The exploration of the search space is not systematic, i.e. results can vary a lot from run to run

+ It parallelizes very well

+ It can stop training early. This is very useful, e.g. for neural networks and can speed up tuning

+ A well-chosen prior can lead to very good results slightly faster

- If the wrong prior is chosen, tuning can take long

- It is not parameter-free, i.e. some tuning of the tuner can be required (the defaults are pretty good though)

- LIPO can converge faster when it is close to the minimum using a quadratic approximation

- The exploration of the search space is not systematic, i.e. results can vary a lot from run to run