optml 0.2.3

A library for hyperparameter optimization of ML models

This package offers implementations of several black-box optimisation methods to tune hyperparameters of machine learning models. Its purpose is to enable data scientists to use optimization techniques for rapid protyping. Simply import OptML and supply it with a model and the parameters to optimize.

OptML offers a unified interface for models built with Scikit-Learn, Keras, XGBoost (and hopefully soon Statsmodels).

Prerequisites

This package requires scikit-learn with version 0.19.0 or higher. If Scikit-Learn is not yet installed run pip install scikit-learn==0.19.0. If you want to make use of the HyperoptOptimizer then you also need to install hyperopt (e.g. by pip install hyperopt).

In order to run with Keras and XGBoost models these libraries have to be install as well, of course.

Installation

If Scikit-Learn is version 0.19 or higher simply install optml using pip install optml and you’re ready to go.

Usage

Specify your ML model and the parameters you want to optimize over. For the parameters you have to choose the type (such as integer, categorical, boolean, etc.) and the range of values it can take.

model = SomeMLModel()
params = [Parameter(name='param1', param_type='continuous', lower=0.1, upper=5),
          Parameter(name='param2', param_type='integer', lower=1, upper=5),
          Parameter(name='param3', param_type='categorical', possible_values=['val1','val2','val3'])]

Then define the evaluation function. This can be anything from RMSE to crossentropy to custom functions. The first argument of the evaluation function is the array of true labels and the second argument is an array with model predictions.

def clf_score(y_true,y_pred):
    return np.sum(y_true==y_pred)/float(len(y_true))

Import and initialize an optimizer and optimize the model for some training data.

from optml.bayesian_optimizer import BayesianOptimizer
bayesOpt = BayesianOptimizer(model=model,
                             hyperparams=params,
                             eval_func=clf_score)
bayes_best_params, bayes_best_model = bayesOpt.fit(X_train=X_train, y_train=y_train, n_iters=50)

Features

At the moment this library includes: * Random Search * Parallelized Gridsearch * A simple Genetic Algorithm * Bayesian Optimisation (also supporting categorical parameters) * Hyperopt (using hyperopt)

How to Choose an Optimizer

OptML implements several optimization methods to address a range of requirements that can arise in data science problems. One of the main concerns is the effort required to evaluate a model for a set of parameters: If a model takes a long time to train we should choose an optimizer that maximises the potential improvement with every new set of parameters. In this case Bayesian Optimization and Hyperopt are more applicable. If a model is cheap to train then we can seek to parallelise the evaluations.

Also consider the number of parameters and their ranges. Clearly, it is more difficult to optimize over a large search space. It is advised to only include parameters in the optimization if they are expected to improve the final model.

Please also note that all of OptML’s optimizers require parameters to be bounded.

	number of evaluations	works with large search space	can use training in parallel	handles categorical parameters	stochastic optimisation
Gridsearch	high	no	yes	yes	no
Random Search	high	yes	yes	yes	yes
Genetic Algorithm	high	yes	not implemented	yes	yes
Bayesian Optimizer	low	yes	not implemented	yes	yes
Hyperopt	low	yes	yes	yes	yes

TODOs

1. algorithms:

• implement more options for genetic algorithms

• meta heuristics/swarm optimisation

2. functionality

• early stopping if there is no significant improvement after x iterations

3. usability

• better documenation

Author

• Johannes Petrat - Initial Release - johannespetrat

License

This project is licensed under the MIT License - see the LICENSE.md file for details