Polaris : A library for hyperparameter optimization

Polaris is a hyperparamter optimization library.
We plan to support every state-of-art tuning method as follows.

• Random Search
• Bayesian Optimization
• Tree-structured Parzen Estimators (TPE)

Documentation

Polaris' documentation can be found at https://polaris.readthedocs.io/

Prerequisites

• Python >= 3.6
• RabbitMQ (Only for parallel execution)

Installation

$ pip install polaris-py

Examples

Sequential Execution

from polaris import Polaris, STATUS_SUCCESS, Trials, Bounds


def pseudo_train(params, exp_info):
    lr_squared = (params['lr'] - 0.006) ** 2
    weight_decay_squared = (params['weight_decay'] - 0.02) ** 2
    loss = lr_squared + weight_decay_squared
    return {
        'loss':  loss,
        'status': STATUS_SUCCESS,
    }


if __name__ == '__main__':
    bounds = [
        Bounds('lr', 0.001, 0.01),
        Bounds('weight_decay', 0.0002, 0.04),
    ]
    trials = Trials()
    polaris = Polaris(
            pseudo_train, bounds, 'bayesian_opt',
            trials, max_evals=100, exp_key='this_is_test')
    best_params = polaris.run()
    print(best_params)

Parallel Execution

Single Process

1. Run rabbitmq-server
2. Set RABBITMQ_URL (ampq://guest:guest@localhost//)
3. Run polaris-worker --exp-key this_is_test
4. Run codes as follows

Multiple Processes (Use MPI)

1. Run rabbitmq-server
2. Set RABBITMQ_URL (ampq://guest:guest@localhost//)
3. Run mpirun -n 4 polaris-worker --mpi --exp-key this_is_test
4. Run codes as follows

...

if __name__ == '__main__':
    bounds = [
        Bounds('lr', 0.001, 0.01),
        Bounds('weight_decay', 0.0002, 0.04),
    ]
    trials = Trials()
    polaris = Polaris(
            pseudo_train, bounds, 'bayesian_opt',
            trials, max_evals=100, exp_key='this_is_test')
    best_params = polaris.run_parallel()
    print(best_params)