ppe-config based optuna

1_orig_optuna_example.py

import optuna

# https://optuna.org/


def objective(trial):
    x = trial.suggest_uniform('x', -10, 10)
    return (x - 2) ** 2


def main():
    study = optuna.create_study()
    study.optimize(objective, n_trials=100)

    print(study.best_params)


if __name__ == '__main__':
    main()

2_base_code.py

import yaml
import optuna
import pytorch_pfn_extras as ppe

class Net:
    def __init__(self, value):
        self.value = value

    def get_value(self):
        return (self.value - 2) ** 2


cfg_yml = """
net:
  type: Net
  value: 0.4
"""


types = {
    'Net': Net,
}

def run_train(net):
    # Run a training and return the final accuracy of this run
    return net.get_value()


def main():
    cfg = yaml.safe_load(cfg_yml)
    cfg = ppe.config.Config(cfg, types)

    net = cfg['/net']
    acc = run_train(net)
    print(acc)


if __name__ == "__main__":
    main()

3_ppe_config_based_example.py

import yaml
import optuna
import pytorch_pfn_extras as ppe
import pytorch_pfn_extras.config_types as config_types


class Net:
    def __init__(self, value):
        self.value = value

    def get_value(self):
        return (self.value - 2) ** 2


cfg_yml = """
net:
  type: Net
  value:
    type: optuna_suggest_float
    name: x
    low: -10.0
    high: 10.0
"""

types = {
    'Net': Net,
}

def run_train(net):
    # Run a training and return the final accuracy of this run
    acc = net.get_value()
    return acc


def main():
    # These params cannot be included in the config yaml
    n_trials = 100
    sampler_type = optuna.samplers.TPESampler
    sampler_kwargs = dict()

    def objective(trial):
        cfg = yaml.safe_load(cfg_yml)
        cfg = ppe.config.Config(cfg, {**types, **config_types.optuna_types(trial)})

        net = cfg['/net']
        return run_train(net)

    if True:
        # Use SQLite (it creates (or loads if already exists) my_study.db in the cwd)
        storage = 'sqlite:///my_study.db'
    else:
        # Use memory
        storage = None
    study = optuna.create_study(sampler=sampler_type(**sampler_kwargs), storage=storage)
    study.optimize(objective, n_trials=n_trials)
    print(study.best_params)


if __name__ == "__main__":
    main()