def experiment(work_dir,
problem_cls=None,
solver_cls=None,
num_rounds=None,
batch_size=None,
seed=None):
"""Run experiment."""
# Must be specified via Gin.
assert problem_cls
assert solver_cls
assert num_rounds
assert batch_size
tf.gfile.MakeDirs(work_dir)
if seed is not None:
utils.set_seed(seed)
print('Running experiment with %s on %s' % (problem_cls, solver_cls))
problem = problem_cls()
solver = solver_cls(problem.domain)
population = controller.run(problem,
solver,
num_rounds=num_rounds,
batch_size=batch_size)
print('Writing output to %s/population.csv' % work_dir)
population.to_csv(os.path.join(work_dir, 'population.csv'))
return population
def __init__(self,
problem,
solver,
initial_population=None,
callbacks=None,
output_dir=None,
verbose=False,
seed=None):
"""Creates an instance of this class.
Args:
problem: An instance of a `BaseProblem`.
solver: An instance of a `BaseSolver`.
initial_population: An instance of a `Population` that is used to
initialize the optimization.
callbacks: A list of callables that take a population as input argument
and that are called at the end of each round.
output_dir: An output directory for storing data.
verbose: Whether to log messages to the console.
seed: The global seed of random number generators. If `None`, no seed is
set explicitly.
"""
if seed is not None:
utils.set_seed(seed)
self._callbacks = utils.to_list(callbacks)
self._output_dir = output_dir
self._problem = problem
self._solver = solver
self._population = (
data.Population()
if initial_population is None else initial_population.copy())
self._verbose = verbose
self._log('Initial population size: %d', len(self._population))
if self._population:
for callback in self._callbacks:
callback(initial_population)