if __name__ == '__main__':
log_dir, run_args = read_arguments_run()
log_id = 'run_' + str(run_args['seed'])
agent_builder = BenchmarkLogger._load_pickle(
os.path.join(log_dir, 'agent_builder.pkl'))
env_builder = BenchmarkLogger._load_pickle(
os.path.join(log_dir, 'environment_builder.pkl'))
logger = BenchmarkLogger(log_dir=log_dir,
log_id=log_id,
use_timestamp=False)
logger.info('Starting experiment.')
result = exec_run(agent_builder, env_builder, **run_args)
logger.info('Saving result.')
cmp_E = agent_builder.compute_policy_entropy
logger.save_Js([result['Js']])
logger.save_Rs([result['Rs']])
logger.save_Qs([result['Qs']])
if cmp_E:
logger.save_policy_entropies([result['Es']])
new_score = result['score']
new_agent = result['builders']
stats = dict(best_J=new_score[0], best_R=new_score[1], best_Q=new_score[2])
if __name__ == '__main__':
logger = BenchmarkLogger(log_dir='./logs', log_id='a2c_pendulum')
agent_builder = A2CBuilder.default(actor_lr=7e-4,
critic_lr=7e-4,
n_features=32)
env_name = 'Gym'
env_params = dict(env_id='Pendulum-v0', horizon=200, gamma=.99)
parallel = dict(max_concurrent_runs=10)
env_builder = EnvironmentBuilder(env_name, env_params)
logger.info('Environment is imported')
exp = BenchmarkExperiment(agent_builder, env_builder, logger)
logger.info('BenchmarkExperiment was built successfully')
start_time = time.time()
exp.run(exec_type='parallel',
n_runs=10,
n_epochs=100,
n_steps=30000,
n_episodes_test=5,
parallel=parallel)
end_time = time.time()
logger.info('Execution time: {} SEC'.format(end_time - start_time))
exp.save_plot()