def train(num_timesteps, seed): rank = MPI.COMM_WORLD.Get_rank() sess = U.single_threaded_session() sess.__enter__() if rank == 0: logger.configure() else: logger.configure(format_strs=[]) workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank() set_global_seeds(workerseed) config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)), '..', 'configs', 'ant_gibson_flagrun.yaml') print(config_file) env = AntGibsonFlagRunEnv(is_discrete=False, config = config_file) def mlp_policy_fn(name, sensor_space, ac_space): return mlp_policy.MlpPolicy(name=name, ob_space=sensor_space, ac_space=ac_space, hid_size=64, num_hid_layers=2) env.seed(workerseed) gym.logger.setLevel(logging.WARN) pposgd_sensor.learn(env, mlp_policy_fn, max_timesteps=int(num_timesteps * 1.1 * 5), timesteps_per_actorbatch=6000, clip_param=0.2, entcoeff=0.00, optim_epochs=4, optim_stepsize=1e-4, optim_batchsize=64, gamma=0.99, lam=0.95, schedule='linear', save_per_acts=500 ) env.close()
def train(num_timesteps, seed): rank = MPI.COMM_WORLD.Get_rank() sess = utils.make_gpu_session(args.num_gpu) sess.__enter__() if rank == 0: logger.configure() else: logger.configure(format_strs=[]) workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank() set_global_seeds(workerseed) config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)), '..', 'configs', 'ant_climb.yaml') print(config_file) env = AntClimbEnv(config=config_file) env = Monitor(env, logger.get_dir() and osp.join(logger.get_dir(), str(rank))) env.seed(workerseed) gym.logger.setLevel(logging.WARN) def mlp_policy_fn(name, sensor_space, ac_space): return mlp_policy.MlpPolicy(name=name, ob_space=sensor_space, ac_space=ac_space, hid_size=64, num_hid_layers=2) def fuse_policy_fn(name, ob_space, sensor_space, ac_space): return fuse_policy.FusePolicy(name=name, ob_space=ob_space, sensor_space=sensor_space, hid_size=64, num_hid_layers=2, ac_space=ac_space, save_per_acts=10000, session=sess) if args.mode == "SENSOR": pposgd_sensor.learn(env, mlp_policy_fn, max_timesteps=int(num_timesteps * 1.1 * 5), timesteps_per_actorbatch=6000, clip_param=0.2, entcoeff=0.00, optim_epochs=4, optim_stepsize=1e-3, optim_batchsize=64, gamma=0.99, lam=0.95, schedule='linear', save_per_acts=100, save_name="ant_ppo_mlp" ) env.close() else: pposgd_fuse.learn(env, fuse_policy_fn, max_timesteps=int(num_timesteps * 1.1), timesteps_per_actorbatch=2000, clip_param=0.2, entcoeff=0.01, optim_epochs=4, optim_stepsize=LEARNING_RATE, optim_batchsize=64, gamma=0.99, lam=0.95, schedule='linear', save_per_acts=50, save_name="ant_ppo_fuse", reload_name=args.reload_name ) env.close()