def make_robotics_env(env_id, seed, rank=0, allow_early_resets=True):
"""
Create a wrapped, monitored gym.Env for MuJoCo.
:param env_id: (str) the environment ID
:param seed: (int) the initial seed for RNG
:param rank: (int) the rank of the environment (for logging)
:param allow_early_resets: (bool) allows early reset of the environment
:return: (Gym Environment) The robotic environment
"""
set_global_seeds(seed)
env = gym.make(env_id)
keys = ['observation', 'desired_goal']
# TODO: remove try-except once most users are running modern Gym
try: # for modern Gym (>=0.15.4)
from gym.wrappers import FilterObservation, FlattenObservation
env = FlattenObservation(FilterObservation(env, keys))
except ImportError: # for older gym (<=0.15.3)
from gym.wrappers import FlattenDictWrapper # pytype:disable=import-error
env = FlattenDictWrapper(env, keys)
env = Monitor(env,
logger.get_dir()
and os.path.join(logger.get_dir(), str(rank)),
info_keywords=('is_success', ),
allow_early_resets=allow_early_resets)
env.seed(seed)
return env
def _thunk():
env = make_atari(env_id)
env.seed(seed + rank)
env = Monitor(env,
logger.get_dir()
and os.path.join(logger.get_dir(), str(rank)),
allow_early_resets=allow_early_resets)
return wrap_deepmind(env, **wrapper_kwargs)
def train(env_id, num_timesteps, seed):
"""
Train TRPO model for the atari environment, for testing purposes
:param env_id: (str) Environment ID
:param num_timesteps: (int) The total number of samples
:param seed: (int) The initial seed for training
"""
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
env = make_atari(env_id)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
env = wrap_deepmind(env)
env.seed(workerseed)
model = TRPO(CnnPolicy,
env,
timesteps_per_batch=512,
max_kl=0.001,
cg_iters=10,
cg_damping=1e-3,
entcoeff=0.0,
gamma=0.98,
lam=1,
vf_iters=3,
vf_stepsize=1e-4)
model.learn(total_timesteps=int(num_timesteps * 1.1))
env.close()
# Free memory
del env
def make_mujoco_env(env_id, seed, allow_early_resets=True):
"""
Create a wrapped, monitored gym.Env for MuJoCo.
:param env_id: (str) the environment ID
:param seed: (int) the initial seed for RNG
:param allow_early_resets: (bool) allows early reset of the environment
:return: (Gym Environment) The mujoco environment
"""
set_global_seeds(seed + 10000 * mpi_rank_or_zero())
env = gym.make(env_id)
env = Monitor(env,
os.path.join(logger.get_dir(), '0'),
allow_early_resets=allow_early_resets)
env.seed(seed)
return env
def main():
"""
Run the atari test
"""
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
policy = partial(CnnPolicy, dueling=args.dueling == 1)
model = DQN(
env=env,
policy=policy,
learning_rate=1e-4,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
)
model.learn(total_timesteps=args.num_timesteps)
env.close()
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
"""
run the training of DDPG
:param env_id: (str) the environment ID
:param seed: (int) the initial random seed
:param noise_type: (str) the wanted noises ('adaptive-param', 'normal' or 'ou'), can use multiple noise type by
seperating them with commas
:param layer_norm: (bool) use layer normalization
:param evaluation: (bool) enable evaluation of DDPG training
:param kwargs: (dict) extra keywords for the training.train function
"""
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
if evaluation and rank == 0:
eval_env = gym.make(env_id)
eval_env = bench.Monitor(eval_env,
os.path.join(logger.get_dir(), 'gym_eval'))
env = bench.Monitor(env, None)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(
initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mean=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mean=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
# Seed everything to make things reproducible.
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed,
logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
start_time = 0
if rank == 0:
start_time = time.time()
if layer_norm:
policy = LnMlpPolicy
else:
policy = MlpPolicy
num_timesteps = kwargs['num_timesteps']
del kwargs['num_timesteps']
model = DDPG(policy=policy,
env=env,
eval_env=eval_env,
param_noise=param_noise,
action_noise=action_noise,
buffer_size=int(1e6),
verbose=2,
**kwargs)
model.learn(total_timesteps=num_timesteps)
env.close()
if eval_env is not None:
eval_env.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
def make_env():
env_out = gym.make(env_id)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out