def get_marioenv(world=1, stage=1, version=0, movement = RIGHT_ONLY, max_episode_steps=4500):
mario_env = JoypadSpace(SuperMario_Env(world, stage, version), movement)
mario_env._max_episode_steps = max_episode_steps * 4
mario_env = StickyActionEnv(mario_env)
mario_env = MaxAndSkipEnv(mario_env, skip=4)
mario_env = DummyMontezumaInfoWrapper(mario_env)
mario_env = AddRandomStateToInfo(mario_env)
mario_env = wrap_deepmind(mario_env, frame_stack=True)
return mario_env
def create_env(config):
env = gym.make(config.env)
expt_dir = config.log_path + '/monitor'
env = gym.wrappers.Monitor(env, expt_dir, force=True, video_callable=False)
# If the replay buffer is needed (off-policy algorithms),
# observation is stored in (int) format to efficiently manage the memory
# Therefore, scaling function is passed on to the buffer's encoding function
if config.agent_type in alg_type['off_policy']:
env = wrap_deepmind(env, scale=False)
else:
env = wrap_deepmind(env, scale=True)
# Parallelization is supported if number of actor is more than one
if config.num_actors > 1:
env = ParallelEnv(num_processes=config.num_actors, env=env)
return env
def _thunk():
mario_env = JoypadSpace(SuperMario_Env(world, stage, version),
movement)
if wrap_atari:
mario_env._max_episode_steps = max_episode_steps * 4
mario_env = StickyActionEnv(mario_env)
mario_env = MaxAndSkipEnv(mario_env, skip=4)
mario_env = DummyMontezumaInfoWrapper(mario_env)
mario_env = AddRandomStateToInfo(mario_env)
# mario_env.seed(seed + rank)
mario_env = Monitor(
mario_env,
logger.get_dir()
and os.path.join(logger.get_dir(), str(rank)),
allow_early_resets=True)
if wrap_atari:
mario_env = wrap_deepmind(mario_env)
mario_env = BlocksWrapper(mario_env)
mario_env.seed(seed)
return mario_env