def make(env_name, *make_args, **make_kwargs):
if env_name == "CartPoleSwingUp":
return CartPoleSwingUpEnv()
elif env_name == "CartPoleSwingUpContinuous":
return CartPoleSwingUpContinuousEnv()
else:
return gym_make(env_name, *make_args, **make_kwargs)
def test(path, chkpt, config):
"""
Retrieves a trained PPO agent from the indicated path. While
the specified workload runs in the background, the specified
GYM environment is created and tested with the trained agent,
according to the indicated test configuration. If stated,
test results and environment history log will be saved in the
indicated path.
Parameters
----------
path : str
Path of folder where the trained agent checkpoints are stored.
chkpt : int
Number of the training checkpoint to be used.
config : dict
General configuration of the test environment.
Returns
-------
results : dict(int, int)
Results of test, storing the frequency count of iterations.
history : dict(str, str, dict)
Status history of the environment during test, for each iteration
and step.
"""
import ray
from ray.tune.registry import register_env
env = config['env']
work = config['work']
## REGISTER ENVIRONMENT
Env = load_env_class(gym_spec(env).entry_point)
register_env(env, lambda config: Env(**config))
## TRAINED AGENT RETRIEVAL
ray.init(ignore_reinit_error=True)
chkptpath = path + f"/checkpoint_{chkpt}/checkpoint-{chkpt}"
agent = get_PPOagent(env, config['envconfig'], chkptpath)
## BACKGROUND WORKLOAD INITIALIZATION
workers = start_work(work, config['workconfig'])
## TEST ENVIRONMENT WITH TRAINED AGENT
testenv = gym_make(env, **config['envconfig'])
history = test_env(testenv, agent, config['testconfig'])
## BACKGROUND WORKLOAD KILL
end_work(workers)
## SAVE RESULTS AND STATUS HISTORY
logpath = config['testconfig'].get('logpath', None)
if logpath is not None:
logpath += f'/checkpoint-{chkpt}'
generate_log(history, logpath)
save_config(config, logpath)
return history
def make_environment(env_name):
"""Convert environment with wrappers"""
env = gym_make(env_name)
env = MaxAndSkipEnv(env)
env = FireResetEnv(env)
env = ProcessFrame84(env)
env = ImageToPyTorch(env)
env = BufferWrapper(env, 4)
return ScaledFloatFrame(env)
def __init__(self, ctrl_other=None):
ENV_NAME = 'Pendulum-v0'
# Get the environment and extract the number of actions.
self.env = gym_make(ENV_NAME)
assert len(self.env.action_space.shape) == 1
self.other_controller = ctrl_other # Static Controller for now
self.t = 0.
self.dt = self.env.unwrapped.dt
self.state = self.reset()
self.action_space = self.env.action_space
self.max_torque = self.env.unwrapped.max_torque
high = np.array([np.pi, 1])
self.observation_space = Box(low=-high, high=high, dtype=np.float32)
def __init__(self, ctrl_other=None, max_torque=None, action_space_u1=None, action_space_u2=None, perfident = False):
ENV_NAME = 'Pendulum-v0'
# Get the environment and extract the number of actions.
self.env = gym_make(ENV_NAME).env # last .env enables episodes longer than 200 steps
assert len(self.env.action_space.shape) == 1
self.other_controller = ctrl_other # Can be overwritten by calling set_ctrl_other() afterwards
self.t = 0.
self.dt = self.env.unwrapped.dt
self.state = self.reset()
self.action_space = self.env.action_space
self.action_space_u1 = deepcopy(self.action_space) if action_space_u1 is None else action_space_u1 # limits u1
self.action_space_u2 = deepcopy(self.action_space) if action_space_u2 is None else action_space_u2 # limits u2
self.max_torque = self.env.unwrapped.max_torque if max_torque is None else max_torque # limits the summed u
high = np.array([np.pi, 1])
self.observation_space = Box(low=-high, high=high, dtype=np.float32)
self.perfident = perfident # if the limit of u_other is incorporated
def make_with_gym():
env = gym_make("BreakoutNoFrameskip-v4")
env.seed(0)
env.action_space.np_random.seed(0)
return env
