def __init__(self, spec):
super().__init__(spec)
try_register_env(spec) # register if it's a custom gym env
seed = ps.get(spec, 'meta.random_seed')
episode_life = not util.in_eval_lab_modes()
if self.is_venv: # make vector environment
self.u_env = make_gym_venv(name=self.name,
num_envs=self.num_envs,
seed=seed,
frame_op=self.frame_op,
frame_op_len=self.frame_op_len,
image_downsize=self.image_downsize,
reward_scale=self.reward_scale,
normalize_state=self.normalize_state,
episode_life=episode_life)
else:
self.u_env = make_gym_env(name=self.name,
seed=seed,
frame_op=self.frame_op,
frame_op_len=self.frame_op_len,
image_downsize=self.image_downsize,
reward_scale=self.reward_scale,
normalize_state=self.normalize_state,
episode_life=episode_life)
if self.name.startswith('Unity'):
# Unity is always initialized as singleton gym env, but the Unity runtime can be vec_env
self.num_envs = self.u_env.num_envs
# update variables dependent on num_envs
self._infer_venv_attr()
self._set_clock()
self._set_attr_from_u_env(self.u_env)
self.max_t = self.max_t or self.u_env.spec.max_episode_steps
assert self.max_t is not None
logger.info(util.self_desc(self))
def test_make_gym_env_nostack(name, state_shape, reward_scale):
seed = 0
frame_op = None
frame_op_len = None
env = make_gym_env(name, seed, frame_op, frame_op_len, reward_scale)
env.reset()
for i in range(5):
state, reward, done, info = env.step(env.action_space.sample())
assert isinstance(state, np.ndarray)
assert state.shape == state_shape
assert state.shape == env.observation_space.shape
assert isinstance(reward, float)
assert isinstance(done, bool)
assert isinstance(info, dict)
env.close()
def __init__(self, spec):
super().__init__(spec)
try_register_env(spec) # register if it's a custom gym env
seed = ps.get(spec, 'meta.random_seed')
if self.is_venv: # make vector environment
self.u_env = make_gym_venv(self.name, self.num_envs, seed,
self.frame_op, self.frame_op_len,
self.reward_scale, self.normalize_state)
else:
self.u_env = make_gym_env(self.name, seed, self.frame_op,
self.frame_op_len, self.reward_scale,
self.normalize_state)
self._set_attr_from_u_env(self.u_env)
self.max_t = self.max_t or self.u_env.spec.max_episode_steps
assert self.max_t is not None
logger.info(util.self_desc(self))
def test_make_gym_env_stack(name, state_shape, reward_scale):
seed = 0
frame_op = 'stack' # used for rnn
frame_op_len = 4
env = make_gym_env(name, seed, frame_op, frame_op_len, reward_scale)
env.reset()
for i in range(5):
state, reward, done, info = env.step(env.action_space.sample())
assert isinstance(state, LazyFrames)
state = state.__array__() # realize data
assert isinstance(state, np.ndarray)
# stack creates new dim
stack_shape = (frame_op_len, ) + state_shape
assert state.shape == stack_shape
assert state.shape == env.observation_space.shape
assert isinstance(reward, float)
assert isinstance(done, bool)
assert isinstance(info, dict)
env.close()
def test_make_gym_env_concat(name, state_shape, reward_scale):
seed = 0
frame_op = 'concat' # used for image, or for concat vector
frame_op_len = 4
env = make_gym_env(name, seed, frame_op, frame_op_len, reward_scale)
env.reset()
for i in range(5):
state, reward, done, info = env.step(env.action_space.sample())
assert isinstance(state, LazyFrames)
state = state.__array__() # realize data
assert isinstance(state, np.ndarray)
# concat multiplies first dim
stack_shape = (frame_op_len * state_shape[0], ) + state_shape[1:]
assert state.shape == stack_shape
assert state.shape == env.observation_space.shape
assert isinstance(reward, float)
assert isinstance(done, bool)
assert isinstance(info, dict)
env.close()
