def test_check_spaces_sync_vector_env():
# CartPole-v1 - observation_space: Box(4,), action_space: Discrete(2)
env_fns = [make_env("CartPole-v1", i) for i in range(8)]
# FrozenLake-v1 - Discrete(16), action_space: Discrete(4)
env_fns[1] = make_env("FrozenLake-v1", 1)
with pytest.raises(RuntimeError):
env = SyncVectorEnv(env_fns)
env.close()
def test_check_observations_sync_vector_env():
# CubeCrash-v0 - observation_space: Box(40, 32, 3)
env_fns = [make_env('CubeCrash-v0', i) for i in range(8)]
# MemorizeDigits-v0 - observation_space: Box(24, 32, 3)
env_fns[1] = make_env('MemorizeDigits-v0', 1)
with pytest.raises(RuntimeError):
env = SyncVectorEnv(env_fns)
env.close()
def test_create_sync_vector_env():
env_fns = [make_env("FrozenLake-v1", i) for i in range(8)]
try:
env = SyncVectorEnv(env_fns)
finally:
env.close()
assert env.num_envs == 8
def test_create_sync_vector_env():
env_fns = [make_env('CubeCrash-v0', i) for i in range(8)]
try:
env = SyncVectorEnv(env_fns)
finally:
env.close()
assert env.num_envs == 8
def test_sync_vector_determinism(spec: EnvSpec, seed: int = 123, n: int = 3):
"""Check that for all environments, the sync vector envs produce the same action samples using the same seeds"""
env_1 = SyncVectorEnv([make_env(spec.id, seed=seed) for _ in range(n)])
env_2 = SyncVectorEnv([make_env(spec.id, seed=seed) for _ in range(n)])
assert_rng_equal(env_1.action_space.np_random, env_2.action_space.np_random)
for _ in range(100):
env_1_samples = env_1.action_space.sample()
env_2_samples = env_2.action_space.sample()
assert np.all(env_1_samples == env_2_samples)
def test_reset_sync_vector_env():
env_fns = [make_env('CubeCrash-v0', i) for i in range(8)]
try:
env = SyncVectorEnv(env_fns)
observations = env.reset()
finally:
env.close()
assert isinstance(env.observation_space, Box)
assert isinstance(observations, np.ndarray)
assert observations.dtype == env.observation_space.dtype
assert observations.shape == (8, ) + env.single_observation_space.shape
assert observations.shape == env.observation_space.shape
def test_vector_env_info_concurrent_termination(concurrent_ends):
# envs that need to terminate together will have the same action
actions = [0] * concurrent_ends + [1] * (NUM_ENVS - concurrent_ends)
envs = [make_env(ENV_ID, SEED) for _ in range(NUM_ENVS)]
envs = SyncVectorEnv(envs)
for _ in range(ENV_STEPS):
_, _, dones, infos = envs.step(actions)
if any(dones):
for i, done in enumerate(dones):
if i < concurrent_ends:
assert done
assert infos["_terminal_observation"][i]
else:
assert not infos["_terminal_observation"][i]
assert infos["terminal_observation"][i] is None
return
def test_sync_vector_env_seed():
env = make_env("BipedalWalker-v3", seed=123)()
sync_vector_env = SyncVectorEnv([make_env("BipedalWalker-v3", seed=123)])
assert_rng_equal(env.action_space.np_random, sync_vector_env.action_space.np_random)
for _ in range(100):
env_action = env.action_space.sample()
vector_action = sync_vector_env.action_space.sample()
assert np.all(env_action == vector_action)
def test_set_attr_sync_vector_env():
env_fns = [make_env("CartPole-v1", i) for i in range(4)]
try:
env = SyncVectorEnv(env_fns)
env.set_attr("gravity", [9.81, 3.72, 8.87, 1.62])
gravity = env.get_attr("gravity")
assert gravity == (9.81, 3.72, 8.87, 1.62)
finally:
env.close()
def make(id, num_envs=1, asynchronous=True, wrappers=None, **kwargs):
"""Create a vectorized environment from multiple copies of an environment,
from its id
Parameters
----------
id : str
The environment ID. This must be a valid ID from the registry.
num_envs : int
Number of copies of the environment.
asynchronous : bool (default: `True`)
If `True`, wraps the environments in an `AsyncVectorEnv` (which uses
`multiprocessing` to run the environments in parallel). If `False`,
wraps the environments in a `SyncVectorEnv`.
wrappers : Callable or Iterable of Callables (default: `None`)
If not `None`, then apply the wrappers to each internal
environment during creation.
Returns
-------
env : `gym.vector.VectorEnv` instance
The vectorized environment.
Example
-------
>>> import gym
>>> env = gym.vector.make('CartPole-v1', 3)
>>> env.reset()
array([[-0.04456399, 0.04653909, 0.01326909, -0.02099827],
[ 0.03073904, 0.00145001, -0.03088818, -0.03131252],
[ 0.03468829, 0.01500225, 0.01230312, 0.01825218]],
dtype=float32)
"""
from gym.envs import make as make_
def _make_env():
env = make_(id, **kwargs)
if wrappers is not None:
if callable(wrappers):
env = wrappers(env)
elif isinstance(wrappers, Iterable) and all(
[callable(w) for w in wrappers]
):
for wrapper in wrappers:
env = wrapper(env)
else:
raise NotImplementedError
return env
env_fns = [_make_env for _ in range(num_envs)]
return AsyncVectorEnv(env_fns) if asynchronous else SyncVectorEnv(env_fns)
def make(
id: str,
num_envs: int = 1,
asynchronous: bool = True,
wrappers: Optional[Union[callable, List[callable]]] = None,
disable_env_checker: bool = False,
**kwargs,
) -> VectorEnv:
"""Create a vectorized environment from multiple copies of an environment, from its id.
Example::
>>> import gym
>>> env = gym.vector.make('CartPole-v1', num_envs=3)
>>> env.reset()
array([[-0.04456399, 0.04653909, 0.01326909, -0.02099827],
[ 0.03073904, 0.00145001, -0.03088818, -0.03131252],
[ 0.03468829, 0.01500225, 0.01230312, 0.01825218]],
dtype=float32)
Args:
id: The environment ID. This must be a valid ID from the registry.
num_envs: Number of copies of the environment.
asynchronous: If `True`, wraps the environments in an :class:`AsyncVectorEnv` (which uses `multiprocessing`_ to run the environments in parallel). If ``False``, wraps the environments in a :class:`SyncVectorEnv`.
wrappers: If not ``None``, then apply the wrappers to each internal environment during creation.
disable_env_checker: If to disable the env checker, if True it will only run on the first environment created.
**kwargs: Keywords arguments applied during gym.make
Returns:
The vectorized environment.
"""
def create_env(_disable_env_checker):
"""Creates an environment that can enable or disable the environment checker."""
def _make_env():
env = gym.envs.registration.make(
id, disable_env_checker=_disable_env_checker, **kwargs)
if wrappers is not None:
if callable(wrappers):
env = wrappers(env)
elif isinstance(wrappers, Iterable) and all(
[callable(w) for w in wrappers]):
for wrapper in wrappers:
env = wrapper(env)
else:
raise NotImplementedError
return env
return _make_env
env_fns = [
create_env(disable_env_checker or env_num > 0)
for env_num in range(num_envs)
]
return AsyncVectorEnv(env_fns) if asynchronous else SyncVectorEnv(env_fns)
def test_step_sync_vector_env(use_single_action_space):
env_fns = [make_env('CubeCrash-v0', i) for i in range(8)]
try:
env = SyncVectorEnv(env_fns)
observations = env.reset()
if use_single_action_space:
actions = [env.single_action_space.sample() for _ in range(8)]
else:
actions = env.action_space.sample()
observations, rewards, dones, _ = env.step(actions)
finally:
env.close()
assert isinstance(env.observation_space, Box)
assert isinstance(observations, np.ndarray)
assert observations.dtype == env.observation_space.dtype
assert observations.shape == (8, ) + env.single_observation_space.shape
assert observations.shape == env.observation_space.shape
assert isinstance(rewards, np.ndarray)
assert isinstance(rewards[0], (float, np.floating))
assert rewards.ndim == 1
assert rewards.size == 8
assert isinstance(dones, np.ndarray)
assert dones.dtype == np.bool_
assert dones.ndim == 1
assert dones.size == 8
def test_custom_space_sync_vector_env():
env_fns = [make_custom_space_env(i) for i in range(4)]
try:
env = SyncVectorEnv(env_fns)
reset_observations = env.reset()
assert isinstance(env.single_action_space, CustomSpace)
assert isinstance(env.action_space, Tuple)
actions = ("action-2", "action-3", "action-5", "action-7")
step_observations, rewards, dones, _ = env.step(actions)
finally:
env.close()
assert isinstance(env.single_observation_space, CustomSpace)
assert isinstance(env.observation_space, Tuple)
assert isinstance(reset_observations, tuple)
assert reset_observations == ("reset", "reset", "reset", "reset")
assert isinstance(step_observations, tuple)
assert step_observations == (
"step(action-2)",
"step(action-3)",
"step(action-5)",
"step(action-7)",
)
def test_vector_env_equal(shared_memory):
env_fns = [make_env("CubeCrash-v0", i) for i in range(4)]
num_steps = 100
try:
async_env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
sync_env = SyncVectorEnv(env_fns)
async_env.seed(0)
sync_env.seed(0)
assert async_env.num_envs == sync_env.num_envs
assert async_env.observation_space == sync_env.observation_space
assert async_env.single_observation_space == sync_env.single_observation_space
assert async_env.action_space == sync_env.action_space
assert async_env.single_action_space == sync_env.single_action_space
async_observations = async_env.reset()
sync_observations = sync_env.reset()
assert np.all(async_observations == sync_observations)
for _ in range(num_steps):
actions = async_env.action_space.sample()
assert actions in sync_env.action_space
# fmt: off
async_observations, async_rewards, async_dones, async_infos = async_env.step(
actions)
sync_observations, sync_rewards, sync_dones, sync_infos = sync_env.step(
actions)
# fmt: on
for idx in range(len(sync_dones)):
if sync_dones[idx]:
assert "terminal_observation" in async_infos[idx]
assert "terminal_observation" in sync_infos[idx]
assert sync_dones[idx]
assert np.all(async_observations == sync_observations)
assert np.all(async_rewards == sync_rewards)
assert np.all(async_dones == sync_dones)
finally:
async_env.close()
sync_env.close()
def main():
np.set_printoptions(suppress=True,
formatter={'float_kind': '{:0.2f}'.format})
env_fns = [make_env('MountainCar-v0', i) for i in range(4)]
try:
env = SyncVectorEnv(env_fns)
finally:
env.close()
state_size = env.observation_space.shape[1]
action_size = env.action_space[0].n
NUM_EPISODES = 1000
STEPS_PER_EPISODE = 200
batch_size = 32
eps_mean_reward = [0.0] * NUM_EPISODES
agent = DQNAgent(state_size, action_size)
start_time = datetime.now()
for ep_count in range(NUM_EPISODES):
episode_rew = 0
state = env.reset()
if (ep_count == 0):
print("ep={} state.shape={}".format(ep_count, state.shape))
#state = np.reshape(state, [-1, state_size])
ep_start_time = datetime.now()
for time in range(STEPS_PER_EPISODE):
# env.render()
action = agent.act(state)
next_state, reward, done, _ = env.step(action)
episode_rew += np.sum(reward)
#next_state = np.reshape(next_state, [-1, state_size])
if (time == 0):
print("ep={} time={} action.len={} next_state.shape={} elaps_time={}".format( \
ep_count, time, len(action), next_state.shape, (datetime.now() - ep_start_time)) )
#add to DQN buffer
for idx in range(0, env.num_envs):
agent.memorize(state[idx], action[idx], reward[idx],
next_state[idx], done[idx])
state = next_state
if time >= STEPS_PER_EPISODE - 1:
eps_mean_reward[ep_count] = np.mean(episode_rew) / time
print("ep: {}/{}, mean_avg_reward: {}, exec_time= {}".format( \
ep_count , NUM_EPISODES, eps_mean_reward[ep_count], (datetime.now() - ep_start_time)))
#update DQN model if there are enough samples
if len(agent.memory) > batch_size and time % 8 == 0:
agent.replay(batch_size)
#if ep_count % 2 == 0:
# agent.save(str(os.path.join(save_path,'ma-foraging-dqn.h5')))
print("Finish train DQN Agent with {} episodes in {}".format(
NUM_EPISODES, (datetime.now() - start_time)))
def make_env(
env_id: str,
num_envs: int = 1,
):
def _make():
_env = gym.make(env_id)
return _env
if num_envs == 1:
env = SyncVectorEnv([_make])
if num_envs > 1:
env = AsyncVectorEnv([_make for _ in range(num_envs)])
dummy_env = _make()
setattr(env, "spec", dummy_env.spec)
del dummy_env
else:
env = _make()
return env
def make(id, num_envs=1, asynchronous=True, **kwargs):
"""Create a vectorized environment from multiple copies of an environment,
from its id
Parameters
----------
id : str
The environment ID. This must be a valid ID from the registry.
num_envs : int
Number of copies of the environment. If `1`, then it returns an
unwrapped (i.e. non-vectorized) environment.
asynchronous : bool (default: `True`)
If `True`, wraps the environments in an `AsyncVectorEnv` (which uses
`multiprocessing` to run the environments in parallel). If `False`,
wraps the environments in a `SyncVectorEnv`.
Returns
-------
env : `gym.vector.VectorEnv` instance
The vectorized environment.
Example
-------
>>> import gym
>>> env = gym.vector.make('CartPole-v1', 3)
>>> env.reset()
array([[-0.04456399, 0.04653909, 0.01326909, -0.02099827],
[ 0.03073904, 0.00145001, -0.03088818, -0.03131252],
[ 0.03468829, 0.01500225, 0.01230312, 0.01825218]],
dtype=float32)
"""
from gym.envs import make as make_
def _make_env():
return make_(id, **kwargs)
if num_envs == 1:
return _make_env()
env_fns = [_make_env for _ in range(num_envs)]
return AsyncVectorEnv(env_fns) if asynchronous else SyncVectorEnv(env_fns)
def test_call_sync_vector_env():
env_fns = [make_env("CartPole-v1", i) for i in range(4)]
try:
env = SyncVectorEnv(env_fns)
_ = env.reset()
images = env.call("render", mode="rgb_array")
gravity = env.call("gravity")
finally:
env.close()
assert isinstance(images, tuple)
assert len(images) == 4
for i in range(4):
assert isinstance(images[i], np.ndarray)
assert isinstance(gravity, tuple)
assert len(gravity) == 4
for i in range(4):
assert isinstance(gravity[i], float)
assert gravity[i] == 9.8
def test_vector_env_equal(shared_memory):
env_fns = [make_env('CubeCrash-v0', i) for i in range(4)]
num_steps = 100
try:
async_env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
sync_env = SyncVectorEnv(env_fns)
async_env.seed(0)
sync_env.seed(0)
assert async_env.num_envs == sync_env.num_envs
assert async_env.observation_space == sync_env.observation_space
assert async_env.single_observation_space == sync_env.single_observation_space
assert async_env.action_space == sync_env.action_space
assert async_env.single_action_space == sync_env.single_action_space
async_observations = async_env.reset()
sync_observations = sync_env.reset()
assert np.all(async_observations == sync_observations)
for _ in range(num_steps):
actions = async_env.action_space.sample()
assert actions in sync_env.action_space
async_observations, async_rewards, async_dones, _ = async_env.step(
actions)
sync_observations, sync_rewards, sync_dones, _ = sync_env.step(
actions)
assert np.all(async_observations == sync_observations)
assert np.all(async_rewards == sync_rewards)
assert np.all(async_dones == sync_dones)
finally:
async_env.close()
sync_env.close()
def test_custom_space_sync_vector_env():
env_fns = [make_custom_space_env(i) for i in range(4)]
try:
env = SyncVectorEnv(env_fns)
reset_observations = env.reset()
actions = ('action-2', 'action-3', 'action-5', 'action-7')
step_observations, rewards, dones, _ = env.step(actions)
finally:
env.close()
assert isinstance(env.single_observation_space, CustomSpace)
assert isinstance(env.observation_space, Tuple)
assert isinstance(reset_observations, tuple)
assert reset_observations == ('reset', 'reset', 'reset', 'reset')
assert isinstance(step_observations, tuple)
assert step_observations == ('step(action-2)', 'step(action-3)',
'step(action-5)', 'step(action-7)')
def test_vector_env_equal(shared_memory):
env_fns = [make_env("CartPole-v1", i) for i in range(4)]
num_steps = 100
try:
async_env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
sync_env = SyncVectorEnv(env_fns)
assert async_env.num_envs == sync_env.num_envs
assert async_env.observation_space == sync_env.observation_space
assert async_env.single_observation_space == sync_env.single_observation_space
assert async_env.action_space == sync_env.action_space
assert async_env.single_action_space == sync_env.single_action_space
async_observations = async_env.reset(seed=0)
sync_observations = sync_env.reset(seed=0)
assert np.all(async_observations == sync_observations)
for _ in range(num_steps):
actions = async_env.action_space.sample()
assert actions in sync_env.action_space
# fmt: off
async_observations, async_rewards, async_dones, async_infos = async_env.step(
actions)
sync_observations, sync_rewards, sync_dones, sync_infos = sync_env.step(
actions)
# fmt: on
if any(sync_dones):
assert "terminal_observation" in async_infos
assert "_terminal_observation" in async_infos
assert "terminal_observation" in sync_infos
assert "_terminal_observation" in sync_infos
assert np.all(async_observations == sync_observations)
assert np.all(async_rewards == sync_rewards)
assert np.all(async_dones == sync_dones)
finally:
async_env.close()
sync_env.close()
def test_reset_sync_vector_env():
env_fns = [make_env("CartPole-v1", i) for i in range(8)]
try:
env = SyncVectorEnv(env_fns)
observations = env.reset()
finally:
env.close()
assert isinstance(env.observation_space, Box)
assert isinstance(observations, np.ndarray)
assert observations.dtype == env.observation_space.dtype
assert observations.shape == (8,) + env.single_observation_space.shape
assert observations.shape == env.observation_space.shape
del observations
try:
env = SyncVectorEnv(env_fns)
observations = env.reset(return_info=False)
finally:
env.close()
assert isinstance(env.observation_space, Box)
assert isinstance(observations, np.ndarray)
assert observations.dtype == env.observation_space.dtype
assert observations.shape == (8,) + env.single_observation_space.shape
assert observations.shape == env.observation_space.shape
del observations
env_fns = [make_env("CartPole-v1", i) for i in range(8)]
try:
env = SyncVectorEnv(env_fns)
observations, infos = env.reset(return_info=True)
finally:
env.close()
assert isinstance(env.observation_space, Box)
assert isinstance(observations, np.ndarray)
assert observations.dtype == env.observation_space.dtype
assert observations.shape == (8,) + env.single_observation_space.shape
assert observations.shape == env.observation_space.shape
assert isinstance(infos, list)
assert all([isinstance(info, dict) for info in infos])
