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_check_spaces_async_vector_env(shared_memory):
# 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 = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
env.close(terminate=True)
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_async_vector_env(shared_memory):
# 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 = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
env.close(terminate=True)
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_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_step_sync_vector_env(use_single_action_space):
env_fns = [make_env("FrozenLake-v1", i) for i in range(8)]
try:
env = SyncVectorEnv(env_fns)
observations = env.reset()
assert isinstance(env.single_action_space, Discrete)
assert isinstance(env.action_space, MultiDiscrete)
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, MultiDiscrete)
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_step_async_vector_env(shared_memory, use_single_action_space):
env_fns = [make_env("CubeCrash-v0", i) for i in range(8)]
try:
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
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_step_out_of_order_async_vector_env(shared_memory):
env_fns = [make_env("CartPole-v1", i) for i in range(4)]
with pytest.raises(NoAsyncCallError):
try:
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
actions = env.action_space.sample()
observations = env.reset()
observations, rewards, dones, infos = env.step_wait()
except AlreadyPendingCallError as exception:
assert exception.name == "step"
raise
finally:
env.close(terminate=True)
with pytest.raises(AlreadyPendingCallError):
try:
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
actions = env.action_space.sample()
env.reset_async()
env.step_async(actions)
except AlreadyPendingCallError as exception:
assert exception.name == "reset"
raise
finally:
env.close(terminate=True)
def test_create_async_vector_env(shared_memory):
env_fns = [make_env("CartPole-v1", i) for i in range(8)]
try:
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
finally:
env.close()
assert env.num_envs == 8
def test_no_copy_async_vector_env(shared_memory):
env_fns = [make_env("CartPole-v1", i) for i in range(8)]
try:
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory, copy=False)
observations = env.reset()
observations[0] = 0
finally:
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_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 test_no_copy_async_vector_env(shared_memory):
env_fns = [make_env("CubeCrash-v0", i) for i in range(8)]
try:
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory, copy=False)
observations = env.reset()
observations[0] = 128
assert np.all(env.observations[0] == 128)
finally:
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])
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_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 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("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_already_closed_async_vector_env(shared_memory):
env_fns = [make_env("CubeCrash-v0", i) for i in range(4)]
with pytest.raises(ClosedEnvironmentError):
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
env.close()
observations = env.reset()
def test_already_closed_async_vector_env(shared_memory):
env_fns = [make_env("CartPole-v1", i) for i in range(4)]
with pytest.raises(ClosedEnvironmentError):
env = AsyncVectorEnv(env_fns, shared_memory=shared_memory)
env.close()
env.reset()
