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_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_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 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 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_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_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_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()