def test_pickleable(self):
env = GridWorldEnv(desc="8x8")
round_trip = pickle.loads(pickle.dumps(env))
assert round_trip
assert round_trip.start_state == env.start_state
step_env(round_trip)
round_trip.close()
env.close()
def test_does_not_modify_action(self):
env = GridWorldEnv(desc='8x8')
a = env.action_space.sample()
a_copy = a
env.reset()
env.step(a)
assert a == a_copy
env.close()
def test_does_not_modify_action(self):
env = GridWorldEnv(desc="8x8")
a = env.action_space.sample()
a_copy = a
env.reset()
env.step(a)
self.assertEqual(a, a_copy)
env.close()
def setup_method(self):
self.env = TfEnv(GridWorldEnv(desc='4x4'))
self.policy = ScriptedPolicy(
scripted_actions=[2, 2, 1, 0, 3, 1, 1, 1, 2, 2, 1, 1, 1, 2, 2, 1])
self.algo = Mock(env_spec=self.env.spec,
policy=self.policy,
max_path_length=16)
def setup_method(self):
ray.init(local_mode=True, ignore_reinit_error=True)
self.env = TfEnv(GridWorldEnv(desc='4x4'))
self.policy = ScriptedPolicy(
scripted_actions=[2, 2, 1, 0, 3, 1, 1, 1, 2, 2, 1, 1, 1, 2, 2, 1])
self.algo = Mock(
env_spec=self.env.spec, policy=self.policy, max_path_length=16)
def other_envs():
descs = [
['FFFS', 'FHFH', 'FFFH', 'HFFG'],
['FFSF', 'FFFH', 'FHFH', 'HFFG'],
['FFFF', 'FFSH', 'FHFH', 'FFFG'],
['FFFF', 'FFFF', 'FSFF', 'FFFF'],
['HHFF', 'HHHF', 'HSHF', 'HHHF'],
]
return [TfEnv(GridWorldEnv(desc=desc)) for desc in descs]
def envs():
descs = [
['SFFF', 'FHFH', 'FFFH', 'HFFG'],
['SFFF', 'FFFH', 'FHFH', 'HFFG'],
['SFFF', 'FFFH', 'FHFH', 'FFFG'],
['SFFF', 'FFFF', 'FFFF', 'FFFF'],
['SHFF', 'HHFF', 'FFFF', 'FFFF'],
]
return [TfEnv(GridWorldEnv(desc=desc)) for desc in descs]
def test_obtain_samples(ray_local_session_fixture):
del ray_local_session_fixture
env = GridWorldEnv(desc='4x4')
policy = ScriptedPolicy(
scripted_actions=[2, 2, 1, 0, 3, 1, 1, 1, 2, 2, 1, 1, 1, 2, 2, 1])
algo = Mock(env_spec=env.spec, policy=policy, max_episode_length=16)
assert ray.is_initialized()
workers = WorkerFactory(seed=100,
max_episode_length=algo.max_episode_length,
n_workers=8)
sampler1 = RaySampler.from_worker_factory(workers, policy, env)
sampler2 = LocalSampler.from_worker_factory(workers, policy, env)
eps1 = sampler1.obtain_samples(0, 1000,
tuple(algo.policy.get_param_values()))
eps2 = sampler2.obtain_samples(0, 1000,
tuple(algo.policy.get_param_values()))
assert eps1.observations.shape[0] >= 1000
assert eps1.actions.shape[0] >= 1000
assert (sum(eps1.rewards[:eps1.lengths[0]]) == sum(
eps2.rewards[:eps2.lengths[0]]) == 1)
true_obs = np.array([0, 1, 2, 6, 10, 14])
true_actions = np.array([2, 2, 1, 1, 1, 2])
true_rewards = np.array([0, 0, 0, 0, 0, 1])
start = 0
for length in eps1.lengths:
observations = eps1.observations[start:start + length]
actions = eps1.actions[start:start + length]
rewards = eps1.rewards[start:start + length]
assert np.array_equal(observations, true_obs)
assert np.array_equal(actions, true_actions)
assert np.array_equal(rewards, true_rewards)
start += length
sampler1.shutdown_worker()
sampler2.shutdown_worker()
env.close()
def test_obtain_samples():
env = GarageEnv(GridWorldEnv(desc='4x4'))
policy = ScriptedPolicy(
scripted_actions=[2, 2, 1, 0, 3, 1, 1, 1, 2, 2, 1, 1, 1, 2, 2, 1])
algo = Mock(env_spec=env.spec, policy=policy, max_episode_length=16)
workers = WorkerFactory(seed=100,
max_episode_length=algo.max_episode_length,
n_workers=8)
sampler1 = MultiprocessingSampler.from_worker_factory(workers, policy, env)
sampler2 = LocalSampler.from_worker_factory(workers, policy, env)
trajs1 = sampler1.obtain_samples(0, 1000,
tuple(algo.policy.get_param_values()))
trajs2 = sampler2.obtain_samples(0, 1000,
tuple(algo.policy.get_param_values()))
# pylint: disable=superfluous-parens
assert trajs1.observations.shape[0] >= 1000
assert trajs1.actions.shape[0] >= 1000
assert (sum(trajs1.rewards[:trajs1.lengths[0]]) == sum(
trajs2.rewards[:trajs2.lengths[0]]) == 1)
true_obs = np.array([0, 1, 2, 6, 10, 14])
true_actions = np.array([2, 2, 1, 1, 1, 2])
true_rewards = np.array([0, 0, 0, 0, 0, 1])
start = 0
for length in trajs1.lengths:
observations = trajs1.observations[start:start + length]
actions = trajs1.actions[start:start + length]
rewards = trajs1.rewards[start:start + length]
assert np.array_equal(observations, true_obs)
assert np.array_equal(actions, true_actions)
assert np.array_equal(rewards, true_rewards)
start += length
sampler1.shutdown_worker()
sampler2.shutdown_worker()
env.close()
def test_ray_batch_sampler(ray_local_session_fixture):
del ray_local_session_fixture
env = TfEnv(GridWorldEnv(desc='4x4'))
policy = ScriptedPolicy(
scripted_actions=[2, 2, 1, 0, 3, 1, 1, 1, 2, 2, 1, 1, 1, 2, 2, 1])
algo = Mock(env_spec=env.spec, policy=policy, max_path_length=16)
assert ray.is_initialized()
workers = WorkerFactory(seed=100, max_path_length=algo.max_path_length)
sampler1 = RaySampler(workers, policy, env)
sampler1.start_worker()
sampler2 = OnPolicyVectorizedSampler(algo, env)
sampler2.start_worker()
trajs1 = sampler1.obtain_samples(0, 1000,
tuple(algo.policy.get_param_values()))
trajs2 = sampler2.obtain_samples(0, 1000)
# pylint: disable=superfluous-parens
assert trajs1.observations.shape[0] >= 1000
assert trajs1.actions.shape[0] >= 1000
assert (sum(trajs1.rewards[:trajs1.lengths[0]]) == sum(
trajs2[0]['rewards']) == 1)
true_obs = np.array([0, 1, 2, 6, 10, 14])
true_actions = np.array([2, 2, 1, 1, 1, 2])
true_rewards = np.array([0, 0, 0, 0, 0, 1])
start = 0
for length in trajs1.lengths:
observations = trajs1.observations[start:start + length]
actions = trajs1.actions[start:start + length]
rewards = trajs1.rewards[start:start + length]
assert np.array_equal(observations, true_obs)
assert np.array_equal(actions, true_actions)
assert np.array_equal(rewards, true_rewards)
start += length
sampler1.shutdown_worker()
sampler2.shutdown_worker()
env.close()
def env():
return TfEnv(GridWorldEnv(desc='4x4'))
