def test_config(normalization_config):
# init environment
env = GymMazeEnv("CartPole-v0")
# wrap env with observation normalization
env = ObservationNormalizationWrapper(
env,
default_strategy=normalization_config["default_strategy"],
default_strategy_config=normalization_config[
"default_strategy_config"],
default_statistics=normalization_config["default_statistics"],
statistics_dump=normalization_config["statistics_dump"],
sampling_policy=normalization_config['sampling_policy'],
exclude=normalization_config["exclude"],
manual_config=normalization_config["manual_config"])
# check if action space clipping was applied
assert np.alltrue(env.observation_space["observation"].high <= 1.0)
assert np.alltrue(env.observation_space["observation"].low >= 0.0)
# check if stats have been set properly
statistics = env.get_statistics()
assert np.all(statistics["observation"]["mean"] == np.zeros(shape=4))
assert np.all(statistics["observation"]["std"] == np.ones(shape=4))
# test sampling
obs = random_env_steps(env, steps=100)
assert np.min(obs) >= 0 and np.max(obs) <= 1
def test_observation_normalization_manual_default_stats():
""" observation normalization test """
# init environment
env = GymMazeEnv("CartPole-v0")
# normalization config
normalization_config = {
"default_strategy":
"maze.normalization_strategies.MeanZeroStdOneObservationNormalizationStrategy",
"default_strategy_config": {
"clip_range": (0, 1),
"axis": 0
},
"default_statistics": {
"mean": [0, 0, 0, 0],
"std": [1, 1, 1, 1]
},
"statistics_dump": "statistics.pkl",
"sampling_policy": RandomPolicy(env.action_spaces_dict),
"exclude": None,
"manual_config": None,
}
# wrap env with observation normalization
env = ObservationNormalizationWrapper(
env,
default_strategy=normalization_config["default_strategy"],
default_strategy_config=normalization_config[
"default_strategy_config"],
default_statistics=normalization_config["default_statistics"],
statistics_dump=normalization_config["statistics_dump"],
sampling_policy=normalization_config['sampling_policy'],
exclude=normalization_config["exclude"],
manual_config=normalization_config["manual_config"])
# check if action space clipping was applied
assert np.alltrue(env.observation_space["observation"].high <= 1.0)
assert np.alltrue(env.observation_space["observation"].low >= 0.0)
# check if stats have been set properly
statistics = env.get_statistics()
assert np.all(statistics["observation"]["mean"] == np.zeros(shape=4))
assert np.all(statistics["observation"]["std"] == np.ones(shape=4))
# test sampling
obs = random_env_steps(env, steps=100)
assert np.min(obs) >= 0 and np.max(obs) <= 1
def test_observation_normalization_init_from_yaml_config():
""" observation normalization test """
# load config
config = load_env_config(test_observation_normalization_module,
"dummy_config_file.yml")
# init environment
env = GymMazeEnv("CartPole-v0")
env = ObservationNormalizationWrapper(
env, **config["observation_normalization_wrapper"])
assert isinstance(env, ObservationNormalizationWrapper)
stats = env.get_statistics()
assert "stat_1" in stats["observation"] and "stat_2" in stats["observation"]
norm_strategies = getattr(env, "_normalization_strategies")
strategy = norm_strategies["observation"]
assert isinstance(strategy, ObservationNormalizationStrategy)
assert strategy._clip_min == 0
assert strategy._clip_max == 1
assert np.all(strategy._statistics["stat_1"] == np.asarray([0, 0, 0, 0]))
assert np.all(strategy._statistics["stat_2"] == np.asarray([1, 1, 1, 1]))
def test_observation_normalization_pipeline():
""" observation normalization test """
# wrap env with observation normalization
env = GymMazeEnv("CartPole-v0")
# normalization config
normalization_config = {
"default_strategy":
"maze.normalization_strategies.RangeZeroOneObservationNormalizationStrategy",
"default_strategy_config": {
"clip_range": (None, None),
"axis": 0
},
"default_statistics": None,
"sampling_policy": RandomPolicy(env.action_spaces_dict),
"statistics_dump": "statistics.pkl",
"exclude": None,
"manual_config": None
}
env = ObservationNormalizationWrapper(
env,
default_strategy=normalization_config["default_strategy"],
default_strategy_config=normalization_config[
"default_strategy_config"],
default_statistics=normalization_config["default_statistics"],
statistics_dump=normalization_config["statistics_dump"],
sampling_policy=normalization_config['sampling_policy'],
exclude=normalization_config["exclude"],
manual_config=normalization_config["manual_config"])
# check statistics
statistics = env.get_statistics()
assert statistics["observation"] is None, statistics
# check that assertion is thrown
with pytest.raises(AssertionError):
random_env_steps(env, steps=1)
# estimate normalization statistics
statistics = obtain_normalization_statistics(env, n_samples=1000)
# check statistics
for sub_step_key in env.observation_spaces_dict:
for obs_key in env.observation_spaces_dict[sub_step_key].spaces:
assert obs_key in statistics
for stats_key in statistics[obs_key]:
stats = statistics[obs_key][stats_key]
assert isinstance(stats, np.ndarray)
# test normalization
random_env_steps(env, steps=100)
# test file dump and loading
statistics_copy = copy.deepcopy(env.get_statistics())
assert os.path.exists("statistics.pkl")
# wrap env with observation normalization
env = GymMazeEnv("CartPole-v0")
env = ObservationNormalizationWrapper(
env,
default_strategy=normalization_config["default_strategy"],
default_strategy_config=normalization_config[
"default_strategy_config"],
default_statistics=normalization_config["default_statistics"],
statistics_dump=normalization_config["statistics_dump"],
sampling_policy=normalization_config['sampling_policy'],
exclude=normalization_config["exclude"],
manual_config=normalization_config["manual_config"])
# check if stats loading worked properly
statistics = env.get_statistics()
for _ in env.observation_spaces_dict:
for obs_key in statistics:
for stats_key in statistics[obs_key]:
assert np.all(statistics[obs_key][stats_key] ==
statistics_copy[obs_key][stats_key])
# check if stepping works
random_env_steps(env, steps=100)