Files
Gymnasium/tests/wrappers/test_record_episode_statistics.py

77 lines
2.7 KiB
Python
Raw Normal View History

New `info` API for vectorized environments #2657 (#2773) * WIP refactor info API sync vector. * Add missing untracked file. * Add info strategy to reset_wait. * Add interface and docstring. * info with strategy pattern on async vector env. * Add default to async vecenv. * episode statistics for asyncvecnev. * Add tests info strategy format. * Add info strategy to reset_wait. * refactor and cleanup. * Code cleanup. Add tests. * Add tests for video recording with new info format. * fix test case. * fix camelcase. * rename enum. * update tests, docstrings, cleanup. * Changes brax strategy to numpy. add_strategy method in StrategyFactory. Add tests. * fix docstring and logging format. * Set Brax info format as default. Remove classic info format. Update tests. * breaking the wrong loop. * WIP: wrapper. * Add wrapper for brax to classic info. * WIP: wrapper with nested RecordEpisodeStatistic. * Add tests. Refactor docstrings. Cleanup. * cleanup. * patch conflicts. * rebase and conflicts. * new pre-commit conventions. * docstring. * renaming. * incorporate info_processor in vecEnv. * renaming. Create info dict only if needed. * remove all brax references. update docstring. Update duplicate test. * reviews. * pre-commit. * reviews. * docstring. * cleanup blank lines. * add support for numpy dtypes. * docstring fix. * formatting. * naming. * assert correct info from wrappers chaining. Test correct wrappers chaining. naming. * simplify episode_statistics. * change args orer. * update tests. * wip: refactor episode_statistics. * Add test for add_vecore_episode_statistics.
2022-05-24 16:36:35 +02:00
import numpy as np
import pytest
import gymnasium as gym
2022-09-08 10:10:07 +01:00
from gymnasium.wrappers import RecordEpisodeStatistics, VectorListInfo
Reduces warnings produced by pytest from ~1500 to 13 (#2660) * Updated cartpole-v0 to v1 to prevent warning and added pytest.mark.filterwarnings for tests where warnings are unavoidable * Change np.bool to bool as numpy raises a warning and bool is the suggested solution * Seeding randint is deprecated in the future, integers is new solution * Fixed errors thrown when the video recorder is deleted but not closed * spaces.Box expects a floating array, updated all cases where this was not true and modified float32 to float64 as float array default to float64. Otherwise space.Box raises warning that dtype precision (float32) is lower than array precision (float64). * Added pytest.mark.filterwarnings to preventing the raising of an intended warning * Added comment to explain why a warning is raised that can't be prevented without version update to the environment * Added comment to explain why warning is raised * Changed values to float as expected by the box which default to float64 * Removed --forked from pytest as the pytest-forked project is no being maintained and was not raising warnings as expected * When AsyncVectorEnv has shared_memory=True then a ValueError is raised before _state is initialised. Therefore, on the destruction on the env an error is thrown in .close_extra as _state does not exist * Possible fix that was causing an error in test_call_async_vector_env by ensuring that pygame resources are released * Pygame throws an error with ALSA when closed, using a fix from PettingZoo (https://github.com/Farama-Foundation/PettingZoo/blob/master/pettingzoo/__init__.py). We use the dsp audiodriver to prevent this issue * Modification due to running pre-commit locally * Updated cartpole-v0 to v1 to prevent warning and added pytest.mark.filterwarnings for tests where warnings are unavoidable * Change np.bool to bool as numpy raises a warning and bool is the suggested solution * Seeding randint is deprecated in the future, integers is new solution * Fixed errors thrown when the video recorder is deleted but not closed * spaces.Box expects a floating array, updated all cases where this was not true and modified float32 to float64 as float array default to float64. Otherwise space.Box raises warning that dtype precision (float32) is lower than array precision (float64). * Added pytest.mark.filterwarnings to preventing the raising of an intended warning * Added comment to explain why a warning is raised that can't be prevented without version update to the environment * Added comment to explain why warning is raised * Changed values to float as expected by the box which default to float64 * Removed --forked from pytest as the pytest-forked project is no being maintained and was not raising warnings as expected * When AsyncVectorEnv has shared_memory=True then a ValueError is raised before _state is initialised. Therefore, on the destruction on the env an error is thrown in .close_extra as _state does not exist * Possible fix that was causing an error in test_call_async_vector_env by ensuring that pygame resources are released * Pygame throws an error with ALSA when closed, using a fix from PettingZoo (https://github.com/Farama-Foundation/PettingZoo/blob/master/pettingzoo/__init__.py). We use the dsp audiodriver to prevent this issue * Modification due to running pre-commit locally
2022-03-14 14:27:03 +00:00
@pytest.mark.parametrize("env_id", ["CartPole-v1", "Pendulum-v1"])
2021-07-29 02:26:34 +02:00
@pytest.mark.parametrize("deque_size", [2, 5])
def test_record_episode_statistics(env_id, deque_size):
env = gym.make(env_id, disable_env_checker=True)
env = RecordEpisodeStatistics(env, deque_size)
for n in range(5):
env.reset()
assert env.episode_returns is not None and env.episode_lengths is not None
assert env.episode_returns[0] == 0.0
assert env.episode_lengths[0] == 0
assert env.spec is not None
for t in range(env.spec.max_episode_steps):
_, _, terminated, truncated, info = env.step(env.action_space.sample())
if terminated or truncated:
2021-07-29 02:26:34 +02:00
assert "episode" in info
assert all([item in info["episode"] for item in ["r", "l", "t"]])
break
assert len(env.return_queue) == deque_size
assert len(env.length_queue) == deque_size
def test_record_episode_statistics_reset_info():
env = gym.make("CartPole-v1", disable_env_checker=True)
env = RecordEpisodeStatistics(env)
ob_space = env.observation_space
obs, info = env.reset()
assert ob_space.contains(obs)
assert isinstance(info, dict)
@pytest.mark.parametrize(
("num_envs", "asynchronous"), [(1, False), (1, True), (4, False), (4, True)]
)
def test_record_episode_statistics_with_vectorenv(num_envs, asynchronous):
envs = gym.vector.make(
2022-06-16 14:29:13 +01:00
"CartPole-v1",
render_mode=None,
num_envs=num_envs,
asynchronous=asynchronous,
disable_env_checker=True,
Render API (#2671) * add pygame GUI for frozen_lake.py env * add new line at EOF * pre-commit reformat * improve graphics * new images and dynamic window size * darker tile borders and fix ICC profile * pre-commit hook * adjust elf and stool size * Update frozen_lake.py * reformat * fix #2600 * #2600 * add rgb_array support * reformat * test render api change on FrozenLake * add render support for reset on frozenlake * add clock on pygame render * new render api for blackjack * new render api for cliffwalking * new render api for Env class * update reset method, lunar and Env * fix wrapper * fix reset lunar * new render api for box2d envs * new render api for mujoco envs * fix bug * new render api for classic control envs * fix tests * add render_mode None for CartPole * new render api for test fake envs * pre-commit hook * fix FrozenLake * fix FrozenLake * more render_mode to super - frozenlake * remove kwargs from frozen_lake new * pre-commit hook * add deprecated render method * add backwards compatibility * fix test * add _render * move pygame.init() (avoid pygame dependency on init) * fix pygame dependencies * remove collect_render() maintain multi-behaviours .render() * add type hints * fix renderer * don't call .render() with None * improve docstring * add single_rgb_array to all envs * remove None from metadata["render_modes"] * add type hints to test_env_checkers * fix lint * add comments to renderer * add comments to single_depth_array and single_state_pixels * reformat * add deprecation warnings and env.render_mode declaration * fix lint * reformat * fix tests * add docs * fix car racing determinism * remove warning test envs, customizable modes on renderer * remove commments and add todo for env_checker * fix car racing * replace render mode check with assert * update new mujoco * reformat * reformat * change metaclass definition * fix tests * implement mark suggestions (test, docs, sets) * check_render Co-authored-by: J K Terry <jkterry0@gmail.com>
2022-06-08 00:20:56 +02:00
)
envs = RecordEpisodeStatistics(envs)
max_episode_step = (
envs.env_fns[0]().spec.max_episode_steps
if asynchronous
else envs.env.envs[0].spec.max_episode_steps
)
envs.reset()
for _ in range(max_episode_step + 1):
_, _, terminateds, truncateds, infos = envs.step(envs.action_space.sample())
if any(terminateds) or any(truncateds):
New `info` API for vectorized environments #2657 (#2773) * WIP refactor info API sync vector. * Add missing untracked file. * Add info strategy to reset_wait. * Add interface and docstring. * info with strategy pattern on async vector env. * Add default to async vecenv. * episode statistics for asyncvecnev. * Add tests info strategy format. * Add info strategy to reset_wait. * refactor and cleanup. * Code cleanup. Add tests. * Add tests for video recording with new info format. * fix test case. * fix camelcase. * rename enum. * update tests, docstrings, cleanup. * Changes brax strategy to numpy. add_strategy method in StrategyFactory. Add tests. * fix docstring and logging format. * Set Brax info format as default. Remove classic info format. Update tests. * breaking the wrong loop. * WIP: wrapper. * Add wrapper for brax to classic info. * WIP: wrapper with nested RecordEpisodeStatistic. * Add tests. Refactor docstrings. Cleanup. * cleanup. * patch conflicts. * rebase and conflicts. * new pre-commit conventions. * docstring. * renaming. * incorporate info_processor in vecEnv. * renaming. Create info dict only if needed. * remove all brax references. update docstring. Update duplicate test. * reviews. * pre-commit. * reviews. * docstring. * cleanup blank lines. * add support for numpy dtypes. * docstring fix. * formatting. * naming. * assert correct info from wrappers chaining. Test correct wrappers chaining. naming. * simplify episode_statistics. * change args orer. * update tests. * wip: refactor episode_statistics. * Add test for add_vecore_episode_statistics.
2022-05-24 16:36:35 +02:00
assert "episode" in infos
assert "_episode" in infos
assert all(infos["_episode"] == np.bitwise_or(terminateds, truncateds))
New `info` API for vectorized environments #2657 (#2773) * WIP refactor info API sync vector. * Add missing untracked file. * Add info strategy to reset_wait. * Add interface and docstring. * info with strategy pattern on async vector env. * Add default to async vecenv. * episode statistics for asyncvecnev. * Add tests info strategy format. * Add info strategy to reset_wait. * refactor and cleanup. * Code cleanup. Add tests. * Add tests for video recording with new info format. * fix test case. * fix camelcase. * rename enum. * update tests, docstrings, cleanup. * Changes brax strategy to numpy. add_strategy method in StrategyFactory. Add tests. * fix docstring and logging format. * Set Brax info format as default. Remove classic info format. Update tests. * breaking the wrong loop. * WIP: wrapper. * Add wrapper for brax to classic info. * WIP: wrapper with nested RecordEpisodeStatistic. * Add tests. Refactor docstrings. Cleanup. * cleanup. * patch conflicts. * rebase and conflicts. * new pre-commit conventions. * docstring. * renaming. * incorporate info_processor in vecEnv. * renaming. Create info dict only if needed. * remove all brax references. update docstring. Update duplicate test. * reviews. * pre-commit. * reviews. * docstring. * cleanup blank lines. * add support for numpy dtypes. * docstring fix. * formatting. * naming. * assert correct info from wrappers chaining. Test correct wrappers chaining. naming. * simplify episode_statistics. * change args orer. * update tests. * wip: refactor episode_statistics. * Add test for add_vecore_episode_statistics.
2022-05-24 16:36:35 +02:00
assert all([item in infos["episode"] for item in ["r", "l", "t"]])
break
else:
assert "episode" not in infos
assert "_episode" not in infos
def test_wrong_wrapping_order():
envs = gym.vector.make("CartPole-v1", num_envs=3, disable_env_checker=True)
New `info` API for vectorized environments #2657 (#2773) * WIP refactor info API sync vector. * Add missing untracked file. * Add info strategy to reset_wait. * Add interface and docstring. * info with strategy pattern on async vector env. * Add default to async vecenv. * episode statistics for asyncvecnev. * Add tests info strategy format. * Add info strategy to reset_wait. * refactor and cleanup. * Code cleanup. Add tests. * Add tests for video recording with new info format. * fix test case. * fix camelcase. * rename enum. * update tests, docstrings, cleanup. * Changes brax strategy to numpy. add_strategy method in StrategyFactory. Add tests. * fix docstring and logging format. * Set Brax info format as default. Remove classic info format. Update tests. * breaking the wrong loop. * WIP: wrapper. * Add wrapper for brax to classic info. * WIP: wrapper with nested RecordEpisodeStatistic. * Add tests. Refactor docstrings. Cleanup. * cleanup. * patch conflicts. * rebase and conflicts. * new pre-commit conventions. * docstring. * renaming. * incorporate info_processor in vecEnv. * renaming. Create info dict only if needed. * remove all brax references. update docstring. Update duplicate test. * reviews. * pre-commit. * reviews. * docstring. * cleanup blank lines. * add support for numpy dtypes. * docstring fix. * formatting. * naming. * assert correct info from wrappers chaining. Test correct wrappers chaining. naming. * simplify episode_statistics. * change args orer. * update tests. * wip: refactor episode_statistics. * Add test for add_vecore_episode_statistics.
2022-05-24 16:36:35 +02:00
wrapped_env = RecordEpisodeStatistics(VectorListInfo(envs))
wrapped_env.reset()
with pytest.raises(AssertionError):
wrapped_env.step(wrapped_env.action_space.sample())