Stable-Baselines3 v2.4.0: New algorithm (CrossQ in SB3-Contrib) and Gymnasium v1.0 support

Warning

Stable-Baselines3 (SB3) v2.4.0 will be the last one supporting Python 3.8 (end of life in October 2024)
and PyTorch < 2.3.
We highly recommended you to upgrade to Python >= 3.9 and PyTorch >= 2.3 (compatible with NumPy v2).

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo
Stable-Baselines Jax (SBX): https://github.com/araffin/sbx

To upgrade:

pip install stable_baselines3 sb3_contrib rl_zoo3 --upgrade

Note

DQN (and QR-DQN) models saved with SB3 < 2.4.0 will show a warning about truncation of optimizer state when loaded with SB3 >= 2.4.0.
To suppress the warning, simply save the model again.
You can find more info in PR #1963

Breaking Changes:

• Increased minimum required version of Gymnasium to 0.29.1

New Features:

• Added support for pre_linear_modules and post_linear_modules in create_mlp (useful for adding normalization layers, like in DroQ or CrossQ)
• Enabled np.ndarray logging for TensorBoardOutputFormat as histogram (see GH#1634) (@iwishwasaneagle)
• Updated env checker to warn users when using multi-dim array to define MultiDiscrete spaces
• Added support for Gymnasium v1.0

Bug Fixes:

• Fixed memory leak when loading learner from storage, set_parameters() does not try to load the object data anymore
  and only loads the PyTorch parameters (@peteole)
• Cast type in compute gae method to avoid error when using torch compile (@amjames)
• CallbackList now sets the .parent attribute of child callbacks to its own .parent. (will-maclean)
• Fixed error when loading a model that has net_arch manually set to None (@jak3122)
• Set requirement numpy<2.0 until PyTorch is compatible (pytorch/pytorch#107302)
• Updated DQN optimizer input to only include q_network parameters, removing the target_q_network ones (@corentinlger)
• Fixed test_buffers.py::test_device which was not actually checking the device of tensors (@rhaps0dy)

SB3-Contrib

• Added CrossQ algorithm, from "Batch Normalization in Deep Reinforcement Learning" paper (@danielpalen)
• Added BatchRenorm PyTorch layer used in CrossQ (@danielpalen)
• Updated QR-DQN optimizer input to only include quantile_net parameters (@corentinlger)
• Fixed loading QRDQN changes target_update_interval (@jak3122)

RL Zoo

• Updated defaults hyperparameters for TQC/SAC for Swimmer-v4 (decrease gamma for more consistent results)

SBX (SB3 + Jax)

• Added CNN support for DQN
• Bug fix for SAC and related algorithms, optimize log of ent coeff to be consistent with SB3

Others:

• Fixed various typos (@cschindlbeck)
• Remove unnecessary SDE noise resampling in PPO update (@brn-dev)
• Updated PyTorch version on CI to 2.3.1
• Added a warning to recommend using CPU with on policy algorithms (A2C/PPO) and MlpPolicy
• Switched to uv to download packages faster on GitHub CI
• Updated dependencies for read the doc
• Removed unnecessary copy_obs_dict method for SubprocVecEnv, remove the use of ordered dict and rename flatten_obs to stack_obs

Documentation:

• Updated PPO doc to recommend using CPU with MlpPolicy
• Clarified documentation about planned features and citing software
• Added a note about the fact we are optimizing log of ent coeff for SAC

New Contributors

• @amjames made their first contribution in #1922
• @cschindlbeck made their first contribution in #1926
• @peteole made their first contribution in #1908
• @jak3122 made their first contribution in #1937
• @will-maclean made their first contribution in #1939
• @brn-dev made their first contribution in #1933
• @chsahit made their first contribution in #1962
• @Dev1nW made their first contribution in #2017

Full Changelog: v2.3.2...v2.4.0

Stable-Baselines3 v2.3.2: Hotfix for PyTorch 1.13

Bug fixes

• Reverted torch.load() to be called weights_only=False as it caused loading issue with old version of PyTorch. #1913
• Cast learning_rate to float lambda for pickle safety when doing model.load by @markscsmith in #1901

Documentation

• Fix typo in changelog by @araffin in #1882
• Fixed broken link in ppo.rst by @chaitanyabisht in #1884
• Adding ER-MRL to community project by @corentinlger in #1904
• Fix tensorboad video slow numpy->torch conversion by @NickLucche in #1910

New Contributors

• @chaitanyabisht made their first contribution in #1884
• @markscsmith made their first contribution in #1901
• @NickLucche made their first contribution in #1910

Full Changelog: v2.3.0...v2.3.2

Stable-Baselines3 v2.3.0: New defaults hyperparameters for DDPG, TD3 and DQN

Warning

Because of weights_only=True, this release breaks loading of policies when using PyTorch 1.13.
Please upgrade to PyTorch >= 2.0 or upgrade SB3 version (we reverted the change in SB3 2.3.2)

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo
Stable-Baselines Jax (SBX): https://github.com/araffin/sbx

To upgrade:

pip install stable_baselines3 sb3_contrib --upgrade

or simply (rl zoo depends on SB3 and SB3 contrib):

pip install rl_zoo3 --upgrade

Breaking Changes:

• The defaults hyperparameters of TD3 and DDPG have been changed to be more consistent with SAC

  # SB3 < 2.3.0 default hyperparameters
  # model = TD3("MlpPolicy", env, train_freq=(1, "episode"), gradient_steps=-1, batch_size=100)
  # SB3 >= 2.3.0:
  model = TD3("MlpPolicy", env, train_freq=1, gradient_steps=1, batch_size=256)

Note

Two inconsistencies remain: the default network architecture for TD3/DDPG is [400, 300] instead of [256, 256] for SAC (for backward compatibility reasons, see report on the influence of the network size ) and the default learning rate is 1e-3 instead of 3e-4 for SAC (for performance reasons, see W&B report on the influence of the lr )

• The default learning_starts parameter of DQN have been changed to be consistent with the other offpolicy algorithms

  # SB3 < 2.3.0 default hyperparameters, 50_000 corresponded to Atari defaults hyperparameters
  # model = DQN("MlpPolicy", env, learning_starts=50_000)
  # SB3 >= 2.3.0:
  model = DQN("MlpPolicy", env, learning_starts=100)

• For safety, torch.load() is now called with weights_only=True when loading torch tensors,
  policy load() still uses weights_only=False as gymnasium imports are required for it to work
• When using huggingface_sb3, you will now need to set TRUST_REMOTE_CODE=True when downloading models from the hub, as pickle.load is not safe.

New Features:

• Log success rate rollout/success_rate when available for on policy algorithms (@corentinlger)

Bug Fixes:

• Fixed monitor_wrapper argument that was not passed to the parent class, and dones argument that wasn't passed to _update_into_buffer (@corentinlger)

SB3-Contrib

• Added rollout_buffer_class and rollout_buffer_kwargs arguments to MaskablePPO
• Fixed train_freq type annotation for tqc and qrdqn (@Armandpl)
• Fixed sb3_contrib/common/maskable/*.py type annotations
• Fixed sb3_contrib/ppo_mask/ppo_mask.py type annotations
• Fixed sb3_contrib/common/vec_env/async_eval.py type annotations
• Add some additional notes about MaskablePPO (evaluation and multi-process) (@icheered)

RL Zoo

• Updated defaults hyperparameters for TD3/DDPG to be more consistent with SAC
• Upgraded MuJoCo envs hyperparameters to v4 (pre-trained agents need to be updated)
• Added test dependencies to setup.py (@power-edge)
• Simplify dependencies of requirements.txt (remove duplicates from setup.py)

SBX (SB3 + Jax)

• Added support for MultiDiscrete and MultiBinary action spaces to PPO
• Added support for large values for gradient_steps to SAC, TD3, and TQC
• Fix train() signature and update type hints
• Fix replay buffer device at load time
• Added flatten layer
• Added CrossQ

Others:

• Updated black from v23 to v24
• Updated ruff to >= v0.3.1
• Updated env checker for (multi)discrete spaces with non-zero start.

Documentation:

• Added a paragraph on modifying vectorized environment parameters via setters (@fracapuano)
• Updated callback code example
• Updated export to ONNX documentation, it is now much simpler to export SB3 models with newer ONNX Opset!
• Added video link to "Practical Tips for Reliable Reinforcement Learning" video
• Added render_mode="human" in the README example (@marekm4)
• Fixed docstring signature for sum_independent_dims (@StagOverflow)
• Updated docstring description for log_interval in the base class (@rushitnshah).

Full Changelog: v2.2.1...v2.3.0

Stable-Baselines3 v2.2.1: Support for options at reset, bug fixes and better error messages

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo
Stable-Baselines Jax (SBX): https://github.com/araffin/sbx

To upgrade:

pip install stable_baselines3 sb3_contrib --upgrade

or simply (rl zoo depends on SB3 and SB3 contrib):

pip install rl_zoo3 --upgrade

Note

Stable-Baselines3 (SB3) v2.2.0 was yanked after a breaking change was found in GH#1751.
Please use SB3 v2.2.1 and not v2.2.0.

Breaking Changes:

• Switched to ruff for sorting imports (isort is no longer needed), black and ruff version now require a minimum version
• Dropped x is False in favor of not x, which means that callbacks that wrongly returned None (instead of a boolean) will cause the training to stop (@iwishiwasaneagle)

New Features:

• Improved error message of the env_checker for env wrongly detected as GoalEnv (compute_reward() is defined)
• Improved error message when mixing Gym API with VecEnv API (see GH#1694)
• Add support for setting options at reset with VecEnv via the set_options() method. Same as seeds logic, options are reset at the end of an episode (@ReHoss)
• Added rollout_buffer_class and rollout_buffer_kwargs arguments to on-policy algorithms (A2C and PPO)

Bug Fixes:

• Prevents using squash_output and not use_sde in ActorCritcPolicy (@PatrickHelm)
• Performs unscaling of actions in collect_rollout in OnPolicyAlgorithm (@PatrickHelm)
• Moves VectorizedActionNoise into _setup_learn() in OffPolicyAlgorithm (@PatrickHelm)
• Prevents out of bound error on Windows if no seed is passed (@PatrickHelm)
• Calls callback.update_locals() before callback.on_rollout_end() in OnPolicyAlgorithm (@PatrickHelm)
• Fixed replay buffer device after loading in OffPolicyAlgorithm (@PatrickHelm)
• Fixed render_mode which was not properly loaded when using VecNormalize.load()
• Fixed success reward dtype in SimpleMultiObsEnv (@NixGD)
• Fixed check_env for Sequence observation space (@corentinlger)
• Prevents instantiating BitFlippingEnv with conflicting observation spaces (@kylesayrs)
• Fixed ResourceWarning when loading and saving models (files were not closed), please note that only path are closed automatically,
  the behavior stay the same for tempfiles (they need to be closed manually),
  the behavior is now consistent when loading/saving replay buffer

SB3-Contrib

• Added set_options for AsyncEval
• Added rollout_buffer_class and rollout_buffer_kwargs arguments to TRPO

RL Zoo

• Removed gym dependency, the package is still required for some pretrained agents.
• Added --eval-env-kwargs to train.py (@Quentin18)
• Added ppo_lstm to hyperparams_opt.py (@technocrat13)
• Upgraded to pybullet_envs_gymnasium>=0.4.0
• Removed old hacks (for instance limiting offpolicy algorithms to one env at test time)
• Updated docker image, removed support for X server
• Replaced deprecated optuna.suggest_uniform(...) by optuna.suggest_float(..., low=..., high=...)

SBX (SB3 + Jax)

• Added DDPG and TD3 algorithms

Others:

• Fixed stable_baselines3/common/callbacks.py type hints
• Fixed stable_baselines3/common/utils.py type hints
• Fixed stable_baselines3/common/vec_envs/vec_transpose.py type hints
• Fixed stable_baselines3/common/vec_env/vec_video_recorder.py type hints
• Fixed stable_baselines3/common/save_util.py type hints
• Updated docker images to Ubuntu Jammy using micromamba 1.5
• Fixed stable_baselines3/common/buffers.py type hints
• Fixed stable_baselines3/her/her_replay_buffer.py type hints
• Buffers do no call an additional .copy() when storing new transitions
• Fixed ActorCriticPolicy.extract_features() signature by adding an optional features_extractor argument
• Update dependencies (accept newer Shimmy/Sphinx version and remove sphinx_autodoc_typehints)
• Fixed stable_baselines3/common/off_policy_algorithm.py type hints
• Fixed stable_baselines3/common/distributions.py type hints
• Fixed stable_baselines3/common/vec_env/vec_normalize.py type hints
• Fixed stable_baselines3/common/vec_env/__init__.py type hints
• Switched to PyTorch 2.1.0 in the CI (fixes type annotations)
• Fixed stable_baselines3/common/policies.py type hints
• Switched to mypy only for checking types
• Added tests to check consistency when saving/loading files

Documentation:

• Updated RL Tips and Tricks (include recommendation for evaluation, added links to DroQ, ARS and SBX).
• Fixed various typos and grammar mistakes

Full changelog: v2.1.0...v2.2.1

Stable-Baselines3 v2.1.0: Float64 actions, Gymnasium 0.29 support and bug fixes

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo
Stable-Baselines Jax (SBX): https://github.com/araffin/sbx

To upgrade:

pip install stable_baselines3 sb3_contrib --upgrade

or simply (rl zoo depends on SB3 and SB3 contrib):

pip install rl_zoo3 --upgrade

Breaking Changes:

• Removed Python 3.7 support
• SB3 now requires PyTorch >= 1.13

New Features:

• Added Python 3.11 support
• Added Gymnasium 0.29 support (@pseudo-rnd-thoughts)

SB3-Contrib

• Fixed MaskablePPO ignoring stats_window_size argument
• Added Python 3.11 support

RL Zoo

• Upgraded to Huggingface-SB3 >= 2.3
• Added Python 3.11 support

Bug Fixes:

• Relaxed check in logger, that was causing issue on Windows with colorama
• Fixed off-policy algorithms with continuous float64 actions (see #1145) (@tobirohrer)
• Fixed env_checker.py warning messages for out of bounds in complex observation spaces (@Gabo-Tor)

Others:

• Updated GitHub issue templates
• Fix typo in gym patch error message (@lukashass)
• Refactor test_spaces.py tests

Documentation:

• Fixed callback example (@BertrandDecoster)
• Fixed policy network example (@kyle-he)
• Added mobile-env as new community project (@stefanbschneider)
• Added DeepNetSlice to community projects (@AlexPasqua)

Full Changelog: v2.0.0...v2.1.0

Stable-Baselines3 v2.0.0: Gymnasium Support

Warning

Stable-Baselines3 (SB3) v2.0 will be the last one supporting python 3.7 (end of life in June 2023).
We highly recommended you to upgrade to Python >= 3.8.

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo
Stable-Baselines Jax (SBX): https://github.com/araffin/sbx

To upgrade:

pip install stable_baselines3 sb3_contrib rl_zoo3 --upgrade

or simply (rl zoo depends on SB3 and SB3 contrib):

pip install rl_zoo3 --upgrade

Breaking Changes:

• Switched to Gymnasium as primary backend, Gym 0.21 and 0.26 are still supported via the shimmy package (@carlosluis, @arjun-kg, @tlpss)
• The deprecated online_sampling argument of HerReplayBuffer was removed
• Removed deprecated stack_observation_space method of StackedObservations
• Renamed environment output observations in evaluate_policy to prevent shadowing the input observations during callbacks (@npit)
• Upgraded wrappers and custom environment to Gymnasium
• Refined the HumanOutputFormat file check: now it verifies if the object is an instance of io.TextIOBase instead of only checking for the presence of a write method.
• Because of new Gym API (0.26+), the random seed passed to vec_env.seed(seed=seed) will only be effective after then env.reset() call.

New Features:

• Added Gymnasium support (Gym 0.21 and 0.26 are supported via the shimmy package)

SB3-Contrib

• Fixed QRDQN update interval for multi envs

RL Zoo

• Gym 0.26+ patches to continue working with pybullet and TimeLimit wrapper
• Renamed CarRacing-v1 to CarRacing-v2 in hyperparameters
• Huggingface push to hub now accepts a --n-timesteps argument to adjust the length of the video
• Fixed record_video steps (before it was stepping in a closed env)
• Dropped Gym 0.21 support

Bug Fixes:

• Fixed VecExtractDictObs does not handle terminal observation (@WeberSamuel)
• Set NumPy version to >=1.20 due to use of numpy.typing (@troiganto)
• Fixed loading DQN changes target_update_interval (@tobirohrer)
• Fixed env checker to properly reset the env before calling step() when checking
  for Inf and NaN (@lutogniew)
• Fixed HER truncate_last_trajectory() (@lbergmann1)
• Fixed HER desired and achieved goal order in reward computation (@JonathanKuelz)

Others:

• Fixed stable_baselines3/a2c/*.py type hints
• Fixed stable_baselines3/ppo/*.py type hints
• Fixed stable_baselines3/sac/*.py type hints
• Fixed stable_baselines3/td3/*.py type hints
• Fixed stable_baselines3/common/base_class.py type hints
• Fixed stable_baselines3/common/logger.py type hints
• Fixed stable_baselines3/common/envs/*.py type hints
• Fixed stable_baselines3/common/vec_env/vec_monitor|vec_extract_dict_obs|util.py type hints
• Fixed stable_baselines3/common/vec_env/base_vec_env.py type hints
• Fixed stable_baselines3/common/vec_env/vec_frame_stack.py type hints
• Fixed stable_baselines3/common/vec_env/dummy_vec_env.py type hints
• Fixed stable_baselines3/common/vec_env/subproc_vec_env.py type hints
• Upgraded docker images to use mamba/micromamba and CUDA 11.7
• Updated env checker to reflect what subset of Gymnasium is supported and improve GoalEnv checks
• Improve type annotation of wrappers
• Tests envs are now checked too
• Added render test for VecEnv and VecEnvWrapper
• Update issue templates and env info saved with the model
• Changed seed() method return type from List to Sequence
• Updated env checker doc and requirements for tuple spaces/goal envs

Documentation:

• Added Deep RL Course link to the Deep RL Resources page
• Added documentation about VecEnv API vs Gym API
• Upgraded tutorials to Gymnasium API
• Make it more explicit when using VecEnv vs Gym env
• Added UAV_Navigation_DRL_AirSim to the project page (@heleidsn)
• Added EvalCallback example (@sidney-tio)
• Update custom env documentation
• Added pink-noise-rl to projects page
• Fix custom policy example, ortho_init was ignored
• Added SBX page

Full Changelog: v1.8.0...v2.0.0

Stable-Baselines3 v1.8.0: Multi-env HerReplayBuffer, Open RL Benchmark, Improved env checker

Warning

Stable-Baselines3 (SB3) v1.8.0 will be the last one to use Gym as a backend.
Starting with v2.0.0, Gymnasium will be the default backend (though SB3 will have compatibility layers for Gym envs).
You can find a migration guide here.
If you want to try the SB3 v2.0 alpha version, you can take a look at PR #1327.

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo

To upgrade:

pip install stable_baselines3 sb3_contrib rl_zoo3 --upgrade

or simply (rl zoo depends on SB3 and SB3 contrib):

pip install rl_zoo3 --upgrade

Breaking Changes:

• Removed shared layers in mlp_extractor (@AlexPasqua)
• Refactored StackedObservations (it now handles dict obs, StackedDictObservations was removed)
• You must now explicitely pass a features_extractor parameter when calling extract_features()
• Dropped offline sampling for HerReplayBuffer
• As HerReplayBuffer was refactored to support multiprocessing, previous replay buffer are incompatible with this new version
• HerReplayBuffer doesn't require a max_episode_length anymore

New Features:

• Added repeat_action_probability argument in AtariWrapper.
• Only use NoopResetEnv and MaxAndSkipEnv when needed in AtariWrapper
• Added support for dict/tuple observations spaces for VecCheckNan, the check is now active in the env_checker() (@DavyMorgan)
• Added multiprocessing support for HerReplayBuffer
• HerReplayBuffer now supports all datatypes supported by ReplayBuffer
• Provide more helpful failure messages when validating the observation_space of custom gym environments using check_env (@FieteO)
• Added stats_window_size argument to control smoothing in rollout logging (@jonasreiher)

SB3-Contrib

• Added warning about potential crashes caused by check_env in the MaskablePPO docs (@AlexPasqua)
• Fixed sb3_contrib/qrdqn/*.py type hints
• Removed shared layers in mlp_extractor (@AlexPasqua)

RL Zoo

• Open RL Benchmark
• Upgraded to new HerReplayBuffer implementation that supports multiple envs
• Removed TimeFeatureWrapper for Panda and Fetch envs, as the new replay buffer should handle timeout.
• Tuned hyperparameters for RecurrentPPO on Swimmer
• Documentation is now built using Sphinx and hosted on read the doc
• Removed use_auth_token for push to hub util
• Reverted from v3 to v2 for HumanoidStandup, Reacher, InvertedPendulum and InvertedDoublePendulum since they were not part of the mujoco refactoring (see openai/gym#1304)
• Fixed gym-minigrid policy (from MlpPolicy to MultiInputPolicy)
• Replaced deprecated optuna.suggest_loguniform(...) by optuna.suggest_float(..., log=True)
• Switched to ruff and pyproject.toml
• Removed online_sampling and max_episode_length argument when using HerReplayBuffer

Bug Fixes:

• Fixed Atari wrapper that missed the reset condition (@luizapozzobon)
• Added the argument dtype (default to float32) to the noise for consistency with gym action (@sidney-tio)
• Fixed PPO train/n_updates metric not accounting for early stopping (@adamfrly)
• Fixed loading of normalized image-based environments
• Fixed DictRolloutBuffer.add with multidimensional action space (@younik)

Deprecations:

Others:

• Fixed tests/test_tensorboard.py type hint
• Fixed tests/test_vec_normalize.py type hint
• Fixed stable_baselines3/common/monitor.py type hint
• Added tests for StackedObservations
• Removed Gitlab CI file
• Moved from setup.cg to pyproject.toml configuration file
• Switched from flake8 to ruff
• Upgraded AutoROM to latest version
• Fixed stable_baselines3/dqn/*.py type hints
• Added extra_no_roms option for package installation without Atari Roms

Documentation:

• Renamed load_parameters to set_parameters (@DavyMorgan)
• Clarified documentation about subproc multiprocessing for A2C (@Bonifatius94)
• Fixed typo in A2C docstring (@AlexPasqua)
• Renamed timesteps to episodes for log_interval description (@theSquaredError)
• Removed note about gif creation for Atari games (@harveybellini)
• Added information about default network architecture
• Update information about Gymnasium support

Stable-Baselines3 v1.7.0 : non-shared features extractor, bug fixes and quality of life improvements

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo

To upgrade:

pip install stable_baselines3 sb3_contrib rl_zoo3 --upgrade

or simply (rl zoo depends on SB3 and SB3 contrib):

pip install rl_zoo3 --upgrade

Warning
Shared layers in MLP policy (mlp_extractor) are now deprecated for PPO, A2C and TRPO.
This feature will be removed in SB3 v1.8.0 and the behavior of net_arch=[64, 64]
will create separate networks with the same architecture, to be consistent with the off-policy algorithms.

Note
A2C and PPO models saved with SB3 < 1.7.0 will show a warning about
missing keys in the state dict when loaded with SB3 >= 1.7.0.
To suppress the warning, simply save the model again.
You can find more info in issue #1233

Breaking Changes:

• Removed deprecated create_eval_env, eval_env, eval_log_path, n_eval_episodes and eval_freq parameters,
  please use an EvalCallback instead
• Removed deprecated sde_net_arch parameter
• Removed ret attributes in VecNormalize, please use returns instead
• VecNormalize now updates the observation space when normalizing images

New Features:

• Introduced mypy type checking
• Added option to have non-shared features extractor between actor and critic in on-policy algorithms (@AlexPasqua)
• Added with_bias argument to create_mlp
• Added support for multidimensional spaces.MultiBinary observations
• Features extractors now properly support unnormalized image-like observations (3D tensor)
  when passing normalize_images=False
• Added normalized_image parameter to NatureCNN and CombinedExtractor
• Added support for Python 3.10

SB3-Contrib

• Fixed a bug in RecurrentPPO where the lstm states where incorrectly reshaped for n_lstm_layers > 1 (thanks @kolbytn)
• Fixed RuntimeError: rnn: hx is not contiguous while predicting terminal values for RecurrentPPO when n_lstm_layers > 1

RL Zoo

• Added support for python file for configuration
• Added monitor_kwargs parameter

Bug Fixes:

• Fixed ProgressBarCallback under-reporting (@dominicgkerr)
• Fixed return type of evaluate_actions in ActorCritcPolicy to reflect that entropy is an optional tensor (@Rocamonde)
• Fixed type annotation of policy in BaseAlgorithm and OffPolicyAlgorithm
• Allowed model trained with Python 3.7 to be loaded with Python 3.8+ without the custom_objects workaround
• Raise an error when the same gym environment instance is passed as separate environments when creating a vectorized environment with more than one environment. (@Rocamonde)
• Fix type annotation of model in evaluate_policy
• Fixed Self return type using TypeVar
• Fixed the env checker, the key was not passed when checking images from Dict observation space
• Fixed normalize_images which was not passed to parent class in some cases
• Fixed load_from_vector that was broken with newer PyTorch version when passing PyTorch tensor

Deprecations:

• You should now explicitely pass a features_extractor parameter when calling extract_features()
• Deprecated shared layers in MlpExtractor (@AlexPasqua)

Others:

• Used issue forms instead of issue templates
• Updated the PR template to associate each PR with its peer in RL-Zoo3 and SB3-Contrib
• Fixed flake8 config to be compatible with flake8 6+
• Goal-conditioned environments are now characterized by the availability of the compute_reward method, rather than by their inheritance to gym.GoalEnv
• Replaced CartPole-v0 by CartPole-v1 is tests
• Fixed tests/test_distributions.py type hints
• Fixed stable_baselines3/common/type_aliases.py type hints
• Fixed stable_baselines3/common/torch_layers.py type hints
• Fixed stable_baselines3/common/env_util.py type hints
• Fixed stable_baselines3/common/preprocessing.py type hints
• Fixed stable_baselines3/common/atari_wrappers.py type hints
• Fixed stable_baselines3/common/vec_env/vec_check_nan.py type hints
• Exposed modules in __init__.py with the __all__ attribute (@ZikangXiong)
• Upgraded GitHub CI/setup-python to v4 and checkout to v3
• Set tensors construction directly on the device (~8% speed boost on GPU)
• Monkey-patched np.bool = bool so gym 0.21 is compatible with NumPy 1.24+
• Standardized the use of from gym import spaces
• Modified get_system_info to avoid issue linked to copy-pasting on GitHub issue

Documentation:

• Updated Hugging Face Integration page (@simoninithomas)
• Changed env to vec_env when environment is vectorized
• Updated custom policy docs to better explain the mlp_extractor's dimensions (@AlexPasqua)
• Updated custom policy documentation (@athatheo)
• Improved tensorboard callback doc
• Clarify doc when using image-like input
• Added RLeXplore to the project page (@yuanmingqi)

SB3 v1.6.2: Progress bar and RL Zoo3 package

SB3 Contrib: https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3: https://github.com/DLR-RM/rl-baselines3-zoo

New Features:

• Added progress_bar argument in the learn() method, displayed using TQDM and rich packages
• Added progress bar callback

RL Zoo3

• The RL Zoo can now be installed as a package (pip install rl_zoo3)

Bug Fixes:

• self.num_timesteps was initialized properly only after the first call to on_step() for callbacks
• Set importlib-metadata version to ~=4.13 to be compatible with gym=0.21

Deprecations:

• Added deprecation warning if parameters eval_env, eval_freq or create_eval_env are used (see #925) (@tobirohrer)

Others:

• Fixed type hint of the env_id parameter in make_vec_env and make_atari_env (@AlexPasqua)

Documentation:

• Extended docstring of the wrapper_class parameter in make_vec_env (@AlexPasqua)

SB3 v1.6.1: Bug fix release

SB3 Contrib: https://github.com/Stable-Baselines-Team/stable-baselines3-contrib

Breaking Changes:

• Switched minimum tensorboard version to 2.9.1

New Features:

• Support logging hyperparameters to tensorboard (@timothe-chaumont)
• Added checkpoints for replay buffer and VecNormalize statistics (@anand-bala)
• Added option for Monitor to append to existing file instead of overriding (@sidney-tio)
• The env checker now raises an error when using dict observation spaces and observation keys don't match observation space keys

SB3-Contrib

• Fixed the issue of wrongly passing policy arguments when using CnnLstmPolicy or MultiInputLstmPolicy with RecurrentPPO (@mlodel)

Bug Fixes:

• Fixed issue where PPO gives NaN if rollout buffer provides a batch of size 1 (@hughperkins)
• Fixed the issue that predict does not always return action as np.ndarray (@qgallouedec)
• Fixed division by zero error when computing FPS when a small number of time has elapsed in operating systems with low-precision timers.
• Added multidimensional action space support (@qgallouedec)
• Fixed missing verbose parameter passing in the EvalCallback constructor (@BurakDmb)
• Fixed the issue that when updating the target network in DQN, SAC, TD3, the running_mean and running_var properties of batch norm layers are not updated (@honglu2875)
• Fixed incorrect type annotation of the replay_buffer_class argument in common.OffPolicyAlgorithm initializer, where an instance instead of a class was required (@Rocamonde)
• Fixed loading saved model with different number of envrionments
• Removed forward() abstract method declaration from common.policies.BaseModel (already defined in torch.nn.Module) to fix type errors in subclasses (@Rocamonde)
• Fixed the return type of .load() and .learn() methods in BaseAlgorithm so that they now use TypeVar (@Rocamonde)
• Fixed an issue where keys with different tags but the same key raised an error in common.logger.HumanOutputFormat (@Rocamonde and @AdamGleave)

Others:

• Fixed DictReplayBuffer.next_observations typing (@qgallouedec)
• Added support for device="auto" in buffers and made it default (@qgallouedec)
• Updated ResultsWriter` (used internally by Monitorwrapper) to automatically create missing directories whenfilename`` is a path (@dominicgkerr)

Documentation:

• Added an example of callback that logs hyperparameters to tensorboard. (@timothe-chaumont)
• Fixed typo in docstring "nature" -> "Nature" (@Melanol)
• Added info on split tensorboard logs into (@Melanol)
• Fixed typo in ppo doc (@francescoluciano)
• Fixed typo in install doc(@jlp-ue)
• Clarified and standardized verbosity documentation
• Added link to a GitHub issue in the custom policy documentation (@AlexPasqua)
• Fixed typos (@Akhilez)