Merge branch 'master' into feat/mps-support

.github/workflows/ci.yml

@@ -31,18 +31,21 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
+          # Use uv for faster downloads
+          pip install uv
           # cpu version of pytorch
-          pip install torch==2.3.1 --index-url https://download.pytorch.org/whl/cpu
+          # See https://github.com/astral-sh/uv/issues/1497
+          uv pip install --system torch==2.3.1+cpu --index https://download.pytorch.org/whl/cpu
 
           # Install Atari Roms
-          pip install autorom
+          uv pip install --system autorom
           wget https://gist.githubusercontent.com/jjshoots/61b22aefce4456920ba99f2c36906eda/raw/00046ac3403768bfe45857610a3d333b8e35e026/Roms.tar.gz.b64
           base64 Roms.tar.gz.b64 --decode &> Roms.tar.gz
           AutoROM --accept-license --source-file Roms.tar.gz
 
-          pip install .[extra_no_roms,tests,docs]
+          uv pip install --system .[extra_no_roms,tests,docs]
           # Use headless version
-          pip install opencv-python-headless
+          uv pip install --system opencv-python-headless
       - name: Lint with ruff
         run: |
           make lint

.readthedocs.yml

@@ -16,6 +16,6 @@ conda:
   environment: docs/conda_env.yml
 
 build:
-  os: ubuntu-22.04
+  os: ubuntu-24.04
   tools:
-    python: "mambaforge-22.9"
+    python: "mambaforge-23.11"

CONTRIBUTING.md

@@ -6,7 +6,7 @@ into two categories:
     - Create an issue about your intended feature, and we shall discuss the design and
     implementation. Once we agree that the plan looks good, go ahead and implement it.
 2. You want to implement a feature or bug-fix for an outstanding issue
-    - Look at the outstanding issues here: https://github.com/DLR-RM/stable-baselines3/issues
+    - Look at the outstanding issues here: https://github.com/DLR-RM/stable-baselines3/labels/help%20wanted
     - Pick an issue or feature and comment on the task that you want to work on this feature.
     - If you need more context on a particular issue, please ask, and we shall provide.
 

README.md

@@ -1,6 +1,6 @@
 <!-- [![pipeline status](https://gitlab.com/araffin/stable-baselines3/badges/master/pipeline.svg)](https://gitlab.com/araffin/stable-baselines3/-/commits/master) -->
-![CI](https://github.com/DLR-RM/stable-baselines3/workflows/CI/badge.svg)
-[![Documentation Status](https://readthedocs.org/projects/stable-baselines/badge/?version=master)](https://stable-baselines3.readthedocs.io/en/master/?badge=master) [![coverage report](https://gitlab.com/araffin/stable-baselines3/badges/master/coverage.svg)](https://gitlab.com/araffin/stable-baselines3/-/commits/master)
+[![CI](https://github.com/DLR-RM/stable-baselines3/workflows/CI/badge.svg)](https://github.com/DLR-RM/stable-baselines3/actions/workflows/ci.yml)
+[![Documentation Status](https://readthedocs.org/projects/stable-baselines/badge/?version=master)](https://stable-baselines3.readthedocs.io/en/master/?badge=master) [![coverage report](https://gitlab.com/araffin/stable-baselines3/badges/master/coverage.svg)](https://github.com/DLR-RM/stable-baselines3/actions/workflows/ci.yml)
 [![codestyle](https://img.shields.io/badge/code%20style-black-000000.svg)](https://github.com/psf/black)
 
 
@@ -22,6 +22,8 @@ These algorithms will make it easier for the research community and industry to
 **The performance of each algorithm was tested** (see *Results* section in their respective page),
 you can take a look at the issues [#48](https://github.com/DLR-RM/stable-baselines3/issues/48) and [#49](https://github.com/DLR-RM/stable-baselines3/issues/49) for more details.
 
+We also provide detailed logs and reports on the [OpenRL Benchmark](https://wandb.ai/openrlbenchmark/sb3) platform.
+
 
 | **Features**                | **Stable-Baselines3** |
 | --------------------------- | ----------------------|
@@ -41,7 +43,13 @@ you can take a look at the issues [#48](https://github.com/DLR-RM/stable-baselin
 
 ### Planned features
 
-Please take a look at the [Roadmap](https://github.com/DLR-RM/stable-baselines3/issues/1) and [Milestones](https://github.com/DLR-RM/stable-baselines3/milestones).
+Since most of the features from the [original roadmap](https://github.com/DLR-RM/stable-baselines3/issues/1) have been implemented, there are no major changes planned for SB3, it is now *stable*.
+If you want to contribute, you can search in the issues for the ones where [help is welcomed](https://github.com/DLR-RM/stable-baselines3/labels/help%20wanted) and the other [proposed enhancements](https://github.com/DLR-RM/stable-baselines3/labels/enhancement).
+
+While SB3 development is now focused on bug fixes and maintenance (doc update, user experience, ...), there is more active development going on in the associated repositories:
+- newer algorithms are regularly added to the [SB3 Contrib](https://github.com/Stable-Baselines-Team/stable-baselines3-contrib) repository
+- faster variants are developed in the [SBX (SB3 + Jax)](https://github.com/araffin/sbx) repository
+- the training framework for SB3, the RL Zoo, has an active [roadmap](https://github.com/DLR-RM/rl-baselines3-zoo/issues/299)
 
 ## Migration guide: from Stable-Baselines (SB2) to Stable-Baselines3 (SB3)
 
@@ -79,7 +87,7 @@ Documentation: https://rl-baselines3-zoo.readthedocs.io/en/master/
 
 We implement experimental features in a separate contrib repository: [SB3-Contrib](https://github.com/Stable-Baselines-Team/stable-baselines3-contrib)
 
-This allows SB3 to maintain a stable and compact core, while still providing the latest features, like Recurrent PPO (PPO LSTM), Truncated Quantile Critics (TQC), Quantile Regression DQN (QR-DQN) or PPO with invalid action masking (Maskable PPO).
+This allows SB3 to maintain a stable and compact core, while still providing the latest features, like Recurrent PPO (PPO LSTM), CrossQ, Truncated Quantile Critics (TQC), Quantile Regression DQN (QR-DQN) or PPO with invalid action masking (Maskable PPO).
 
 Documentation is available online: [https://sb3-contrib.readthedocs.io/](https://sb3-contrib.readthedocs.io/)
 
@@ -97,17 +105,16 @@ It provides a minimal number of features compared to SB3 but can be much faster
 ### Prerequisites
 Stable Baselines3 requires Python 3.8+.
 
-#### Windows 10
+#### Windows
 
 To install stable-baselines on Windows, please look at the [documentation](https://stable-baselines3.readthedocs.io/en/master/guide/install.html#prerequisites).
 
 
 ### Install using pip
 Install the Stable Baselines3 package:
+```sh
+pip install 'stable-baselines3[extra]'
 ```
-pip install stable-baselines3[extra]
-```
-**Note:** Some shells such as Zsh require quotation marks around brackets, i.e. `pip install 'stable-baselines3[extra]'` ([More Info](https://stackoverflow.com/a/30539963)).
 
 This includes an optional dependencies like Tensorboard, OpenCV or `ale-py` to train on atari games. If you do not need those, you can use:
 ```sh
@@ -177,6 +184,7 @@ All the following examples can be executed online using Google Colab notebooks:
 | ------------------- | ------------------ | ------------------ | ------------------ | ------------------- | ------------------ | --------------------------------- |
 | ARS<sup>[1](#f1)</sup>   | :x: | :heavy_check_mark: | :heavy_check_mark: | :x: | :x: | :heavy_check_mark: |
 | A2C   | :x: | :heavy_check_mark: | :heavy_check_mark: | :heavy_check_mark: | :heavy_check_mark: | :heavy_check_mark: |
+| CrossQ<sup>[1](#f1)</sup>   | :x: | :heavy_check_mark: | :x:                | :x:                 | :x:                | :heavy_check_mark: |
 | DDPG  | :x: | :heavy_check_mark: | :x:                | :x:                 | :x:                | :heavy_check_mark: |
 | DQN   | :x: | :x: | :heavy_check_mark: | :x:                 | :x:                | :heavy_check_mark: |
 | HER   | :x: | :heavy_check_mark: | :heavy_check_mark: | :x: | :x: | :heavy_check_mark: |
@@ -191,7 +199,7 @@ All the following examples can be executed online using Google Colab notebooks:
 
 <b id="f1">1</b>: Implemented in [SB3 Contrib](https://github.com/Stable-Baselines-Team/stable-baselines3-contrib) GitHub repository.
 
-Actions `gym.spaces`:
+Actions `gymnasium.spaces`:
  * `Box`: A N-dimensional box that contains every point in the action space.
  * `Discrete`: A list of possible actions, where each timestep only one of the actions can be used.
  * `MultiDiscrete`: A list of possible actions, where each timestep only one action of each discrete set can be used.
@@ -218,9 +226,9 @@ To run a single test:
 python3 -m pytest -v -k 'test_check_env_dict_action'
 ```
 
-You can also do a static type check using `pytype` and `mypy`:
+You can also do a static type check using `mypy`:
 ```sh
-pip install pytype mypy
+pip install mypy
 make type
 ```
 
@@ -252,6 +260,8 @@ To cite this repository in publications:
 }
 ```
 
+Note: If you need to refer to a specific version of SB3, you can also use the [Zenodo DOI](https://doi.org/10.5281/zenodo.8123988).
+
 ## Maintainers
 
 Stable-Baselines3 is currently maintained by [Ashley Hill](https://github.com/hill-a) (aka @hill-a), [Antonin Raffin](https://araffin.github.io/) (aka [@araffin](https://github.com/araffin)), [Maximilian Ernestus](https://github.com/ernestum) (aka @ernestum), [Adam Gleave](https://github.com/adamgleave) (@AdamGleave), [Anssi Kanervisto](https://github.com/Miffyli) (@Miffyli) and [Quentin Gallouédec](https://gallouedec.com/) (@qgallouedec).

docs/conda_env.yml

@@ -1,18 +1,18 @@
 name: root
 channels:
   - pytorch
-  - defaults
+  - conda-forge
 dependencies:
   - cpuonly=1.0=0
-  - pip=22.3.1
-  - python=3.8
-  - pytorch=1.13.0=py3.8_cpu_0
+  - pip=24.2
+  - python=3.11
+  - pytorch=2.5.0=py3.11_cpu_0
   - pip:
-    - gymnasium
+    - gymnasium>=0.28.1,<0.30
     - cloudpickle
     - opencv-python-headless
     - pandas
-    - numpy
+    - numpy>=1.20,<2.0
     - matplotlib
     - sphinx>=5,<8
     - sphinx_rtd_theme>=1.3.0

docs/guide/algos.rst

@@ -10,6 +10,7 @@ Name                 ``Box``     ``Discrete`` ``MultiDiscrete`` ``MultiBinary``
 ===================  =========== ============ ================= =============== ================
 ARS [#f1]_           ✔️           ✔️            ❌                 ❌              ✔️
 A2C                  ✔️           ✔️            ✔️                 ✔️               ✔️
+CrossQ [#f1]_        ✔️           ❌            ❌                ❌               ✔️
 DDPG                 ✔️           ❌            ❌                ❌               ✔️
 DQN                  ❌           ✔️            ❌                ❌               ✔️
 HER                  ✔️           ✔️            ❌                ❌               ✔️

docs/guide/sb3_contrib.rst

@@ -42,6 +42,7 @@ See documentation for the full list of included features.
 - `PPO with recurrent policy (RecurrentPPO aka PPO LSTM) <https://ppo-details.cleanrl.dev//2021/11/05/ppo-implementation-details/>`_
 - `Truncated Quantile Critics (TQC)`_
 - `Trust Region Policy Optimization (TRPO) <https://arxiv.org/abs/1502.05477>`_
+- `Batch Normalization in Deep Reinforcement Learning (CrossQ) <https://openreview.net/forum?id=PczQtTsTIX>`_
 
 
 **Gym Wrappers**:

docs/index.rst

@@ -113,12 +113,14 @@ To cite this project in publications:
     url     = {http://jmlr.org/papers/v22/20-1364.html}
   }
 
+Note: If you need to refer to a specific version of SB3, you can also use the `Zenodo DOI <https://doi.org/10.5281/zenodo.8123988>`_.
+
 Contributing
 ------------
 
 To any interested in making the rl baselines better, there are still some improvements
 that need to be done.
-You can check issues in the `repo <https://github.com/DLR-RM/stable-baselines3/issues>`_.
+You can check issues in the `repository <https://github.com/DLR-RM/stable-baselines3/labels/help%20wanted>`_.
 
 If you want to contribute, please read `CONTRIBUTING.md <https://github.com/DLR-RM/stable-baselines3/blob/master/CONTRIBUTING.md>`_ first.
 

docs/misc/changelog.rst

@@ -3,9 +3,11 @@
 Changelog
 ==========
 
-Release 2.4.0a9 (WIP)
+Release 2.4.0a10 (WIP)
 --------------------------
 
+**New algorithm: CrossQ in SB3 Contrib**
+
 .. note::
 
   DQN (and QR-DQN) models saved with SB3 < 2.4.0 will show a warning about
@@ -43,6 +45,10 @@ Bug Fixes:
 
 `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`_
 ^^^^^^^^^
@@ -60,12 +66,17 @@ 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
 
 Bug Fixes:
 ^^^^^^^^^^
 
 Documentation:
 ^^^^^^^^^^^^^^
+- Updated PPO doc to recommend using CPU with ``MlpPolicy``
+- Clarified documentation about planned features and citing software
 
 Release 2.3.2 (2024-04-27)
 --------------------------

docs/modules/ppo.rst

@@ -88,6 +88,23 @@ Train a PPO agent on ``CartPole-v1`` using 4 environments.
       vec_env.render("human")
 
 
+.. note::
+
+  PPO is meant to be run primarily on the CPU, especially when you are not using a CNN. To improve CPU utilization, try turning off the GPU and using ``SubprocVecEnv`` instead of the default ``DummyVecEnv``:
+
+  .. code-block::
+
+    from stable_baselines3 import PPO
+    from stable_baselines3.common.env_util import make_vec_env
+    from stable_baselines3.common.vec_env import SubprocVecEnv
+
+    if __name__=="__main__":
+        env = make_vec_env("CartPole-v1", n_envs=8, vec_env_cls=SubprocVecEnv)
+        model = PPO("MlpPolicy", env, device="cpu")
+        model.learn(total_timesteps=25_000)
+  
+  For more information, see :ref:`Vectorized Environments <vec_env>`, `Issue #1245 <https://github.com/DLR-RM/stable-baselines3/issues/1245#issuecomment-1435766949>`_ or the `Multiprocessing notebook <https://colab.research.google.com/github/Stable-Baselines-Team/rl-colab-notebooks/blob/sb3/multiprocessing_rl.ipynb>`_.
+
 Results
 -------
 

stable_baselines3/common/on_policy_algorithm.py

@@ -1,5 +1,6 @@
 import sys
 import time
+import warnings
 from typing import Any, Dict, List, Optional, Tuple, Type, TypeVar, Union
 
 import numpy as np
@@ -135,6 +136,28 @@ def _setup_model(self) -> None:
             self.observation_space, self.action_space, self.lr_schedule, use_sde=self.use_sde, **self.policy_kwargs
         )
         self.policy = self.policy.to(self.device)
+        # Warn when not using CPU with MlpPolicy
+        self._maybe_recommend_cpu()
+
+    def _maybe_recommend_cpu(self, mlp_class_name: str = "ActorCriticPolicy") -> None:
+        """
+        Recommend to use CPU only when using A2C/PPO with MlpPolicy.
+
+        :param: The name of the class for the default MlpPolicy.
+        """
+        policy_class_name = self.policy_class.__name__
+        if self.device != th.device("cpu") and policy_class_name == mlp_class_name:
+            warnings.warn(
+                f"You are trying to run {self.__class__.__name__} on the GPU, "
+                "but it is primarily intended to run on the CPU when not using a CNN policy "
+                f"(you are using {policy_class_name} which should be a MlpPolicy). "
+                "See https://github.com/DLR-RM/stable-baselines3/issues/1245 "
+                "for more info. "
+                "You can pass `device='cpu'` or `export CUDA_VISIBLE_DEVICES=` to force using the CPU."
+                "Note: The model will train, but the GPU utilization will be poor and "
+                "the training might take longer than on CPU.",
+                UserWarning,
+            )
 
     def collect_rollouts(
         self,

stable_baselines3/common/utils.py

@@ -46,7 +46,7 @@ def set_random_seed(seed: int, using_cuda: bool = False) -> None:
 
 
 # From stable baselines
-def explained_variance(y_pred: np.ndarray, y_true: np.ndarray) -> np.ndarray:
+def explained_variance(y_pred: np.ndarray, y_true: np.ndarray) -> float:
     """
     Computes fraction of variance that ypred explains about y.
     Returns 1 - Var[y-ypred] / Var[y]
@@ -62,7 +62,7 @@ def explained_variance(y_pred: np.ndarray, y_true: np.ndarray) -> np.ndarray:
     """
     assert y_true.ndim == 1 and y_pred.ndim == 1
     var_y = np.var(y_true)
-    return np.nan if var_y == 0 else 1 - np.var(y_true - y_pred) / var_y
+    return np.nan if var_y == 0 else float(1 - np.var(y_true - y_pred) / var_y)
 
 
 def update_learning_rate(optimizer: th.optim.Optimizer, learning_rate: float) -> None:

stable_baselines3/version.txt

@@ -1 +1 @@
-2.4.0a9
+2.4.0a10

tests/test_run.py

@@ -1,6 +1,7 @@
 import gymnasium as gym
 import numpy as np
 import pytest
+import torch as th
 
 from stable_baselines3 import A2C, DDPG, DQN, PPO, SAC, TD3
 from stable_baselines3.common.env_util import make_vec_env
@@ -211,8 +212,11 @@ def test_warn_dqn_multi_env():
 
 
 def test_ppo_warnings():
-    """Test that PPO warns and errors correctly on
-    problematic rollout buffer sizes"""
+    """
+    Test that PPO warns and errors correctly on
+    problematic rollout buffer sizes,
+    and recommend using CPU.
+    """
 
     # Only 1 step: advantage normalization will return NaN
     with pytest.raises(AssertionError):
@@ -234,3 +238,9 @@ def test_ppo_warnings():
     loss = model.logger.name_to_value["train/loss"]
     assert loss > 0
     assert not np.isnan(loss)  # check not nan (since nan does not equal nan)
+
+    with pytest.warns(UserWarning, match="You are trying to run PPO on the GPU"):
+        model = PPO("MlpPolicy", "Pendulum-v1")
+        # Pretend to be on the GPU
+        model.device = th.device("cuda")
+        model._maybe_recommend_cpu()

tests/test_utils.py

@@ -177,7 +177,7 @@ def test_custom_vec_env(tmp_path):
 
 
 @pytest.mark.parametrize("direct_policy", [False, True])
-def test_evaluate_policy(direct_policy: bool):
+def test_evaluate_policy(direct_policy):
     model = A2C("MlpPolicy", "Pendulum-v1", seed=0)
     n_steps_per_episode, n_eval_episodes = 200, 2