Merge branch 'master' into feat/mps-support

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -67,7 +67,7 @@ body:
           required: true
         - label: I have read the [documentation](https://stable-baselines3.readthedocs.io/en/master/)
           required: true
-        - label: I have provided a minimal working example to reproduce the bug
+        - label: I have provided a [minimal and working](https://github.com/DLR-RM/stable-baselines3/issues/982#issuecomment-1197044014) example to reproduce the bug
           required: true
         - label: I've used the [markdown code blocks](https://help.github.com/en/articles/creating-and-highlighting-code-blocks) for both code and stack traces.
           required: true

.github/ISSUE_TEMPLATE/custom_env.yml

@@ -28,7 +28,7 @@ body:
     attributes:
       label: Code example
       description: |
-        Please try to provide a minimal example to reproduce the bug.
+        Please try to provide a [minimal example](https://github.com/DLR-RM/stable-baselines3/issues/982#issuecomment-1197044014) to reproduce the bug.
         For a custom environment, you need to give at least the observation space, action space, `reset()` and `step()` methods (see working example below).
         Error messages and stack traces are also helpful.
         Please use the [markdown code blocks](https://help.github.com/en/articles/creating-and-highlighting-code-blocks) for both code and stack traces.
@@ -101,7 +101,7 @@ body:
           required: true
         - label: I have read the [documentation](https://stable-baselines3.readthedocs.io/en/master/)
           required: true
-        - label: I have provided a minimal working example to reproduce the bug
+        - label: I have provided a [minimal and working](https://github.com/DLR-RM/stable-baselines3/issues/982#issuecomment-1197044014) example to reproduce the bug
           required: true
         - label: I have checked my env using the env checker
           required: true

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -42,3 +42,5 @@ body:
       options:
         - label: I have checked that there is no similar [issue](https://github.com/DLR-RM/stable-baselines3/issues) in the repo
           required: true
+        - label: If I'm requesting a new feature, I have proposed alternatives
+          required: true

.github/ISSUE_TEMPLATE/question.yml

@@ -26,7 +26,7 @@ body:
           required: true
         - label: I have read the [documentation](https://stable-baselines3.readthedocs.io/en/master/)
           required: true
-        - label: If code there is, it is minimal and working
+        - label: If code there is, it is [minimal and working](https://github.com/DLR-RM/stable-baselines3/issues/982#issuecomment-1197044014)
           required: true
         - label: If code there is, it is formatted using the [markdown code blocks](https://help.github.com/en/articles/creating-and-highlighting-code-blocks) for both code and stack traces.
           required: true

.github/workflows/ci.yml

@@ -20,7 +20,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10"]
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
 
     steps:
     - uses: actions/checkout@v3
@@ -32,7 +32,7 @@ jobs:
       run: |
         python -m pip install --upgrade pip
         # cpu version of pytorch
-        pip install torch==1.11+cpu -f https://download.pytorch.org/whl/torch_stable.html
+        pip install torch==1.13+cpu -f https://download.pytorch.org/whl/torch_stable.html
 
         # Install Atari Roms
         pip install autorom
@@ -55,8 +55,8 @@ jobs:
     - name: Type check
       run: |
         make type
-      # skip mypy type check for python3.7 (result is different to all other versions)
+      # skip PyType, doesn't support 3.11 yet
-      if: "!(matrix.python-version == '3.7')"
+      if: "!(matrix.python-version == '3.11')"
     - name: Test with pytest
       run: |
         make pytest

README.md

@@ -92,10 +92,10 @@ It provides a minimal number of features compared to SB3 but can be much faster
 
 ## Installation
 
-**Note:** Stable-Baselines3 supports PyTorch >= 1.11
+**Note:** Stable-Baselines3 supports PyTorch >= 1.13
 
 ### Prerequisites
-Stable Baselines3 requires Python 3.7+.
+Stable Baselines3 requires Python 3.8+.
 
 #### Windows 10
 

docs/README.md

@@ -8,7 +8,7 @@ This folder contains documentation for the RL baselines.
 #### Install Sphinx and Theme
 Execute this command in the project root:
 ```
-pip install -e .[docs]
+pip install -e ".[docs]"
 ```
 
 #### Building the Docs

docs/conda_env.yml

@@ -5,8 +5,8 @@ channels:
 dependencies:
   - cpuonly=1.0=0
   - pip=22.3.1
-  - python=3.7
+  - python=3.8
-  - pytorch=1.11.0=py3.7_cpu_0
+  - pytorch=1.13.0=py3.8_cpu_0
   - pip:
     - gymnasium
     - cloudpickle
@@ -15,6 +15,6 @@ dependencies:
     - numpy
     - matplotlib
     - sphinx_autodoc_typehints
-    - sphinx>=4.2
+    - sphinx>=5.3,<7.0
     - sphinx_rtd_theme>=1.0
     - sphinx_copybutton

docs/guide/custom_policy.rst

@@ -215,7 +215,7 @@ downsampling and "vector" with a single linear layer.
   from stable_baselines3.common.torch_layers import BaseFeaturesExtractor
 
   class CustomCombinedExtractor(BaseFeaturesExtractor):
-      def __init__(self, observation_space: spaces.Dict):
+      def __init__(self, observation_space: gym.spaces.Dict):
           # We do not know features-dim here before going over all the items,
           # so put something dummy for now. PyTorch requires calling
           # nn.Module.__init__ before adding modules

docs/guide/examples.rst

@@ -319,7 +319,7 @@ You can control the evaluation frequency with ``eval_freq`` to monitor your agen
 
   from stable_baselines3 import SAC
   from stable_baselines3.common.callbacks import EvalCallback
-  from stable-baselines3.common.env_util import make_vec_env
+  from stable_baselines3.common.env_util import make_vec_env
 
   env_id = "Pendulum-v1"
   n_training_envs = 1
@@ -330,7 +330,7 @@ You can control the evaluation frequency with ``eval_freq`` to monitor your agen
   os.makedirs(eval_log_dir, exist_ok=True)
 
   # Initialize a vectorized training environment with default parameters
-  train_env = make_vec_env(env_id, n_env=n_training_envs, seed=0)
+  train_env = make_vec_env(env_id, n_envs=n_training_envs, seed=0)
 
   # Separate evaluation env, with different parameters passed via env_kwargs
   # Eval environments can be vectorized to speed up evaluation.

docs/guide/install.rst

@@ -7,7 +7,7 @@ Installation
 Prerequisites
 -------------
 
-Stable-Baselines3 requires python 3.7+ and PyTorch >= 1.11
+Stable-Baselines3 requires python 3.8+ and PyTorch >= 1.13
 
 Windows 10
 ~~~~~~~~~~

docs/misc/changelog.rst

@@ -3,6 +3,49 @@
 Changelog
 ==========
 
+
+Release 2.1.0a4 (WIP)
+--------------------------
+
+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`_
+^^^^^^^^^^^^^^
+
+`RL Zoo`_
+^^^^^^^^^
+
+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)
+ 
+Deprecations:
+^^^^^^^^^^^^^
+
+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](https://github.com/AlexPasqua/DeepNetSlice) to community projects (@AlexPasqua)
+
+
 Release 2.0.0 (2023-06-22)
 --------------------------
 
@@ -1359,8 +1402,8 @@ And all the contributors:
 @eleurent @ac-93 @cove9988 @theDebugger811 @hsuehch @Demetrio92 @thomasgubler @IperGiove @ScheiklP
 @simoninithomas @armandpl @manuel-delverme @Gautam-J @gianlucadecola @buoyancy99 @caburu @xy9485
 @Gregwar @ycheng517 @quantitative-technologies @bcollazo @git-thor @TibiGG @cool-RR @MWeltevrede
-@carlosluis @arjun-kg @tlpss @JonathanKuelz
+@carlosluis @arjun-kg @tlpss @JonathanKuelz @Gabo-Tor
 @Melanol @qgallouedec @francescoluciano @jlp-ue @burakdmb @timothe-chaumont @honglu2875
 @anand-bala @hughperkins @sidney-tio @AlexPasqua @dominicgkerr @Akhilez @Rocamonde @tobirohrer @ZikangXiong
 @DavyMorgan @luizapozzobon @Bonifatius94 @theSquaredError @harveybellini @DavyMorgan @FieteO @jonasreiher @npit @WeberSamuel @troiganto
-@lutogniew @lbergmann1
+@lutogniew @lbergmann1 @lukashass @BertrandDecoster @pseudo-rnd-thoughts @stefanbschneider @kyle-he

docs/misc/projects.rst

@@ -197,3 +197,35 @@ A simple library for pink noise exploration with deterministic (DDPG / TD3) and
 | Authors: Onno Eberhard, Jakob Hollenstein, Cristina Pinneri, Georg Martius
 | Github: https://github.com/martius-lab/pink-noise-rl
 | Paper: https://openreview.net/forum?id=hQ9V5QN27eS (Oral at ICLR 2023)
+
+
+mobile-env
+----------
+
+An open, minimalist Gymnasium environment for autonomous coordination in wireless mobile networks.
+It allows simulating various scenarios with moving users in a cellular network with multiple base stations.
+
+- Written in pure Python, easy to modify and extend, and can be installed directly via PyPI.
+- Implements the standard Gymnasium interface such that it can be used with all common frameworks for reinforcement learning.
+- There are examples for both single-agent and multi-agent RL using either `stable-baselines3` or Ray RLlib.
+
+| Authors: Stefan Schneider, Stefan Werner
+| Github: https://github.com/stefanbschneider/mobile-env
+| Paper: https://ris.uni-paderborn.de/download/30236/30237 (2022 IEEE/IFIP Network Operations and Management Symposium (NOMS))
+
+
+DeepNetSlice
+------------
+
+A Deep Reinforcement Learning Open-Source Toolkit for Network Slice Placement (NSP).
+
+NSP is the problem of deciding which physical servers in a network should host the virtual network functions (VNFs) that make up a network slice, as well as managing the mapping of the virtual links between the VNFs onto the physical infrastructure.
+It is a complex optimization problem, as it involves considering the requirements of the network slice and the available resources on the physical network.
+The goal is generally to maximize the utilization of the physical resources while ensuring that the network slices meet their performance requirements.
+
+The toolkit includes a customizable simulation environments, as well as some ready-to-use demos for training
+intelligent agents to perform network slice placement.
+
+| Author: Alex Pasquali
+| Github: https://github.com/AlexPasqua/DeepNetSlice
+| Paper: **under review** (citation instructions on the project's README.md) -> see this Master's Thesis for the moment: https://etd.adm.unipi.it/theses/available/etd-01182023-110038/unrestricted/Tesi_magistrale_Pasquali_Alex.pdf

pyproject.toml

@@ -1,8 +1,8 @@
 [tool.ruff]
 # Same as Black.
 line-length = 127
-# Assume Python 3.7
+# Assume Python 3.8
-target-version = "py37"
+target-version = "py38"
 # See https://beta.ruff.rs/docs/rules/
 select = ["E", "F", "B", "UP", "C90", "RUF"]
 # B028: Ignore explicit stacklevel`

setup.py

@@ -76,9 +76,7 @@ model = PPO("MlpPolicy", "CartPole-v1").learn(10_000)
 extra_no_roms = [
     # For render
     "opencv-python",
-    'pygame; python_version >= "3.8.0"',
+    "pygame",
-    # See https://github.com/pygame/pygame/issues/3572
-    'pygame>=2.0,<2.1.3; python_version < "3.8.0"',
     # Tensorboard support
     "tensorboard>=2.9.1",
     # Checking memory taken by replay buffer
@@ -87,13 +85,13 @@ extra_no_roms = [
     "tqdm",
     "rich",
     # For atari games,
-    "shimmy[atari]~=0.2.1",
+    "shimmy[atari]~=1.1.0",
     "pillow",
 ]
 
 extra_packages = extra_no_roms + [  # noqa: RUF005
     # For atari roms,
-    "autorom[accept-rom-license]~=0.6.0",
+    "autorom[accept-rom-license]~=0.6.1",
 ]
 
 
@@ -102,10 +100,9 @@ setup(
     packages=[package for package in find_packages() if package.startswith("stable_baselines3")],
     package_data={"stable_baselines3": ["py.typed", "version.txt"]},
     install_requires=[
-        "gymnasium==0.28.1",
+        "gymnasium>=0.28.1,<0.30",
         "numpy>=1.20",
-        "torch>=1.11",
+        "torch>=1.13",
-        'typing_extensions>=4.0,<5; python_version < "3.8.0"',
         # For saving models
         "cloudpickle",
         # For reading logs
@@ -154,7 +151,7 @@ setup(
     long_description=long_description,
     long_description_content_type="text/markdown",
     version=__version__,
-    python_requires=">=3.7",
+    python_requires=">=3.8",
     # PyPI package information.
     project_urls={
         "Code": "https://github.com/DLR-RM/stable-baselines3",
@@ -166,10 +163,10 @@ setup(
     },
     classifiers=[
         "Programming Language :: Python :: 3",
-        "Programming Language :: Python :: 3.7",
         "Programming Language :: Python :: 3.8",
         "Programming Language :: Python :: 3.9",
         "Programming Language :: Python :: 3.10",
+        "Programming Language :: Python :: 3.11",
     ],
 )
 

stable_baselines3/common/atari_wrappers.py

@@ -11,7 +11,7 @@ try:
 
     cv2.ocl.setUseOpenCL(False)
 except ImportError:
-    cv2 = None
+    cv2 = None  # type: ignore[assignment]
 
 
 class StickyActionEnv(gym.Wrapper[np.ndarray, int, np.ndarray, int]):
@@ -241,6 +241,7 @@ class WarpFrame(gym.ObservationWrapper[np.ndarray, int, np.ndarray]):
         :param frame: environment frame
         :return: the observation
         """
+        assert cv2 is not None, "OpenCV is not installed, you can do `pip install opencv-python`"
         frame = cv2.cvtColor(frame, cv2.COLOR_RGB2GRAY)
         frame = cv2.resize(frame, (self.width, self.height), interpolation=cv2.INTER_AREA)
         return frame[:, :, None]

stable_baselines3/common/buffers.py

@@ -207,7 +207,9 @@ class ReplayBuffer(BaseBuffer):
         else:
             self.next_observations = np.zeros((self.buffer_size, self.n_envs, *self.obs_shape), dtype=observation_space.dtype)
 
-        self.actions = np.zeros((self.buffer_size, self.n_envs, self.action_dim), dtype=action_space.dtype)
+        self.actions = np.zeros(
+            (self.buffer_size, self.n_envs, self.action_dim), dtype=self._maybe_cast_dtype(action_space.dtype)
+        )
 
         self.rewards = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
         self.dones = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
@@ -311,6 +313,21 @@ class ReplayBuffer(BaseBuffer):
         )
         return ReplayBufferSamples(*tuple(map(self.to_torch, data)))
 
+    @staticmethod
+    def _maybe_cast_dtype(dtype: np.typing.DTypeLike) -> np.typing.DTypeLike:
+        """
+        Cast `np.float64` action datatype to `np.float32`,
+        keep the others dtype unchanged.
+        See GH#1572 for more information.
+
+        :param dtype: The original action space dtype
+        :return: ``np.float32`` if the dtype was float64,
+            the original dtype otherwise.
+        """
+        if dtype == np.float64:
+            return np.float32
+        return dtype
+
 
 class RolloutBuffer(BaseBuffer):
     """
@@ -543,7 +560,9 @@ class DictReplayBuffer(ReplayBuffer):
             for key, _obs_shape in self.obs_shape.items()
         }
 
-        self.actions = np.zeros((self.buffer_size, self.n_envs, self.action_dim), dtype=action_space.dtype)
+        self.actions = np.zeros(
+            (self.buffer_size, self.n_envs, self.action_dim), dtype=self._maybe_cast_dtype(action_space.dtype)
+        )
         self.rewards = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
         self.dones = np.zeros((self.buffer_size, self.n_envs), dtype=np.float32)
 

stable_baselines3/common/env_checker.py

@@ -203,18 +203,24 @@ def _check_obs(obs: Union[tuple, dict, np.ndarray, int], observation_space: spac
             f"Expected: {observation_space.dtype}, actual dtype: {obs.dtype}"
         )
         if isinstance(observation_space, spaces.Box):
-            assert np.all(obs >= observation_space.low), (
+            lower_bounds, upper_bounds = observation_space.low, observation_space.high
-                f"The observation returned by the `{method_name}()` method does not match the lower bound "
+            # Expose all invalid indices at once
-                f"of the given observation space {observation_space}."
+            invalid_indices = np.where(np.logical_or(obs < lower_bounds, obs > upper_bounds))
-                f"Expected: obs >= {np.min(observation_space.low)}, "
+            if (obs > upper_bounds).any() or (obs < lower_bounds).any():
-                f"actual min value: {np.min(obs)} at index {np.argmin(obs)}"
+                message = (
-            )
+                    f"The observation returned by the `{method_name}()` method does not match the bounds "
-            assert np.all(obs <= observation_space.high), (
+                    f"of the given observation space {observation_space}. \n"
-                f"The observation returned by the `{method_name}()` method does not match the upper bound "
+                )
-                f"of the given observation space {observation_space}. "
+                message += f"{len(invalid_indices[0])} invalid indices: \n"
-                f"Expected: obs <= {np.max(observation_space.high)}, "
+
-                f"actual max value: {np.max(obs)} at index {np.argmax(obs)}"
+                for index in zip(*invalid_indices):
-            )
+                    index_str = ",".join(map(str, index))
+                    message += (
+                        f"Expected: {lower_bounds[index]} <= obs[{index_str}] <= {upper_bounds[index]}, "
+                        f"actual value: {obs[index]} \n"
+                    )
+
+                raise AssertionError(message)
 
     assert observation_space.contains(obs), (
         f"The observation returned by the `{method_name}()` method "

stable_baselines3/common/logger.py

@@ -164,8 +164,10 @@ class HumanOutputFormat(KVWriter, SeqWriter):
         if isinstance(filename_or_file, str):
             self.file = open(filename_or_file, "w")
             self.own_file = True
-        elif isinstance(filename_or_file, TextIOBase):
+        elif isinstance(filename_or_file, TextIOBase) or hasattr(filename_or_file, "write"):
-            self.file = filename_or_file
+            # Note: in theory `TextIOBase` check should be sufficient,
+            # in practice, libraries don't always inherit from it, see GH#1598
+            self.file = filename_or_file  # type: ignore[assignment]
             self.own_file = False
         else:
             raise ValueError(f"Expected file or str, got {filename_or_file}")

stable_baselines3/common/monitor.py

@@ -193,7 +193,9 @@ class ResultsWriter:
         mode = "w" if override_existing else "a"
         # Prevent newline issue on Windows, see GH issue #692
         self.file_handler = open(filename, f"{mode}t", newline="\n")
-        self.logger = csv.DictWriter(self.file_handler, fieldnames=("r", "l", "t", *extra_keys))
+        self.logger = csv.DictWriter(
+            self.file_handler, fieldnames=("r", "l", "t", *extra_keys)
+        )  # pytype: disable=wrong-arg-types
         if override_existing:
             self.file_handler.write(f"#{json.dumps(header)}\n")
             self.logger.writeheader()

stable_baselines3/common/type_aliases.py

@@ -1,18 +1,12 @@
 """Common aliases for type hints"""
 
-import sys
 from enum import Enum
-from typing import Any, Callable, Dict, List, NamedTuple, Optional, SupportsFloat, Tuple, Union
+from typing import Any, Callable, Dict, List, NamedTuple, Optional, Protocol, SupportsFloat, Tuple, Union
 
 import gymnasium as gym
 import numpy as np
 import torch as th
 
-if sys.version_info >= (3, 8):
-    from typing import Protocol
-else:
-    from typing_extensions import Protocol
-
 from stable_baselines3.common import callbacks, vec_env
 
 GymEnv = Union[gym.Env, vec_env.VecEnv]

stable_baselines3/common/vec_env/patch_gym.py

@@ -40,7 +40,7 @@ def _patch_env(env: Union["gym.Env", gymnasium.Env]) -> gymnasium.Env:  # pragma
         import shimmy  # pytype: disable=import-error
     except ImportError as e:
         raise ImportError(
-            "Missing shimmy installation. You an OpenAI Gym environment. "
+            "Missing shimmy installation. You provided an OpenAI Gym environment. "
             "Stable-Baselines3 (SB3) has transitioned to using Gymnasium internally. "
             "In order to use OpenAI Gym environments with SB3, you need to "
             "install shimmy (`pip install 'shimmy>=0.2.1'`)."

stable_baselines3/dqn/dqn.py

@@ -245,7 +245,7 @@ class DQN(OffPolicyAlgorithm):
         if not deterministic and np.random.rand() < self.exploration_rate:
             if self.policy.is_vectorized_observation(observation):
                 if isinstance(observation, dict):
-                    n_batch = observation[list(observation.keys())[0]].shape[0]
+                    n_batch = observation[next(iter(observation.keys()))].shape[0]
                 else:
                     n_batch = observation.shape[0]
                 action = np.array([self.action_space.sample() for _ in range(n_batch)])

stable_baselines3/version.txt

@@ -1 +1 @@
-2.0.0
+2.1.0a4

tests/test_env_checker.py

@@ -42,15 +42,28 @@ def test_check_env_dict_action():
     [
         # Above upper bound
         (
-            spaces.Box(low=0.0, high=1.0, shape=(3,), dtype=np.float32),
+            spaces.Box(low=np.array([0.0, 0.0, 0.0]), high=np.array([2.0, 1.0, 1.0]), shape=(3,), dtype=np.float32),
             np.array([1.0, 1.5, 0.5], dtype=np.float32),
-            r"Expected: obs <= 1\.0, actual max value: 1\.5 at index 1",
+            r"Expected: 0\.0 <= obs\[1] <= 1\.0, actual value: 1\.5",
+        ),
+        # Above upper bound (multi-dim)
+        (
+            spaces.Box(low=-1.0, high=2.0, shape=(2, 3, 3, 1), dtype=np.float32),
+            3.0 * np.ones((2, 3, 3, 1), dtype=np.float32),
+            # Note: this is one of the 18 invalid indices
+            r"Expected: -1\.0 <= obs\[1,2,1,0\] <= 2\.0, actual value: 3\.0",
         ),
         # Below lower bound
         (
-            spaces.Box(low=0.0, high=2.0, shape=(3,), dtype=np.float32),
+            spaces.Box(low=np.array([0.0, -10.0, 0.0]), high=np.array([2.0, 1.0, 1.0]), shape=(3,), dtype=np.float32),
             np.array([-1.0, 1.5, 0.5], dtype=np.float32),
-            r"Expected: obs >= 0\.0, actual min value: -1\.0 at index 0",
+            r"Expected: 0\.0 <= obs\[0] <= 2\.0, actual value: -1\.0",
+        ),
+        # Below lower bound (multi-dim)
+        (
+            spaces.Box(low=-1.0, high=2.0, shape=(2, 3, 3, 1), dtype=np.float32),
+            -2 * np.ones((2, 3, 3, 1), dtype=np.float32),
+            r"18 invalid indices:",
         ),
         # Wrong dtype
         (
@@ -111,7 +124,7 @@ def test_check_env_detailed_error(obs_tuple, method):
 
     test_env = TestEnv()
     with pytest.raises(AssertionError, match=error_message):
-        check_env(env=test_env)
+        check_env(env=test_env, warn=False)
 
 
 class LimitedStepsTestEnv(gym.Env):

tests/test_envs.py

@@ -156,8 +156,6 @@ def test_non_default_spaces(new_obs_space):
         spaces.Box(low=-1000, high=1000, shape=(3,), dtype=np.float32),
         # Too small range
         spaces.Box(low=-0.1, high=0.1, shape=(2,), dtype=np.float32),
-        # Inverted boundaries
-        spaces.Box(low=1, high=-1, shape=(2,), dtype=np.float32),
         # Same boundaries
         spaces.Box(low=1, high=1, shape=(2,), dtype=np.float32),
         # Unbounded action space

tests/test_logger.py

@@ -437,8 +437,9 @@ def test_ep_buffers_stats_window_size(algo, stats_window_size):
     assert model.ep_success_buffer.maxlen == stats_window_size
 
 
-def test_human_output_format_custom_test_io():
+@pytest.mark.parametrize("base_class", [object, TextIOBase])
-    class DummyTextIO(TextIOBase):
+def test_human_output_format_custom_test_io(base_class):
+    class DummyTextIO(base_class):
         def __init__(self) -> None:
             super().__init__()
             self.lines = [[]]

tests/test_spaces.py

@@ -1,63 +1,67 @@
+from dataclasses import dataclass
 from typing import Dict, Optional
 
 import gymnasium as gym
 import numpy as np
 import pytest
 from gymnasium import spaces
+from gymnasium.spaces.space import Space
 
 from stable_baselines3 import A2C, DDPG, DQN, PPO, SAC, TD3
 from stable_baselines3.common.env_checker import check_env
 from stable_baselines3.common.env_util import make_vec_env
 from stable_baselines3.common.evaluation import evaluate_policy
 
+BOX_SPACE_FLOAT64 = spaces.Box(low=-1, high=1, shape=(2,), dtype=np.float64)
+BOX_SPACE_FLOAT32 = spaces.Box(low=-1, high=1, shape=(2,), dtype=np.float32)
 
-class DummyMultiDiscreteSpace(gym.Env):
+
-    def __init__(self, nvec):
+@dataclass
-        super().__init__()
+class DummyEnv(gym.Env):
-        self.observation_space = spaces.MultiDiscrete(nvec)
+    observation_space: Space
-        self.action_space = spaces.Box(low=-1, high=1, shape=(2,), dtype=np.float32)
+    action_space: Space
+
+    def step(self, action):
+        return self.observation_space.sample(), 0.0, False, False, {}
 
     def reset(self, *, seed: Optional[int] = None, options: Optional[Dict] = None):
         if seed is not None:
             super().reset(seed=seed)
         return self.observation_space.sample(), {}
 
-    def step(self, action):
-        return self.observation_space.sample(), 0.0, False, False, {}
 
-
+class DummyMultidimensionalAction(DummyEnv):
-class DummyMultiBinary(gym.Env):
-    def __init__(self, n):
-        super().__init__()
-        self.observation_space = spaces.MultiBinary(n)
-        self.action_space = spaces.Box(low=-1, high=1, shape=(2,), dtype=np.float32)
-
-    def reset(self, *, seed: Optional[int] = None, options: Optional[Dict] = None):
-        if seed is not None:
-            super().reset(seed=seed)
-        return self.observation_space.sample(), {}
-
-    def step(self, action):
-        return self.observation_space.sample(), 0.0, False, False, {}
-
-
-class DummyMultidimensionalAction(gym.Env):
     def __init__(self):
-        super().__init__()
+        super().__init__(
-        self.observation_space = spaces.Box(low=-1, high=1, shape=(2,), dtype=np.float32)
+            BOX_SPACE_FLOAT32,
-        self.action_space = spaces.Box(low=-1, high=1, shape=(2, 2), dtype=np.float32)
+            spaces.Box(low=-1, high=1, shape=(2, 2), dtype=np.float32),
+        )
 
-    def reset(self, *, seed: Optional[int] = None, options: Optional[Dict] = None):
-        if seed is not None:
-            super().reset(seed=seed)
-        return self.observation_space.sample(), {}
 
-    def step(self, action):
+class DummyMultiBinary(DummyEnv):
-        return self.observation_space.sample(), 0.0, False, False, {}
+    def __init__(self, n):
+        super().__init__(
+            spaces.MultiBinary(n),
+            BOX_SPACE_FLOAT32,
+        )
+
+
+class DummyMultiDiscreteSpace(DummyEnv):
+    def __init__(self, nvec):
+        super().__init__(
+            spaces.MultiDiscrete(nvec),
+            BOX_SPACE_FLOAT32,
+        )
 
 
 @pytest.mark.parametrize(
-    "env", [DummyMultiDiscreteSpace([4, 3]), DummyMultiBinary(8), DummyMultiBinary((3, 2)), DummyMultidimensionalAction()]
+    "env",
+    [
+        DummyMultiDiscreteSpace([4, 3]),
+        DummyMultiBinary(8),
+        DummyMultiBinary((3, 2)),
+        DummyMultidimensionalAction(),
+    ],
 )
 def test_env(env):
     # Check the env used for testing
@@ -127,3 +131,40 @@ def test_discrete_obs_space(model_class, env):
     else:
         kwargs = dict(n_steps=256)
     model_class("MlpPolicy", env, **kwargs).learn(256)
+
+
+@pytest.mark.parametrize("model_class", [SAC, TD3, PPO, DDPG, A2C])
+@pytest.mark.parametrize(
+    "obs_space",
+    [
+        BOX_SPACE_FLOAT32,
+        BOX_SPACE_FLOAT64,
+        spaces.Dict({"a": BOX_SPACE_FLOAT32, "b": BOX_SPACE_FLOAT32}),
+        spaces.Dict({"a": BOX_SPACE_FLOAT32, "b": BOX_SPACE_FLOAT64}),
+    ],
+)
+@pytest.mark.parametrize(
+    "action_space",
+    [
+        BOX_SPACE_FLOAT32,
+        BOX_SPACE_FLOAT64,
+    ],
+)
+def test_float64_action_space(model_class, obs_space, action_space):
+    env = DummyEnv(obs_space, action_space)
+    env = gym.wrappers.TimeLimit(env, max_episode_steps=200)
+    if isinstance(env.observation_space, spaces.Dict):
+        policy = "MultiInputPolicy"
+    else:
+        policy = "MlpPolicy"
+
+    if model_class in [PPO, A2C]:
+        kwargs = dict(n_steps=64, policy_kwargs=dict(net_arch=[12]))
+    else:
+        kwargs = dict(learning_starts=60, policy_kwargs=dict(net_arch=[12]))
+
+    model = model_class(policy, env, **kwargs)
+    model.learn(64)
+    initial_obs, _ = env.reset()
+    action, _ = model.predict(initial_obs, deterministic=False)
+    assert action.dtype == env.action_space.dtype