multi-agent-rlenv 3.5.2__tar.gz → 3.5.5__tar.gz

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/.github/workflows/ci.yaml

@@ -46,7 +46,7 @@ jobs:
           uv-version: 0.6.4
       - name: Install dependencies and run pytest
         run: |
-          uv sync --extra overcooked --extra gym --extra pettingzoo
+          uv sync --extra overcooked --extra gym --extra pettingzoo --extra torch
           uv run pytest
 
   build:

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: multi-agent-rlenv
-Version: 3.5.2
+Version: 3.5.5
 Summary: A strongly typed Multi-Agent Reinforcement Learning framework
 Project-URL: repository, https://github.com/yamoling/multi-agent-rlenv
 Author-email: Yannick Molinghen <yannick.molinghen@ulb.be>
@@ -19,6 +19,7 @@ Requires-Dist: pymunk>=6.0; extra == 'all'
 Requires-Dist: pysc2; extra == 'all'
 Requires-Dist: scipy>=1.10; extra == 'all'
 Requires-Dist: smac; extra == 'all'
+Requires-Dist: torch>=2.0; extra == 'all'
 Provides-Extra: gym
 Requires-Dist: gymnasium>=0.29.1; extra == 'gym'
 Provides-Extra: overcooked
@@ -31,6 +32,8 @@ Requires-Dist: scipy>=1.10; extra == 'pettingzoo'
 Provides-Extra: smac
 Requires-Dist: pysc2; extra == 'smac'
 Requires-Dist: smac; extra == 'smac'
+Provides-Extra: torch
+Requires-Dist: torch>=2.0; extra == 'torch'
 Description-Content-Type: text/markdown
 
 # `marlenv` - A unified framework for muti-agent reinforcement learning

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/pyproject.toml

@@ -20,6 +20,7 @@ gym = ["gymnasium>=0.29.1"]
 smac = ["smac", "pysc2"]
 pettingzoo = ["pettingzoo>=1.20", "pymunk>=6.0", "scipy>=1.10"]
 overcooked = ["overcooked-ai>=1.1.0", "scipy>=1.10"]
+torch = ["torch>=2.0"]
 all = [
     "gymnasium>0.29.1",
     "pettingzoo>=1.20",
@@ -28,6 +29,7 @@ all = [
     "pysc2",
     "pymunk>=6.0",
     "scipy>=1.10",
+    "torch>=2.0",
 ]
 
 [build-system]

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/src/marlenv/__init__.py

@@ -62,7 +62,7 @@ print(env.extras_shape) # (1, )
 If you want to create a new environment, you can simply create a class that inherits from `MARLEnv`. If you want to create a wrapper around an existing `MARLEnv`, you probably want to subclass `RLEnvWrapper` which implements a default behaviour for every method.
 """
 
-__version__ = "3.5.2"
+__version__ = "3.5.5"
 
 from . import models
 from .models import (

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/src/marlenv/models/episode.py

@@ -2,30 +2,32 @@ from dataclasses import dataclass
 from functools import cached_property
 from typing import Any, Callable, Optional, Sequence, overload
 
+import cv2
 import numpy as np
 import numpy.typing as npt
-import cv2
 
+from marlenv.exceptions import EnvironmentMismatchException, ReplayMismatchException
+from marlenv.utils import CachedPropertyInvalidator
+
+from .env import MARLEnv
 from .observation import Observation
 from .state import State
 from .step import Step
 from .transition import Transition
-from .env import MARLEnv
-from marlenv.exceptions import EnvironmentMismatchException, ReplayMismatchException
 
 
 @dataclass
-class Episode:
+class Episode(CachedPropertyInvalidator):
     """Episode model made of observations, actions, rewards, ..."""
 
     all_observations: list[npt.NDArray[np.float32]]
     all_extras: list[npt.NDArray[np.float32]]
     actions: list[npt.NDArray]
     rewards: list[npt.NDArray[np.float32]]
-    all_available_actions: list[npt.NDArray[np.bool_]]
+    all_available_actions: list[npt.NDArray[np.bool]]
     all_states: list[npt.NDArray[np.float32]]
     all_states_extras: list[npt.NDArray[np.float32]]
-    metrics: dict[str, float]
+    metrics: dict[str, Any]
     episode_len: int
     other: dict[str, list[Any]]
     is_done: bool = False
@@ -33,7 +35,7 @@ class Episode:
     """Whether the episode did reach a terminal state (different from truncated)"""
 
     @staticmethod
-    def new(obs: Observation, state: State, metrics: Optional[dict[str, float]] = None) -> "Episode":
+    def new(obs: Observation, state: State, metrics: Optional[dict[str, Any]] = None) -> "Episode":
         if metrics is None:
             metrics = {}
         return Episode(
@@ -153,12 +155,12 @@ class Episode:
         """Get the next extra features"""
         return self.all_extras[1:]
 
-    @cached_property
+    @property
     def n_agents(self):
         """The number of agents in the episode"""
         return self.all_extras[0].shape[0]
 
-    @cached_property
+    @property
     def n_actions(self):
         """The number of actions"""
         return len(self.all_available_actions[0][0])
@@ -267,7 +269,7 @@ class Episode:
     def __len__(self):
         return self.episode_len
 
-    @cached_property
+    @property
     def score(self) -> list[float]:
         """The episode score (sum of all rewards across all objectives)"""
         score = []
@@ -363,51 +365,6 @@ class Episode:
             for i, s in enumerate(scores):
                 self.metrics[f"score-{i}"] = float(s)
 
-    # def add_data(
-    #     self,
-    #     new_obs: Observation,
-    #     new_state: State,
-    #     action: A,
-    #     reward: np.ndarray,
-    #     done: bool,
-    #     truncated: bool,
-    #     info: dict[str, Any],
-    #     **kwargs,
-    # ):
-    #     """Add a new transition to the episode"""
-    #     self.episode_len += 1
-    #     self.all_observations.append(new_obs.data)
-    #     self.all_extras.append(new_obs.extras)
-    #     self.all_available_actions.append(new_obs.available_actions)
-    #     self.all_states.append(new_state.data)
-    #     self.all_states_extras.append(new_state.extras)
-    #     match action:
-    #         case np.ndarray() as action:
-    #             self.actions.append(action)
-    #         case other:
-    #             self.actions.append(np.array(other))
-    #     self.rewards.append(reward)
-    #     for key, value in kwargs.items():
-    #         current = self.other.get(key, [])
-    #         current.append(value)
-    #         self.other[key] = current
-
-    #     if done:
-    #         # Only set the truncated flag if the episode is not done (both could happen with a time limit)
-    #         self.is_truncated = truncated
-    #         self.is_done = done
-    #         # Add metrics that can be plotted
-    #         for key, value in info.items():
-    #             if isinstance(value, bool):
-    #                 value = int(value)
-    #             self.metrics[key] = value
-    #         self.metrics["episode_len"] = self.episode_len
-
-    #         rewards = np.array(self.rewards)
-    #         scores = np.sum(rewards, axis=0)
-    #         for i, s in enumerate(scores):
-    #             self.metrics[f"score-{i}"] = float(s)
-
-    def add_metrics(self, metrics: dict[str, float]):
+    def add_metrics(self, metrics: dict[str, Any]):
         """Add metrics to the episode"""
         self.metrics.update(metrics)

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/src/marlenv/models/observation.py

@@ -87,3 +87,13 @@ class Observation:
             if not np.array_equal(self.data, other.data):
                 return False
         return np.array_equal(self.extras, other.extras) and np.array_equal(self.available_actions, other.available_actions)
+
+    def as_tensors(self, device=None):
+        """
+        Convert the observation to a tuple of tensors of shape (1, n_agents, <dim>).
+        """
+        import torch
+
+        data = torch.from_numpy(self.data).unsqueeze(0).to(device, non_blocking=True)
+        extras = torch.from_numpy(self.extras).unsqueeze(0).to(device, non_blocking=True)
+        return data, extras

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/src/marlenv/models/state.py

@@ -52,3 +52,11 @@ class State(Generic[StateType]):
         if not np.array_equal(self.extras, value.extras):
             return False
         return True
+
+    def as_tensors(self, device=None):
+        """Convert the state to a tuple of tensors of shape (1, <dim>)."""
+        import torch
+
+        data = torch.from_numpy(self.data).unsqueeze(0).to(device, non_blocking=True)
+        extras = torch.from_numpy(self.extras).unsqueeze(0).to(device, non_blocking=True)
+        return data, extras

multi_agent_rlenv-3.5.5/src/marlenv/utils/__init__.py

@@ -0,0 +1,12 @@
+from .cached_property_collector import CachedPropertyCollector, CachedPropertyInvalidator
+from .schedule import ExpSchedule, LinearSchedule, MultiSchedule, RoundedSchedule, Schedule
+
+__all__ = [
+    "Schedule",
+    "LinearSchedule",
+    "ExpSchedule",
+    "MultiSchedule",
+    "RoundedSchedule",
+    "CachedPropertyCollector",
+    "CachedPropertyInvalidator",
+]

multi_agent_rlenv-3.5.5/src/marlenv/utils/cached_property_collector.py

@@ -0,0 +1,17 @@
+from functools import cached_property
+
+
+class CachedPropertyCollector(type):
+    def __init__(cls, name: str, bases: tuple, namespace: dict):
+        super().__init__(name, bases, namespace)
+        cls.CACHED_PROPERTY_NAMES = [key for key, value in namespace.items() if isinstance(value, cached_property)]
+
+
+class CachedPropertyInvalidator(metaclass=CachedPropertyCollector):
+    def __init__(self):
+        super().__init__()
+
+    def invalidate_cached_properties(self):
+        for key in self.__class__.CACHED_PROPERTY_NAMES:
+            if hasattr(self, key):
+                delattr(self, key)

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/src/marlenv/wrappers/__init__.py

@@ -11,6 +11,7 @@ from .centralised import Centralized
 from .available_actions_mask import AvailableActionsMask
 from .delayed_rewards import DelayedReward
 from .potential_shaping import PotentialShaping
+from .action_randomizer import ActionRandomizer
 
 __all__ = [
     "RLEnvWrapper",
@@ -28,4 +29,5 @@ __all__ = [
     "Centralized",
     "DelayedReward",
     "PotentialShaping",
+    "ActionRandomizer",
 ]

multi_agent_rlenv-3.5.5/src/marlenv/wrappers/action_randomizer.py

@@ -0,0 +1,17 @@
+from .rlenv_wrapper import RLEnvWrapper, AS, MARLEnv
+import numpy as np
+
+
+class ActionRandomizer(RLEnvWrapper[AS]):
+    def __init__(self, env: MARLEnv[AS], p: float):
+        super().__init__(env)
+        self.p = p
+
+    def step(self, action):
+        if np.random.rand() < self.p:
+            action = self.action_space.sample()
+        return super().step(action)
+
+    def seed(self, seed_value: int):
+        np.random.seed(seed_value)
+        super().seed(seed_value)

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/tests/test_episode.py

@@ -1,7 +1,8 @@
 import numpy as np
-from marlenv.models import Transition, Episode
+
+from marlenv import DiscreteMockEnv, Episode, Transition, wrappers
 from marlenv.exceptions import EnvironmentMismatchException, ReplayMismatchException
-from marlenv import wrappers, DiscreteMockEnv
+
 from .utils import generate_episode
 
 
@@ -219,3 +220,31 @@ def test_get_images():
     images = episode.get_images(env, seed=0)
     # +1 because there is one image for the initial state (after reset)
     assert len(images) == len(episode) + 1
+
+
+def test_episode_invalidate_cached_properties():
+    env = DiscreteMockEnv(end_game=10)
+    obs, state = env.reset()
+    episode = Episode.new(obs, state)
+    for _ in range(5):
+        action = env.sample_action()
+        step = env.step(action)
+        episode.add(Transition.from_step(obs, state, action, step))
+        obs = step.obs
+        state = step.state
+
+    # Access the cached properties
+    assert len(episode.states) == 5
+    assert len(episode.obs) == 5
+
+    # Invalidate the cached properties
+    episode.invalidate_cached_properties()
+    for _ in range(3):
+        action = env.sample_action()
+        step = env.step(action)
+        episode.add(Transition.from_step(obs, state, action, step))
+        obs = step.obs
+        state = step.state
+
+    # Access the cached properties again
+    assert len(episode.obs) == 8

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/tests/test_models.py

@@ -1,8 +1,11 @@
 from marlenv import Observation, Transition, DiscreteMockEnv, DiscreteMOMockEnv, Builder, State, Episode, MARLEnv, DiscreteSpace
 import numpy as np
-
+import pytest
+from importlib.util import find_spec
 from .utils import generate_episode
 
+HAS_PYTORCH = find_spec("torch") is not None
+
 
 def test_obs_eq():
     obs1 = Observation(
@@ -422,3 +425,33 @@ def test_wrong_extras_meanings_length():
         assert False, "This should raise a ValueError because the length of extras_meanings is different from the actual number of extras"
     except ValueError:
         pass
+
+
+@pytest.mark.skipif(not HAS_PYTORCH, reason="torch is not installed")
+def test_observation_as_tensor():
+    import torch
+
+    env = DiscreteMockEnv(4)
+    obs = env.reset()[0]
+    data, extras = obs.as_tensors()
+    assert isinstance(data, torch.Tensor)
+    assert data.shape == (1, env.n_agents, *env.observation_shape)
+    assert data.dtype == torch.float32
+    assert isinstance(extras, torch.Tensor)
+    assert extras.shape == (1, env.n_agents, *env.extras_shape)
+    assert extras.dtype == torch.float32
+
+
+@pytest.mark.skipif(not HAS_PYTORCH, reason="torch is not installed")
+def test_state_as_tensor():
+    import torch
+
+    env = DiscreteMockEnv(4)
+    state = env.reset()[1]
+    data, extras = state.as_tensors()
+    assert isinstance(data, torch.Tensor)
+    assert data.shape == (1, *env.state_shape)
+    assert data.dtype == torch.float32
+    assert isinstance(extras, torch.Tensor)
+    assert extras.shape == (1, *env.state_extra_shape)
+    assert extras.dtype == torch.float32

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/tests/test_schedules.py

@@ -1,4 +1,4 @@
-from marlenv.utils import Schedule, MultiSchedule
+from marlenv.utils import MultiSchedule, Schedule
 
 
 def is_close(a: float, b: float, tol: float = 1e-6) -> bool:

{multi_agent_rlenv-3.5.2 → multi_agent_rlenv-3.5.5}/tests/test_wrappers.py

@@ -1,6 +1,6 @@
 import numpy as np
 from marlenv import Builder, DiscreteMOMockEnv, DiscreteMockEnv, MARLEnv
-from marlenv.wrappers import Centralized, AvailableActionsMask, TimeLimit, LastAction, DelayedReward
+from marlenv.wrappers import Centralized, AvailableActionsMask, TimeLimit, LastAction, DelayedReward, ActionRandomizer
 import marlenv
 
 

multi_agent_rlenv-3.5.2/src/marlenv/utils/__init__.py

@@ -1,10 +0,0 @@
-from .schedule import Schedule, MultiSchedule, RoundedSchedule, LinearSchedule, ExpSchedule
-
-
-__all__ = [
-    "Schedule",
-    "LinearSchedule",
-    "ExpSchedule",
-    "MultiSchedule",
-    "RoundedSchedule",
-]