rlgym-learn-algos 0.2.1__cp38-cp38-musllinux_1_2_i686.whl → 0.2.2__cp38-cp38-musllinux_1_2_i686.whl

rlgym_learn_algos/ppo/experience_buffer.py

@@ -7,6 +7,7 @@ import numpy as np
 import torch
 from pydantic import BaseModel, Field, model_validator
 from rlgym.api import ActionType, AgentID, ObsType, RewardType
+
 from rlgym_learn_algos.util.torch_functions import get_device
 from rlgym_learn_algos.util.torch_pydantic import PydanticTorchDevice
 
@@ -24,6 +25,7 @@ EXPERIENCE_BUFFER_FILE = "experience_buffer.pkl"
 class ExperienceBufferConfigModel(BaseModel, extra="forbid"):
     max_size: int = 100000
     device: PydanticTorchDevice = "auto"
+    save_experience_buffer_in_checkpoint: bool = True
     trajectory_processor_config: Dict[str, Any] = Field(default_factory=dict)
 
     @model_validator(mode="before")
@@ -40,8 +42,9 @@ class ExperienceBufferConfigModel(BaseModel, extra="forbid"):
                     data["trajectory_processor_config"] = data[
                         "trajectory_processor_config"
                     ].model_dump()
-            if "device" not in data or data["device"] == "auto":
-                data["device"] = get_device("auto")
+            if "device" not in data:
+                data["device"] = "auto"
+            data["device"] = get_device(data["device"])
         return data
 
 
@@ -165,21 +168,22 @@ class ExperienceBuffer(
 
     def save_checkpoint(self, folder_path):
         os.makedirs(folder_path, exist_ok=True)
-        with open(
-            os.path.join(folder_path, EXPERIENCE_BUFFER_FILE),
-            "wb",
-        ) as f:
-            pickle.dump(
-                {
-                    "agent_ids": self.agent_ids,
-                    "observations": self.observations,
-                    "actions": self.actions,
-                    "log_probs": self.log_probs,
-                    "values": self.values,
-                    "advantages": self.advantages,
-                },
-                f,
-            )
+        if self.config.experience_buffer_config.save_experience_buffer_in_checkpoint:
+            with open(
+                os.path.join(folder_path, EXPERIENCE_BUFFER_FILE),
+                "wb",
+            ) as f:
+                pickle.dump(
+                    {
+                        "agent_ids": self.agent_ids,
+                        "observations": self.observations,
+                        "actions": self.actions,
+                        "log_probs": self.log_probs,
+                        "values": self.values,
+                        "advantages": self.advantages,
+                    },
+                    f,
+                )
         self.trajectory_processor.save_checkpoint(folder_path)
 
     # TODO: update docs

rlgym_learn_algos/ppo/ppo_learner.py

@@ -39,17 +39,25 @@ class PPOLearnerConfigModel(BaseModel, extra="forbid"):
     clip_range: float = 0.2
     actor_lr: float = 3e-4
     critic_lr: float = 3e-4
+    advantage_normalization: bool = True
     device: PydanticTorchDevice = "auto"
+    cudnn_benchmark_mode: bool = True
 
     @model_validator(mode="before")
     @classmethod
     def set_device(cls, data):
-        if isinstance(data, dict) and (
-            "device" not in data or data["device"] == "auto"
-        ):
-            data["device"] = get_device("auto")
+        if isinstance(data, dict):
+            if "device" not in data:
+                data["device"] = "auto"
+            data["device"] = get_device(data["device"])
         return data
 
+    @model_validator(mode="after")
+    def validate_cudnn_benchmark(self):
+        if self.device.type != "cuda":
+            self.cudnn_benchmark_mode = False
+        return self
+
 
 @dataclass
 class DerivedPPOLearnerConfig:
@@ -107,6 +115,12 @@ class PPOLearner(
     def load(self, config: DerivedPPOLearnerConfig):
         self.config = config
 
+        if (
+            config.learner_config.cudnn_benchmark_mode
+            and config.learner_config.device.type == "cuda"
+        ):
+            torch.backends.cudnn.benchmark = True
+
         self.actor = self.actor_factory(
             config.obs_space, config.action_space, config.learner_config.device
         )
@@ -292,6 +306,10 @@ class PPOLearner(
                     advantages = batch_advantages[start:stop].to(
                         self.config.learner_config.device
                     )
+                    if self.config.learner_config.advantage_normalization:
+                        advantages = (advantages - torch.mean(advantages)) / (
+                            torch.std(advantages) + 1e-8
+                        )
                     old_probs = batch_old_probs[start:stop].to(
                         self.config.learner_config.device
                     )

rlgym_learn_algos/util/torch_functions.py

@@ -13,7 +13,6 @@ import torch.nn as nn
 
 def get_device(device: str):
     if device in ["auto", "gpu"] and torch.cuda.is_available():
-        torch.backends.cudnn.benchmark = True
         return "cuda:0"
     elif device == "auto" and not torch.cuda.is_available():
         return "cpu"

rlgym_learn_algos/util/torch_pydantic.py

@@ -42,7 +42,7 @@ device_str_regex = (
             "privateuseone",
         ]
     )
-    + ")(:\d+)"
+    + ")(:\d+)?"
 )
 
 

{rlgym_learn_algos-0.2.1.dist-info → rlgym_learn_algos-0.2.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rlgym-learn-algos
-Version: 0.2.1
+Version: 0.2.2
 Classifier: Programming Language :: Rust
 Classifier: Programming Language :: Python :: Implementation :: CPython
 Requires-Dist: pydantic>=2.8.2

{rlgym_learn_algos-0.2.1.dist-info → rlgym_learn_algos-0.2.2.dist-info}/RECORD

@@ -1,6 +1,6 @@
-rlgym_learn_algos-0.2.1.dist-info/METADATA,sha256=u9tM8KxVpm3nQ86a01tlDy5csGAeTmCMDZdccXgrUrg,2403
-rlgym_learn_algos-0.2.1.dist-info/WHEEL,sha256=X1CAAG5PtBwuCUbvZLy-66J-Jfin3nupluhqSslkPtI,103
-rlgym_learn_algos-0.2.1.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+rlgym_learn_algos-0.2.2.dist-info/METADATA,sha256=9m49XNjxVc1H0Hj27PV8BszTk9DnOBNBNIjdxb0uwcQ,2403
+rlgym_learn_algos-0.2.2.dist-info/WHEEL,sha256=zrTr349f3UPOfsn1HRNexsaeA05ZzRsCL0d94NnDDZc,103
+rlgym_learn_algos-0.2.2.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
 rlgym_learn_algos.libs/libgcc_s-b5472b99.so.1,sha256=wh8CpjXz9IccAyeERcB7YDEx7NH2jF-PykwOyYNeRRI,453841
 rlgym_learn_algos/__init__.py,sha256=dZeTgNro6qG1Hu0l0UBhgHOYiyeCwPWndC84dJAp__U,203
 rlgym_learn_algos/conversion/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -16,18 +16,18 @@ rlgym_learn_algos/ppo/continuous_actor.py,sha256=A4FQ0lKqlB47AeSrDdrPXMvNKXhl5to
 rlgym_learn_algos/ppo/critic.py,sha256=XPleWDO8uM25zlzptWDvZQpUpKvib5kRs9JpmWTVPuY,669
 rlgym_learn_algos/ppo/discrete_actor.py,sha256=TZC7b7ss16giobPC1oz-maOSDX-SrNBUzS1wIV2Rzgw,2547
 rlgym_learn_algos/ppo/env_trajectories.py,sha256=PaO6dmpNkQ3yDLaHIRc0ipn45t5zAjE5U1D_N-LQtgY,3684
-rlgym_learn_algos/ppo/experience_buffer.py,sha256=f_baTo18JKBQjTQoO0FJoCOWlZFV8vJgMXqevg-TSi8,10552
+rlgym_learn_algos/ppo/experience_buffer.py,sha256=4wWSfq1tobXv7lmBbkM5sbTVuPJxrdAbxi5rNHc04g4,10769
 rlgym_learn_algos/ppo/experience_buffer_numpy.py,sha256=6AOGQjDn_dHLS9bmxJW_cGEjBUbe8u5VWS0LVlpIdmY,4617
 rlgym_learn_algos/ppo/gae_trajectory_processor.py,sha256=gv5kxvvPnK7SyQIAq6MbOFILIMdPlzoLZwM8TRmtNWw,5302
 rlgym_learn_algos/ppo/gae_trajectory_processor_pure_python.py,sha256=cq7qbK0mcLDXRzA6-pKW0OC50X52XhT5himcOTD6Ei4,6657
 rlgym_learn_algos/ppo/multi_discrete_actor.py,sha256=Mik0X79dUy2ZRIMol4RMTZE9qzsOk6f_6bDaOl5ghxs,3039
 rlgym_learn_algos/ppo/ppo_agent_controller.py,sha256=iVmCvN7H1IcKX7VrZnMParQ148EnPqs9yZ9CCgtdsq8,24524
-rlgym_learn_algos/ppo/ppo_learner.py,sha256=oRDBQszbgFvoVmGjRqpklQWTOI2NazOLzhDU_c9-SAU,14827
+rlgym_learn_algos/ppo/ppo_learner.py,sha256=utEWkikXCpC6Xc1D3asohO0HsIaq3tLyoTlb7fXLOw4,15522
 rlgym_learn_algos/ppo/ppo_metrics_logger.py,sha256=iUyUc2GPwDIIjZeJPZWxoeRrzUWV_qLOac0vApQBkp0,2803
 rlgym_learn_algos/ppo/trajectory.py,sha256=_xyS9ueU6iVvqMUpFr-kb42wEHHZy4zCse7_r660n5E,690
 rlgym_learn_algos/ppo/trajectory_processor.py,sha256=3XRsXXexHWp6UV5nAeBLYvWqvQ9EbNHSN3Yooi4cezo,2031
 rlgym_learn_algos/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rlgym_learn_algos/rlgym_learn_algos.cpython-38-i386-linux-gnu.so,sha256=VL0hdEQVoAI5zvHCFLRz6jYIARLjUJwoGNHQwhBgHyA,719349
+rlgym_learn_algos/rlgym_learn_algos.cpython-38-i386-linux-gnu.so,sha256=DseEGF0EbwC1IS58CCtlI3cmsPJhtH-Cob_e1wQpHX8,707061
 rlgym_learn_algos/rlgym_learn_algos.pyi,sha256=B9Kt9uK8xCqASRxWvzLdV501TSCMO4vTNqvZ0MhOHyo,1164
 rlgym_learn_algos/stateful_functions/__init__.py,sha256=OAVy6cQIS85Utyp18jjHgdmascX_8nkwk3A0OpFJxT4,230
 rlgym_learn_algos/stateful_functions/batch_reward_type_numpy_converter.py,sha256=a3q2l5SIgDI36ImF_kYoa684pghnFnlV2vGYvV2zcV0,743
@@ -35,6 +35,6 @@ rlgym_learn_algos/stateful_functions/numpy_obs_standardizer.py,sha256=Xa_fuJCSGu
 rlgym_learn_algos/stateful_functions/obs_standardizer.py,sha256=m2nw1JUg2MKYthn6tWrv2HYIuQ-GfUm48RX9s99jXF4,589
 rlgym_learn_algos/util/__init__.py,sha256=hq7M00Q7zAfyQmIGmXOif0vI40aj_FQ5SqI5dnuGvb0,46
 rlgym_learn_algos/util/running_stats.py,sha256=KtzdKKT75-5ZC58JRqaDXk6sBqa3ZSjQQZrRajAw3Yk,4339
-rlgym_learn_algos/util/torch_functions.py,sha256=6esZL8FeVwWWQWvwLEhkh_B0WqfIWzIWkCDbuhmRlnk,3366
-rlgym_learn_algos/util/torch_pydantic.py,sha256=zXllJoV8HgqJxguPKJ4Y3DIWEwDeJlBW9CIps-yxM44,3412
-rlgym_learn_algos-0.2.1.dist-info/RECORD,,
+rlgym_learn_algos/util/torch_functions.py,sha256=ImgDw4I3ZixGDi17YRkW6UbaiaQTbvOCUCS7N0QVSsU,3320
+rlgym_learn_algos/util/torch_pydantic.py,sha256=khPGA6kWh4_WHoploDkl_SCIGX8SkKkFT40RE06PImc,3413
+rlgym_learn_algos-0.2.2.dist-info/RECORD,,

{rlgym_learn_algos-0.2.1.dist-info → rlgym_learn_algos-0.2.2.dist-info}/WHEEL

@@ -1,4 +1,4 @@
 Wheel-Version: 1.0
-Generator: maturin (1.8.6)
+Generator: maturin (1.9.0)
 Root-Is-Purelib: false
 Tag: cp38-cp38-musllinux_1_2_i686