rlgym-learn-algos 0.2.1__cp311-cp311-win32.whl → 0.2.2__cp311-cp311-win32.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=pRplMtq88vWNms7sVhWJfwf-W7GATsQH3617hrkNl3s,2431
-rlgym_learn_algos-0.2.1.dist-info/WHEEL,sha256=1cEMGItI5ycdglW0xuhYFX4p-uaMeYRsFVmN9jJO6VY,92
-rlgym_learn_algos-0.2.1.dist-info/licenses/LICENSE,sha256=HrhfyXIkWY2tGFK11kg7vPCqhgh5DcxleloqdhrpyMY,11558
+rlgym_learn_algos-0.2.2.dist-info/METADATA,sha256=4wwr9xqqVWvZ7HYM4cumHiRdz79gkixfpe11b4MyvSU,2431
+rlgym_learn_algos-0.2.2.dist-info/WHEEL,sha256=zfc_r7GoDDc6Hz8pkKR77dEQzJMJDgktYOCKUorzovQ,92
+rlgym_learn_algos-0.2.2.dist-info/licenses/LICENSE,sha256=HrhfyXIkWY2tGFK11kg7vPCqhgh5DcxleloqdhrpyMY,11558
 rlgym_learn_algos/__init__.py,sha256=C7cRdL4lZrpk3ge_4_lGAbGodqWJXM56FfgO0keRPAY,207
 rlgym_learn_algos/conversion/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rlgym_learn_algos/conversion/convert_rlgym_ppo_checkpoint.py,sha256=A9nvzjp3DQNRNL5TAt-u3xE80JDIpYEDqAGNReHvFG0,908
@@ -15,18 +15,18 @@ rlgym_learn_algos/ppo/continuous_actor.py,sha256=1vdBUw2mQNFNu6A6ZrAztBjd4DmwjGk
 rlgym_learn_algos/ppo/critic.py,sha256=RB89WtiN52BEq5QCpGAPrASUnasac-Bpg7B0lM3UXHw,689
 rlgym_learn_algos/ppo/discrete_actor.py,sha256=Nuc3EndIQud3NGrkBIQgy-Z-okhXVrj6p6okSGD1KNY,2620
 rlgym_learn_algos/ppo/env_trajectories.py,sha256=gzQBRkzwZhlZeSvWL50cc8AOgBfsg5zUys0aTJj6aZU,3775
-rlgym_learn_algos/ppo/experience_buffer.py,sha256=QdyFMMM8YpEYrmtFaeaHXvFlNT2pCZwQKBEqsrv4v2I,10838
+rlgym_learn_algos/ppo/experience_buffer.py,sha256=xDm8NIMdErpv3GyWUBcTvzkLBQa8tW1TXb7OrKRDIu4,11059
 rlgym_learn_algos/ppo/experience_buffer_numpy.py,sha256=Apk4x-pfRnitKJPW6LBZyOPIhgeJs_5EG7BbTCqMwjk,4761
 rlgym_learn_algos/ppo/gae_trajectory_processor.py,sha256=JK958vasIIiuf3ALcFNlvBgGNhFshK8MhQJjwvxhrAM,5453
 rlgym_learn_algos/ppo/gae_trajectory_processor_pure_python.py,sha256=RpyDR6GQ1JXvwtoKkx5V3z3WvU9ElJdzfNtpPiZDaTc,6831
 rlgym_learn_algos/ppo/multi_discrete_actor.py,sha256=zSYeBBirjguSv_wO-peo06hioHiVhZQjnd-NYwJxmag,3127
 rlgym_learn_algos/ppo/ppo_agent_controller.py,sha256=h0UR-o2k-_LyeFTzvII3HQHHWyeMJewqLlca8ThtyfA,25105
-rlgym_learn_algos/ppo/ppo_learner.py,sha256=3YTfs7LhjiJ0u3-k84rYWcmAQxKIf2yp1i1UVY4v8Oc,15229
+rlgym_learn_algos/ppo/ppo_learner.py,sha256=Cbbuz0AMwPCmkQ1YPDdZLkbgZOdyrOLEx89Camn-nGE,15942
 rlgym_learn_algos/ppo/ppo_metrics_logger.py,sha256=niW8xgQLEBCGgTaVyiE_JqsU6RTjV6h-JzM-7c3JT38,2868
 rlgym_learn_algos/ppo/trajectory.py,sha256=IIH_IG8B_HkyxRPf-YsCyF1jQqNjDx752hgzAehG25I,719
 rlgym_learn_algos/ppo/trajectory_processor.py,sha256=5eY_mNGjqIkhqnbKeaqDvqIWPdg6wD6Ai3fXH2WoXbw,2091
 rlgym_learn_algos/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rlgym_learn_algos/rlgym_learn_algos.cp311-win32.pyd,sha256=TQ1jNEan9yh_o1CkLLdtdke1q8DvOPsosdz3Fcdt1PE,337920
+rlgym_learn_algos/rlgym_learn_algos.cp311-win32.pyd,sha256=wSNbyQLW2s8XROJJ3CwpPBEPaYWqpAsZJ8syylePLNE,337920
 rlgym_learn_algos/rlgym_learn_algos.pyi,sha256=NwY-sDZWM06TUiKPzxpfH1Td6G6E8TdxtRPgBSh-PPE,1203
 rlgym_learn_algos/stateful_functions/__init__.py,sha256=QS0KYjuzagNkYiYllXQmjoJn14-G7KZawq1Zvwh8alY,236
 rlgym_learn_algos/stateful_functions/batch_reward_type_numpy_converter.py,sha256=1yte5qYyl9LWdClHZ_YsF7R9dJqQeYfINMdgNF_59Gs,767
@@ -34,6 +34,6 @@ rlgym_learn_algos/stateful_functions/numpy_obs_standardizer.py,sha256=OgtwCaxBGT
 rlgym_learn_algos/stateful_functions/obs_standardizer.py,sha256=qPPc3--J_3mpJJ-QHJjta6dbWWBobL7SYdK5MUP-XMw,606
 rlgym_learn_algos/util/__init__.py,sha256=VPM6SN4T_625H9t30s9EiLeXiEEWgcyRVHa-LLVNrn4,47
 rlgym_learn_algos/util/running_stats.py,sha256=0tiGFpKtHWzMa1CxM_ueBzd_ryX4bJBriC8MXcSLg8w,4479
-rlgym_learn_algos/util/torch_functions.py,sha256=CTTHzTIi7u1O9HyX0cVJOrnYVbAtnlVs0g1fO9s3ano,3458
-rlgym_learn_algos/util/torch_pydantic.py,sha256=pgj3I-3q8iW9qtOCv1fgjNkZgA00G_Rdkb4qJPk5gxo,3530
-rlgym_learn_algos-0.2.1.dist-info/RECORD,,
+rlgym_learn_algos/util/torch_functions.py,sha256=_uAXhq1YYPneWI3_XXRYsSA3Hn1a8wGjUnI3m9UojdU,3411
+rlgym_learn_algos/util/torch_pydantic.py,sha256=5AbXQcfQtVgLRBSgCj0Hvi_H42WHLu4Oty4l_i22nAo,3531
+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: cp311-cp311-win32