feat: abstracted checkpoint eval configs

configs/evaluation/default.yaml

@@ -0,0 +1,8 @@
+# Default Evaluation Configuration
+# Settings used for checkpoint evaluation during training.
+
+evaluate_checkpoints: false
+# Max number of control steps during evaluation rollout.
+eval_max_steps: 5000
+# Seed for deterministic evaluation reset.
+eval_seed: 0

configs/logging/default.yaml

@@ -11,6 +11,3 @@ checkpoint_frequency: 100
 upload_final_model: false
 upload_checkpoints: false
 hf_entity: ""
-evaluate_checkpoints: false
-eval_max_steps: 5000
-eval_seed: 0

configs/main_config.yaml

@@ -6,6 +6,7 @@ defaults:
   - brittle_star_config
   - experiment: base
   - logging: default
+  - evaluation: default
   - ppo: default
   - architecture: centralized
   - morphology: 5_arms_full

src/brittle_star_project/configs/config_evaluation.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+
+@dataclass
+class EvaluationConfig:
+    """Evaluation settings.
+
+    Currently used for synchronous checkpoint evaluation during training.
+    """
+
+    # When enabled, each saved checkpoint is evaluated headlessly and the results
+    # are appended to a CSV in the run's metrics/ folder.
+    evaluate_checkpoints: bool = False
+    eval_max_steps: int = 5000
+    eval_seed: int = 0
+
+    def __post_init__(self) -> None:
+        if self.evaluate_checkpoints and self.eval_max_steps <= 0:
+            raise ValueError(
+                "Configuration Error: 'eval_max_steps' must be > 0 when "
+                "'evaluate_checkpoints' is enabled."
+            )

src/brittle_star_project/configs/main_config.py

@@ -2,6 +2,7 @@ from dataclasses import dataclass, field
 
 from experiment_logger.config_logger import LoggingConfig
 from brittle_star_project.configs.config_experiment import ExperimentConfig
+from brittle_star_project.configs.config_evaluation import EvaluationConfig
 from brittle_star_project.configs.config_ppo import PPOConfig
 from brittle_star_project.configs.config_architecture import ArchitectureConfig
 from brittle_star_project.configs.config_simulation import SimulationSettings
@@ -23,6 +24,7 @@ class BrittleStarConfig:
 
     experiment: ExperimentConfig = field(default_factory=ExperimentConfig)
     logging: LoggingConfig = field(default_factory=LoggingConfig)
+    evaluation: EvaluationConfig = field(default_factory=EvaluationConfig)
     ppo: PPOConfig = field(default_factory=PPOConfig)
     # This field is polymorphic; defaults to the base class to allow subclasses
     # (CentralizedConfig, DecentralizedConfig) to be merged in via Hydra.

src/brittle_star_project/configs/register_configs.py

@@ -2,6 +2,7 @@ from hydra.core.config_store import ConfigStore
 
 from experiment_logger.config_logger import LoggingConfig
 from brittle_star_project.configs.config_experiment import ExperimentConfig
+from brittle_star_project.configs.config_evaluation import EvaluationConfig
 from brittle_star_project.configs.config_ppo import PPOConfig
 from brittle_star_project.configs.config_architecture import (
     CentralizedConfig,
@@ -32,6 +33,7 @@ def register_configs() -> None:
     # Sub-config groups — each group corresponds to a configs/ subdirectory.
     cs.store(group="experiment", name="base_experiment", node=ExperimentConfig)
     cs.store(group="logging", name="base_logging", node=LoggingConfig)
+    cs.store(group="evaluation", name="base_evaluation", node=EvaluationConfig)
     cs.store(group="ppo", name="base_ppo", node=PPOConfig)
 
     # Architecture variants — swap via CLI: architecture=decentralized

src/brittle_star_project/trainers/PPOTrainer.py

@@ -263,6 +263,7 @@ class PPOTrainer:
         self.ppo = cfg.ppo
         self.experiment = cfg.experiment
         self.logging_cfg = cfg.logging
+        self.evaluation_cfg = cfg.evaluation
         self.env = env
         self.run_dir = run_dir
         self.run_name = run_name
@@ -710,11 +711,22 @@ class PPOTrainer:
         return csv_path
 
     def _evaluate_checkpoint(self, iteration: int, *, trained_timesteps: int) -> None:
-        if not self.logging_cfg.evaluate_checkpoints:
+        if not self.evaluation_cfg.evaluate_checkpoints:
             return
 
-        max_steps = int(self.logging_cfg.eval_max_steps)
+        max_steps = int(self.evaluation_cfg.eval_max_steps)
-        seed = int(self.logging_cfg.eval_seed)
+        seed = int(self.evaluation_cfg.eval_seed)
+
+        if max_steps <= 0:
+            self.logger.warning("[EVAL]: eval_max_steps must be > 0; skipping evaluation")
+            return
+
+        if not self.logging_cfg.save_checkpoints or self.logging_cfg.checkpoint_frequency <= 0:
+            self.logger.warning(
+                "[EVAL]: evaluate_checkpoints is enabled but checkpoint saving is disabled; "
+                "skipping evaluation"
+            )
+            return
 
         # Run evaluation best-effort; never fail training because evaluation failed.
         try:

src/experiment_logger/config_logger.py

@@ -18,13 +18,6 @@ class LoggingConfig:
     upload_final_model: bool = False
     upload_checkpoints: bool = False
 
-    # Checkpoint evaluation (synchronous, in-process)
-    # When enabled, each saved checkpoint is evaluated headlessly and the results
-    # are appended to a CSV in the run's metrics/ folder.
-    evaluate_checkpoints: bool = False
-    eval_max_steps: int = 5000
-    eval_seed: int = 0
-
     hf_entity: str = ""
 
     def __post_init__(self):
@@ -39,19 +32,5 @@ class LoggingConfig:
                 "both 'track' and 'save_checkpoints' to also be True."
             )
 
-        if self.evaluate_checkpoints:
+        # NOTE: Checkpoint evaluation settings live under the project's
-            if not self.save_checkpoints:
+        # `evaluation` config group (see brittle_star_project.configs).
-                raise ValueError(
-                    "Configuration Error: 'evaluate_checkpoints' is True, but it requires "
-                    "'save_checkpoints' to also be True."
-                )
-            if self.checkpoint_frequency <= 0:
-                raise ValueError(
-                    "Configuration Error: 'evaluate_checkpoints' is True, but it requires "
-                    "'checkpoint_frequency' to be > 0."
-                )
-            if self.eval_max_steps <= 0:
-                raise ValueError(
-                    "Configuration Error: 'eval_max_steps' must be > 0 when "
-                    "'evaluate_checkpoints' is enabled."
-                )