Deep CFR checkpoint CLI override 추가

.gitignore

@@ -7,6 +7,7 @@ dist/
 wheels/
 *.egg-info
 runs/
+/runs
 
 # Cython-generated sources
 src/coolrl_lost_cities/games/classic/game.c

src/coolrl_lost_cities/games/classic/deep_cfr/cli.py

@@ -60,6 +60,8 @@ def _resolve_resume_path(config: DeepCFRConfig, resume: str | None) -> str | Non
 
 def _train_overrides_from_args(args: argparse.Namespace) -> dict[str, Any]:
     overrides: dict[str, Any] = {}
+    if args.no_save and (args.save_latest_only or args.save_iteration_interval is not None):
+        raise ValueError("--no-save cannot be combined with checkpoint save overrides")
     run_overrides = overrides.setdefault("run", {})
     if args.iterations is not None:
         run_overrides["iterations"] = args.iterations
@@ -85,7 +87,19 @@ def _train_overrides_from_args(args: argparse.Namespace) -> dict[str, Any]:
     if args.eval_games is not None:
         overrides.setdefault("evaluation", {})["games"] = args.eval_games
     if args.no_save:
-        overrides.setdefault("checkpoint", {})["save_every_iteration"] = False
+        checkpoint_overrides = overrides.setdefault("checkpoint", {})
+        checkpoint_overrides["save_latest"] = False
+        checkpoint_overrides["save_every_iteration"] = False
+        checkpoint_overrides["save_iteration_interval"] = 0
+    if args.save_latest_only:
+        checkpoint_overrides = overrides.setdefault("checkpoint", {})
+        checkpoint_overrides["save_latest"] = True
+        checkpoint_overrides["save_latest_only"] = True
+        checkpoint_overrides["save_every_iteration"] = False
+    if args.save_iteration_interval is not None:
+        overrides.setdefault("checkpoint", {})["save_iteration_interval"] = (
+            args.save_iteration_interval
+        )
     if args.exact_resume:
         overrides.setdefault("checkpoint", {})["exact_resume"] = True
     return overrides
@@ -206,6 +220,8 @@ def main(argv: list[str] | None = None) -> None:
     train.add_argument("--eval-games", type=int)
     train.add_argument("--seed", type=int)
     train.add_argument("--no-save", action="store_true")
+    train.add_argument("--save-latest-only", action="store_true")
+    train.add_argument("--save-iteration-interval", type=int)
     train.set_defaults(func=train_command)
 
     evaluate = subparsers.add_parser("eval")

src/coolrl_lost_cities/games/classic/deep_cfr/config.py

@@ -214,6 +214,7 @@ class MemoryConfig(StrictModel):
 
 class CheckpointConfig(StrictModel):
     directory: str = "runs/deep_cfr/default"
+    save_latest: bool = True
     save_every_iteration: bool = True
     save_iteration_interval: int = 0
     save_latest_only: bool = False

src/coolrl_lost_cities/games/classic/deep_cfr/trainer.py

@@ -456,6 +456,7 @@ class DeepCFRTrainer:
         checkpoint_dir = self.run_dir
         if self._should_save_iteration(iteration) and not self.config.checkpoint.save_latest_only:
             self.save_checkpoint(checkpoint_dir / f"iteration_{iteration:05d}.pt", item)
+        if self.config.checkpoint.save_latest:
             self.save_checkpoint(checkpoint_dir / "latest.pt", item)
 
     def _start_run_logging(self) -> None:

tests/games/classic/test_deep_cfr_trainer.py

@@ -80,6 +80,8 @@ def test_deep_cfr_train_cli_count_overrides_disable_duration_limits() -> None:
             "eval_every": None,
             "eval_games": None,
             "no_save": True,
+            "save_latest_only": False,
+            "save_iteration_interval": None,
             "exact_resume": False,
         },
     )()
@@ -94,6 +96,36 @@ def test_deep_cfr_train_cli_count_overrides_disable_duration_limits() -> None:
     assert overridden.traversal.resolved_traversals_per_player() == 1
     assert overridden.traversal.resolved_num_workers() == 0
     assert overridden.checkpoint.save_every_iteration is False
+    assert overridden.checkpoint.save_latest is False
+
+
+def test_deep_cfr_train_cli_checkpoint_save_overrides() -> None:
+    args = type(
+        "Args",
+        (),
+        {
+            "iterations": None,
+            "max_hours": None,
+            "max_iterations": None,
+            "seed": None,
+            "traversals_per_iteration": None,
+            "num_workers": None,
+            "checkpoint_dir": None,
+            "eval_every": None,
+            "eval_games": None,
+            "no_save": False,
+            "save_latest_only": True,
+            "save_iteration_interval": 1,
+            "exact_resume": False,
+        },
+    )()
+
+    overridden = _with_overrides(DeepCFRConfig(), _train_overrides_from_args(args))
+
+    assert overridden.checkpoint.save_latest is True
+    assert overridden.checkpoint.save_latest_only is True
+    assert overridden.checkpoint.save_every_iteration is False
+    assert overridden.checkpoint.save_iteration_interval == 1
 
 
 def test_deep_cfr_resume_latest_resolution_uses_config_checkpoint_dir(tmp_path) -> None: