Generator runner improvement

README.md

@@ -123,10 +123,3 @@ The default policy in `pipeline/defaults/vast.json` now targets:
 - `<= $0.20/hour`
 - sorted by `dlperf` descending
 - uses `vastai/pytorch:latest` as the default image
-
-## Diagnostics
-
-```bash
-python3 classifier/tools/analyze.py classifier/configs/phase2/p2_resnet18_facecrop.json
-python3 classifier/tools/ensemble.py classifier/configs/phase4/p4_ensemble.json
-```

generator/run.py

@@ -37,7 +37,7 @@ def main(config_path, *, data_dir_override=None, output_root="generator/outputs"
 
     cfg = load_config(config_path)
 
-    run_name    = cfg["run_name"]
+    run_name    = cfg.get("run_name", Path(config_path).stem)
     device      = "cuda" if torch.cuda.is_available() else "cpu"
     data_dir    = data_dir_override or cfg.get("data_dir", "data")
     output_root = Path(output_root)

pipeline/cli.py

@@ -7,7 +7,7 @@ from pipeline.orchestrator import EphemeralVastRunner, RunOptions
 # Accept one or more config files, or a single directory (all *.json inside, sorted)
 def _resolve_configs(raw: list[str]) -> list[Path]:
     if len(raw) == 1 and Path(raw[0]).is_dir():
-        configs = sorted(Path(raw[0]).glob("*.json"))
+        configs = sorted(p for p in Path(raw[0]).glob("*.json") if not p.name.startswith("_"))
         if not configs:
             raise ValueError(f"No JSON configs found in directory: {raw[0]}")
         return configs

pipeline/orchestrator.py

@@ -429,21 +429,22 @@ class EphemeralVastRunner:
         seen.remove(resolved)
         return self._deep_merge_dicts(base_cfg, cfg)
 
-    # Return a stable signature used to detect duplicate training configs
+    # Resolve a config the same way the runners do: extends chain first,
-    def _normalized_config_signature(self, config_path: Path) -> str:
+    # then shared.json overlaid underneath. Mirrors load_config() in
-        run_script, _ = self._module_for_config(config_path)
+    # classifier/src/utils/config.py and generator/src/utils/config.py — keep
-        # Generator configs do not currently use shared/extends inheritance.
+    # in sync if merge semantics ever change.
-        if run_script.startswith("generator/"):
+    def _load_config_merged(self, config_path: Path) -> dict[str, Any]:
-            with open(config_path, encoding="utf-8") as fh:
-                cfg = json.load(fh)
-        else:
         cfg = self._load_config_with_extends(config_path)
         shared_path = config_path.parent.parent / "shared.json"
         if shared_path.exists():
             with open(shared_path, encoding="utf-8") as fh:
                 shared_cfg = json.load(fh)
             cfg = self._deep_merge_dicts(shared_cfg, cfg)
+        return cfg
 
+    # Return a stable signature used to detect duplicate training configs
+    def _normalized_config_signature(self, config_path: Path) -> str:
+        cfg = self._load_config_merged(config_path)
         # run_name should not influence whether two configs are equivalent to train.
         cfg.pop("run_name", None)
         return json.dumps(cfg, sort_keys=True, separators=(",", ":"))