saved_prediction_path always runs decoding (treat as raw prediction)

No fake TTA suffix needed — test pipeline checks saved_prediction_path
directly to decide whether to run decoding.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Donglai Wei

connectomics/training/lightning/model.py

@@ -604,9 +604,7 @@ def _load_cached_predictions(
                 pred = read_volume(str(pred_file), dataset="main")
                 if pred.ndim < 4:
                     pred = pred[np.newaxis, ...]
-                # Return a TTA-like suffix so the pipeline treats this as
-                # intermediate predictions (runs decoding), not final.
-                return pred, True, "_tta_x1_prediction.h5"
+                return pred, True, "_prediction.h5"
             else:
                 raise FileNotFoundError(
                     f"inference.saved_prediction_path not found: {pred_file}"

connectomics/training/lightning/test_pipeline.py

@@ -977,10 +977,15 @@ def run_test_step(module, batch: Dict[str, torch.Tensor], batch_idx: int) -> STE
         mode,
     )
 
-    # *_decoding*.h5 = already decoded (skip decoding)
-    # *_prediction*.h5 with TTA suffix = intermediate (run decoding)
+    # Determine whether loaded predictions need decoding:
+    # - saved_prediction_path → always run decoding (it's raw affinity)
+    # - *_decoding*.h5 → already decoded, skip decoding
+    # - *_prediction*.h5 with TTA suffix → intermediate, run decoding
+    # - other → final, skip decoding
+    _saved_pred = getattr(getattr(module.cfg, "inference", None), "saved_prediction_path", "")
+    _from_saved_path = bool(loaded_from_file and _saved_pred)
     _is_decoding_file = loaded_from_file and "_decoding" in (loaded_suffix or "")
-    loaded_final_predictions = loaded_from_file and (
+    loaded_final_predictions = loaded_from_file and not _from_saved_path and (
         _is_decoding_file or not is_tta_cache_suffix(loaded_suffix)
     )
     loaded_intermediate_predictions = loaded_from_file and not loaded_final_predictions