Validate source paths in six-method manifest builder

benchmarks/edit/build_six_method_manifest.py

@@ -21,18 +21,53 @@ def read_jsonl(path: Path) -> list[dict]:
     return rows
 
 
-def build_split(repo_root: Path, split: str, samples_path: Path, outputs_root: Path, output_path: Path) -> None:
+def resolve_source_video(repo_root: Path, raw_path: str, source_roots: list[Path]) -> Path:
+    source_video = Path(raw_path)
+    if not source_video.is_absolute():
+        source_video = repo_root / source_video
+    if source_video.exists():
+        return source_video
+
+    # Some machines only have downloaded/evaluated outputs, not the original
+    # absolute data tree. Let callers provide roots that contain source mp4s.
+    for root in source_roots:
+        root = root if root.is_absolute() else repo_root / root
+        candidate = root / source_video.name
+        if candidate.exists():
+            return candidate
+    for root in source_roots:
+        root = root if root.is_absolute() else repo_root / root
+        if not root.exists():
+            continue
+        matches = list(root.rglob(source_video.name))
+        if matches:
+            return matches[0]
+    return source_video
+
+
+def build_split(
+    repo_root: Path,
+    split: str,
+    samples_path: Path,
+    outputs_root: Path,
+    output_path: Path,
+    source_roots: list[Path],
+) -> None:
     samples = read_jsonl(samples_path)
     output_path.parent.mkdir(parents=True, exist_ok=True)
     count = 0
+    missing_sources: set[str] = set()
+    missing_edits: list[str] = []
     with output_path.open("w", encoding="utf-8") as handle:
         for sample in samples:
             sample_id = str(sample["id"])
-            source_video = Path(str(sample["control_video"]))
-            if not source_video.is_absolute():
-                source_video = repo_root / source_video
+            source_video = resolve_source_video(repo_root, str(sample["control_video"]), source_roots)
+            if not source_video.exists():
+                missing_sources.add(str(source_video))
             for method in METHODS:
                 edited_video = outputs_root / method / f"{sample_id}.mp4"
+                if not edited_video.exists():
+                    missing_edits.append(str(edited_video))
                 row = {
                     "split": split,
                     "sample_id": sample_id,
@@ -46,12 +81,27 @@ def build_split(repo_root: Path, split: str, samples_path: Path, outputs_root: P
                 handle.write(json.dumps(row, ensure_ascii=False) + "\n")
                 count += 1
     print(f"{split}: wrote {count} rows -> {output_path}")
+    if missing_sources:
+        print(f"{split}: warning missing source videos: {len(missing_sources)} unique paths")
+        for path in sorted(missing_sources)[:20]:
+            print(f"  missing source: {path}")
+    if missing_edits:
+        print(f"{split}: warning missing edited videos: {len(missing_edits)} rows")
+        for path in missing_edits[:20]:
+            print(f"  missing edited: {path}")
 
 
 def main() -> None:
     parser = argparse.ArgumentParser()
     parser.add_argument("--repo-root", type=Path, default=Path.cwd())
     parser.add_argument("--output-dir", type=Path, default=Path("out/edit_model_face_stage1/traditional_eval_manifests"))
+    parser.add_argument(
+        "--source-root",
+        type=Path,
+        action="append",
+        default=[],
+        help="Optional directory to search for original source mp4s by basename when eval_samples control_video is missing.",
+    )
     args = parser.parse_args()
 
     repo_root = args.repo_root.resolve()
@@ -75,7 +125,7 @@ def main() -> None:
     for split, samples_path, outputs_root, output_path in jobs:
         if not samples_path.exists():
             raise FileNotFoundError(samples_path)
-        build_split(repo_root, split, samples_path, outputs_root, output_path)
+        build_split(repo_root, split, samples_path, outputs_root, output_path, args.source_root)
 
 
 if __name__ == "__main__":

benchmarks/edit/traditional_eval_notes.md

@@ -500,6 +500,8 @@ PY
 
 ### 1. 生成 manifest
 
+`edited_video` 固定来自 `out/edit_model_face_stage1/eval_outputs*`。`source_video` 应该是真实原视频，默认来自 `eval_samples/*.jsonl` 里的 `control_video`；不能用某个 method 的 edited 输出冒充 source，否则 `PSNR/SSIM/LPIPS/source-edit CLIP` 会变成错误比较。
+
 ```bash
 cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/low-high-new
 
@@ -508,6 +510,35 @@ python3 reference/benchmarks/edit/build_six_method_manifest.py \
   --output-dir out/edit_model_face_stage1/traditional_eval_manifests
 ```
 
+如果真实源视频在另一份目录里，给 builder 加 `--source-root`，脚本会按 `control_video` 的文件名递归搜索并重写 manifest：
+
+```bash
+python3 reference/benchmarks/edit/build_six_method_manifest.py \
+  --repo-root . \
+  --output-dir out/edit_model_face_stage1/traditional_eval_manifests \
+  --source-root /path/to/source/videos
+```
+
+生成后快速检查路径：
+
+```bash
+python3 - <<'PY'
+import json
+from pathlib import Path
+
+for manifest in [
+    Path("out/edit_model_face_stage1/traditional_eval_manifests/val20.jsonl"),
+    Path("out/edit_model_face_stage1/traditional_eval_manifests/val100.jsonl"),
+]:
+    rows = [json.loads(x) for x in manifest.read_text().splitlines() if x.strip()]
+    missing_sources = sorted({r["source_video"] for r in rows if not Path(r["source_video"]).exists()})
+    missing_edits = [r["edited_video"] for r in rows if not Path(r["edited_video"]).exists()]
+    print(manifest, "rows", len(rows), "missing_sources", len(missing_sources), "missing_edits", len(missing_edits))
+    for path in missing_sources[:10]:
+        print("  missing source:", path)
+PY
+```
+
 ### 2. 主环境一次性跑 CLIP / DINO / LAION / MUSIQ / NIQE / LPIPS
 
 这些指标可以在主环境跑。如果 `pyiqa` 或 `lpips` 没装，先安装：