Support source maps for renamed source videos

benchmarks/edit/build_six_method_manifest.py

@@ -21,7 +21,54 @@ def read_jsonl(path: Path) -> list[dict]:
     return rows
 
 
-def resolve_source_video(repo_root: Path, raw_path: str, source_roots: list[Path]) -> Path:
+def load_source_maps(paths: list[Path], repo_root: Path) -> tuple[dict[str, Path], dict[str, Path]]:
+    """Load optional source maps keyed by sample_id and by original path/basename."""
+    by_id: dict[str, Path] = {}
+    by_key: dict[str, Path] = {}
+    for path in paths:
+        path = path if path.is_absolute() else repo_root / path
+        rows = read_jsonl(path)
+        for row in rows:
+            video = (
+                row.get("path")
+                or row.get("video")
+                or row.get("source_video")
+                or row.get("source_path")
+                or row.get("control_video")
+            )
+            if not video:
+                continue
+            video_path = Path(str(video))
+            if not video_path.is_absolute():
+                video_path = repo_root / video_path
+            for key_name in ("sample_id", "id"):
+                if row.get(key_name) is not None:
+                    by_id[str(row[key_name])] = video_path
+            for key_name in ("old_path", "old_video", "control_video", "source_video"):
+                if row.get(key_name) is not None:
+                    key = str(row[key_name])
+                    by_key[key] = video_path
+                    by_key[Path(key).name] = video_path
+            by_key[video_path.name] = video_path
+    return by_id, by_key
+
+
+def resolve_source_video(
+    repo_root: Path,
+    sample_id: str,
+    raw_path: str,
+    source_roots: list[Path],
+    source_by_id: dict[str, Path],
+    source_by_key: dict[str, Path],
+) -> Path:
+    if sample_id in source_by_id:
+        return source_by_id[sample_id]
+    if raw_path in source_by_key:
+        return source_by_key[raw_path]
+    raw_name = Path(raw_path).name
+    if raw_name in source_by_key:
+        return source_by_key[raw_name]
+
     source_video = Path(raw_path)
     if not source_video.is_absolute():
         source_video = repo_root / source_video
@@ -52,6 +99,8 @@ def build_split(
     outputs_root: Path,
     output_path: Path,
     source_roots: list[Path],
+    source_by_id: dict[str, Path],
+    source_by_key: dict[str, Path],
 ) -> None:
     samples = read_jsonl(samples_path)
     output_path.parent.mkdir(parents=True, exist_ok=True)
@@ -61,7 +110,14 @@ def build_split(
     with output_path.open("w", encoding="utf-8") as handle:
         for sample in samples:
             sample_id = str(sample["id"])
-            source_video = resolve_source_video(repo_root, str(sample["control_video"]), source_roots)
+            source_video = resolve_source_video(
+                repo_root,
+                sample_id,
+                str(sample["control_video"]),
+                source_roots,
+                source_by_id,
+                source_by_key,
+            )
             if not source_video.exists():
                 missing_sources.add(str(source_video))
             for method in METHODS:
@@ -102,11 +158,22 @@ def main() -> None:
         default=[],
         help="Optional directory to search for original source mp4s by basename when eval_samples control_video is missing.",
     )
+    parser.add_argument(
+        "--source-map",
+        type=Path,
+        action="append",
+        default=[],
+        help=(
+            "Optional JSONL map. Each row may contain sample_id/id and path/video/source_video/source_path. "
+            "It can also contain old_path/old_video/control_video to map old names to current hashed paths."
+        ),
+    )
     args = parser.parse_args()
 
     repo_root = args.repo_root.resolve()
     base = repo_root / "out/edit_model_face_stage1"
     output_dir = args.output_dir if args.output_dir.is_absolute() else repo_root / args.output_dir
+    source_by_id, source_by_key = load_source_maps(args.source_map, repo_root)
 
     jobs = [
         (
@@ -125,7 +192,16 @@ 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, args.source_root)
+        build_split(
+            repo_root,
+            split,
+            samples_path,
+            outputs_root,
+            output_path,
+            args.source_root,
+            source_by_id,
+            source_by_key,
+        )
 
 
 if __name__ == "__main__":

benchmarks/edit/run_traditional_metrics.py

@@ -101,6 +101,25 @@ def align_frame_arrays(source: np.ndarray, edited: np.ndarray) -> tuple[np.ndarr
     return source[:n], edited[:n]
 
 
+def append_error(existing: Any, message: str) -> str:
+    existing_text = str(existing) if existing else ""
+    return f"{existing_text} | {message}" if existing_text else message
+
+
+def tensor_from_model_output(output: Any) -> Any:
+    """Handle transformers versions that return model outputs instead of tensors."""
+    if hasattr(output, "norm"):
+        return output
+    for attr in ("image_embeds", "text_embeds", "pooler_output"):
+        value = getattr(output, attr, None)
+        if value is not None:
+            return value
+    hidden = getattr(output, "last_hidden_state", None)
+    if hidden is not None:
+        return hidden[:, 0]
+    raise TypeError(f"cannot extract tensor from model output {type(output).__name__}")
+
+
 def global_ssim_gray(a: np.ndarray, b: np.ndarray) -> float:
     """Simple global SSIM over all sampled grayscale pixels, no skimage dependency."""
     a_gray = 0.299 * a[..., 0] + 0.587 * a[..., 1] + 0.114 * a[..., 2]
@@ -172,7 +191,7 @@ class ClipMetrics:
             for start in range(0, len(frames), self.batch_size):
                 batch = frames[start : start + self.batch_size]
                 inputs = self.processor(images=batch, return_tensors="pt").to(self.device)
-                feats = self.model.get_image_features(**inputs)
+                feats = tensor_from_model_output(self.model.get_image_features(**inputs))
                 feats = feats / feats.norm(dim=-1, keepdim=True).clamp_min(1e-12)
                 chunks.append(feats)
         return self.torch.cat(chunks, dim=0)
@@ -180,7 +199,7 @@ class ClipMetrics:
     def text_feature(self, text: str):
         with self.torch.inference_mode():
             inputs = self.processor(text=[text], return_tensors="pt", padding=True, truncation=True).to(self.device)
-            feat = self.model.get_text_features(**inputs)
+            feat = tensor_from_model_output(self.model.get_text_features(**inputs))
             feat = feat / feat.norm(dim=-1, keepdim=True).clamp_min(1e-12)
             return feat[0]
 
@@ -288,7 +307,7 @@ class LaionAestheticMetrics:
             for start in range(0, len(edited_frames), self.batch_size):
                 batch = edited_frames[start : start + self.batch_size]
                 inputs = self.processor(images=batch, return_tensors="pt").to(self.device)
-                feats = self.clip.get_image_features(**inputs)
+                feats = tensor_from_model_output(self.clip.get_image_features(**inputs))
                 feats = feats / feats.norm(dim=-1, keepdim=True).clamp_min(1e-12)
                 if feats.shape[-1] != self.linear.in_features:
                     raise RuntimeError(
@@ -378,9 +397,11 @@ def init_optional_scorer(name: str, factory):
     return scorer, ""
 
 
-def append_error(existing: Any, message: str) -> str:
-    existing_text = str(existing) if existing else ""
-    return f"{existing_text} | {message}" if existing_text else message
+def compute_optional(result: dict[str, Any], name: str, callback) -> None:
+    try:
+        result.update(callback())
+    except Exception as exc:
+        result["error"] = append_error(result.get("error"), f"{name}_failed={type(exc).__name__}: {exc}")
 
 
 def configure_cache_dirs(torch_home: Path | None, hf_home: Path | None, xdg_cache_home: Path | None) -> None:
@@ -518,15 +539,19 @@ def main() -> None:
                     )
 
                 if clip is not None:
-                    result.update(clip.compute(source_frames, edited_frames, str(row["instruction"])))
+                    compute_optional(
+                        result,
+                        "clip",
+                        lambda: clip.compute(source_frames, edited_frames, str(row["instruction"])),
+                    )
                 if dino is not None:
-                    result.update(dino.compute(edited_frames))
+                    compute_optional(result, "dino", lambda: dino.compute(edited_frames))
                 if aesthetic is not None:
-                    result.update(aesthetic.compute(edited_frames))
+                    compute_optional(result, "laion_aesthetic", lambda: aesthetic.compute(edited_frames))
                 if pyiqa_metrics is not None:
-                    result.update(pyiqa_metrics.compute(edited_frames))
+                    compute_optional(result, "pyiqa", lambda: pyiqa_metrics.compute(edited_frames))
                 if lpips_metric is not None and source_frames is not None:
-                    result.update(lpips_metric.compute(source_frames, edited_frames))
+                    compute_optional(result, "lpips", lambda: lpips_metric.compute(source_frames, edited_frames))
             except Exception as exc:
                 row_error = repr(exc)
                 result["error"] = append_error(result.get("error"), row_error)

benchmarks/edit/traditional_eval_notes.md

@@ -519,6 +519,29 @@ python3 reference/benchmarks/edit/build_six_method_manifest.py \
   --source-root /path/to/source/videos
 ```
 
+如果当前机器上的源视频已经被重命名成 hash 文件名，需要提供一个 JSONL 映射文件，至少包含 `sample_id` 和当前视频路径：
+
+```jsonl
+{"sample_id": "val_0000", "path": "/inspire/hdd/project/.../datas/ditto_face/low/0a1ec4e5fd6d9079ad1be0ea03cbcfed.mp4"}
+```
+
+然后生成 manifest：
+
+```bash
+python3 reference/benchmarks/edit/build_six_method_manifest.py \
+  --repo-root . \
+  --output-dir out/edit_model_face_stage1/traditional_eval_manifests \
+  --source-map out/edit_model_face_stage1/source_video_map.jsonl
+```
+
+`--source-map` 也支持这些字段名：
+
+- 样本键：`sample_id` 或 `id`
+- 当前视频路径：`path`、`video`、`source_video`、`source_path` 或 `control_video`
+- 旧路径键：`old_path`、`old_video`、`control_video` 或 `source_video`
+
+注意：如果只有一个目录里一堆 hash mp4，但没有 `sample_id -> hash mp4` 或 `旧文件名 -> hash mp4` 映射，脚本不能安全自动匹配。此时要先从生成数据时的 records/metadata 找回映射。
+
 生成后快速检查路径：
 
 ```bash