Add ROMA frozen backbone extraction support

extract_ego4d_batch.py

@@ -87,6 +87,11 @@ def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
         "--model-id",
         default="/data/wangx/Models/Qwen2.5-Omni-3B",
     )
+    parser.add_argument(
+        "--backbone-loader",
+        choices=("qwen", "roma"),
+        default="qwen",
+    )
     parser.add_argument(
         "--start-index",
         type=int,
@@ -283,6 +288,8 @@ def clip_argv(
     values = [
         "--model-id",
         str(args.model_id),
+        "--backbone-loader",
+        args.backbone_loader,
         "--video",
         str(args.video_root / f"{record.uid}.mp4"),
         "--target",
@@ -373,6 +380,7 @@ def main(argv: Sequence[str] | None = None) -> None:
         dtype=args.dtype,
         device_map={"": args.device},
         use_flash_attention_2=args.flash_attention_2,
+        backbone_loader=args.backbone_loader,
     )
     model.requires_grad_(False)
     print(