more updates

annotation_example/gradio_ui/sv_sam.py

@@ -307,7 +307,7 @@ def _preprocess_video(
     first_bgr: np.ndarray = cv2.imread(str(first_frame_path))
 
     progress(0.5, desc="Initializing SAM")
-    VIDEO_SAM_PREDICTOR: SAM2VideoPredictor = SAM2VideoPredictor.from_pretrained("facebook/sam2-hiera-tiny", device="cpu")
+    VIDEO_SAM_PREDICTOR: SAM2VideoPredictor = SAM2VideoPredictor.from_pretrained("facebook/sam2-hiera-tiny")
     with torch.inference_mode():
         inference_state = VIDEO_SAM_PREDICTOR.init_state(video_path=tmp_frames_dir)
         VIDEO_SAM_PREDICTOR.reset_state(inference_state)
@@ -382,6 +382,7 @@ def get_initial_mask(
         f"Include points: {keypoint_container.include_points.shape}, Exclude points: {keypoint_container.exclude_points.shape}"
     )
 
+    VIDEO_SAM_PREDICTOR: SAM2VideoPredictor = SAM2VideoPredictor.from_pretrained("facebook/sam2-hiera-tiny")
     with torch.inference_mode():
         frame_idx: int
         object_ids: list
@@ -439,7 +440,8 @@ def propagate_mask(
         f"{log_paths['pinhole_path']}/exclude",
         rr.Clear(recursive=True),
     )
-
+    
+    VIDEO_SAM_PREDICTOR: SAM2VideoPredictor = SAM2VideoPredictor.from_pretrained("facebook/sam2-hiera-tiny")
     with torch.inference_mode():
         frame_idx: int
         object_ids: list