gg

app.py

@@ -70,9 +70,12 @@ with gr.Blocks(title="VideoLLM-online", css=css) as demo:
         gr_video_time = gr.Number(value=0, visible=False)
         gr_liveinfer_queue_refresher = gr.Number(value=False, visible=False)
 
-        def gr_video_input(src_video_path, history, video_time, gate):
+        def gr_video_change(src_video_path, history, video_time, gate):
             name, ext = os.path.splitext(src_video_path)
             ffmpeg_video_path = os.path.join('demo/assets/cache', name + f'_{liveinfer.frame_fps}fps_{liveinfer.frame_resolution}' + ext)
+            if ffmpeg_video_path == liveinfer.video_path:
+                return 
+            liveinfer.video_path = ffmpeg_video_path
             if not os.path.exists(ffmpeg_video_path):
                 os.makedirs(os.path.dirname(ffmpeg_video_path), exist_ok=True)
                 ffmpeg_once(src_video_path, ffmpeg_video_path, fps=liveinfer.frame_fps, resolution=liveinfer.frame_resolution)
@@ -83,14 +86,13 @@ with gr.Blocks(title="VideoLLM-online", css=css) as demo:
             if query or response:
                 history.append((query, response))
             return history, video_time + 1 / liveinfer.frame_fps, not gate
-        gr_video.input(
-            gr_video_input, inputs=[gr_video, gr_chat_interface.chatbot, gr_video_time, gr_liveinfer_queue_refresher], 
+        gr_video.change(
+            gr_video_change, inputs=[gr_video, gr_chat_interface.chatbot, gr_video_time, gr_liveinfer_queue_refresher], 
             outputs=[gr_chat_interface.chatbot, gr_video_time, gr_liveinfer_queue_refresher]
         )
         
         def gr_video_time_change(_, video_time):
-            time.sleep(1/30)
-            video_time += 1/30
+            video_time += 1 / liveinfer.frame_fps
             liveinfer.input_video_stream(video_time)
             print(video_time)
             return video_time

inference.py

@@ -85,6 +85,7 @@ class LiveInfer:
     def reset(self, ):
         self.query_queue = collections.deque()
         self.frame_embeds_queue = collections.deque()
+        self.video_path = None
         self.video_time = 0
         self.last_frame_idx = -1
         self.video_tensor = None