import os, torchvision, transformers, subprocess, huggingface_hub, time from functools import partial import gradio as gr from inference import LiveInfer logger = transformers.logging.get_logger('liveinfer') huggingface_hub.login(os.getenv('HF_TOKEN')) # python -m demo.app --resume_from_checkpoint ... liveinfer = LiveInfer() def ffmpeg_once(src_path: str, dst_path: str, *, fps: int = None, resolution: int = None, pad: str = '#000000', mode='bicubic'): os.makedirs(os.path.dirname(dst_path), exist_ok=True) command = [ './ffmpeg/ffmpeg', '-y', '-sws_flags', mode, '-i', src_path, '-an', '-threads', '10', ] if fps is not None: command += ['-r', str(fps)] if resolution is not None: command += ['-vf', f"scale='if(gt(iw\\,ih)\\,{resolution}\\,-2)':'if(gt(iw\\,ih)\\,-2\\,{resolution})',pad={resolution}:{resolution}:(ow-iw)/2:(oh-ih)/2:color='{pad}'"] command += [dst_path] subprocess.run(command, check=True) css = """ #gr_title {text-align: center;} #gr_video {max-height: 480px;} #gr_chatbot {max-height: 480px;} """ with gr.Blocks(title="VideoLLM-online", css=css) as demo: gr.Markdown("# VideoLLM-online: Online Video Large Language Model for Streaming Video", elem_id='gr_title') with gr.Row(): with gr.Column(): gr_video = gr.Video(label="video stream", elem_id="gr_video", visible=True, sources=['upload'], autoplay=True) gr_examples = gr.Examples( examples=[["cooking.mp4"], ["bicycle.mp4"]], inputs=gr_video, outputs=gr_video, label="Examples" ) gr.Markdown("## Tips:") gr.Markdown("- When you upload/click a video, the model starts processing the video stream. You can input a query before or after that, at any point during the video as you like.") gr.Markdown("- **Gradio refreshes the chatbot box to update the answer, which will delay the program. If you want to enjoy faster demo as we show in teaser video, please use https://github.com/showlab/videollm-online/blob/main/demo/cli.py.**") gr.Markdown("- This work is primarily done at a university, and our resources are limited. Our model is trained with limited data, so it cannot solve very complicated questions and may produce hallucination. However, we have seen the potential of 'learning in streaming'. We are working on new data method to scale streaming dialogue data to our next model.") with gr.Column(): gr_chat_interface = gr.ChatInterface( fn=liveinfer.input_query_stream, chatbot=gr.Chatbot( elem_id="gr_chatbot", label='chatbot', avatar_images=('user_avatar.png', 'assistant_avatar.png'), render=False ), examples=['Please narrate the video in real time.', 'Please describe what I am doing.', 'Could you summarize what have been done?', 'Hi, guide me the next step.'], ) def gr_frame_token_interval_threshold_change(frame_token_interval_threshold): liveinfer.frame_token_interval_threshold = frame_token_interval_threshold gr_frame_token_interval_threshold = gr.Slider(minimum=0, maximum=1, step=0.05, value=liveinfer.frame_token_interval_threshold, interactive=True, label="Streaming Threshold") gr_frame_token_interval_threshold.change(gr_frame_token_interval_threshold_change, inputs=[gr_frame_token_interval_threshold]) gr_video_time = gr.Number(value=0, visible=False) gr_liveinfer_queue_refresher = gr.Number(value=False, visible=False) 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) logger.warning(f'{src_video_path} -> {ffmpeg_video_path}, {liveinfer.frame_fps} FPS, {liveinfer.frame_resolution} Resolution') liveinfer.load_video(ffmpeg_video_path) liveinfer.input_video_stream(0) query, response = liveinfer() if query or response: history.append((query, response)) return history, video_time + 1 / liveinfer.frame_fps, not gate 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): video_time += 1 / liveinfer.frame_fps liveinfer.input_video_stream(video_time) print(video_time) return video_time gr_video_time.change(gr_video_time_change, [gr_video, gr_video_time], [gr_video_time]) def gr_liveinfer_queue_refresher_change(history): while True: query, response = liveinfer() if query or response: history[-1][1] += f'\n{response}' print(history) yield history gr_liveinfer_queue_refresher.change(gr_liveinfer_queue_refresher_change, inputs=[gr_chat_interface.chatbot], outputs=[gr_chat_interface.chatbot]) demo.queue() demo.launch(share=False, debug=True)