app.py

###  -----------------  ###
# Standard library imports
import os
import re
import sys
import copy
import warnings
from typing import Optional

# Third-party imports
import numpy as np
import torch
import torch.distributed as dist
import uvicorn
import librosa
import whisper
import requests
from fastapi import FastAPI
from pydantic import BaseModel
from decord import VideoReader, cpu
from transformers import AutoModelForCausalLM, AutoTokenizer

import gradio as gr
import spaces

# Local imports
from egogpt.model.builder import load_pretrained_model
from egogpt.mm_utils import get_model_name_from_path, process_images
from egogpt.constants import (
    IMAGE_TOKEN_INDEX, 
    DEFAULT_IMAGE_TOKEN, 
    IGNORE_INDEX,
    SPEECH_TOKEN_INDEX,
    DEFAULT_SPEECH_TOKEN
)
from egogpt.conversation import conv_templates, SeparatorStyle
import subprocess
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
from huggingface_hub import snapshot_download

# 下载整个模型文件夹到本地 ./llava-onevision-qwen2-7b-ov
# snapshot_download(
#     repo_id="lmms-lab/llava-onevision-qwen2-7b-ov",
#     local_dir="./llava-onevision-qwen2-7b-ov",  # 指定本地存储目录
#     ignore_patterns=["*.md", "*.txt"]  # 可以忽略一些不必要的文件（可选）
# )

from huggingface_hub import hf_hub_download

# Download the model checkpoint file (large-v3.pt)
ego_gpt_path = hf_hub_download(
    repo_id="EgoLife-v1/EgoGPT",
    filename="large-v3.pt",
    local_dir="./"
)


# pretrained = "/mnt/sfs-common/jkyang/EgoGPT/checkpoints/EgoGPT-llavaov-7b-EgoIT-109k-release"
# pretrained = "/mnt/sfs-common/jkyang/EgoGPT/checkpoints/EgoGPT-llavaov-7b-EgoIT-EgoLife-Demo"
pretrained = 'EgoLife-v1/EgoGPT'
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
device_map = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Add this initialization code before loading the model
def setup(rank, world_size):
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = '12377'

    # initialize the process group
    dist.init_process_group("gloo", rank=rank, world_size=world_size)

setup(0,1)
tokenizer, model, max_length = load_pretrained_model(pretrained,device_map=device_map)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device).eval()

title_markdown = """
<div style="display: flex; justify-content: space-between; align-items: center; background: linear-gradient(90deg, rgba(72,219,251,0.1), rgba(29,209,161,0.1)); border-radius: 20px; box-shadow: 0 4px 6px rgba(0,0,0,0.1); padding: 20px; margin-bottom: 20px;">
    <div style="display: flex; align-items: center;">
        <a href="https://egolife-ntu.github.io/" style="margin-right: 20px; text-decoration: none; display: flex; align-items: center;">
            <img src="https://egolife-ntu.github.io/egolife.png" alt="EgoLife" style="max-width: 100px; height: auto; border-radius: 15px; box-shadow: 0 2px 4px rgba(0,0,0,0.1);">
        </a>
        <div>
            <h1 style="margin: 0; background: linear-gradient(90deg, #48dbfb, #1dd1a1); -webkit-background-clip: text; -webkit-text-fill-color: transparent; font-size: 2.5em; font-weight: 700;">EgoLife</h1>
            <h2 style="margin: 10px 0; color: #2d3436; font-weight: 500;">Towards Egocentric Life Assistant</h2>
            <div style="display: flex; gap: 15px; margin-top: 10px;">
                <a href="https://egolife-ntu.github.io/" style="text-decoration: none; color: #48dbfb; font-weight: 500; transition: color 0.3s;">Project Page</a> |
                <a href="https://github.com/egolife-ntu/EgoGPT" style="text-decoration: none; color: #48dbfb; font-weight: 500; transition: color 0.3s;">Github</a> |
                <a href="https://huggingface.co/lmms-lab" style="text-decoration: none; color: #48dbfb; font-weight: 500; transition: color 0.3s;">Huggingface</a> |
                <a href="https://arxiv.org/" style="text-decoration: none; color: #48dbfb; font-weight: 500; transition: color 0.3s;">Paper</a> |
                <a href="https://x.com/" style="text-decoration: none; color: #48dbfb; font-weight: 500; transition: color 0.3s;">Twitter (X)</a>
            </div>
        </div>
    </div>
    <div style="text-align: right; margin-left: 20px;">
        <h1 style="margin: 0; background: linear-gradient(90deg, #48dbfb, #1dd1a1); -webkit-background-clip: text; -webkit-text-fill-color: transparent; font-size: 2.5em; font-weight: 700;">EgoGPT</h1>
        <h2 style="margin: 10px 0; background: linear-gradient(90deg, #48dbfb, #1dd1a1); -webkit-background-clip: text; -webkit-text-fill-color: transparent; font-size: 1.8em; font-weight: 600;">An Egocentric Video-Audio-Text Model<br>from EgoLife Project</h2>
    </div>
</div>
"""
notice_html = """
<div style="background-color: #f9f9f9; border-left: 5px solid #48dbfb; padding: 20px; margin-top: 20px; border-radius: 10px; box-shadow: 0 2px 4px rgba(0, 0, 0, 0.1);">
    <ul style="list-style-type: none; padding-left: 0; font-size: 1.1em; color: #555;">
        <li>- Due to hardware limitations on this demo page, we recommend users only try 10-second videos.</li>
        <li>- The demo model is used for the egocentric video captioning step for the EgoRAG framework. The recommended prompt includes:</li>
        <ul style="padding-left: 20px; margin-top: 10px; color: #333;">
            <li>Can you help me log everything I do and the key things I see, like a personal journal? Describe them in a natural style. 
            <li>Please provide your response using the first person, with "I" as the subject. Make sure the descriptions are detailed and natural.</li>
            <li>Can you write down important things I notice or interact with? Please respond in the first person, using "I" as the subject. Describe them in a natural style.</li>
        </ul>
    </ul>
</div>
"""

bibtext = """
### Citation
```
@article{yang2025egolife,
  title={EgoLife\: Towards Egocentric Life Assistant},
  author={The EgoLife Team},
  journal={arXiv preprint arXiv:25xxx},
  year={2025}
  }
```
"""

# cur_dir = os.path.dirname(os.path.abspath(__file__))
cur_dir = '.'


def time_to_frame_idx(time_int: int, fps: int) -> int:
    """
    Convert time in HHMMSSFF format (integer or string) to frame index.
    :param time_int: Time in HHMMSSFF format, e.g., 10483000 (10:48:30.00) or "10483000".
    :param fps: Frames per second of the video.
    :return: Frame index corresponding to the given time.
    """
    # Ensure time_int is a string for slicing
    time_str = str(time_int).zfill(
        8)  # Pad with zeros if necessary to ensure it's 8 digits

    hours = int(time_str[:2])
    minutes = int(time_str[2:4])
    seconds = int(time_str[4:6])
    frames = int(time_str[6:8])

    total_seconds = hours * 3600 + minutes * 60 + seconds
    total_frames = total_seconds * fps + frames  # Convert to total frames

    return total_frames

def split_text(text, keywords):
    # 创建一个正则表达式模式，将所有关键词用 | 连接，并使用捕获组
    pattern = '(' + '|'.join(map(re.escape, keywords)) + ')'
    # 使用 re.split 保留分隔符
    parts = re.split(pattern, text)
    # 去除空字符串
    parts = [part for part in parts if part]
    return parts

warnings.filterwarnings("ignore")

# Create FastAPI instance
app = FastAPI()
def load_video(
    video_path: Optional[str] = None,
    max_frames_num: int = 16,
    fps: int = 1,
    video_start_time: Optional[float] = None,
    start_time: Optional[float] = None,
    end_time: Optional[float] = None,
    time_based_processing: bool = False
) -> tuple:
    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    target_sr = 16000
    
    # Add new time-based processing logic
    if time_based_processing:
        # Initialize video reader
        vr = decord.VideoReader(video_path, ctx=decord.cpu(0), num_threads=1)
        total_frame_num = len(vr)

        # Get the actual FPS of the video
        video_fps = vr.get_avg_fps()

        # Convert time to frame index based on the actual video FPS
        video_start_frame = int(time_to_frame_idx(video_start_time, video_fps))
        start_frame = int(time_to_frame_idx(start_time, video_fps))
        end_frame = int(time_to_frame_idx(end_time, video_fps))

        print("start frame", start_frame)
        print("end frame", end_frame)

        # Ensure the end time does not exceed the total frame number
        if end_frame - start_frame > total_frame_num:
            end_frame = total_frame_num + start_frame

        # Adjust start_frame and end_frame based on video start time
        start_frame -= video_start_frame
        end_frame -= video_start_frame
        start_frame = max(0, int(round(start_frame)))  # 确保不会小于0
        end_frame = min(total_frame_num, int(round(end_frame))) # 确保不会超过总帧数
        start_frame = int(round(start_frame))
        end_frame = int(round(end_frame))

        # Sample frames based on the provided fps (e.g., 1 frame per second)
        frame_idx = [i for i in range(start_frame, end_frame) if (i - start_frame) % int(video_fps / fps) == 0]

        # Get the video frames for the sampled indices
        video = vr.get_batch(frame_idx).asnumpy()
        target_sr = 16000  # Set target sample rate to 16kHz
    
        # Load audio from video with resampling
        y, _ = librosa.load(video_path, sr=target_sr)
        
        # Convert time to audio samples (using 16kHz sample rate)
        start_sample = int(start_time * target_sr)
        end_sample = int(end_time * target_sr)
        
        # Extract audio segment
        speech = y[start_sample:end_sample]
    else:
        # Original processing logic
        speech, _ = librosa.load(video_path, sr=target_sr)
        total_frame_num = len(vr)
        avg_fps = round(vr.get_avg_fps() / fps)
        frame_idx = [i for i in range(0, total_frame_num, avg_fps)]
        
        if max_frames_num > 0:
            if len(frame_idx) > max_frames_num:
                uniform_sampled_frames = np.linspace(0, total_frame_num - 1, max_frames_num, dtype=int)
                frame_idx = uniform_sampled_frames.tolist()
        
        video = vr.get_batch(frame_idx).asnumpy()

    # Process audio
    speech = whisper.pad_or_trim(speech.astype(np.float32))
    speech = whisper.log_mel_spectrogram(speech, n_mels=128).permute(1, 0)
    speech_lengths = torch.LongTensor([speech.shape[0]])
    
    return video, speech, speech_lengths

class PromptRequest(BaseModel):
    prompt: str
    video_path: str = None
    max_frames_num: int = 16
    fps: int = 1
    video_start_time: float = None
    start_time: float = None
    end_time: float = None
    time_based_processing: bool = False

# @spaces.GPU(duration=120)
def generate_text(video_path, audio_track, prompt):
    max_frames_num = 30
    fps = 1
    # model.eval()

    # Video + speech branch
    conv_template = "qwen_1_5"  # Make sure you use correct chat template for different models
    question = f"<image>\n{prompt}"
    conv = copy.deepcopy(conv_templates[conv_template])
    conv.append_message(conv.roles[0], question)
    conv.append_message(conv.roles[1], None)
    prompt_question = conv.get_prompt()

    video, speech, speech_lengths = load_video(
        video_path=video_path,
        max_frames_num=max_frames_num,
        fps=fps,
    )
    speech=torch.stack([speech]).to("cuda").half()
    processor = model.get_vision_tower().image_processor
    processed_video = processor.preprocess(video, return_tensors="pt")["pixel_values"]
    image = [(processed_video, video[0].size, "video")]

    print(prompt_question)
    parts=split_text(prompt_question,["<image>","<speech>"])
    input_ids=[]
    for part in parts:
        if "<image>"==part:
            input_ids+=[IMAGE_TOKEN_INDEX]
        elif "<speech>"==part:
            input_ids+=[SPEECH_TOKEN_INDEX]
        else:
            input_ids+=tokenizer(part).input_ids
            
    input_ids = torch.tensor(input_ids,dtype=torch.long).unsqueeze(0).to(device)
    image_tensor = [image[0][0].half()]
    image_sizes = [image[0][1]]

    generate_kwargs={"eos_token_id":tokenizer.eos_token_id}
    print(input_ids)
    cont = model.generate(
        input_ids,
        images=image_tensor,
        image_sizes=image_sizes,
        speech=speech,
        speech_lengths=speech_lengths,
        do_sample=False,
        temperature=0.5,
        max_new_tokens=4096,
        modalities=["video"],
        **generate_kwargs
    )
    
    text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
    
    return text_outputs[0]

def extract_audio_from_video(video_path, audio_path=None):
    if audio_path:
        try:
            y, sr = librosa.load(audio_path, sr=8000, mono=True, res_type='kaiser_fast')
            return (sr, y)
        except Exception as e:
            print(f"Error loading audio from {audio_path}: {e}")
            return None
    if video_path is None:
        return None
    try:
        y, sr = librosa.load(video_path, sr=8000, mono=True, res_type='kaiser_fast')
        return (sr, y)
    except Exception as e:
        print(f"Error extracting audio from video: {e}")
        return None

head = """
<style>
/* Submit按钮默认和悬停效果 */
button.lg.secondary.svelte-1gz44hr {
    background-color: #ff9933 !important;
    transition: background-color 0.3s ease !important;
}

button.lg.secondary.svelte-1gz44hr:hover {
    background-color: #ff7777 !important;  /* 悬停时颜色加深 */
}

/* 确保按钮文字始终清晰可见 */
button.lg.secondary.svelte-1gz44hr span {
    color: white !important;
}

/* 隐藏表头中的第二列 */
.table-wrap .svelte-p5q82i th:nth-child(2) {
    display: none;
}

/* 隐藏表格内容中的第二列 */
.table-wrap .svelte-p5q82i td:nth-child(2) {
    display: none;
}

.table-wrap {
    max-height: 300px;
    overflow-y: auto;
}

</style>

<script>
// 新版同步控制代码
function syncMediaElements() {
    // 获取视频和音频元素
    const video = document.querySelector('[data-testid="Video-player"] video');
    const waveform = document.querySelector('#waveform');
    const audio = waveform?.querySelector('audio') || waveform?.shadowRoot?.querySelector('audio');

    // 如果任一元素不存在，则退出
    if (!video || !audio) return;

    // 解除旧的事件监听（避免重复绑定）
    video.removeEventListener('play', syncPlay);
    audio.removeEventListener('play', syncPlay);
    video.removeEventListener('timeupdate', syncVideoTime);
    audio.removeEventListener('timeupdate', syncAudioTime);

    // 定义同步函数
    function syncPlay(e) {
        if(e.target === video && audio.paused) audio.play();
        if(e.target === audio && video.paused) video.play();
    }

    function syncVideoTime() {
        if(!audio.seeking && Math.abs(video.currentTime - audio.currentTime) > 0.1){
            audio.currentTime = video.currentTime;
        }
    }

    function syncAudioTime() {
        if(!video.seeking && Math.abs(audio.currentTime - video.currentTime) > 0.1){
            video.currentTime = audio.currentTime;
        }
    }

    // 绑定新的事件监听
    video.addEventListener('play', syncPlay);
    audio.addEventListener('play', syncPlay);
    video.addEventListener('timeupdate', syncVideoTime);
    audio.addEventListener('timeupdate', syncAudioTime);

    // 同步暂停事件
    video.addEventListener('pause', () => audio.pause());
    audio.addEventListener('pause', () => video.pause());

    console.log('Media elements synced successfully!');
}

// 智能DOM观察器
const observer = new MutationObserver((mutations) => {
    mutations.forEach((mutation) => {
        if (mutation.addedNodes.length) {
            mutation.addedNodes.forEach((node) => {
                // 深度检查新增节点
                if (node.nodeType === 1) { // Element node
                    // 检查是否包含视频组件
                    if (node.querySelector?.('[data-testid="Video-player"]')) {
                        // 当视频组件出现时，开始查找音频
                        const audioObserver = new MutationObserver(() => {
                            if(document.querySelector('#waveform audio')) {
                                audioObserver.disconnect();
                                setTimeout(syncMediaElements, 500); // 等待组件完全加载
                            }
                        });
                        audioObserver.observe(document.body, {
                            childList: true,
                            subtree: true
                        });
                    }
                }
            });
        }
    });
});

// 开始观察整个文档
observer.observe(document.body, {
    childList: true,
    subtree: true
});

// 初始检查（应对组件已存在的情况）
setTimeout(() => {
    if(document.querySelector('[data-testid="Video-player"] video') && 
       document.querySelector('#waveform audio')){
        syncMediaElements();
    }
}, 1000);
</script>
"""

with gr.Blocks(head=head) as demo:
    gr.HTML(title_markdown)
    gr.HTML(notice_html)
    
    with gr.Row():
        with gr.Column():
            video_input = gr.Video(label="Video", autoplay=True, loop=True, format="mp4", width=600, height=400, show_label=False, elem_id='video')
            # Audio input synchronized with video playback
            audio_display = gr.Audio(label="Video Audio Track", autoplay=False, show_label=True, visible=True, interactive=False, elem_id="audio")
            text_input = gr.Textbox(label="Question", placeholder="Enter your message here...")
        
        with gr.Column():  # Create a separate column for output and examples
            output_text = gr.Textbox(label="Response", lines=14, max_lines=14)
            gr.Examples(
                examples=[
                    [f"{cur_dir}/bike.mp4", f"{cur_dir}/bike.mp3", "Can you tell me what I'm doing in short words. Describe them in a natural style."],
                    [f"{cur_dir}/bike.mp4", f"{cur_dir}/bike.mp3", "Can you tell me what I'm doing in short words. Describe them in a natural style."],
                    [f"{cur_dir}/bike.mp4", f"{cur_dir}/bike.mp3", "Can you tell me what I'm doing in short words. Describe them in a natural style."],
                    [f"{cur_dir}/bike.mp4", f"{cur_dir}/bike.mp3", "Can you tell me what I'm doing in short words. Describe them in a natural style."]
                ],
                inputs=[video_input, audio_display, text_input],
                outputs=[output_text]
            )

    # Add event handler for video changes
    video_input.change(
        fn=lambda video_path: extract_audio_from_video(video_path, audio_path=None),
        inputs=[video_input],
        outputs=[audio_display]
    )

    # Add event handler for video clear/delete
    def clear_outputs(video):
        if video is None:  # Video is cleared/deleted
            return ""
        return gr.skip()  # Keep existing text if video exists
    
    video_input.change(
        fn=clear_outputs,
        inputs=[video_input],
        outputs=[output_text]
    )

    # Add submit button and its event handler
    submit_btn = gr.Button("Submit")
    submit_btn.click(
        fn=generate_text,
        inputs=[video_input, audio_display, text_input],
        outputs=[output_text]
    )

    gr.Markdown(bibtext)
# Launch the Gradio app
if __name__ == "__main__":
    demo.launch(share=True)