data_prepare/data.py

import os
import yt_dlp
from datasets import load_dataset
from concurrent.futures import ThreadPoolExecutor, as_completed
from tqdm import tqdm
import time
import random

# ================= 配置区域 =================
DATASET_PATH = "/data/shuimu.chen/videomarathon"
SAVE_ROOT = "/data/shuimu.chen/videomarathon/downloaded_videos"

# 🔥 核心安全配置：挂 Cookie 时并发不能太高！建议 2，最高不要超过 4
MAX_WORKERS = 2  

# 你的 Cookie 路径
COOKIE_PATH = "/data/shuimu.chen/TimeSearch-R/data_prepare/cookies.txt" 

# Node 路径设置
NODE_BIN_DIR = "/data/shuimu.chen/TimeSearch-R/data_prepare/node-v24.12.0-linux-x64/bin"
os.environ["PATH"] = NODE_BIN_DIR + os.pathsep + os.environ["PATH"]
# ===========================================

def process_video(item):
    video_rel_path = item['rel_path']
    url = item['url']
    
    base_name = os.path.splitext(video_rel_path)[0]
    full_save_path_no_ext = os.path.join(SAVE_ROOT, base_name)
    final_file_path = full_save_path_no_ext + ".mp4"

    # 1. 检查已存在文件 (跳过 > 1KB 的正常文件)
    if os.path.exists(final_file_path):
        if os.path.getsize(final_file_path) > 1024:
            return "skipped"
        else:
            try:
                os.remove(final_file_path) # 删除 0KB 的坏文件
            except:
                pass

    os.makedirs(os.path.dirname(final_file_path), exist_ok=True)

    # 2. 核心配置：复刻命令行成功的环境
# ... (前面的路径配置保持不变) ...

    # 4. 终极配置：使用 Aria2 接管下载
    ydl_opts = {
        'format': 'bestvideo[height<=480][ext=mp4]+bestaudio[ext=m4a]/best[height<=480]/best',
        'merge_output_format': 'mp4',
        'outtmpl': f"{full_save_path_no_ext}.%(ext)s",
        
        # 🔥 核心修改：启用外部下载器 aria2c
        'external_downloader': 'aria2c',
        
        # 🔥 aria2c 的专用参数 (伪装性更强，断点续传更稳)
        'external_downloader_args': [
            '-c', '-j', '3', '-x', '3',  # 3个连接，3个线程 (不宜过多，否则触发IP限流)
            '-s', '3', '-k', '1M',       # 分块配置
            '--user-agent="Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"'
        ],

        # 这里的参数依然保留，用于辅助
        # 'cookiefile': COOKIE_PATH,
        
        # 强制 IPv4 (很多时候 IPv6 线路拥堵或风控更严)
        'source_address': '0.0.0.0',
        
        # 'retries': 2,
        # 'fragment_retries': 2,
        'quiet': False,
        'no_warnings': False,
        'ignoreerrors': False,
    }

    # ... (后面的 try-catch 逻辑保持不变) ...

    try:
        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
            ydl.download([url])
        
        # 3. 下载后校验
        if os.path.exists(final_file_path):
            if os.path.getsize(final_file_path) > 1024:
                return "success"
            else:
                # 下载了但是空文件 (可能是 IP 临时限流返回了空包)
                os.remove(final_file_path)
                return "failed_empty"
        else:
            return "failed_unavailable" # 视频不存在
            
    except Exception as e:
        # 这里仅捕获严重异常，普通的 "Video unavailable" 会被 ignoreerrors 处理
        # 如果想看报错可以把下面注释打开
        # print(f"Error: {e}")
        return "error"

def main():
    # 0. 环境自检
    print(f"🛠️ 环境检查...")
    if os.system("node -v") != 0:
        print("❌ 错误: Node 环境未生效，请检查路径！")
        return
    if not os.path.exists(COOKIE_PATH):
        print("❌ 错误: Cookie 文件不存在！")
        return

    # 1. 加载数据
    print(f"📂 正在加载数据集: {DATASET_PATH} ...")
    try:
        dataset = load_dataset(DATASET_PATH)
    except Exception as e:
        print(f"❌ 加载失败: {e}")
        return
        
    unique_videos = {}
    print("🔍 正在筛选 ActivityNet 数据...")
    
    for item in tqdm(dataset['train'], desc="Filtering"):
        source = item.get('data_source')
        q_type = item.get('question_type')
        video_rel_path = item.get('video')
        url = item.get('URL')
        
        if source == 'ActivityNet' and q_type and str(q_type).endswith('mc'):
            if video_rel_path and url:
                if video_rel_path not in unique_videos:
                    unique_videos[video_rel_path] = url

    tasks = [{'rel_path': k, 'url': v} for k, v in unique_videos.items()]
    print(f"✅ 筛选完成！共 {len(tasks)} 个唯一视频任务。")

    # 2. 开始下载
    print(f"🚀 启动下载 (Workers={MAX_WORKERS}, Cookie 已启用)...")
    results = {"success": 0, "skipped": 0, "failed_unavailable": 0, "failed_empty": 0, "error": 0}

    with ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor:
        future_to_video = {executor.submit(process_video, task): task for task in tasks}
        
        progress = tqdm(as_completed(future_to_video), total=len(tasks), desc="Downloading")
        for future in progress:
            status = future.result()
            
            # 简单的状态归类
            if status in results:
                results[status] += 1
            else:
                results["error"] += 1
                
            progress.set_postfix(
                ok=results['success'], 
                skip=results['skipped'], 
                fail=results['failed_unavailable'] + results['failed_empty'] + results['error']
            )

    print("\n" + "=" * 30)
    print(f"🎉 处理完毕")
    print(f"✅ 成功下载: {results['success']}")
    print(f"⏩ 跳过存在: {results['skipped']}")
    print(f"❌ 视频失效(死链): {results['failed_unavailable']}")
    print(f"⚠️ 下载为空(网络/限流): {results['failed_empty']}")
    print(f"🚫 其他错误: {results['error']}")
    print(f"保存在: {SAVE_ROOT}")

if __name__ == "__main__":
    main()