dataprocess_code/split_clip_from_s3.py

import os
import json
import subprocess
import re
import boto3
from botocore.config import Config
from botocore.exceptions import ClientError
import tempfile
import shutil
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed
from multiprocessing import Manager, cpu_count
from tqdm import tqdm
import threading


part_id = 2


# S3/Tigris 配置
S3_ENDPOINT_URL = "https://t3.storage.dev"
AWS_ACCESS_KEY_ID = "tid_cqKPLHboixMUUQxq_ImANLFwrehWmWZHlEaPZXzXNbKxf_fugg"
AWS_SECRET_ACCESS_KEY = "tsec_CXLclBpmOD2blVqdL+smpI52cOxQiXs-pH-INnfU6yfhc1MAajUTpI7xWO+5YAyLwyXjpq"
AWS_REGION = "auto"
S3_BUCKET = "youtube-downloads"
S3_PREFIX = "VideoRaw"

# 路径配置
MERGED_ANNO_DIR = f"/share/eason/SpeakerVid/SpeakerVid-5M-Dataset/merged_anno_1_6/split1-6/split_{part_id}/"
MISSING_VIDEOS_FILE = f"/share/eason/SpeakerVid/missing_videos_{part_id}.txt"

# S3 输出配置
S3_OUTPUT_PREFIX = "SenceCutV3"

# 视频处理配置
MAX_VIDEO_RES = 1920

# 并发配置 (160 CPU)
NUM_VIDEO_WORKERS = 32      # 并行处理的视频数量
NUM_CLIP_WORKERS = 4        # 每个视频内并行处理的clip数量

# 线程本地存储，用于每个线程/进程持有自己的S3客户端
_thread_local = threading.local()


def get_s3_client():
    """获取线程/进程本地的 S3 客户端"""
    if not hasattr(_thread_local, 's3_client'):
        _thread_local.s3_client = boto3.client(
            's3',
            endpoint_url=S3_ENDPOINT_URL,
            aws_access_key_id=AWS_ACCESS_KEY_ID,
            aws_secret_access_key=AWS_SECRET_ACCESS_KEY,
            region_name=AWS_REGION,
            config=Config(signature_version='s3v4')
        )
    return _thread_local.s3_client


def check_s3_file_exists(bucket, key):
    """检查 S3 文件是否存在"""
    s3_client = get_s3_client()
    try:
        s3_client.head_object(Bucket=bucket, Key=key)
        return True
    except ClientError as e:
        if e.response['Error']['Code'] == '404':
            return False
        raise


def download_from_s3(bucket, key, local_path):
    """从 S3 下载文件"""
    s3_client = get_s3_client()
    try:
        s3_client.download_file(bucket, key, local_path)
        return True
    except ClientError as e:
        print(f"下载失败: {e}")
        return False


def upload_to_s3(bucket, local_path, s3_key):
    """上传文件到 S3"""
    s3_client = get_s3_client()
    try:
        s3_client.upload_file(local_path, bucket, s3_key)
        return True
    except ClientError as e:
        print(f"上传失败: {e}")
        return False


def organize_videos_from_folder(folder_path: str) -> dict:
    """从文件夹组织视频数据"""
    video_data = {}

    try:
        files = os.listdir(folder_path)
        
        for file in tqdm(files, desc="扫描 JSON 文件"):
            if not file.endswith(".json"):
                continue
                
            file_path = os.path.join(folder_path, file)
            filename = os.path.basename(file_path)

            if '_full_video_' not in filename:
                continue
            
            parts = filename.split('_full_video_')
            video_id_with_res = parts[0]
            
            # 去掉分辨率部分 (如 _1920x1080, _1280x720 等)
            video_id = re.sub(r'_\d+x\d+$', '', video_id_with_res)

            suffix_parts = parts[1].split('_')
            if len(suffix_parts) < 3:
                continue
            
            if video_id not in video_data:
                video_data[video_id] = []
            video_data[video_id].append(file_path)

    except FileNotFoundError:
        print(f"错误: 文件夹 '{folder_path}' 不存在")
        return {}
    except Exception as e:
        print(f"处理文件时发生错误: {e}")
        return {}

    return video_data


def process_single_clip(args):
    """处理单个视频片段并上传到 S3（用于并行处理）"""
    org_video_path, json_file, temp_dir, max_video_res = args
    json_filename = os.path.basename(json_file).split('.')[0]
    
    try:
        with open(json_file, 'r') as f:
            clip_info = json.load(f)
        
        box = clip_info['bbox']
        height = clip_info['raw_video_height']
        width = clip_info['raw_video_width']
        box = [box[0] * width, box[1] * height, box[2] * width, box[3] * height]
        start_time = clip_info['start_seconds']  # 修正字段名
        duration = clip_info['video_total_duration']

        # 计算裁剪参数
        crop_w = int(box[2] - box[0])
        crop_h = int(box[3] - box[1])
        crop_x = int(box[0])
        crop_y = int(box[1])
        
        # 确保尺寸为偶数
        if crop_w % 2 != 0:
            crop_w -= 1
        if crop_h % 2 != 0:
            crop_h -= 1
        if crop_x % 2 != 0:
            crop_x -= 1
        if crop_y % 2 != 0:
            crop_y -= 1

        # 临时输出路径（使用唯一文件名避免冲突）
        video_name = os.path.basename(org_video_path).split('.')[0]
        temp_video_path = os.path.join(temp_dir, f"{json_filename}_{os.getpid()}.mp4")
        temp_audio_path = os.path.join(temp_dir, f"{json_filename}_{os.getpid()}.wav")
        
        # S3 输出路径
        s3_video_key = f"{S3_OUTPUT_PREFIX}/{video_name}/{json_filename}.mp4"
        s3_audio_key = f"{S3_OUTPUT_PREFIX}/{video_name}/{json_filename}.wav"

        # 计算缩放尺寸
        max_scale = max(crop_w, crop_h)
        if max_scale > max_video_res:
            scale_w = int(crop_w / max_scale * max_video_res)
            scale_h = int(crop_h / max_scale * max_video_res)
        else:
            scale_w = crop_w
            scale_h = crop_h
            
        if scale_w % 2 != 0:
            scale_w -= 1
        if scale_h % 2 != 0:
            scale_h -= 1

        # 处理视频
        video_command = [
            '/usr/bin/ffmpeg',
            '-y',
            '-ss', str(start_time),
            '-i', org_video_path,
            '-t', str(duration),
            '-vf', f'crop={crop_w}:{crop_h}:{crop_x}:{crop_y},scale={scale_w}:{scale_h}',
            '-an',
            '-q:v', '2',
            '-r', '25',
            '-vcodec', 'libx264',
            temp_video_path,
        ]

        subprocess.run(video_command, check=True, capture_output=True, text=True)

        # 处理音频
        audio_command = [
            '/usr/bin/ffmpeg',
            '-y',
            '-i', org_video_path,
            '-ss', str(start_time),
            '-t', str(duration),
            '-vn',
            '-acodec', 'pcm_s16le',
            '-ar', '16000',
            '-ac', '1',
            temp_audio_path
        ]

        subprocess.run(audio_command, check=True, capture_output=True, text=True)
        
        # 上传到 S3
        if not upload_to_s3(S3_BUCKET, temp_video_path, s3_video_key):
            return False, f"{json_filename}: 视频上传失败"
        
        if not upload_to_s3(S3_BUCKET, temp_audio_path, s3_audio_key):
            return False, f"{json_filename}: 音频上传失败"
        
        # 删除临时文件
        if os.path.exists(temp_video_path):
            os.remove(temp_video_path)
        if os.path.exists(temp_audio_path):
            os.remove(temp_audio_path)
        
        return True, json_filename
        
    except subprocess.CalledProcessError as e:
        return False, f"{json_filename}: FFmpeg 错误 - {e.stderr[:200] if e.stderr else 'unknown'}"
    except Exception as e:
        return False, f"{json_filename}: 处理错误 - {str(e)}"


def process_video(args):
    """处理单个视频的所有clips（用于并行处理）"""
    video_id, json_list, temp_base_dir = args
    
    video_name = video_id + '.mp4'
    s3_key = f"{S3_PREFIX}/{video_name}"
    
    # 检查 S3 上是否存在视频
    if not check_s3_file_exists(S3_BUCKET, s3_key):
        return video_name, "missing", 0, 0
    
    # 创建该视频专用的临时目录
    temp_dir = os.path.join(temp_base_dir, f"video_{os.getpid()}_{video_id}")
    os.makedirs(temp_dir, exist_ok=True)
    
    try:
        # 下载视频到临时目录
        local_video_path = os.path.join(temp_dir, video_name)
        
        if not download_from_s3(S3_BUCKET, s3_key, local_video_path):
            return video_name, "download_failed", 0, 0
        
        # 并行处理该视频的所有clips
        success_count = 0
        fail_count = 0
        
        clip_args = [
            (local_video_path, json_file, temp_dir, MAX_VIDEO_RES)
            for json_file in json_list
        ]
        
        # 使用线程池处理clips（同一视频的clips共享已下载的视频文件）
        with ThreadPoolExecutor(max_workers=NUM_CLIP_WORKERS) as executor:
            futures = [executor.submit(process_single_clip, arg) for arg in clip_args]
            for future in futures:
                try:
                    success, msg = future.result()
                    if success:
                        success_count += 1
                    else:
                        fail_count += 1
                except Exception as e:
                    fail_count += 1
        
        return video_name, "processed", success_count, fail_count
        
    finally:
        # 清理临时目录
        if os.path.exists(temp_dir):
            shutil.rmtree(temp_dir, ignore_errors=True)


def main():
    print(f"CPU 数量: {cpu_count()}")
    print(f"视频并行数: {NUM_VIDEO_WORKERS}")
    print(f"Clip 并行数: {NUM_CLIP_WORKERS}")
    print(f"预计同时处理: {NUM_VIDEO_WORKERS * NUM_CLIP_WORKERS} 个 FFmpeg 进程")
    print()
    
    # 组织视频数据
    print(f"正在扫描 annotation 文件... {MERGED_ANNO_DIR}")
    json_data = organize_videos_from_folder(MERGED_ANNO_DIR)
    print(f"共找到 {len(json_data)} 个视频的 annotation 数据")
    print(f"输出目标: s3://{S3_BUCKET}/{S3_OUTPUT_PREFIX}/")
    
    # 创建临时基础目录
    os.makedirs("/data/tmp", exist_ok=True)
    temp_base_dir = tempfile.mkdtemp(prefix="s3_video_parallel_", dir="/data/tmp")
    print(f"临时目录: {temp_base_dir}")
    
    # 统计
    missing_videos = []
    download_failed = []
    total_success = 0
    total_fail = 0
    
    try:
        # 准备任务参数
        video_args = [
            (video_id, json_list, temp_base_dir)
            for video_id, json_list in json_data.items()
        ]
        
        # 使用进程池并行处理视频
        with ProcessPoolExecutor(max_workers=NUM_VIDEO_WORKERS) as executor:
            futures = {executor.submit(process_video, arg): arg[0] for arg in video_args}
            
            with tqdm(total=len(futures), desc="处理视频") as pbar:
                for future in as_completed(futures):
                    video_id = futures[future]
                    try:
                        video_name, status, success, fail = future.result()
                        
                        if status == "missing":
                            missing_videos.append(video_name)
                            pbar.set_postfix_str(f"缺失: {video_name}")
                        elif status == "download_failed":
                            download_failed.append(video_name)
                            pbar.set_postfix_str(f"下载失败: {video_name}")
                        else:
                            total_success += success
                            total_fail += fail
                            pbar.set_postfix_str(f"完成: {video_name} ({success}✓/{fail}✗)")
                            
                    except Exception as e:
                        print(f"\n处理 {video_id} 时发生错误: {e}")
                        
                    pbar.update(1)
                    
    finally:
        # 清理临时目录
        if os.path.exists(temp_base_dir):
            shutil.rmtree(temp_base_dir, ignore_errors=True)
            print(f"\n已清理临时目录: {temp_base_dir}")
    
    # 保存缺失视频列表
    all_failed = missing_videos + download_failed
    if all_failed:
        with open(MISSING_VIDEOS_FILE, 'w') as f:
            for video in all_failed:
                f.write(video + '\n')
        print(f"\n缺失/失败视频已保存到: {MISSING_VIDEOS_FILE}")
    
    # 打印统计
    print("\n" + "=" * 50)
    print("处理完成统计:")
    print(f"  成功处理 clips: {total_success}")
    print(f"  失败 clips: {total_fail}")
    print(f"  缺失视频: {len(missing_videos)}")
    print(f"  下载失败: {len(download_failed)}")
    print("=" * 50)


if __name__ == '__main__':
    main()