PaddleX/zhch/ppstructurev3_dual_gpu_optimized.py

# zhch/ppstructurev3_dual_gpu_optimized.py
import json
import time
import os
import glob
import traceback
from pathlib import Path
from typing import List, Dict, Any, Tuple
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor, as_completed
from multiprocessing import Queue, Manager, Process
import cv2
import numpy as np
from paddlex import create_pipeline
from tqdm import tqdm
import threading
import queue
import paddle

from dotenv import load_dotenv
load_dotenv(override=True)

class PPStructureV3DualGPUPredictor:
    """
    PP-StructureV3双GPU并行预测器
    """

    def __init__(self, pipeline_config_path: str = "PP-StructureV3", output_path: str = "output", gpu_id: int = 0):
        """
        初始化预测器
        
        Args:
            pipeline_config_path: PaddleX pipeline配置文件路径
            output_path: 输出路径
            gpu_id: GPU设备ID (0 或 1)
        """
        self.pipeline_config = pipeline_config_path
        self.pipeline = None  # 延迟初始化
        self.output_path = output_path
        self.gpu_id = gpu_id
        self.device = f"gpu:{gpu_id}"

    def _ensure_pipeline(self):
        """确保pipeline已初始化（线程安全）"""
        if self.pipeline is None:
            # 设置当前GPU
            paddle.device.set_device(f"gpu:{self.gpu_id}")
            self.pipeline = create_pipeline(pipeline=self.pipeline_config)
            print(f"Pipeline初始化完成 - GPU:{self.gpu_id}")

    def process_single_image(self, image_path: str) -> Dict[str, Any]:
        """
        处理单张图像
        
        Args:
            image_path: 图像路径
            
        Returns:
            处理结果{"image_path": str, "success": bool, "processing_time": float, "error": str}
        """
        try:
            # 确保pipeline已初始化
            self._ensure_pipeline()
            
            # 读取图像获取尺寸信息
            image = cv2.imread(image_path)
            if image is None:
                return {
                    "image_path": Path(image_path).name,
                    "error": "无法读取图像",
                    "success": False,
                    "processing_time": 0,
                    "gpu_id": self.gpu_id
                }
                
            height, width = image.shape[:2]
            
            # 运行PaddleX pipeline
            start_time = time.time()
            
            output = self.pipeline.predict(
                input=image_path,
                device=self.device,
                use_doc_orientation_classify=True,
                use_doc_unwarping=False,
                use_seal_recognition=True,
                use_chart_recognition=True,
                use_table_recognition=True,
                use_formula_recognition=True,
            )
            
            # 保存结果
            for res in output:
                res.save_to_json(save_path=self.output_path)
                res.save_to_markdown(save_path=self.output_path)
            
            process_time = time.time() - start_time
            
            # 返回处理结果
            return {
                "image_path": Path(image_path).name,
                "processing_time": process_time,
                "success": True,
                "gpu_id": self.gpu_id
            }
            
        except Exception as e:
            return {
                "image_path": Path(image_path).name,
                "error": str(e),
                "success": False,
                "processing_time": 0,
                "gpu_id": self.gpu_id
            }
    
    def process_batch(self, image_paths: List[str]) -> List[Dict[str, Any]]:
        """
        批处理图像
        
        Args:
            image_paths: 图像路径列表
            
        Returns:
            结果列表
        """
        results = []
        
        for image_path in image_paths:
            result = self.process_single_image(image_path)
            results.append(result)
        
        return results

class DualGPUThreadWorker:
    """双GPU线程工作器 - 每个线程维护自己的pipeline实例"""
    
    def __init__(self, pipeline_config: str, output_path: str, gpu_id: int, worker_id: int):
        self.worker_id = worker_id
        self.gpu_id = gpu_id
        self.predictor = PPStructureV3DualGPUPredictor(
            pipeline_config, 
            output_path=f"{output_path}/gpu{gpu_id}_worker_{worker_id}", 
            gpu_id=gpu_id
        )
        self.task_queue = queue.Queue()
        self.result_queue = queue.Queue()
        self.running = True
        
    def add_batch(self, batch: List[str]):
        """添加批处理任务"""
        self.task_queue.put(batch)
    
    def get_results(self) -> List[Dict[str, Any]]:
        """获取处理结果"""
        results = []
        while not self.result_queue.empty():
            try:
                result = self.result_queue.get_nowait()
                results.extend(result)
            except queue.Empty:
                break
        return results
    
    def worker_loop(self):
        """工作循环"""
        print(f"GPU{self.gpu_id} Worker{self.worker_id} 开始工作")
        
        while self.running:
            try:
                batch = self.task_queue.get(timeout=1.0)
                if batch is None:  # 结束信号
                    break
                    
                # 处理批次
                batch_results = self.predictor.process_batch(batch)
                self.result_queue.put(batch_results)
                
            except queue.Empty:
                continue
            except Exception as e:
                print(f"GPU{self.gpu_id} Worker{self.worker_id} 处理出错: {e}")
    
    def stop(self):
        """停止工作线程"""
        self.running = False
        self.task_queue.put(None)  # 发送结束信号

def parallel_process_with_dual_gpu(image_paths: List[str],
                                 batch_size: int = 4,
                                 workers_per_gpu: int = 2,  # 每个GPU的worker数量
                                 pipeline_config: str = "PP-StructureV3",
                                 output_path: str = "./output") -> List[Dict[str, Any]]:
    """
    使用双GPU优化的多线程并行处理
    
    Args:
        image_paths: 图像路径列表
        batch_size: 批处理大小
        workers_per_gpu: 每个GPU的worker数量（推荐2个）
        pipeline_config: pipeline配置
        output_path: 输出路径
        
    Returns:
        处理结果列表
    """
    # 确保输出目录存在
    os.makedirs(output_path, exist_ok=True)
    
    # 检查可用GPU
    try:
        gpu_count = paddle.device.cuda.device_count()
        print(f"检测到 {gpu_count} 个GPU")
        
        if gpu_count < 2:
            print("警告：检测到的GPU数量少于2个，建议检查CUDA配置")
            available_gpus = list(range(gpu_count))
        else:
            available_gpus = [0, 1]  # 使用GPU 0和1
            
    except Exception as e:
        print(f"GPU检测失败: {e}")
        available_gpus = [0]  # 降级为单GPU
    
    total_workers = len(available_gpus) * workers_per_gpu
    print(f"使用GPU: {available_gpus}")
    print(f"每GPU Worker数: {workers_per_gpu}")
    print(f"总Worker数: {total_workers}")
    
    # 将图像路径分批
    batches = [image_paths[i:i + batch_size] for i in range(0, len(image_paths), batch_size)]
    
    # 创建工作线程
    workers = []
    threads = []
    
    worker_id = 0
    for gpu_id in available_gpus:
        for i in range(workers_per_gpu):
            worker = DualGPUThreadWorker(pipeline_config, output_path, gpu_id, worker_id)
            workers.append(worker)
            
            thread = threading.Thread(target=worker.worker_loop, name=f"GPU{gpu_id}_Worker{worker_id}")
            thread.daemon = True
            thread.start()
            threads.append(thread)
            worker_id += 1
    
    print(f"启动了 {len(workers)} 个工作线程，分布在 {len(available_gpus)} 个GPU上")
    
    # 分发任务
    all_results = []
    total_images = len(image_paths)
    completed_count = 0
    
    try:
        with tqdm(total=total_images, desc="双GPU处理图像", unit="张") as pbar:
            # 轮流分发批次到不同的worker
            for i, batch in enumerate(batches):
                worker_id = i % len(workers)
                workers[worker_id].add_batch(batch)
            
            # 等待所有任务完成
            while completed_count < total_images:
                time.sleep(0.1)  # 短暂等待
                
                # 收集结果
                for worker in workers:
                    batch_results = worker.get_results()
                    if batch_results:
                        all_results.extend(batch_results)
                        completed_count += len(batch_results)
                        pbar.update(len(batch_results))
                        
                        # 更新进度条
                        success_count = sum(1 for r in batch_results if r.get('success', False))
                        
                        # 按GPU统计
                        gpu_stats = {}
                        for r in all_results:
                            gpu_id = r.get('gpu_id', 'unknown')
                            if gpu_id not in gpu_stats:
                                gpu_stats[gpu_id] = {'success': 0, 'total': 0}
                            gpu_stats[gpu_id]['total'] += 1
                            if r.get('success', False):
                                gpu_stats[gpu_id]['success'] += 1
                        
                        gpu_info = ', '.join([f"GPU{k}:{v['success']}/{v['total']}" for k, v in gpu_stats.items()])
                        
                        pbar.set_postfix({
                            'recent_success': f"{success_count}/{len(batch_results)}",
                            'gpu_distribution': gpu_info
                        })
    
    finally:
        # 停止所有工作线程
        for worker in workers:
            worker.stop()
        
        # 等待线程结束
        for thread in threads:
            thread.join(timeout=3.0)
    
    return all_results

def monitor_gpu_memory():
    """监控GPU内存使用情况"""
    try:
        for gpu_id in [0, 1]:
            paddle.device.set_device(f"gpu:{gpu_id}")
            allocated = paddle.device.cuda.memory_allocated() / 1024**3
            reserved = paddle.device.cuda.memory_reserved() / 1024**3
            print(f"GPU {gpu_id} - 已分配: {allocated:.2f}GB, 已预留: {reserved:.2f}GB")
    except Exception as e:
        print(f"GPU内存监控失败: {e}")

def main():
    """主函数 - 双GPU优化的并行处理"""
    
    # 配置参数
    dataset_path = "../../OmniDocBench/OpenDataLab___OmniDocBench/images"
    output_dir = "./OmniDocBench_Results_DualGPU"
    pipeline_config = "PP-StructureV3"
    
    # 双GPU处理参数
    batch_size = 4              # 批处理大小
    workers_per_gpu = 1         # 每个GPU的worker数量（24GB GPU推荐2个）
    
    # 确保输出目录存在
    print(f"输出目录: {Path(output_dir).absolute()}")
    os.makedirs(output_dir, exist_ok=True)
    
    dataset_path = Path(dataset_path).resolve()
    output_dir = Path(output_dir).resolve()
    
    print("="*60)
    print("OmniDocBench 双GPU优化并行处理开始")
    print("="*60)
    print(f"数据集路径: {dataset_path}")
    print(f"输出目录: {output_dir}")
    print(f"批处理大小: {batch_size}")
    print(f"每GPU Worker数: {workers_per_gpu}")
    print(f"总Worker数: {workers_per_gpu * 2}")
    
    # 监控初始GPU状态
    print("\n初始GPU内存状态:")
    monitor_gpu_memory()
    
    # 查找所有图像文件
    image_extensions = ['*.jpg', '*.jpeg', '*.png', '*.bmp', '*.tiff']
    image_files = []
    
    for ext in image_extensions:
        image_files.extend(glob.glob(os.path.join(dataset_path, ext)))
    
    print(f"\n找到 {len(image_files)} 个图像文件")
    
    if not image_files:
        print("未找到任何图像文件，程序终止")
        return
    
    # 限制处理数量用于测试
    # image_files = image_files[:40]  # 取消注释以限制处理数量
    
    # 开始处理
    start_time = time.time()
    
    try:
        print("\n使用双GPU优化并行处理...")
        results = parallel_process_with_dual_gpu(
            image_files, 
            batch_size, 
            workers_per_gpu, 
            pipeline_config, 
            str(output_dir)
        )
        
        total_time = time.time() - start_time
        
        # 统计信息
        success_count = sum(1 for r in results if r.get('success', False))
        error_count = len(results) - success_count
        total_processing_time = sum(r.get('processing_time', 0) for r in results if r.get('success', False))
        avg_processing_time = total_processing_time / success_count if success_count > 0 else 0
        
        # 按GPU统计
        gpu_stats = {}
        for r in results:
            gpu_id = r.get('gpu_id', 'unknown')
            if gpu_id not in gpu_stats:
                gpu_stats[gpu_id] = {'success': 0, 'total': 0, 'total_time': 0}
            gpu_stats[gpu_id]['total'] += 1
            if r.get('success', False):
                gpu_stats[gpu_id]['success'] += 1
                gpu_stats[gpu_id]['total_time'] += r.get('processing_time', 0)
        
        # 保存结果统计
        stats = {
            "total_files": len(image_files),
            "success_count": success_count,
            "error_count": error_count,
            "success_rate": success_count / len(image_files),
            "total_time": total_time,
            "avg_processing_time": avg_processing_time,
            "throughput": len(image_files) / total_time,
            "batch_size": batch_size,
            "workers_per_gpu": workers_per_gpu,
            "total_workers": workers_per_gpu * 2,
            "gpu_stats": gpu_stats,
            "optimization": "双GPU多线程并行"
        }
        
        # 保存最终结果
        output_file = os.path.join(output_dir, f"OmniDocBench_DualGPU_batch{batch_size}_workers{workers_per_gpu}.json")
        final_results = {
            "results": results,
            "stats": stats
        }
        
        with open(output_file, 'w', encoding='utf-8') as f:
            json.dump(final_results, f, ensure_ascii=False, indent=2)
        
        print("\n" + "="*60)
        print("双GPU优化并行处理完成!")
        print("="*60)
        print(f"总文件数: {len(image_files)}")
        print(f"成功处理: {success_count}")
        print(f"失败数量: {error_count}")
        print(f"成功率: {success_count / len(image_files) * 100:.2f}%")
        print(f"总耗时: {total_time:.2f}秒")
        print(f"平均处理时间: {avg_processing_time:.2f}秒/张")
        print(f"吞吐量: {len(image_files) / total_time:.2f}张/秒")
        print(f"Worker数: {workers_per_gpu * 2} (每GPU {workers_per_gpu}个)")
        
        # GPU统计
        print(f"\nGPU分布统计:")
        for gpu_id, stat in gpu_stats.items():
            if stat['total'] > 0:
                gpu_success_rate = stat['success'] / stat['total'] * 100
                gpu_avg_time = stat['total_time'] / stat['success'] if stat['success'] > 0 else 0
                print(f"  GPU {gpu_id}: {stat['success']}/{stat['total']} 成功 "
                      f"({gpu_success_rate:.1f}%), 平均 {gpu_avg_time:.2f}s/张")
        
        print(f"\n结果保存至: {output_file}")
        
        # 监控最终GPU状态
        print("\n最终GPU内存状态:")
        monitor_gpu_memory()
        
    except Exception as e:
        print(f"处理过程中发生错误: {str(e)}")
        traceback.print_exc()

if __name__ == "__main__":
    main()