fastdeploy部署多线程/进程paddle ocr(python flask框架 )

部署参考:https://github.com/PaddlePaddle/FastDeploy/blob/develop/tutorials/multi_thread/python/pipeline/README_CN.md

安装

cpu: pip install fastdeploy-python

gpu :pip install fastdeploy-gpu-python

python 复制代码
#下载部署示例代码
git clone https://github.com/PaddlePaddle/FastDeploy.git
cd  FastDeploy/tutorials/multi_thread/python/pipeline

# 下载模型,图片和字典文件
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar
tar xvf ch_PP-OCRv3_det_infer.tar

wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar
tar -xvf ch_ppocr_mobile_v2.0_cls_infer.tar

wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar
tar xvf ch_PP-OCRv3_rec_infer.tar

wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/doc/imgs/12.jpg

wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/ppocr/utils/ppocr_keys_v1.txt

命令:

多线程

python 复制代码
python multi_thread_process_ocr.py --det_model ch_PP-OCRv3_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv3_rec_infer --rec_label_file ppocr_keys_v1.txt --image_path xxx/xxx --device gpu --thread_num 3

多进程

python 复制代码
python multi_thread_process_ocr.py --det_model ch_PP-OCRv3_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv3_rec_infer --rec_label_file ppocr_keys_v1.txt --image_path xxx/xxx --device gpu --use_multi_process True --process_num 3

问题

多进程图片分配有bug

文件:multi_thread_process_ocr.py

原始代码:270行

修改为如下,去掉1

ModuleNotFoundError: No module named 'example'

因为安装包不对,fastdeploy与fastdeploy-python不是同一个包

CUDA error(3), initialization error.

bash 复制代码
  ----------------------
  Error Message Summary:
  ----------------------
  ExternalError: CUDA error(3), initialization error. 
    [Hint: Please search for the error code(3) on website (https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038) to get Nvidia's official solution and advice about CUDA Error.] (at /home/fastdeploy/develop/paddle_build/v0.0.0/Paddle/paddle/phi/backends/gpu/cuda/cuda_info.cc:251)

参考:
PaddlePaddle------问题解决:使用Python multiprocessing时报错:CUDA error(3), initialization error.
https://github.com/PaddlePaddle/PaddleDetection/issues/2241

paddle 相关模块只在方法里面引用,要在多进程外有 import 这些模块

flask部署

发送列表类型的图片base64编码,返回列表类型的字符串

注意server端文件放在FastDeploy/tutorials/multi_thread/python/pipeline目录下

创建server端

python 复制代码
from threading import Thread

import cv2
import os
from multiprocessing import Pool
import sys
 
import fastdeploy as fd
import numpy as np
import base64
from PIL import Image
from io import BytesIO
from sqlalchemy import create_engine, text

from flask import Flask, request, jsonify
import argparse
import ast
 
# watch -n 0.1 nvidia-smi

def parse_arguments():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--det_model",
        # required=True,
        type=str,
        default='ch_PP-OCRv3_det_infer',
        help="Path of Detection model of PPOCR.")
    parser.add_argument(
        "--cls_model",
        # required=True,
        type=str,
        default='ch_ppocr_mobile_v2.0_cls_infer',
        help="Path of Classification model of PPOCR.")
    parser.add_argument(
        "--rec_model",
        # required=True,
        type=str,
        default='ch_PP-OCRv3_rec_infer',
        help="Path of Recognization model of PPOCR.")
    parser.add_argument(
        "--rec_label_file",
        # required=True,
        type=str,
        default='ppocr_keys_v1.txt',
        help="Path of Recognization model of PPOCR.")
    # parser.add_argument(
    #     "--image_path",
    #     type=str,
    #     required=True,
    #     help="The directory or path or file list of the images to be predicted."
    # )
    parser.add_argument(
        "--device",
        type=str,
        default='gpu', # cpu
        help="Type of inference device, support 'cpu', 'kunlunxin' or 'gpu'.")
    parser.add_argument(
        "--backend",
        type=str,
        default="default",
        help="Type of inference backend, support ort/trt/paddle/openvino, default 'openvino' for cpu, 'tensorrt' for gpu"
    )
    parser.add_argument(
        "--device_id",
        type=int,
        default=0,
        help="Define which GPU card used to run model.")
    parser.add_argument(
        "--cpu_thread_num",
        type=int,
        default=9,
        help="Number of threads while inference on CPU.")
    parser.add_argument(
        "--cls_bs",
        type=int,
        default=1,
        help="Classification model inference batch size.")
    parser.add_argument(
        "--rec_bs",
        type=int,
        default=6,
        help="Recognition model inference batch size")
    parser.add_argument("--thread_num", type=int, default=1, help="thread num")
    parser.add_argument(
        "--use_multi_process",
        type=ast.literal_eval,
        default=True,
        help="Wether to use multi process.")
    parser.add_argument(
        "--process_num", type=int, default=5, help="process num")
    return parser.parse_args()


def get_image_list(image_path):
    image_list = []
    if os.path.isfile(image_path):
        image_list.append(image_path)
    # load image in a directory
    elif os.path.isdir(image_path):
        for root, dirs, files in os.walk(image_path):
            for f in files:
                image_list.append(os.path.join(root, f))
    else:
        raise FileNotFoundError(
            '{} is not found. it should be a path of image, or a directory including images.'.
            format(image_path))

    if len(image_list) == 0:
        raise RuntimeError(
            'There are not image file in `--image_path`={}'.format(image_path))

    return image_list


def build_option(args):
    option = fd.RuntimeOption()
    if args.device.lower() == "gpu":
        option.use_gpu(args.device_id)

    option.set_cpu_thread_num(args.cpu_thread_num)

    if args.device.lower() == "kunlunxin":
        option.use_kunlunxin()
        return option

    if args.backend.lower() == "trt":
        assert args.device.lower(
        ) == "gpu", "TensorRT backend require inference on device GPU."
        option.use_trt_backend()
    elif args.backend.lower() == "pptrt":
        assert args.device.lower(
        ) == "gpu", "Paddle-TensorRT backend require inference on device GPU."
        option.use_trt_backend()
        option.enable_paddle_trt_collect_shape()
        option.enable_paddle_to_trt()
    elif args.backend.lower() == "ort":
        option.use_ort_backend()
    elif args.backend.lower() == "paddle":
        option.use_paddle_infer_backend()
    elif args.backend.lower() == "openvino":
        assert args.device.lower(
        ) == "cpu", "OpenVINO backend require inference on device CPU."
        option.use_openvino_backend()
    return option


def load_model(args, runtime_option):
    # Detection模型, 检测文字框
    det_model_file = os.path.join(args.det_model, "inference.pdmodel")
    det_params_file = os.path.join(args.det_model, "inference.pdiparams")
    # Classification模型,方向分类,可选
    cls_model_file = os.path.join(args.cls_model, "inference.pdmodel")
    cls_params_file = os.path.join(args.cls_model, "inference.pdiparams")
    # Recognition模型,文字识别模型
    rec_model_file = os.path.join(args.rec_model, "inference.pdmodel")
    rec_params_file = os.path.join(args.rec_model, "inference.pdiparams")
    rec_label_file = args.rec_label_file

    # PPOCR的cls和rec模型现在已经支持推理一个Batch的数据
    # 定义下面两个变量后, 可用于设置trt输入shape, 并在PPOCR模型初始化后, 完成Batch推理设置
    cls_batch_size = 1
    rec_batch_size = 6

    # 当使用TRT时,分别给三个模型的runtime设置动态shape,并完成模型的创建.
    # 注意: 需要在检测模型创建完成后,再设置分类模型的动态输入并创建分类模型, 识别模型同理.
    # 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
    det_option = runtime_option
    det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
                                   [1, 3, 960, 960])
    # 用户可以把TRT引擎文件保存至本地
    #det_option.set_trt_cache_file(args.det_model  + "/det_trt_cache.trt")
    global det_model
    det_model = fd.vision.ocr.DBDetector(
        det_model_file, det_params_file, runtime_option=det_option)

    cls_option = runtime_option
    cls_option.set_trt_input_shape("x", [1, 3, 48, 10],
                                   [cls_batch_size, 3, 48, 320],
                                   [cls_batch_size, 3, 48, 1024])
    # 用户可以把TRT引擎文件保存至本地
    # cls_option.set_trt_cache_file(args.cls_model  + "/cls_trt_cache.trt")
    global cls_model
    cls_model = fd.vision.ocr.Classifier(
        cls_model_file, cls_params_file, runtime_option=cls_option)

    rec_option = runtime_option
    rec_option.set_trt_input_shape("x", [1, 3, 48, 10],
                                   [rec_batch_size, 3, 48, 320],
                                   [rec_batch_size, 3, 48, 2304])
    # 用户可以把TRT引擎文件保存至本地
    #rec_option.set_trt_cache_file(args.rec_model  + "/rec_trt_cache.trt")
    global rec_model
    rec_model = fd.vision.ocr.Recognizer(
        rec_model_file,
        rec_params_file,
        rec_label_file,
        runtime_option=rec_option)

    # 创建PP-OCR,串联3个模型,其中cls_model可选,如无需求,可设置为None
    global ppocr_v3
    ppocr_v3 = fd.vision.ocr.PPOCRv3(
        det_model=det_model, cls_model=cls_model, rec_model=rec_model)

    # 给cls和rec模型设置推理时的batch size
    # 此值能为-1, 和1到正无穷
    # 当此值为-1时, cls和rec模型的batch size将默认和det模型检测出的框的数量相同
    ppocr_v3.cls_batch_size = cls_batch_size
    ppocr_v3.rec_batch_size = rec_batch_size


def predict(model, img_list):
    result_list = []
    # predict ppocr result
    for image in img_list:
        im = cv2.imread(image)
        result = model.predict(im)
        result_list.append(result)
    return result_list

def process_predict(image):
    # predict ppocr result
    im = cv2.imread(image)
    result = ppocr_v3.predict(im)
    print(result)


def process_predict_text(base64_str):
    image = base64_to_bgr(base64_str)
    result = ppocr_v3.predict(image)
    # print(result)
    return ''.join(result.text) #不能直接返回OCR对象序列化会失败

def cv_show(img):
    '''
    展示图片
    @param img:
    @param name:
    @return:
    '''
    cv2.namedWindow('name', cv2.WINDOW_KEEPRATIO)  # cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO
    cv2.imshow('name', img)
    cv2.waitKey(0)
    cv2.destroyAllWindows()


def base64_to_bgr(base64_str):
    base64_hex = base64.b64decode(base64_str)
    image = BytesIO(base64_hex)
    img = Image.open(image)

    if img.mode=='RGBA':
        width = img.width
        height = img.height
        img2 = Image.new('RGB', size=(width, height), color=(255, 255, 255))
        img2.paste(img, (0, 0), mask=img)
        image_array = np.array(img2)
    else:
        image_array = np.array(img)
    image = cv2.cvtColor(image_array, cv2.COLOR_RGB2BGR)
    return image



class WrapperThread(Thread):
    def __init__(self, func, args):
        super(WrapperThread, self).__init__()
        self.func = func
        self.args = args
        # self.result = self.func(*self.args)

    def run(self):
        self.result = self.func(*self.args)

    def get_result(self):
        return self.result

def ocr_image_list(imgs_list):
    args = parse_arguments()

    # 对于三个模型,均采用同样的部署配置
    # 用户也可根据自行需求分别配置
    runtime_option = build_option(args)

    if args.use_multi_process:
        process_num = args.process_num
        with Pool(
                process_num,
                initializer=load_model,
                initargs=(args, runtime_option)) as pool:
            #
            results = pool.map(process_predict_text, imgs_list)
            # pool.map(process_predict, imgs_list)
            # 进一步处理结果
            for i, result in enumerate(results):
                print(i, result)
    else:
        load_model(args, runtime_option)
        threads = []
        thread_num = args.thread_num
        image_num_each_thread = int(len(imgs_list) / thread_num)
        # unless you want independent model in each thread, actually model.clone()
        # is the same as model when creating thead because of the existence of
        # GIL(Global Interpreter Lock) in python. In addition, model.clone() will consume
        # additional memory to store independent member variables
        for i in range(thread_num):
            if i == thread_num - 1:
                t = WrapperThread(
                    predict,
                    args=(ppocr_v3.clone(),
                          imgs_list[i * image_num_each_thread:]))
            else:
                t = WrapperThread(
                    predict,
                    args=(ppocr_v3.clone(),
                          imgs_list[i * image_num_each_thread:(i + 1) *
                                                              image_num_each_thread]))  # - 1
            threads.append(t)
            t.start()

        for i in range(thread_num):
            threads[i].join()

        for i in range(thread_num):
            for result in threads[i].get_result():
                print('thread:', i, ', result: ', result)


@app.route('/ocr/submit', methods=['POST'])
def ocr():
    args = parse_arguments()
    process_num = 1#args.process_num
    runtime_option = build_option(args)
    data = request.get_json()

    # 获取 Base64 数据
    base64_str = data['img_base64']

    with Pool(
        process_num, initializer=load_model, initargs=(args, runtime_option)) as pool:
            results = pool.map(process_predict_text, base64_str)
    # 返回响应
    response = {'message': 'Data received', 'result': results}
    return jsonify(response)

import json
import pandas as pd
import time

if __name__ == '__main__':
    app.run(host='192.168.xxx.xxx', port=5000)
    

client 端

python 复制代码
import base64
import sys

import requests
import json
# 读取图像文件
with open('./pic/img.png', 'rb') as image_file:
    # 将图像文件内容读取为字节流
    image_data = image_file.read()

# 将图像字节流进行 Base64 编码
img_base64 = base64.b64encode(image_data)
 
data = {
    'img_base64': [img_base64.decode('utf-8')] 
   
}


headers = {
    'Content-Type': 'application/json'
}

 
response = requests.post("http://192.168.xxx.xxx:5000/ocr/submit",  data=json.dumps(data),headers = headers)

if response.status_code == 200:
    result = response.json()
    print(result['result'])
else:
    print('Error:', response.status_code)
相关推荐
TF男孩44 分钟前
ARQ:一款低成本的消息队列,实现每秒万级吞吐
后端·python·消息队列
该用户已不存在6 小时前
Mojo vs Python vs Rust: 2025年搞AI,该学哪个?
后端·python·rust
站大爷IP8 小时前
Java调用Python的5种实用方案:从简单到进阶的全场景解析
python
用户83562907805113 小时前
从手动编辑到代码生成:Python 助你高效创建 Word 文档
后端·python
c8i13 小时前
python中类的基本结构、特殊属性于MRO理解
python
liwulin050614 小时前
【ESP32-CAM】HELLO WORLD
python
Doris_202314 小时前
Python条件判断语句 if、elif 、else
前端·后端·python
Doris_202314 小时前
Python 模式匹配match case
前端·后端·python
这里有鱼汤15 小时前
Python量化实盘踩坑指南:分钟K线没处理好,小心直接亏钱!
后端·python·程序员
大模型真好玩15 小时前
深入浅出LangGraph AI Agent智能体开发教程(五)—LangGraph 数据分析助手智能体项目实战
人工智能·python·mcp