1 基于PaddleServing的服务部署
本文档将介绍如何使用PaddleServing工具部署PPOCR
动态图模型的pipeline在线服务。
相比较于hubserving部署,PaddleServing具备以下优点:
- 支持客户端和服务端之间高并发和高效通信
- 支持 工业级的服务能力 例如模型管理,在线加载,在线A/B测试等
- 支持 多种编程语言 开发客户端,例如C++, Python和Java
更多有关PaddleServing服务化部署框架介绍和使用教程参考文档。
2 环境准备
需要准备PaddleOCR的运行环境和Paddle Serving的运行环境。
可以参考:https://blog.csdn.net/weixin_43702146/article/details/122229227
-
准备PaddleOCR的运行环境链接
根据环境下载对应的paddle whl包,推荐安装2.0.1版本 -
准备PaddleServing的运行环境
2.1 准备PaddleOCR的运行环境
2.1.1 docker基础镜像拉取
docker镜像地址:
https://github.com/PaddlePaddle/Serving/blob/develop/doc/Docker_Images_CN.md
# 首次运行需创建一个docker容器,再次运行时不需要运行当前命令
# 如果使用CUDA10,请运行以下命令拉取容器
# 如果是CUDA11+CUDNN8,推荐使用镜像registry.baidubce.com/paddlepaddle/paddle:2.1.3-gpu-cuda11.2-cudnn8
docker pull registry.baidubce.com/paddlepaddle/paddle:2.1.3-gpu-cuda10.2-cudnn7
2.1.2 运行镜像并赋予显卡资源
–gpus=“device=6”: 这里让其感知宿主机的6号卡。
–storage-opt:因为生成的容器即将大于10g,这里给30g。
-it:代表终端交互
/bin/bash: 一个守护进程命令,保证容器运行不退出
docker run --name ppocr -p 8075:8868 -it --gpus="device=6" -e NVIDIA_DRIVER_CAPABILITIES=compute,utility -e NVIDIA_VISIBLE_DEVICES=all --storage-opt size=30G registry.baidubce.com/paddlepaddle/paddle:2.1.3-gpu-cuda10.2-cudnn7 /bin/bash
运行此命令后,便进入了ppocr容器内部。
2.1.3 检查容器内部基础环境
python环境
终端输入python,显示3.4版本
cuda和cudnn环境
注意宿主机需要安装好cuda驱动
λ 1e7c36db9018 /usr/local/cuda nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:24:38_PDT_2019
Cuda compilation tools, release 10.2, V10.2.89
glibc环境
λ 1e7c36db9018 /usr/local/cuda ldd --version
ldd (Ubuntu GLIBC 2.23-0ubuntu11.2) 2.23
Copyright (C) 2016 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
Written by Roland McGrath and Ulrich Drepper.
2.1.4 配置CUDA_VISIBLE_DEVICES环境变量
如果不配置cuda环境变量会报错:
RuntimeError: Environment Variable CUDA_VISIBLE_DEVICES is not set correctly. If you wanna use gpu, please set CUDA_VISIBLE_DEVICES via export CUDA_VISIBLE_DEVICES=cuda_device_id.
vim ~/.bashrc
在底部添加
export CUDA_VISIBLE_DEVICES=0
wq保存
source ~/.bashrc
为什么我们的环境变量是0号卡,因为我们容器内部只有一个显卡,所以就是0号卡
2.1.5 拉取工程代码
默认在home目录下
【推荐】git clone https://github.com/PaddlePaddle/PaddleOCR
如果因为网络问题无法pull成功,也可选择使用码云上的托管:
git clone https://gitee.com/paddlepaddle/PaddleOCR
注:码云托管代码可能无法实时同步本github项目更新,存在3~5天延时,请优先使用推荐方式。
然后进入到PaddleOCR目录下
2.1.6 安装第三方库
cd PaddleOCR
pip3 install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
2.2 准备PaddleServing的运行环境
cd PaddleOCR/deploy/pdserving
# 安装serving,用于启动服务
wget https://paddle-serving.bj.bcebos.com/test-dev/whl/paddle_serving_server_gpu-0.7.0.post102-py3-none-any.whl && pip3 install paddle_serving_server_gpu-0.7.0.post102-py3-none-any.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
# 如果是cuda10.1环境,可以使用下面的命令安装paddle-serving-server
# wget https://paddle-serving.bj.bcebos.com/test-dev/whl/paddle_serving_server_gpu-0.7.0.post101-py3-none-any.whl
# pip3 install paddle_serving_server_gpu-0.7.0.post101-py3-none-any.whl
# 安装client,用于向服务发送请求
wget https://paddle-serving.bj.bcebos.com/test-dev/whl/paddle_serving_client-0.7.0-cp37-none-any.whl && pip3 install paddle_serving_client-0.7.0-cp37-none-any.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
# 安装serving-app
wget https://paddle-serving.bj.bcebos.com/test-dev/whl/paddle_serving_app-0.7.0-py3-none-any.whl && pip3 install paddle_serving_app-0.7.0-py3-none-any.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
Note: 如果要安装最新版本的PaddleServing参考链接。
3 模型转换
使用PaddleServing做服务化部署时,需要将保存的inference模型转换为serving易于部署的模型。
首先,下载PPOCR的det和rec inference模型
# 下载并解压 OCR 文本检测模型
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar -O ch_PP-OCRv2_det_infer.tar && tar -xf ch_PP-OCRv2_det_infer.tar
# 下载并解压 OCR 文本识别模型
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar -O ch_PP-OCRv2_rec_infer.tar && tar -xf ch_PP-OCRv2_rec_infer.tar
接下来,用安装的paddle_serving_client把下载的inference模型转换成易于server部署的模型格式。
# 转换检测模型
python3 -m paddle_serving_client.convert --dirname ./ch_PP-OCRv2_det_infer/ \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--serving_server ./ppocrv2_det_serving/ \
--serving_client ./ppocrv2_det_client/
# 转换识别模型
python3 -m paddle_serving_client.convert --dirname ./ch_PP-OCRv2_rec_infer/ \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--serving_server ./ppocrv2_rec_serving/ \
--serving_client ./ppocrv2_rec_client/
这里会出现一个问题:'Program' object has no attribute '_remove_training_info'
Traceback (most recent call last):
File "/usr/local/python3.7.0/lib/python3.7/runpy.py", line 193, in _run_module_as_main
"__main__", mod_spec)
File "/usr/local/python3.7.0/lib/python3.7/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/usr/local/python3.7.0/lib/python3.7/site-packages/paddle_serving_client/convert.py", line 87, in <module>
fetch_alias_names=args.fetch_alias_names)
File "/usr/local/python3.7.0/lib/python3.7/site-packages/paddle_serving_client/io/__init__.py", line 334, in inference_model_to_serving
model_filename, params_filename, show_proto, feed_alias_names, fetch_alias_names)
File "/usr/local/python3.7.0/lib/python3.7/site-packages/paddle_serving_client/io/__init__.py", line 214, in save_model
new_model_file.write(main_program._remove_training_info(False).desc.serialize_to_string())
AttributeError: 'Program' object has no attribute '_remove_training_info'
修改源码即可
vim /usr/local/python3.7.0/lib/python3.7/site-packages/paddle_serving_client/io/__init__.py
在214行
new_model_file.write(main_program._remove_training_info(False).desc.serialize_to_string())
删除._remove_training_info(False)并保存
删除生成的错误目录
rm -rf ppocrv2_det_serving
然后重新进行模型转换
检测模型转换完成后,会在当前文件夹多出ppocrv2_det_serving
和ppocrv2_det_client
的文件夹,具备如下格式:
|- ppocrv2_det_serving/
|- __model__
|- __params__
|- serving_server_conf.prototxt
|- serving_server_conf.stream.prototxt
|- ppocrv2_det_client
|- serving_client_conf.prototxt
|- serving_client_conf.stream.prototxt
识别模型同理。
4 Paddle Serving pipeline部署
-
下载PaddleOCR代码,若已下载可跳过此步骤
git clone https://github.com/PaddlePaddle/PaddleOCR # 进入到工作目录 cd PaddleOCR/deploy/pdserving/
pdserver目录包含启动pipeline服务和发送预测请求的代码,包括:
__init__.py config.yml # 启动服务的配置文件 ocr_reader.py # OCR模型预处理和后处理的代码实现 pipeline_http_client.py # 发送pipeline预测请求的脚本 web_service.py # 启动pipeline服务端的脚本
-
启动服务可运行如下命令:
# 启动服务,运行日志保存在log.txt python3 web_service.py &>log.txt &
成功启动服务后,log.txt中会打印类似如下日志
-
发送服务请求:
python3 pipeline_http_client.py
调整 config.yml 中的并发个数获得最大的QPS, 一般检测和识别的并发数为2:1
det: #并发数,is_thread_op=True时,为线程并发;否则为进程并发 concurrency: 8 ... rec: #并发数,is_thread_op=True时,为线程并发;否则为进程并发 concurrency: 4 ...
有需要的话可以同时发送多个服务请求
预测性能数据会被自动写入
PipelineServingLogs/pipeline.tracer
文件中。在200张真实图片上测试,把检测长边限制为960。T4 GPU 上 QPS 均值可达到23左右:
2021-05-13 03:42:36,895 ==================== TRACER ====================== 2021-05-13 03:42:36,975 Op(rec): 2021-05-13 03:42:36,976 in[14.472382882882883 ms] 2021-05-13 03:42:36,976 prep[9.556855855855856 ms] 2021-05-13 03:42:36,976 midp[59.921905405405404 ms] 2021-05-13 03:42:36,976 postp[15.345945945945946 ms] 2021-05-13 03:42:36,976 out[1.9921216216216215 ms] 2021-05-13 03:42:36,976 idle[0.16254943864471572] 2021-05-13 03:42:36,976 Op(det): 2021-05-13 03:42:36,976 in[315.4468035714286 ms] 2021-05-13 03:42:36,976 prep[69.5980625 ms] 2021-05-13 03:42:36,976 midp[18.989535714285715 ms] 2021-05-13 03:42:36,976 postp[18.857803571428573 ms] 2021-05-13 03:42:36,977 out[3.1337544642857145 ms] 2021-05-13 03:42:36,977 idle[0.7477961159203756] 2021-05-13 03:42:36,977 DAGExecutor: 2021-05-13 03:42:36,977 Query count[224] 2021-05-13 03:42:36,977 QPS[22.4 q/s] 2021-05-13 03:42:36,977 Succ[0.9910714285714286] 2021-05-13 03:42:36,977 Error req[169, 170] 2021-05-13 03:42:36,977 Latency: 2021-05-13 03:42:36,977 ave[535.1678348214285 ms] 2021-05-13 03:42:36,977 .50[172.651 ms] 2021-05-13 03:42:36,977 .60[187.904 ms] 2021-05-13 03:42:36,977 .70[245.675 ms] 2021-05-13 03:42:36,977 .80[526.684 ms] 2021-05-13 03:42:36,977 .90[854.596 ms] 2021-05-13 03:42:36,977 .95[1722.728 ms] 2021-05-13 03:42:36,977 .99[3990.292 ms] 2021-05-13 03:42:36,978 Channel (server worker num[10]): 2021-05-13 03:42:36,978 chl0(In: ['@DAGExecutor'], Out: ['det']) size[0/0] 2021-05-13 03:42:36,979 chl1(In: ['det'], Out: ['rec']) size[6/0] 2021-05-13 03:42:36,979 chl2(In: ['rec'], Out: ['@DAGExecutor']) size[0/0]
FAQ
Q1: 发送请求后没有结果返回或者提示输出解码报错
A1: 启动服务和发送请求时不要设置代理,可以在启动服务前和发送请求前关闭代理,关闭代理的命令是:
unset https_proxy
unset http_proxy