官方参考：
https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5/deploy/pdserving
一、Paddle Serving pipeline部署 python
1、下拉镜像
docker pull registry.baidubce.com/paddlepaddle/serving:0.9.0-cuda10.1-cudnn7-devel
2、进入镜像
nvidia-docker run -it --entrypoint=/bin/bash registry.baidubce.com/paddlepaddle/serving:0.9.0-cuda10.1-cudnn7-devel
3、安装特定PaddlePaddle
wget https://paddle-inference-lib.bj.bcebos.com/2.3.0/python/Linux/GPU/x86-64_gcc8.2_avx_mkl_cuda10.1_cudnn7.6.5_trt6.0.1.5/paddlepaddle_gpu-2.3.0.post101-cp37-cp37m-linux_x86_64.whl
pip3.7 install paddlepaddle_gpu-2.3.0.post101-cp37-cp37m-linux_x86_64.whl
3.1、直接在线安装
python3.7 -m pip install paddlepaddle-gpu==2.3.0.post101 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
4、安装相关的包
pip3.7 install paddle-serving-client==0.9.0 paddle-serving-app==0.9.0 paddle-serving-server-gpu==0.9.0.post101 -i https://mirror.baidu.com/pypi/simple
5、下载OCR repo
git clone https://github.com/PaddlePaddle/PaddleOCR
cd PaddleOCR/deploy/pdserving/
6、下载并解压 OCR 文本检测模型
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar -O ch_PP-OCRv3_det_infer.tar && tar -xf ch_PP-OCRv3_det_infer.tar
7、下载并解压 OCR 文本识别模型
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar -O ch_PP-OCRv3_rec_infer.tar && tar -xf ch_PP-OCRv3_rec_infer.tar
8、转换检测模型
python3.7 -m paddle_serving_client.convert --dirname ./ch_PP-OCRv3_det_infer/ \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--serving_server ./ppocr_det_v3_serving/ \
--serving_client ./ppocr_det_v3_client/
9、转换识别模型
python3.7 -m paddle_serving_client.convert --dirname ./ch_PP-OCRv3_rec_infer/ \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--serving_server ./ppocr_rec_v3_serving/ \
--serving_client ./ppocr_rec_v3_client/
10、调整模型服务配置文件config.yml
减少并发数，比如调整为2:1防止资源不足
#并发数，is_thread_op=True时，为线程并发；否则为进程并发 concurrency: 8
#并发数，is_thread_op=True时，为线程并发；否则为进程并发 concurrency: 4
vim config.yml
i
:wq
11、启动服务端
python3.7 web_service.py &>log.txt &
12、启动客户端
python3.7 pipeline_http_client.py
二、Paddle Serving C++ 部署
1、下拉镜像
docker pull registry.baidubce.com/paddlepaddle/serving:0.9.0-cuda10.1-cudnn7-devel
2、进入镜像
nvidia-docker run -it --entrypoint=/bin/bash registry.baidubce.com/paddlepaddle/serving:0.9.0-cuda10.1-cudnn7-devel
3、安装特定PaddlePaddle
wget https://paddle-inference-lib.bj.bcebos.com/2.3.0/python/Linux/GPU/x86-64_gcc8.2_avx_mkl_cuda10.1_cudnn7.6.5_trt6.0.1.5/paddlepaddle_gpu-2.3.0.post101-cp37-cp37m-linux_x86_64.whl
pip3.7 install paddlepaddle_gpu-2.3.0.post101-cp37-cp37m-linux_x86_64.whl
3.1、直接在线安装
python3.7 -m pip install paddlepaddle-gpu==2.3.0.post101 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
4、安装相关的包
pip3.7 install paddle-serving-client==0.9.0 paddle-serving-app==0.9.0 paddle-serving-server-gpu==0.9.0.post101 -i https://mirror.baidu.com/pypi/simple
5、下载OCR repo
git clone https://github.com/PaddlePaddle/PaddleOCR
cd PaddleOCR/deploy/pdserving/
6、下载并解压 OCR 文本检测模型
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar -O ch_PP-OCRv3_det_infer.tar && tar -xf ch_PP-OCRv3_det_infer.tar
7、下载并解压 OCR 文本识别模型
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar -O ch_PP-OCRv3_rec_infer.tar && tar -xf ch_PP-OCRv3_rec_infer.tar
8、转换检测模型
python3.7 -m paddle_serving_client.convert --dirname ./ch_PP-OCRv3_det_infer/ \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--serving_server ./ppocr_det_v3_serving/ \
--serving_client ./ppocr_det_v3_client/
9、转换识别模型
python3.7 -m paddle_serving_client.convert --dirname ./ch_PP-OCRv3_rec_infer/ \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--serving_server ./ppocr_rec_v3_serving/ \
--serving_client ./ppocr_rec_v3_client/
10、调整模型服务配置文件config.yml
减少并发数，比如调整为2:1防止资源不足
#并发数，is_thread_op=True时，为线程并发；否则为进程并发 concurrency: 8
#并发数，is_thread_op=True时，为线程并发；否则为进程并发 concurrency: 4
vim config.yml
i
:wq
11、启动服务端
python3.7 -m paddle_serving_server.serve --model ppocr_det_v3_serving ppocr_rec_v3_serving --op GeneralDetectionOp GeneralInferOp --port 9293 &>log.txt &
12、客户端配置修改
由于需要在C++Server部分进行前后处理，为了加速传入C++Server的仅仅是图片的base64编码的字符串，故需要手动修改 ppocr_det_v3_client/serving_client_conf.prototxt 中 feed_type 字段 和 shape 字段，修改成如下内容：
vim ppocr_det_v3_client/serving_client_conf.prototxt
i
:wq
feed_var {
name: "x"
alias_name: "x"
is_lod_tensor: false
feed_type: 20
shape: 1
}
13、启动客户端
python3.7 ocr_cpp_client.py ppocr_det_v3_client ppocr_rec_v3_client