PaddleOCR 2.0.2 is slower than 1.1 using CUDA11.2

[indrasweb]

I finally managed to get the latest version running without segfaults by doing the following:

2) Update to cuda11.2 https://developer.nvidia.com/cuda-downloads 3) sudo docker run --name ppocr --gpus all -v $PWD:/paddle --shm-size=32G --network=host -it paddlepaddle/paddle:2.0.0rc1-gpu-cuda11.0-cudnn8 /bin/bash 4) python3.8 -m pip install paddleocr (no need to install paddlepaddle-gpu because the docker already has it)

However, inference speed is about 3x slower than using paddleocr==1.1 + paddlepaddle-gpu 1.8.5.post107

• CUDA10 (though this model is more accurate). Any ideas how I could speed it up?

Also, how could I install outside of docker? It seems that the version of paddlepaddle-gpu installed in paddle:2.0.0rc1-gpu-cuda11.0-cudnn8 docker is not available on pypi - I've lost the output now but it was a version with post110 at the end, I think.

[Xreki]

There are some performance issue on CUDA11.0. We are fixing it.

[LDOUBLEV]

I guess your environment is CUDNN8 + CUDA11.2. We found that there is a performance degradation problem when using GPU to predict with cudnn8. It is recommended to use cudnn7 + cuda10

Also, how could I install outside of docker?

# reference https://www.paddlepaddle.org.cn/
python -m pip install paddlepaddle-gpu==2.0.0.post101 -f https://paddlepaddle.org.cn/whl/stable.html

BTW, paddle2.0rc1 version does not support cuda11.2. If U have more question about PaddleOCR, you can open an issue in OCR Issue