HITSZ-NRSL / yolox_for_cann_atlas200dk

yolox_opencv_python

This is a project to deploy YOLOX on Atlas200DK using CANN.

在Atlas200dk中使用CANN部署yolox模型推理

Environments

You should have set up the CANN environments on Atlas200DK,
some other needed packages are as belows

opencv_python (>=4.3 only for opencv dnn inference)  
opencv_contrib_python>=4.3 (only for opencv dnn inference) 
numpy
pyACL (CANN environments have set this)

Usage

First, git clone this code, yolox_nano.onnx has been on the 'model' dir

if you want other models, you can download them on the origin repo: https://github.com/Megvii-BaseDetection/YOLOX.git

and put the downloaded onnx into the ./model dir:

git clone https://github.com/stunback/yolox_for_cann_atlas200dk.git
# if you have downloaded yolox_s.onnx
cd yolox_for_cann_atlas200dk
mv onnx_path model/

Second, remove the focus layer on the onnx model

change the ONNX_MODEL_PATH on ./script/yolo_onnx_opt.py

then run the script:

cd script
python yolo_onnx_opt.py

Third, use atc tool to export the onnx model into cann model

Use yolox_nano_simple.onnx for example:

cd ../model
atc --model=./yolox_nano_simple.onnx --framework=5 --output=yolox_nano_simple --input_format=NCHW --soc_version=Ascend310

At last, run the inference demo

change the model path on src/acl_yolox.py, and run:

cd ../src
python acl_yolox.py 

Additional

An opencv inference demo is also provided:

cd ../src
python main_yolox.py

Model Inference Speed

Hardware: Atlas200dk npu

yolox_nano(416) onnx=308.2ms cann=11.5ms

yolox_tiny(416) onnx=763.8ms cann=12.2ms

yolox_s(640) onnx=2907.3ms cann=16.5ms

yolox_x(640) onnx=24268ms cann=62.8ms (4.49GFLOPs/s)

Others

Blogs about yolov5, yolox and nanodet:
https://blog.csdn.net/qq_41035283/article/details/119150751