Article Directory
1. Preparation
- onnx 1.5
- Setup version 10
- Self-trained darknet model
- yolov3 darknet to onnx script:
#darknet 转 onnx脚本
链接：https://pan.baidu.com/s/1yzk9iiCR21qCh3Q9r6ybGw
提取码：f4xx
- Copy and rename the yolov3 project
- Copy the converted onnx model to:
horizon_x3_tc_1.1.17e/samples/01_common/modelzoo/mapper/detection/yolov3/
- Modify the number and name of classes in the reasoning source code of yolov3_post_process.cc
- Modify the 9 anchors in the anchor source code, where the parameters in the anchor need to be divided by 8.
2. Check the model
cd 04_yolov3_01/mapper
- According to the case of yolov5 using the onnx model, modify 01_chack
- Modify the yolov3_config.yaml configuration file
Only modify the model path and model type here
sh 01_check.sh
- Data calibration
sh 02_preprocess.sh
- Model conversion
It is better to have more than 4GB of memory, otherwise there may be insufficient memory when converting the model.
sh 03_build.sh
There is an error here:
modify the input location of the model in the yolov3_config.yaml configuration file according to the error: the
model_output file will be generated in the directory after the conversion is successful
Three. Run on the board
cd runtime_arm
sh 01_build.sh
sh 02_preprocess.sh
#将构建的包发送到开发板下
sh 03_scp_to_board.sh 192.168.124.103
-
The file will be sent to /userdata/samples/
-
SSH development board to run inference scripts
-
Modify the image path of board_test_image in env.conf
cd /userdata/samples/yolov3
sh dev_board_01_infer.sh
- View inference results
After the inference is over, you can view the output picture under the image_out file.
Four. Reference
https://developer.horizon.ai/resource
https://developer.horizon.ai/forum/id=5f5f19e8cc8b1e59c8582b0a