Model | Size | FPSRK3588, 3 cores int8 batch=1 include all process |
Download |
---|---|---|---|
Tiny-LRELU | 384×640 | 65 | coco model, coco config visdrone model, visdrone config |
Tiny | 384×640 | 24 (too many SiLU activate layers) |
coco model, coco config visdrone model, visdrone config |
- this toolkit only supports python3.6, python3.8 and python3.10
cd toolkit_install
./rknn_toolkit_install 38 # or 36, or 310
- or just use the rknn models given in the above table.
cd /path_to_edgeyolo_project_root_path
python export.py --weights edgeyolo_tiny_lrelu_coco.pth \ # your pth weights
--input-size 384 640 \ # for 16:9, if 4:3, use 480 640
--rknn \ # export rknn model
--dataset cfg/dataset/coco.yaml \ # calib dataset
--num-img 100 \ # number of calib img
# optional but not commend
--rknn-platform rk3588 \ # rk3566 and so on, you can convert model, but our code only support rk3588(and rk3588s)
then it generates 4 files as follows
output/export/edgeyolo_tiny_lrelu_coco/384x640_batch1.rknn # rknn model
output/export/edgeyolo_tiny_lrelu_coco/384x640_batch1_for_rknn.onnx # different from onnx file for tensorrt
output/export/edgeyolo_tiny_lrelu_coco/384x640_batch1.json # json file, not used currently
output/export/edgeyolo_tiny_lrelu_coco/384x640_batch1.yaml # config file to use
- copy dir 'cpp/rknn' to your rk3588 device.
- cd rknn
- copy converted ".yaml" and ".rknn" file to ./model. if rename, rename both file with the same name.
- then
chmod +x ./setup_rk3588.sh
./setup_rk3588.sh
cd install/rknn_edgeyolo_demo_Linux
./rknn_det -? # parser helper
./rknn_det --model model/384x640_batch1.rknn \
--video \ # use video source(include rtsp/rtmp), or --device for camera id, or --picture for single image.
--source /path_to_your_video.mp4 \ # or 0, or /path_to_your_image.jpg
--no-label \ # draw bounding box without label
--loop # play in loop, press esc to quit