1.1 KiB
1.1 KiB
yolov3-spp
For the Pytorch implementation, you can refer to ultralytics/yolov3
Following tricks are used in this yolov3-spp:
- Yololayer plugin is different from the plugin used in yolov3. In this version, I reimplement the calculation of three yololayer plugins into one to improve speed. And the yololayer detect outputs are limited to maxmium 1000. The first number of output is number of targets in current image.
- Batchnorm layer, implemented by scale layer.
Excute:
// 1. generate yolov3-spp_ultralytics68.wts from pytorch implementation with yolov3-spp.cfg and ultralytics68.pt
// 2. put yolov3-spp_ultralytics68.wts into yolov3-spp
// 3. build and run
cd yolov3-spp
mkdir build
cd build
cmake ..
make
sudo ./yolov3-spp -s ../samples // serialize model to plan file i.e. 'yolov3-spp.engine'
sudo ./yolov3-spp -d ../samples // deserialize plan file and run inference
// 4. see if the output is same as pytorch implementation, and see the detect result in build