This repository is a quant implementation demo of PETR. Just for self study, I use the official "petr_vov_p4_800x320.pth" for development.
You can PTQ the model following:
python tools/quant/ptq_bev.pyYou can QAT the model following:
python tools/quant/qat_bev.py You can generate the onnx following:
python tools/quant/export_onnx.py| config | mAP | NDS | Latency |
|---|---|---|---|
| PETR-vov-p4-800x320 | 37.8% | 42.6% | 64.8768 ms |
| PTQ | 32.89% | 30.20% | 31.5722 ms |
| QAT | 30.94% | 27.82% |
- QAT Due to limited personal resources, a single card 3080 single batch trained for 10 epochs, mAP has been rising, there should be room for improvement
- onnx and pth will be pushed later
Many thanks to the authors of Lidar_AI_Solution and PETR .
If you have any questions, feel free to open an issue.