1.4 KiB
1.4 KiB
English | 简体中文
Overall process
1.Pytorch->TensorRT
python export.py --weights "torch's path" --onnx2trt --fp16_trt
2.TensorRT inference
python torch2trt/main.py --trt_path "trt's path"
Image preprocessing -> TensorRT inference -> visualization
Time-consuming comparison
| Backbone | Pytorch(ms) | TensorRT_FP16(ms) |
|---|---|---|
| yolov5n-0.5 | 7.7 | 2.1 |
| yolov5n-face | 7.7 | 2.4 |
| yolov5s-face | 5.6 | 2.2 |
| yolov5m-face | 9.9 | 3.3 |
| yolov5l-face | 15.9 | 4.5 |
Pytorch=1.10.0+cu102 TensorRT=8.2.0.6 Hardware=rtx2080ti
python torch2trt/speed.py --torch_path "torch's path" --trt_path "trt's path"
Visualization
| yolov5n-0.5 | yolov5n-face |
|---|---|
 |
 |
| yolov5s-face | yolov5m-face | yolov5l-face |
|---|---|---|
 |
 |
 |