I am using TensorRT to optimize performance of my model (yolo tiny v3). When I choose INT8 to optimize my model, the performance is OK but the accuracy is not good. I have been added datasets to calibration, but result is not OK.
I have some question, please helf me to solve it:
- Can you show me how to improve the accuracy when running in INT8 mode.?
- Is there any different of performance and accuracy when optimize ONNX model and primitive Tiny-yolo-v3
Thanks & Brgs