I would like to ask how to get a int8 model?

[wilicyy]

i trained a yolov4 model with 1536*1536, but FP16 runs only 15FPS.
how can i get a simple way to change int8 for more fast?
pls give me some instructions about that.

thank you

[AastaLLL]

Hi,

Which framework do you use for inference?
If it is TensorRT, you can create an INT8 engine with the following command directly:

$ /usr/src/tensorrt/bin/trtexec --int8 ...

Thanks.

[wilicyy]

i user yolov4 to onnx to TensorRT, it works well by fp16,
and i can create int8 engine successfully but it can not recognize object from image.
i am not sure it is input problem(608-608??)
my model should be 1536_1536…

thanks

[wilicyy]

not input problem i am sure, and int8.engine can not find object after create from onnx,
my model is yolov4.
does it need a cal_trt.bin file?

thanks

[wilicyy]

anybody to answer me about yolov4 int8 questions?

thanks

[AastaLLL]

Hi,

Sorry for the late update.
If your app works well with the fp16 model, then the issue should come from the calibration file.

Do you use the official YOLOv4 model?
If yes, you can try the calibration cache from our user:

tensorrt_demos/yolo/calib_cache at master · jkjung-avt/tensorrt_demos

Thanks.

[wilicyy]

yes i use official yolov4-tiny, but it is 1536*1536model which be trained,
can i use yolov4-tiny-406-calib_cache directly for int8?

thanks!

[AastaLLL]

Hi,

Since the input size is different, it’s recommended to re-generate a calibration cache.
You can find an example in GitHub as well:

GitHub - jkjung-avt/tensorrt_demos: TensorRT MODNet, YOLOv4, YOLOv3, SSD,...

Thanks.

[wilicyy]

thank you i will study that.