Closed BTVinh0409 closed 2 years ago
Is it possible that your GPU does not support FP16 inference?
I have the same issue with Yolov4, when I change to FP16, the file size is the same like FP32 and the FPS are the same on both.
What GPU are you using?
I am using the Jetson AGX Xavier.
hi @mfoglio, I am using RTX 2080 and it supports FP16 inference. Can you provide the FPS when you inference FP32 and FP16 models? Thank you so much
I will check it
I noticed that the same issue happens when you already have the calib.table and you build the engine on INT8.
Any update? I have the same issue here, but maybe I am missing something. I am running a custom yolor-csp model and get around 20fps on the Jetson AGX Xavier. Changing network-mode
from 0 to 2 does not change the fps at all
I need more days to check, I'm full of work these days.
I fixed this bug by following the link below:
https://forums.developer.nvidia.com/t/deepstream-6-yolo-performance-issue/194238/22
Mainly you have to modify this files: yolo.cpp yolo.h nvdsinfer_yolo_engine.cpp
Mainly you have to modify this files:
If you do that, you will lose all optimizations and new models did by this repo.
Just only put the new line in the file, for example for yolo.cpp you need delete this line and put the new line.
I confirm, after steps mentioned by Pakike engine model size for FP16 decreased which caused FPS increased to normal like in DS5.1 version. Previously FP16 TRT engine model size was the same as FP32.
Thank you @pakike, I confirmed the issue and updated the repo.
Hi, I tried converting yolov5l.pt model to FP32 TensorRT engine. It works well and the FPS is around 60. But when I try to convert to FP16 it doesn't seem to increase the FPS any further. And both exported models have the same size. Maybe I did something wrong?