YOLOX-S works fine, but tiny and nano do not work

NVIDIA-AI-IOT / torch2trt

An easy to use PyTorch to TensorRT converter

MIT License

4.58k stars 675 forks source link

YOLOX-S works fine, but tiny and nano do not work #697

Open tulbureandreit opened 2 years ago

tulbureandreit commented 2 years ago

Hello,

As a summary, I try to run yolox with deepsort on a nvidia jetson xavier AGX dev kit

I tried and managed to run YOLOX-S with deepsort and a nice UI, using the python SDK, but when I try to do the same for yolox-nano or yolox-tiny, it does not work and I get a segmentation fault as a result.

I converted the models to TRT the same way and I use them exactly the same way. YOLOX-s works, but only with 1fps, while nano returns a segmentation fault.

Any ideas to why?

tulbureandreit commented 2 years ago

NVIDIA Jetson AGX Xavier [16GB]
- Jetpack 4.6 [L4T 32.6.1]
- NV Power Mode: MODE_30W_ALL - Type: 3
- jetson_stats.service: active
Libraries:
- CUDA: 10.2.300
- cuDNN: 8.2.1.32
- TensorRT: 8.0.1.6
- Visionworks: 1.6.0.501
- OpenCV: 4.5.1 compiled CUDA: YES
- VPI: ii libnvvpi1 1.1.15 arm64 NVIDIA Vision Programming Interface library
- Vulkan: 1.2.70

dvando commented 2 years ago

Hi @tulbureandreit, most of the time segmentation fault error is caused by either data type or data size. I once had that kind of error too in YOLOX, and it turns out I use the wrong data type for inference, changed it to float16_t and it works fine (I'm using half precision)