YOLOv5 Repo on Xavier?

russ.ferriday · July 15, 2020, 12:56pm

v5 is pytorch, so no, I did not convert. I have not done much related to the network, and i’ve not done any training, yet. I’m trying to use the output of the detector to perform a useful security task.
My frame rate in docker for 640x480 RTP video is about 10 hz on 10W power, and dumping annotated images to ssd. But I can run a separate stream also around the same rate, and I watch the GPU work interleaving nicely.

dusty_nv · July 15, 2020, 2:42pm

Just an aside - if this is using PyTorch, I think this issue is unrelated to TensorRT. There is a PyTorch issue with multi-threaded destruction of resources that causes a segfault when exiting:

C++ API: Crash in cudnnDestroy() when deconstructing · Issue #17658 · pytorch/pytorch · GitHub
Fix a segfault in DeviceThreadHandlePool and PoolWindow by VivekPanyam · Pull Request #36416 · pytorch/pytorch · GitHub

mojtabamir962013193 · July 17, 2020, 6:38am

ty

mojtabamir962013193 · July 17, 2020, 6:41am

is there any difference between installing opencv from source or using pip3 install python3-opencv?
Does opencv from source support cuda? if is so we get higher framerate compared to pip3 install python3-opencv?

mojtabamir962013193 · July 19, 2020, 8:58am

Hi russ.ferriday
I could test it in jetson nano on a offline video. thank you so much.
But how can I test it in live stream with a camera connected to jetson?
python detec.py --source 0 does work in windows no jetson.what should I change?

russ.ferriday · August 4, 2020, 9:03am

If you are using an IP camera your command line might include
…–source rtsp://

hiro.ohnishi · August 16, 2020, 10:24am

I tried the below.

github.com

topiaruss/yolov5/blob/master/Xavier_NV_Notes.md

# Notes for NVIDIA Xavier NV on docker

## Build a base image, extending dusty_nv's image  
 
    sudo docker build -t yolov5-base -f aarch64.Dockerfile .
 
## Build an image with app specifics, based on the base image
 
    sudo docker build -t yolov5-app -f yolov5.app.Dockerfile .
 
## Use a script to invoke the image for detection with mounted volumes. 
   
Default weights will be pulled the first time into ./weights/
and you should put images into ./inference/. Output will be in
./inference/output/
   
    ./run_xavier_detect.sh

Currently there is a kludge in place to make the following script work. It will 
put sample images into the inference/output/ directory.

This file has been truncated. show original

But I found this error “Segmentation fault (core dumped)” when I tried detect.py in the container.

I’m using Jetson Xavier NX.
Please kindly let me know how to solve.