Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Question regarding performance. #69

Open
eeerrrwwwuuu opened this issue May 2, 2019 · 3 comments
Open

Question regarding performance. #69

eeerrrwwwuuu opened this issue May 2, 2019 · 3 comments

Comments

@eeerrrwwwuuu
Copy link

eeerrrwwwuuu commented May 2, 2019

I'm doing bad thing Naisy, You'll hate me, but overall goal is quite simple. Tensor takes 5 gigs or GPU I want to create queue which will use socket.io to read data from multiple cameras and returns boxes.
https://drive.google.com/drive/folders/14d50kFKWnBUaLwL7z06CF2PwHUklaJdA?usp=sharing
I have some problem with performance, I tried to mess with parameters but it makes no difference, it looks like GPU analyse takes a lot of time.
Got any idea why there might be this heart beat ?
https://photos.app.goo.gl/r1srTC1cEHnw8Z4t6
I tasted it previously it was about 20FPS for one source, but now it hangs on gpu queue for half a second. I really have no idea where I should look for improvements.

  • camera_worker.py is just generator which reads frames from file and sends message to socket that frame is ready ( images are exchanged threw RAM DISK )
  • run_socketio_tensor.py is just big script which executes without include of model this big loop
  • socketio.py is equivalent to video or webcam..

ps. be lenient it's my first fight with python ;-) on daily bases I work with DWH/BI .

Overall I have also problem with sending first few images into gpu session, script seems to hang if I restart camera_worker.py it works, is there any special reason why it can't work from first frame?

@naisy
Copy link
Owner

naisy commented May 3, 2019

Hi, @eeerrrwwwuuu,

faster rcnn is really slow. Was 20 fps with faster rcnn? I think that 20 fps with faster rcnn is maximum performance on desktop gpu. If you are using Jetson, it should be slower. It is about 1 fps.

At the first frame, memory initialization also works. Therefore it is slow. If the first frame can not be delayed, you should run with dummy frame before read camera frame.

@eeerrrwwwuuu
Copy link
Author

I guess I removed all what is possible to leave just loop, which takes static image from ram disk, all stats are gone, but still gpu session takes about 1 second per image or RTX 2070 and Ryzen.
When I used Your code without modification I was able to reach 20FPS. Maybe it goes threw CPU instead?

https://drive.google.com/open?id=1d3Fc5H9q1EfKUPIXTnXgRshOlDZQAlKE

@eeerrrwwwuuu
Copy link
Author

@naisy my bad, after rebooting computer it started to fly ! ;-) Thank You as always!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants