Speed Question(Wrong Speed Test) about cgnet HOT 11 CLOSED

Hi! @wutianyiRosun
There are two questions I want to ask:

1. I test your model with 1024 * 2048 as input images. However I only got 9.5 fps on single 1080TI.
   I use the code like the following
   start = time.time()
   out = model(img)
   torch.cuda.synchronize()
   print('Speed: {} fps.'.format(1.0/(time.time()-start)))

2. In your paper, ICnet only 16 fps which is not consistent with his paper. (30 fps)
   I don't know why. The gap between K80 and 1080TI is such large ??
   Is my testing method wrong? I want to know the way you evalute your model for speed.
   Also, I use my test method to test ESP-net which result in nearly the same as your paper report(49 fps)

from cgnet.

I got the same result as you, a single 1080Ti got about 9 FPS. @lxtGH
And if I didn't use torch.cuda.synchronize()
The FPS will increase to ~70Fps, this code is necessary.

from cgnet.

@Reagan1311 Yes, GPU and CPU must synchronize (CPU must wait until the end of GPU forwarding )

from cgnet.

10FPS only! I test it on 1080Ti.
I suggest that you change your arxiv paper.

from cgnet.

Hi , @wutianyiRosun
Did you use this line ?
torch.cuda.synchronize() for speed test?

from cgnet.

I met the same question. I hope I can get an answer. Thank you.

from cgnet.

I think it is wrong speed testing.

from cgnet.

Hi, @lxtGH
A1: The speed we report is the average speed on the verification set. What is your cuda and cudnn version?
A2: The speed of ICNet is tested on K80, The paper of ICNet reported 30 fps, which is tested on 1080TI. The speed between them is about two to three times the difference.

from cgnet.

@lxtGH @wutianyiRosun
I have a problem:
Pytorch DwConv's implement is not good,it's very slow. The CGNet'paper result is based on Pytorch?

from cgnet.

@wutianyiRosun If you didn't use the line of code, your speed reported in your paper is wrong.
@yyfyan What speed results you get ?

from cgnet.

@meteorshowers, @LinZhuoChen We will update in the next few days.

from cgnet.