The problem about FPS about torchseg HOT 10 CLOSED

Hi, @ycszen
Why you use the cityscapes.bisenet.R18.speed folder rather than cityscapes.bisenet.R18 ?
I found that the network.py is different.
In the cityscapes.bisenet.R18.speed/network.py, you drop two srtucture where

        if is_training:
            heads = [BiSeNetHead(conv_channel, out_planes, 2,
                                 True, norm_layer),
                     BiSeNetHead(conv_channel, out_planes, 1,
                                 True, norm_layer),
                     BiSeNetHead(conv_channel * 2, out_planes, 1,
                                 False, norm_layer)]
        else:
            heads = [None, None,
                     BiSeNetHead(conv_channel * 2, out_planes, 1,
                                 False, norm_layer)]

I wonder know why you do that?

@Reagan1311 Besides, I test the cityscapes.bisenet.R18.speed/network.py in TITAN xp, the FPS is ~30 on 1024x2048, and the model output shape is 1x19x128x256.
And in the cityscapes.bisenet.R18/network.py, the FPS is ~25 on 1024x2048 and the output shape is 1x19x1024x2048.

Looking forward to your reply !

from torchseg.

That's still a large gap from the original paper's FPS. @MrLinNing

which is:
Xception39 105.8
Res18 65.5
(Titan XP with 2048x1024 resolution input)

How these papers figure out the accurate FPS? Really curious...

from torchseg.

So currently, are there any good solutions to compute the accurate FPS? @ycszen
Looking forward to your reply！

from torchseg.

Hi,
If you focus on the speed experiments, you should use the experiments with the suffix .speed, like the folder cityscapes.bisenet.R18.speed. Then, as we elaborated in our paper, in the speed experiments, we first resize the 1024x2048 image into 768x1536 input size.
Finally, I have to mention that the speed of Depthwise Conv in PyTorch has a little problem. Its speed is slower than the normal convolution, which results in the speed of our Xception is not accurate in this repo. More details about the depthwise conv in PyTorch can be found here.

from torchseg.

Hi，
I also have the same question and look forward to your reply . @ycszen

from torchseg.

Thanks, @ycszen
By the way, can you share the 960x720 CamVid dataset and usage instruction? I can just find the 480x360 resolution( https://github.com/alexgkendall/SegNet-Tutorial/tree/master/CamVid)

from torchseg.

Thanks for all attention.
In this repo, it is just a pytorch-version reimplementation of our proposed method. Actually, we implement this method with our own framework when I submitted my paper, in which the depthwise conv is correctly optimized.
However, in PyTorch, the depthwise cov is not correctly optimized, which is even slower than the original conv. Therefore, there is a gap between this repo and our paper.
Maybe the official PyTorch will support the optimized depthwise conv.
Or I can implement the optimized depthwise conv in PyTorch if I have time ...

from torchseg.

Hi, @ycszen
Why you use the cityscapes.bisenet.R18.speed folder rather than cityscapes.bisenet.R18 ?
I found that the network.py is different.
In the cityscapes.bisenet.R18.speed/network.py, you drop two srtucture where

        if is_training:
            heads = [BiSeNetHead(conv_channel, out_planes, 2,
                                 True, norm_layer),
                     BiSeNetHead(conv_channel, out_planes, 1,
                                 True, norm_layer),
                     BiSeNetHead(conv_channel * 2, out_planes, 1,
                                 False, norm_layer)]
        else:
            heads = [None, None,
                     BiSeNetHead(conv_channel * 2, out_planes, 1,
                                 False, norm_layer)]

I wonder know why you do that?

@Reagan1311 Besides, I test the cityscapes.bisenet.R18.speed/network.py in TITAN xp, the FPS is ~30 on 1024x2048, and the model output shape is 1x19x128x256.
And in the cityscapes.bisenet.R18/network.py, the FPS is ~25 on 1024x2048 and the output shape is 1x19x1024x2048.

Looking forward to your reply !

We do this to accelerate the inference speed.
Because both structures are auxiliary modules which are designed for better training, we can drop them in the inference phase.

from torchseg.

from torchseg.

Thanks, @ycszen
By the way, can you share the 960x720 CamVid dataset and usage instruction? I can just find the 480x360 resolution( https://github.com/alexgkendall/SegNet-Tutorial/tree/master/CamVid)

Hi! Are you able to get the 960x720 CamVid dataset? Thanks!
@ycszen Could you comment on the resolution of CamVid dataset? Thank you!

from torchseg.