search

WongKinYiu / yolor

implementation of paper - You Only Learn One Representation: Unified Network for Multiple Tasks (https://arxiv.org/abs/2105.04206)
GNU General Public License v3.0
1.99k stars 518 forks source link

transformer yolo is coming！ #1

Closed Zigars closed 3 years ago

Zigars commented 3 years ago

I'm wait for the open!

WongKinYiu commented 3 years ago

Oh, we have already integrated transformer/swin-transformer/... to yolo in our private projects. I think our new transformer based yolo will be released in this year.

Zigars commented 3 years ago

wow, I‘m waiting for the released. And I have a question about yolor, I saw the fps of yolor-d6 can get 30, which part of the net architectural reduce the inference time? I know you add some new implicit and explicit module, but I saw the parameter and Flops did not reduce. can I get the reason for it?

WongKinYiu commented 3 years ago

Compare to YOLOv4-P7, YOLOR-D6 has only 57% FLOPs. This is the main reason it can achieve higher FPS. And we apply the methods proposed in CSPNet, which reduce the memory bandwidth in inference process (Appendix).

image

Zigars commented 3 years ago

oh, sorry about it , maybe I should re-read of the paper and code, I did not notice the appendix, thanks!

RaphaelWag commented 2 years ago

@WongKinYiu are there any updates when you plan to release your transformer/swin-transformer yolo?

Splendon commented 2 years ago

@WongKinYiu What time would you release yolor with transformer/swin-transformer ?