YOLOR - You Only Learn One Representation: Unified Network for Multiple Tasks

[AlexeyAB]

YOLOR - You Only Learn One Representation: Unified Network for Multiple Tasks:

YOLOR is better than YOLOv5 (u5r5), PP-YOLOv2, CenterNet2, EfficientDet and many other

• paper: https://arxiv.org/abs/2105.04206
• Pytorch code: https://github.com/WongKinYiu/yolor/tree/paper
• papers with code comparison: https://paperswithcode.com/sota/real-time-object-detection-on-coco

It is improved Scaled-YOLOv4-P6 (+0.9% AP):

---

---

[Hwijune]

@AlexeyAB hi!

i want to test the YOLOR model on darknet.

1. Can I train the YOLOR model in Darknet? Or can I convert the torch to darknet?

2. I wonder if the four layers below can be replaced with the existing layers of the darknet. [implicit_add] [implicit_mul] [control_channels] [shift_channels]

[IgnacioTsai]

As I know the four new layers: [implicit_add] [implicit_mul]

[control_channels] [shift_channels]

[shift_channels] means create the same tensor shape of the input tensor, making normal distribution with average 0, standard deviation as 0.2 in default setting, and add both tensor value on each element using implicit_add.

[control_channels]means create the same tensor shape of the input tensor, making normal distribution with average 0, standard deviation as 0.2 in default setting, and multiply both tensor value on each element using implicit_mul.

But I did not see the difference between ImplicitA, ImplicitC and ImplicitM. It looks like that they are the same, but get different result by [shift_channels] and [control_channels].

In the paper, should both methods called "feature alignment"?

By the way, YOLOR uses SiLU activate function.