YOLOv3 cfg file

[canyilmaz90]

Hello, firstly thanks for sharing this great work. I want to test this for yolov3 (not tiny), can you share config file for it? Thanks and regards

[WongKinYiu]

@canyilmaz90 Hello,

Do you mean "Darknet53-PRN with YOLOv3 head" or "Darknet53 with YOLOv3-PRN head"?

[canyilmaz90]

@WongKinYiu Hi,

I mean the one you named YOLO-v3-PRN in your paper. I think it's "Darknet53-PRN with YOLOv3 head", but I'm not sure. Btw, what is the difference between two? And is there a possibility of "Darknet53-PRN with YOLOv3-PRN head"?

[WongKinYiu]

Yes, it's "Darknet53-PRN with YOLOv3 head". I did not test "Darknet53-PRN with YOLOv3-PRN head".

I will share the cfg files in few days. I am busy in AVSS 2019 & ICIP 2019 now.

[canyilmaz90]

Thanks a lot! Good luck with the conferences :)

[WongKinYiu]

cfg file for YOLO-v3-PRN in the paper darknet53-prn-yolov3.cfg.txt

[canyilmaz90]

Thank you so much. After I train with these configurations, I will give you feedback.

[WongKinYiu]

The weight file of YOLO-v3-PRN pre-trained model is too big, i can not upload it to github.

[canyilmaz90]

Thanks anyway. Maybe you could put it in a cloud account, such as google drive

[LukeAI]

@WongKinYiu Would you also consider releasing Yolo-v3_FPRN ?

[WongKinYiu]

@LukeAI Hello,

For YOLO-v3-FPRN, I modified shortcut_cpu (blas.c) and shortcut_gpu (blas_kernels.cu) instead of modified the cfg file.

for example, change int minc = (c1 < c2) ? c1 : c2; in shortcut_cpu to int minc = (c1 < c2) ? c1//2 : c2//2;. then the cfg file is totally same as that for YOLO-v3.

[iraadit]

Hi @WongKinYiu

Will you share the YOLO-v3-PRN pre-trained COCO weights? On Google Drive, Dropbox or similar sharing service?

Thank you

[WongKinYiu]

@iraadit hello,

we will release our new model in few days. i ll post the link here.

[WongKinYiu]

@iraadit

we release our new model at https://github.com/WongKinYiu/CrossStagePartialNetworks

[LukeAI]

I'd also like to request any pretrained weights you might have for YOLO-v3-PRN

[reactivetype]

Do you mean "Darknet53-PRN with YOLOv3 head" or "Darknet53 with YOLOv3-PRN head"?

@WongKinYiu Would you please share the cfg of "Darknet53 with YOLOv3-PRN head" to understand the difference?

By the way, what is the main difference between the new CSPDarknet53 backbone and Darknet53-PRN backbone?

[WongKinYiu]

Hello,

We only apply PRN on Darknet53 backbone, which is Darknet53-PRN with YOLOv3 head. We do not have cfg of Darknet53 with YOLOv3-PRN head. https://github.com/WongKinYiu/PartialResidualNetworks/issues/5#issuecomment-532002269

The main difference is that the channel number in each stage of Darknet53-PRN will linear decrease to a fraction.

[reactivetype]

Thanks @WongKinYiu for the clarification. I actually referred to main conceptual differences between Darknet53-PRN and the new CSPNet version (CSDarknet53).

[WongKinYiu]

PRN aims to maximize the combinations of gradient of different layers. CSPNet aims to maximize the combinations of different layers and minimize the mutual information of gradient of different layers.