Darknet: some problems about using Gaussian_yolov3

[EtheneXiang]

this is usage in Gaussian_yolov3_BDD.cfg

I transfer it into my self-made cfg while i use CIOU-LOSS instead of GIOU-LOSS

but when loading data, segmentation fault

[AlexeyAB]

Does it work well - Gaussian_yolo + default Loss?

[EtheneXiang]

let me try.
when i use CIOU without Gaussian_yolo , it works well.

[EtheneXiang]

oh no. not work, Gaussian_yolo + default Loss.

but i run Gaussian_yolov3_BDD.cfg, it works

[AlexeyAB]

oh no. not work, Gaussian_yolo + default Loss.

Set subdivisions=64

[EtheneXiang]

LOL, just now, my GPU is running for other program....
i test again

[EtheneXiang]

I can assure you that my self-made cfg is all right.
i have trained it with CIOU

[EtheneXiang]

and i found something intersting, i trained yolov3.cfg, csresnext50-panet-spp-original-optimal.cfg and myself-made cfg with car detection dataset : UA-DETRAC dataset: http://detrac-db.rit.albany.edu/

80K imgs for train, 50k imgs for valid/test

and all the loss/map curve are funny, map reached its maximum in before 10,000 iterations, and then stay the same or go down a little bit. same data agument, same lr, same bathszie=128 same GPU 4*1080TI

• csresnext50-panet-spp-original-optimal.cfg

• yolov3.cfg
  

• myself-made cfg
  

[AlexeyAB]

Since csresnext50-panet-spp-original-optimal.cfg uses CIoU, so may be during training the AP50 decreases while AP75 and AP50...95 increases. But you see only AP50.

Check AP75 ./darknet detector map ... -iou_thresh 0.75 for weights-files 5000, 10000 and 30000 iterations.

---

Or may be you set bad params in cfg-file.

[EtheneXiang]

params i used in those three cfg

[AlexeyAB]

Check AP75

[EtheneXiang]

OK, it's long time to run map, now i'm running AP75
56168 imgs
Time: 1411 Seconds, about 23.5minutes

-thresh 0.5 -iou_thresh 0.75 -points 101

and maybe there is a liite bug about save weights, i found all saved names are strange, they are not multiples of ten while end with EIGHT 8.

[EtheneXiang]

10k iters 60.10
20K iters 63.12
30k iters 63.66
40k iters 64.48
50K iters 66.24

yeah, ap75 is going up slowly

[AlexeyAB]

@EtheneXiang

So now you know that

• iou_loss=mse is for AP50
• iou_loss=giou / ciou / diou are for AP75 / AP95

[yishunzhijian]

I can assure you that my self-made cfg is all right.
i have trained it with CIOU

What does uc_normalizer mean?How to set the size of this value

[WilburZjh]

Hi @EtheneXiang , are you using the Gaussian_yolov3_BDD.weights for your training?