Custom training goes wrong with "good hyperparameters"

[xevolesi]

Hi @AlexeyAB !

I trained YOLOv3 with custom data set. Classes:

• person;
• animal;
• car;
• truck;
• bus;
• bicycle;
• motorcycle;
• plate (car license plate).

I used YOLOv3 .cfg file with 416x416 input and added "good hyperparameters" from #4430

Here is my chart.png

Here is my yolov3.cfg: yolov3-default-416-mosaic-scale-ciou.cfg.txt

Some image examples from training and validation sets:

I checked my labels - it seems like it's correct, because earlier i trained default YOLOv3 and it gave me about 77% mAP. Could you, please, tell me what could happen with my training?

Thanks!

[AlexeyAB]

Try to:

• remove stopbackward=1

• and add max_delta=10 for each [yolo] layer

---

If it doesn't help try to

• or remove mosaic=1

• or use iou_loss=mse instead of iou_loss=ciou

[xevolesi]

Thanks!

I'll try both options and will notice you.

[xevolesi]

Hi, @AlexeyAB !

Your first suggestion worked well.

Here is final chart:

Do you think this is the effect of max_delta or stopbackward removal?

[AlexeyAB]

Do you think this is the effect of max_delta or stopbackward removal?

I don't know what did you do

[xevolesi]

Hi, @AlexeyAB !

Try to:

* remove `stopbackward=1`

* and add `max_delta=10` for each `[yolo]` layer

i did this