train well map zero with p5 and csp swish

[mnoroozii]

I have implemented yolov4-csp-x-swish.cfg and pre-trained weights: yolov4-csp-swish.conv.164 and also yolov4-p5.cfg with it's corresponding pre-trained weights on my custom dataset

I have tested both cfg and frozen cfg for training.

it got trained well but no ap no map everything 0

Needless to say, I have implemented some previous versions of yolov4 on my custom dataset and it works well. So there is no issue with the dataset

[akashAD98]

maybe a problem with plotting on the graph, but can you check MAP using this command?

To check accuracy mAP@IoU=50: darknet.exe detector map data/obj.data yolo.cfg backup\yolo.weights

[mnoroozii]

I have used your command for map, the results are the same as the image I sent earlier. I have tested some of my test dataset images one by one for detection. but there is no bounding box in the prediction. The train figure is weird and very neat as well. I mean It is not supposed to go to 0 neatly. But it went!

[erikguo]

I had same problem. I resolved this problem by changing "objectness_smooth" to 0 in yolo layer. you can try it. @mnoroozii

[mnoroozii]

I had same problem. I resolved this problem by changing "objectness_smooth" to 0 in yolo layer. you can try it. @mnoroozii

Thanks for your attention. No change. Still have issue with these CFGs.

[midasklr]

I got the same problem

[ashutoshIITK]

same issue

[haxsender]

Same issue. I tried to replace the yolo layer settings with similar ones from yolov4. After that and two or three restarts of training it proceeded normally. Also noticed that on bad attempts, rewritten_bbox was always 0.

[ghost]

I have the same issue