search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.68k stars 7.96k forks source link

strange AP value #4342

Open May-forever opened 4 years ago

May-forever commented 4 years ago

Dear @AlexeyAB ,

I am so sorry to bother you, but could you please do me a favor ?

I once asked you about the problem of how to perform model pruning in darknet.

And you told me I can do as following: Or instead of Model Pruning: It seems that you just can reduce filters= /2 or /4 in the middle layers in cfg-file and train from scratch without pre-trained weights-file.

I follow your suggestion, and train my custom network for one class detection. However, it is very

confusing that：

My custom network achieves AP=89.64, while the pruned network (after reducing filters=/2) achieves AP=92.46.

I am so confused why the AP value been raised after model pruning.

What kind of problem is this situation?

Looking forward to hearing from you. Thanks in advance.

Best regards, May

AlexeyAB commented 4 years ago

There is no problem, may be just mAP fluctuation due to small dataset.

May-forever commented 4 years ago

There is no problem, may be just mAP fluctuation due to small dataset.

Dear @AlexeyAB ,

Thank you very much for your reply.

However, I train the network only for one class detection. Moreover, this class has 4000 images for training and 1300 images for testing.

So I am so confused that the AP value has increased for reducing filters=/2

Could you please give me some advice?

Looking forward to hearing from you, thanks in advance.

Best regards, May