something is wrong with ap=0

[May2333]

after I trained the code followed the readme without any change for 6 epoch, the AP is close to 0. anyone knows what's wrong with it? and how to solve it? any help will appreciate!

[jackroos]

Could you provide the full command you ran? BTW, you could check if your setup is correct by evaluating our pre-trained model.

[lalagood]

Could you provide the full command you ran? BTW, you could check if your setup is correct by evaluating our pre-trained model.

hi, but i found that the pre-trained model could not be downloaded these days

[MaLiXiaoGuo]

我的也是艾

[zhangjing9701]

@May2333 @jackroos @MaLiXiaoGuo

Hi. I met the same problem. Have you solved it? Thanks very much!!!

[MaLiXiaoGuo]

@May2333 @jackroos @MaLiXiaoGuo

Hi. I met the same problem. Have you solved it? Thanks very much!!! I adjested training paraments of "Transformer", and trained more epochs. But the ap is still very low.

[May2333]

I don't know.. So I just fine-tune on official premodel, it seems no problem

[ghost]

Hi @May2333, Can you give me your reference source code for your fine-tune. I tried a lot but still can't work. Thank you in advance.

[102424lmy]

Have you found the reason? I have the same problem as you：No matter how many epoches trained, AP is close to 0. I would appreciate it if you could give me some guidance.

[HanWangSJTU]

There is one possible reason: batchsize is too small, but the learning rate remains unchanged. So the network is hard to reach convergence.

[ishidadadada]

I have the same problem as you：No matter how many epoches trained, AP is close to 0

However, Test is good result of detecting small obstacle. Thank you for your help.

[ishidadadada]

私は、道路上の小さな物体のクラスについて識別実験をしています。　バウンディングボックスを安全に囲むことができるにもかかわらず、APが0になってしまうのはなぜでしょうか？

[ishidadadada]

I am doing an identification experiment for a class of small objects on the road.　Why is it that the AP is 0 even though the bounding box can be enclosed safely?

[ishidadadada]

[ishidadadada]

I have two hypotheses for the cause of the 0 AP.

1) The custom data set is not available in coco format. (of course, it has been converted to JSON format) (2) The official coco-pre trained model is not being used properly.