ap0.5:0.95 is 0.132, which is very low

[Kyle-fang]

Yolox official git down the code, only changed batch_ Size. The data set is in coco format. The image resolution is 2446*1000. There are more than 4000 pieces of training and more than 1000 verifications. I have tried 50 and 300epoch, but ap0.5:0.95 is 0.132, which is very low.total_ Loss also oscillates around 8. At this time, AP is basically unchanged.

yolox官方git下来的代码，只改了batch_size，数据集是coco格式，图片分辨率2446*1000，四千多张训练，一千多验证，尝试过50，300epoch，但是AP0.5:0.95为0.132，很低。total_loss也一直在8左右振荡，这个时候ap也基本不变了

[Kyle-fang]

I do defect detection. Is this problem caused by the fact that my data set is very different from the data set with pre training weight? Because the training weight is obtained on Imagenet

[Zephyr69]

For defect detection, the image space will surely be wildly different from that the pre-trained model was trained on. The pre-trained weight will be useless if not detrimental here.

But there's a wide array of stuff that could lead to low performance here which I would assume small targets are involved; this is just one, albeit a definite one, of them.

[LSH9832]

你自己数据集样本太少了，得先加载预训练的backbone才行