Low accuracy on custom coco format dataset

[JasenWangLab]

你好，感谢你们工作的开源，我把自己的数据转换成coco格式后，在大约20epoch的时候，训练100个epoch准确度还是很低，total_loss基本就不变了，请问下会是什么原因导致的呢 这个是我的训练命令：python tools/train.py -n yolox_tiny -f /YOLOX-main/exps/example/yolox_voc/yolox_my.py -d 2 -b 16 --fp16 -o

[GOATmessi8]

Plz load the pretrained COCO weights using -c /path/to/the/weights

[JasenWangLab]

Plz load the pretrained COCO weights using -c /path/to/the/weights

OK, thank you very much. I'll try this method.But if I want to train from scratch with my own dataset, is there any solution to this problem

[BowieHsu]

I trained yolox tiny on my datasets and meet the same problem, maybe caused by init setting about bn: "m.eps = 1e-3 m.momentum = 0.03", I also try to init weights with coco pretrained model , waiting for eval results now.

[Joker316701882]

@BowieHsu @JasenWangLab Training from scratch is not recommended, unless your training data is as large as COCO dataset.

[TUDelftHao]

你好，感谢你们工作的开源，我把自己的数据转换成coco格式后，在大约20epoch的时候，训练100个epoch准确度还是很低，total_loss基本就不变了，请问下会是什么原因导致的呢 这个是我的训练命令：python tools/train.py -n yolox_tiny -f /YOLOX-main/exps/example/yolox_voc/yolox_my.py -d 2 -b 16 --fp16 -o

I am also suffering from this trouble. Even when I load the pretrained model of yolox-s, the loss stabilizes around 10.0 and the eval map is approaching 0, which is even worse than the beginning (about 0.45). @Joker316701882 @ruinmessi Do you know what causes this phenomenon ? @JasenWangLab Did you manage to solve this problem?

[mikel-brostrom]

Same problem here as @TUDelftHao

Just for comparison. I let YOLOv5 with pretrained weights train one epoch on my custom dataset and got 68AP

[charles-str]

你好，感谢你们工作的开源，我把自己的数据转换成coco格式后，在大约20epoch的时候，训练100个epoch准确度还是很低，total_loss基本就不变了，请问下会是什么原因导致的呢 这个是我的训练命令：python tools/train.py -n yolox_tiny -f /YOLOX-main/exps/example/yolox_voc/yolox_my.py -d 2 -b 16 --fp16 -o

I am also suffering from this trouble. Even when I load the pretrained model of yolox-s, the loss stabilizes around 10.0 and the eval map is approaching 0, which is even worse than the beginning (about 0.45). @Joker316701882 @ruinmessi Do you know what causes this phenomenon ? @JasenWangLab Did you manage to solve this problem?

你是训练的tiny?我也是，map很低，你解决了吗？

[wang1233211]

兄弟，有好的解决方案吗