Xiangchao Yan
**I test the image box:** When the ratio of width/height is small, **the regression loss is 0** ` IMAGES\scratches_152.jpg | 26| 1 | 53 | 200 | scratches -- |...
When I **change 26 to 1** and **change 180 to 100,** **the regression loss is not 0** IMAGES\scratches_152.jpg | 1 | 1 | 53 | 200 | scratches -- |...
OK， I have solved this problem by increasing anchor ratios and nums!
> > 贴一下你的preload.js怎么写的截图？ > > 抱歉，大佬，最近没留意通知，后面我习惯网页版就没使用桌面版了，那天试了好几次，就是改变不了，后面找到一键汉化的脚本就忘记了，再次抱歉！ 可以分享一下脚本吗，我也没法汉化
Can you release the pretraining models on these datasets?
> I tried 8 gpu + 4 batch size. Results are similar. I linearly scale the lr (using 2x) I tried 2 A100 with 2 * 48 batch size，the lr...
> I guess it is too large ( and linear scaling doesn't apply) maybe try some smaller one ? Yeah，I have tried the original lr 0.01 ， and the results...
  