DotWang
DotWang
@youngbaldy 可能是因为kvdiff的代码在上传权重后改过一点,但是这个时间有点久了,我忘了细节了,建议用非kvdiff的权重,或者用现在的代码重新训一下kvdiff的
@youngbaldy 差不多吧,我觉得这还有可能和数据集有关系,我们用的是之前自己做项目时手动裁的数据集,不是用mmseg裁剪的,不同人的数据集裁剪方式和参数,诸如步长之类的可能也不一样,导致了结果有差异,另外,我觉得其实统一在大图上进行滑窗预测是最公平的,按照我们自己的探索,大图滑窗预测的精度也要比一个一个预测小图再评测的高,不过目前大部分人都是裁剪后的小图上测,所以我们这次也还是在小图上测的
@funny000 这个不需要,这次只用到ViT没用到ViTAE,注释掉就行,可能我没删干净
@funny000 ppm里边有bn,batch size不能为1
@martin416 您好,数据集生成代码已发布
@martin416 您好,pretraining and finetuning 代码已发布
@15926273249 We have provided the instruction, please refer to the readme and do corresponding changes as your case.
@15926273249 You mentioned are already segmentation datasets, which can be directly used with the codes of ```Pretraining and Finetuning```.
@15926273249 Our code conducts the prediction by slided windows. It can be used for huge images.
@xulinui please refer to the COCO RLE format, they can be transformed to binary mask