Open jinseok-karl opened 3 years ago
Hi, The only difference is the ImageNet pretrained part. For segmentation, we only change the pretrained checkpoint. We do not apply our loss for segmentation. The reason is that, semantic segmentation itself has already provided dense local labels. We have uploaded some checkpoints for segmentation. Training SWIN on ImageNet uses a similar implementation as DeiT, and we will upload the code.
Yours, Chengyue
Could you explain "The reason is that, semantic segmentation itself has already provided dense local labels." I am confused
Hi, one of our motivations is to provide local labels for each token. For segmentation, the local label is already very dense and therefore we do not add our regularization.
Hi, thanks for sharing code! I'd like to try your code with mmsegmentation. But I can't find which part is the different with original swin. Shortly, I don't know where the diverse part is
Sincerely