Evaluations on different resolutions

[Yeez-lee]

Hi,

Thanks for an interesting idea.

1. I want to double-check evaluations on different resolutions.
   As you mentioned, I just need to replace img_size=256 from vit_rope.rope_mixed_deit_small_patch16_LS(pretrained=True, img_size=256) to others like img_size=192 or img_size=512 so that ViT can receive images of 192 or 512.
   And then --input-size 224 is changed to --input-size 192 or --input-size 512 on evaluation command lines so that the test dataset is interpolated to img_size=192 or img_size=512 without other efforts . Are these correct?

2. What does LS mean in deit_small_patch16_LS?

[bhheo]

Hi

Thank you for your interest in our study

1. Yes. You are correct.
   ViT model loaded with image_size=512 can receive images of 512x512.
   I also did a similar thing to deit evaluation code. So, you can do test with other interpolated resolution with --input-size 512 args.

2. LS means LayerScale, which is added to deit at deit-iii paper.
   You can find the same naming rule here.
   https://github.com/facebookresearch/deit/blob/main/README_revenge.md

If you have any further questions, please let me know