The results of ADE20K about convnext HOT 2 CLOSED

Hi, thanks for your excellent work.

The config files made be slightly modified to slide mode in test_cfg. In this situation, the results of provided models could be aligned.

What's more, comparison between Swin Transformer and ConvNeXt on ADE20K may keep the same config parameters especially same test mode whole or slide. As far as I know, the swin transformer uses whole mode, which may get lower result than slide mode. We hope MMSegmentation could support ConvNext officially to the whole community in the near future.

Best,

from convnext.

Thank you @MengzhangLIand @linfangjian01. You both are right. @HannaMao just pushed a commit that separates the multi-scale testing and single-scale testing to avoid the confusion. For multi-scale testing, the test_mode is 'whole', so it's identical to Swin's strategy and the results are comparable. For single-scale testing, we followed BEiT to use 'slide', and the results are presented in Appendix (but they are not directly comparable to Swin's single-scale numbers due to the difference in test mode).

from convnext.