Thanks again for the great work with ColorMAE! I'm trying to replicate the 800-epoch green-noise ColorMAE downstream result on ADE20K semantic segmentation, and not getting to the reference accuracy—only 47.91 instead of 49.18. Concretely, I'm following the SLURM script for semantic segmentation finetuning of the 300 epoch model, but with model weights converted from the 800-epoch pretraining with green noise. Any insight on if this script wouldn't generalize to the 800-epoch pretraining result, or if something else might be amiss?
Thanks again for the great work with ColorMAE! I'm trying to replicate the 800-epoch green-noise ColorMAE downstream result on ADE20K semantic segmentation, and not getting to the reference accuracy—only 47.91 instead of 49.18. Concretely, I'm following the SLURM script for semantic segmentation finetuning of the 300 epoch model, but with model weights converted from the 800-epoch pretraining with green noise. Any insight on if this script wouldn't generalize to the 800-epoch pretraining result, or if something else might be amiss?