Problems were encountered when conducting high-resolution training !!!

[MyHogsmeade]

I train a vitb model in resolution 224224, then I want to perform high-resolution training in resolution 518518, I meet assertion error when loading statedict.

some details:

What I do:

1. training a vitb model, the log dir is ./log/xxxxx
2. use the same configuration but change the global_crops_size (224->518), use the same log dir ./log/xxxxx
3. it seems that the checkpoint will automatically load the last_checkpoint.rank_0 from the log dir, however the assertion error occurs