Training Cityscapes dataset

[CurryKd7]

Why is this link https://www.rocq.inria.fr/cluster-willow/rstrudel/segmenter/checkpoints/cityscapes/seg_large_mask/log.txt The mlou of the training citycpales dataset in the first round can reach 43%, but I only trained for 9% in the first round, and I used vit_1arge_patch16_384 as the backbone. May I ask why this is?

[LUPIN11]

Maybe the defaulted settings (layers, lr, batchsize) are changed on this variant. I also reproduced this variant on a single gpu, the default settings should be modified.