Performance Gap When Training UPerNet with Swin Backbone

[chamely]

I am trying to reproduce the official performance of UPerNet with Swin-T backbone using the provided configs in MMSegmentation.
However, the reproduced mIoU is significantly lower than the results reported in the model zoo.

tools/train.sh configs/swin/upernet_swin_tiny_patch4_window7_512x512_160k_ade20k_pretrain_224x224_1K.py

20251126_133416.log.json

Official mIoU 44.5
My mIoU 31.58