Metrics discrepancy in training and inference

[WangJuan6]

Dear Author,
I trained a model with a custom dataset, and the checkpoint's name shows miou is 51.62.
Then, I modify the --continue_fpath=checkpoints/proc_DFormerv2_L_20250419-211326/epoch-70_miou_51.62.pth in infer.sh, and C.pretrained_model="checkpoints/proc_DFormerv2_L_20250419-211326/epoch-70_miou_51.62.pth" in DFormerv2_L.py. However, when I use bash infer.sh for inference, the miou is only 24.08. (The same seed is used for training and inference. )
Can you help me solve this problem?
Thanks

[a1037441813]

same problem

[Carbine-P]

我的结果也只有 23.82，但是单张推理的可视化看起来挺好的，可能是验证部分的代码有问题