The results of validation in distill. py and evaluate.py are different

[Rickilous]

There is a significant difference between the validation results obtained during the execution of distill.by during training and the validation results obtained after executing evaluate.py after training.
I am reproducing your code on the Scannet dataset.

The following is the final result of executing evaluate.py
evaluating 49540568 points...
classes IoU

wall : 0.737 (9711412/13169998)
floor : 0.888 (8325308/9374070)
cabinet : 0.458 (798335/1743003)
bed : 0.697 (684456/981554)
chair : 0.710 (2863996/4032050)
sofa : 0.624 (605267/969955)
table : 0.523 (1083707/2070202)
door : 0.446 (909689/2041809)
window : 0.516 (974451/1888683)
bookshelf : 0.675 (975975/1445593)
picture : 0.174 ( 44180/253470)
counter : 0.404 (118225/292895)
desk : 0.451 (404440/896387)
curtain : 0.547 (445784/814689)
refrigerator : 0.413 (108950/264016)
shower curtain: 0.000 ( 0/158208)
toilet : 0.802 ( 99836/124456)
sink : 0.497 ( 61659/124068)
bathtub : 0.601 ( 58872/97972 )
otherfurniture: 0.218 (491901/2252587)
Mean IoU 0.5191339645582693
Mean Acc 0.6294690727698209

The following excerpt is from the training log of distill.by, and the model at the 90th epoch is the best model.
[2024-10-14 01:10:04,873 distill.py line 465] Val result: mIoU/mAcc/allAcc 0.4345/0.5428/0.7595.