validation in inference process has a large difference with validation in train process

[zsc1104]

Thank you for all your works ! Sorry about that I have a question
in train process i get the the following metrics with val datas
2021-10-09 04:18:34,676 INFO [base.py, 84] Performance 0.7751076951173586 -> 0.7788474201729588
2021-10-09 04:18:39,755 INFO [trainer.py, 403] Test Time 16.872s, (0.148) Loss 0.32007823
2021-10-09 04:18:39,755 INFO [base.py, 33] Result for seg
2021-10-09 04:18:39,756 INFO [base.py, 49] Mean IOU: 0.7788474201729588
2021-10-09 04:18:39,756 INFO [base.py, 50] Pixel ACC: 0.9112432591266407

But when i use the same val datas in the inference process , i get the the following metrics
2021-10-11 15:45:30,445 INFO [ade20k_evaluator.py, 46] Evaluate 228 images
2021-10-11 15:45:30,445 INFO [ade20k_evaluator.py, 47] mIOU: 0.47010236649173026
2021-10-11 15:45:30,445 INFO [ade20k_evaluator.py, 48] Pixel ACC: 0.6231233362565961

I feel very strange , could you help me , thank you very much !!!

[faruknane]

@zsc1104 Hi, could you help me about couple of questions below?

Does ade20k dataset need any preprocessing before training? I'm trying to use ade20k dataset. Do I need to make any change in the data structure in the official ade20k dataset? They have multiple labels, so it must be complex?

Thank you