test and val gap in log file

[jerrywyn]

I tried to run the code and found in the log file that the validation result and the test result are very different, and the validation result is much worse than the test result. Why is this?
On the other hand, I run the code to get the result than the result in the paper, why is that?
script:
bs=128, dataset='voc2007', estimator='ours', filter_outlier=False, lr=5e-05, nc=20, nepochs=20, noise_rate_n=0.0343, noise_rate_p=0.4, nworkers=4, out='./results/multi-label-reweight_p0.4n0.0343_voc2007_ours_resnet50/', root='./data/voc/', sample_epoch=10, sample_th=0.5, seed=1, warmup_epoch=20, weight_decay=0
result:

result in paper:

[ShikunLi]

Thanks for your interest in our work.
First, the validation set is noisy in our setting, and the test set is clean. Hence, the validation result is much worse than the test result.
Second, the differences between the results obtained by running the code and the results in the paper may be because the label noise simulation is varied for different random seeds, and we make the implementation simpler when preparing the code.