logs are lost

[goldentimecoolk]

Hi, thanks for sharing your work. But logs that record the process corresponding to results reported in paper are lost. Can you please share them again?

[prlz77]

Hi, thanks for pointing out the issue, I am looking for the logs to put them again. If I have to rerun the code it may take more time.

[goldentimecoolk]

Yea, I agree. It would be better to find out the logs. I'm looking forward to it. -:)

[prlz77]

I have uploaded logs on CIFAR, now I am uploading the logs on ImageNet

[prlz77]

Done

[goldentimecoolk]

Hi, thanks for your quick action. I trained with origional code and default setting in the .sh file. But only get 79% top1. In the paper, this seems to be more than 81%. Then I trained attention version with default hyper parameters, and got top1 acc no more than 80%. Do you have experience in this situation?

[prlz77]

Hi, the results I report are the median of five runs, so maybe if you run more times you will find the correct accuracy. It is also possible that the baselines were run with different number of gpus/batch size, but the sh script was made to be run with minimal resources.

[prlz77]

Ah, maybe you should add 0.3 dropout. The baseline in the sh is the no_dropout baseline I think.