Hi,
I like the idea very much, but I am wondering how this work is going since there have been no updates for 4 months.
I ran the 3 instructions twice mentioned in the readme, and achieved only 27.2% and 26.5% val error on CIFAR-100, substantially higher than the paper.
Is this because the default setting is incorrect, or the variance is huge due to randomness?
Thank you!