Hi, thanks for providing the nice code.
I've tried to reproduce the results in table3 with "bash scripts/table_3_rcifar10.sh wrn_28_10".
The highest test acc is 87.42% among twice trial, which is much lower than the result in the paper.
My tensorflow version is 1.14, ray0.7, cuda10.0, cudnn7.
What would be the reason for the low performance?