Issue reproducing results from the paper for CIFAR-10

[pratik18v]

Hi,
Thank you for making this library available! I was trying to reproduce the results from your paper on the CIFAR-10 dataset. Specifically, I followed the instructions available here and ran the following commands:

Training:

python cifar_training.py --batch_size 256 --lr_decay_milestones 1400 1700 --model ResNeXt_cifar

Evaluation:

python cifar_training.py --verify --model ResNeXt_cifar --load saved_models/ResNeXt_cifar --eps 0.03137254901961

I am getting the following results:

eps=0.031372549020 CE=1.8107 Err=0.6566 Loss=1.9743 Robust_CE=1.9743 Verified_Err=0.7224 Time=0.0736

Do you have any insights on why this discrepancy (from results reported in the paper) is occurring?

Best,
Pratik