Scripts for training adversarially robust classification models

This set of scripts allows train and evaluate robust to PGD attack classification models on ImageNet dataset. Implementations of models are taken from the pytorchcv library, which means robust weights can be substituted to existing scripts that rely on pytorchcv.

Trained models

All the models are trained on adversarial samples obtained after 3 iterations of PGD with step=2/3*eps.

Model	Input Resolution	Params(M)	MACs(G)	eps	dist	Top-1 accuracy	Top-5 accuracy	Top-1 adv accuracy	Top-5 adv accuracy
MobilenetV2 1x	224x224	3.4	0.3	0.02	l2	72.16	90.62	71.72	90.40
MobilenetV2 1x	224x224	3.4	0.3	0.05	l2	72.12	90.34	71.11	89.84
MobilenetV2 1x	224x224	3.4	0.3	0.3	l2	71.38	89.8	68.79	88.50

An example of training command:

python train.py \
 -a mobilenetv2_w1 \
 -b 256 \
 -d $IMAGENET_FOLDER \
 --epochs 150 \
 --lr-decay cos \
 --lr 0.05 \
 --wd 4e-5 \
 -c ./snapshots \
 --input-size 224 \
 --adv-eps 0.3 \
 --euclidean \
 -j 40

An example of an evaluation command:

python3 train.py \
 -a mobilenetv2_w1 \
 -d $IMAGENET_FOLDER  \
 -b 128 \
 --weight ./snapshots/model_best.pth.tar \
 --evaluate \
 --input-size 224 \
 -j 8 \
 --adv-eps 0.3 \
 --euclidean