paper: https://arxiv.org/abs/1706.06083
ResNet50 training on ImageNet with single 8-GPU machine:
python -m torch.distributed.launch --nproc_per_node=8 examples/imageclassification/imagenet/adversarial_training/main.py \
--data_dir=$ImageNetDataDir \
--smoothing=0.0 \
--pin-mem \
--lr=0.2 \
--output=output/adversarial_training \
--experiment=tmp| Model | ImageNet-Val | AutoAttack | Files |
|---|---|---|---|
| ResNet50 | 65.1% | 34.9% | ckpt/args/logs |