paper: https://arxiv.org/abs/1911.09665
ResNet50 training on ImageNet with single 8-GPU machine:
python -m torch.distributed.launch --nproc_per_node=8 examples/imageclassification/imagenet/advprop/main.py \
--data_dir=$ImageNetDataDir \
--lr=0.1 \
--opt=momentum \
--batch-size=32 \
--pin-mem \
--epochs=105 \
--dist-bn='' \
--output=output/advprop \
--experiment=tmp| Training Framework | Method | Model | ImageNet-Val | V2 | C (mCE↓) | R | A | Sketch | Stylized | ObjectNet | Files |
|---|---|---|---|---|---|---|---|---|---|---|---|
| EasyRobust (Ours) | AdvProp | ResNet50 | 76.64% | 64.35% | 77.64 | 37.43% | 2.83% | 24.71% | 7.33% | 16.82% | ckpt/args/logs |