SeongwoongCho / adversarial-autoaugment-pytorch

Unofficial Pytorch Implementation Of AdversarialAutoAugment(ICLR2020)

Adversarial-Autoaugment-Pytorch

Unofficial Pytorch Implementation Of AdversarialAutoAugment(ICLR2020)

Current Issue

I want some help from those who know how to solve these issues.

• Can not reproduce paper's results

• Adversarial Collapsing : See /Examples/Analysis.ipynb

Quick Start

# Training with DistributedDataParallel
$ python -m torch.distributed.launch --nproc_per_node ${NUM_GPUS} main.py \
    --load_conf ${conf_dir} \
    --logdir './logs' \
    --M 8 \ 
    --seed 0 \
    -- amp \ 
    >> output.log

# Training with DataParallel
$ python main.py \
    --load_conf ${conf_dir} \
    --logdir './logs' \
    --M 8 \ 
    --seed 0 \
    -- amp \
    >> output.log

# Evaluate
$ python evaluate.py \
    --load_conf ${conf_dir} \
    --logdir './logs' \
    --seed 0 

Results

CIFAR-10

Model(CIFAR-10)	Paper (direct/transfer)	Ours
Wide-ResNet-28-10	1.90±0.15 / 2.45±0.13	2.35 / -	Download
Shake-Shake(26 2x32d)	2.36±0.10 / -	2.51 / -	Download
Shake-Shake(26 2x96d)	1.85±0.12 / -	2.43 / -	Download
Shake-Shake(26 2x112d)	1.78±0.05 / -	-	Download
PyramidNet+ShakeDrop	1.36±0.06 / -	-	Download

CIFAR-100

Model(CIFAR-100)	Paper (direct/transfer)	Ours
Wide-ResNet-28-10	15.49±0.18 / 16.48±0.15	-	Download
Shake-Shake(26 2x96d)	14.10±0.15 / -	-	Download
PyramidNet+ShakeDrop	10.42±0.20 / -	-	Download

ImageNet

Model(ImageNet)	Paper (top1/top5/ transfer_top1)	Ours
Resnet50	20.60±0.15 / 5.53±0.05 / -	-	Download
Resnet50D	20.00±0.12 / 5.25±0.03 / 20.20±0.05	-	Download
Resnet200	18.68±0.18 / 4.70±0.05 / 19.05±0.10	-	Download

CIFAR-10-C

Model(CIFAR-10-C)	Augmix w/ JSD	Adv AA
Wide-Resnet-40-2	11.2	-	Download
Wide-Resnet-28-10	-	10.41	Download
Shake-Shake(26 2x32d)	-	16.69	Download

Different From The Paper

• I did not include SamplePairing -> NUM_OPS = 15 (16 in the paper)
• Borrow unknown hyperparameter settings from fast-autoaugment

TODO

Observations

References & Open Sources

ENAS

FastAutoAugment

Augmix

OpenReview of AdvAA