fshp971 / robust-unlearnable-examples Goto Github PK

I'm generating error minimizing noise for cifar10 using :
python generate_em.py
--arch resnet18
--dataset cifar10
--train-steps 5000
--batch-size 128
--optim sgd
--lr 0.1
--lr-decay-rate 0.1
--lr-decay-freq 2000
--weight-decay 5e-4
--momentum 0.9
--pgd-radius 8
--pgd-steps 10
--pgd-step-size 1.6
--pgd-random-start
--report-freq 1000
--save-freq 1000
--data-dir ./data
--save-dir ./exp_data/cifar10/noise/em8
--save-name em

And perform standard training on error minimizing noise poisoned dataset using:
python train.py
--arch resnet18
--dataset cifar10
--train-steps 40000
--batch-size 128
--optim sgd
--lr 0.1
--lr-decay-rate 0.1
--lr-decay-freq 16000
--weight-decay 5e-4
--momentum 0.9
--pgd-radius 0
--pgd-steps 10
--pgd-step-size 0.8
--pgd-random-start
--report-freq 1000
--save-freq 100000
--noise-path ./exp_data/cifar10/noise/em8/em-fin-def-noise.pkl
--data-dir ./data
--save-dir ./exp_data/cifar10/train/em8/em
--save-name train

And the trained model accuarcy on clean testset is about 10.10± which is far below the result in your paper (13.20 in Table 1, 16.84 in Table 7 ). I'm quite confused by this huge difference. If I set the parameters wrong or it's due to the randomness of the experiments?

Could you share the imagenet subset that you used?