search

elliothe / CVPR_2019_PNI

pytorch implementation of Parametric Noise Injection for adversarial defense
Apache License 2.0
43 stars 16 forks source link

can't regain the claimed result #1

Open jiaming-lee opened 5 years ago

jiaming-lee commented 5 years ago

I run your code but can't regain the claimed result except the baseline model. Can you upload your pretrained model?

elliothe commented 5 years ago

@yangadan which result you cannot regain?

jiaming-lee commented 5 years ago

@yangadan which result you cannot regain?

noise_resnet20 PNI-W model Test Prec@1 83.090 Prec@5 99.080 Error@1 16.910 PGD Test Prec@1 44.080 Prec@5 93.550 Error@1 55.920 FGSM Test Prec@1 51.000 Prec@5 94.500 Error@1 49.000

elliothe commented 5 years ago

@yangadan Hi, I am sorry for the late reply. The pre-trained models for both layer-wise and channel-wise versions are uploaded at Readme result section.

I think there might be some setting problem. I just trained a new layerwise PNI model, and I can only get 84.11% for the clean accuracy. I will look at the training configuration and give you an update later.

jiaming-lee commented 5 years ago

@yangadan Hi, I am sorry for the late reply. The pre-trained models for both layer-wise and channel-wise versions are uploaded at Readme result section.

I think there might be some setting problem. I just trained a new layerwise PNI model, and I can only get 84.11% for the clean accuracy. I will look at the training configuration and give you an update later.

Thank you