Cant get the reported accuracy by running the training scripts.

aamir-mustafa / pcl-adversarial-defense

Adversarial Defense by Restricting the Hidden Space of Deep Neural Networks, in ICCV 2019

59 stars 14 forks source link

Cant get the reported accuracy by running the training scripts. #10

Closed Amose-Yao closed 4 years ago

Amose-Yao commented 4 years ago

Hi, I am impressed by your work. But when I run the scripts which are softmax_training.py and pcl_training.py, I only got the 91 and 78 respectively. I wonder there could be some differences between your configuration and mine. Could you please share the hardware configuration and the hyperparameters with us? Best wishes and Merry Christmas, Amose.

aamir-mustafa commented 4 years ago

Hi,

Yes I am planning to share a separate file containing the hyper-parameters chosen while training for all 5 datasets used in the paper. Meanwhile, I would suggest reducing weight_prox in pcl_training.py. This should help in improving the clean accuracy of the model.

Hope it helps