Why do you trained a model build from scratch instead of fine-tuning a pretrained model？

ashafahi / free_adv_train

Official TensorFlow Implementation of Adversarial Training for Free! which trains robust models at no extra cost compared to natural training.

https://arxiv.org/abs/1904.12843

173 stars 30 forks source link

Why do you trained a model build from scratch instead of fine-tuning a pretrained model？ #3

Open Sumching opened 5 years ago

Sumching commented 5 years ago

Hi， I notice that you did not use a pretrained model. Is it too hard to train? Or is its performation not good? Thanks.

ashafahi commented 5 years ago

Hi, Correct. We have started with random initialization. However, starting from a pre-trained net also works. It can even increase clean validation accuracy a bit.