MadryLab / robustness

A library for experimenting with, training and evaluating neural networks, with a focus on adversarial robustness.
MIT License
903 stars 181 forks source link

Training times for robust models? #76

Closed MalharJ closed 4 years ago

MalharJ commented 4 years ago

Is it possible to get an estimate of how much time it takes to train a robust ImageNet model, like say ResNet-50 for epsilon=4/255?

Hadisalman commented 4 years ago

Hi @MalharJ,

It mainly depends on the number of PGD steps you use, the batch-size, and your GPU. As an example, a robust ImageNet model trained with a 3-step PGD with a batch-size of 512 on 4 V100 GPUs should take ~1.6 hrs/epoch (~6 days for a full 90 epoch training).

Hope this helps!

MalharJ commented 4 years ago

Sounds good, thanks!