search

dydjw9 / MetaAttack_ICLR2020

44 stars 6 forks source link

Epoch number for mnist meta-training #15

Open abdullahjamal opened 4 years ago

abdullahjamal commented 4 years ago

The default epoch number is 54600 for mnist in meta-training. Is the number correct or should I use a different value? https://github.com/dydjw9/MetaAttack_ICLR2020/blob/master/meta_training/mnist_meta_training/mnist_train.py#L161

dydjw9 commented 4 years ago

Hi, sorry that is another debug code here, we don't need to use so much epoch, because it will autosave the checkpoint so we did not change the maximum epochs here.

Cuz the for loop for epoch in range(args.epoch//100), so you may set 10000 to 30000 would be quite enough.We will modify it sooner. Sorry for that inconvience

abdullahjamal commented 4 years ago

Thanks, i will try the limits

abdullahjamal commented 4 years ago

if you remember, what was your maximum epoch for the experimentation?

dydjw9 commented 4 years ago

Mnist would be very fast. 20000 would be quite enough

abdullahjamal commented 4 years ago

https://github.com/dydjw9/MetaAttack_ICLR2020/blob/master/meta_attack/meta_attack_mnist/load_attacked_and_meta_model.py#L52

Can you please explain, if the model is loaded here or not?

abdullahjamal commented 4 years ago

meta_model.load_state_dict(pretrained_dict) something like this should be added here