search

hendrycks / imagenet-r

ImageNet-R(endition) and DeepAugment (ICCV 2021)
MIT License
254 stars 18 forks source link

noise2net always uses a pre-trained ResNet50 #4

Closed amodas closed 3 years ago

amodas commented 3 years ago

Hi,

It seems that at Line173 of train_noise2net.py the pretrained argument of ResNet50 is always set to True.

Is this simply a bug or something we should take into account? As far as I understand, the results of the paper do not consider an ImageNet pre-trained classifier, right?

Thanks, Apostolos

hendrycks commented 3 years ago

noise2net models were pretrained for a portion of the results at least, following AugMix. (However, this may have been from pre-experimentation.) We hardcoded pretrained=True so that things behaved correctly if we forgot the pretrained flag, though of course such hardcoding is not a best coding practice.