Could you share the pretrained model for imagenet? about swa_gaussian HOT 4 CLOSED

using python experiments/imagenet/run_swag_imagenet.py
--batch_size=256
--pretrained
--parallel
--epochs=10
--save_freq=10
--eval_freq=1
--swa
--swa_start=0
--swa_lr=0.001
--swa_freq=4
can republicate the result

from swa_gaussian.

Hi,

We used pre-trained models, which can be downloaded from https://pytorch.org/docs/stable/torchvision/models.html following our code.

However, we haven't published the SWAG checkpoints at the end of training due to their size --- @timgaripov might be able to provide them if he still has them.

from swa_gaussian.

@wjmaddox could you release the training scripts for imagenet? like lr config, ..
so that we can repulicate the result in the paper

from swa_gaussian.

That sounds about right given the hyper-parameters in Appendix H of https://arxiv.org/pdf/1902.02476.pdf

from swa_gaussian.