search

deeplearning-wisc / gradnorm_ood

On the Importance of Gradients for Detecting Distributional Shifts in the Wild
Apache License 2.0
53 stars 7 forks source link

DenseNet Settings #7

Closed KingJamesSong closed 2 years ago

KingJamesSong commented 2 years ago

Hi,

Thanks for open-sourcing the codes.

Could you please also reveal the code or the implementation details of the DenseNet-121 experiment?

DId you use the same resolution and the same temperature as the experiment of ResNetv2-121?

iurgnauh commented 2 years ago

Hi,

Thanks for being interested in our work!

For DenseNet-121, we used the PyTorch implementation: https://pytorch.org/vision/main/_modules/torchvision/models/densenet.html#densenet121. And we used the pre-trained model provided by PyTorch: https://download.pytorch.org/models/densenet121-a639ec97.pth

The resolution we used here is a 224x224 center crop. Everything else (including temperature) stays the same.

Let me know if there is any issue in reproducing the results!

KingJamesSong commented 2 years ago

Thanks! The results are almost the same when using the resolution of 224.

iurgnauh commented 2 years ago

Awesome! Thanks for the follow up!