facebookresearch / odin

A simple and effective method for detecting out-of-distribution images in neural networks.
Other
530 stars 102 forks source link

Unable to reproduce results when training densenet from scratch #13

Open praveen5733 opened 4 years ago

praveen5733 commented 4 years ago

I am able to reproduce the results reported in the paper when I use the pretrained models provided in the repo. But when I train a densenet from scratch the results are poorer compared to the report. Did anyone face a similar problem?

tangbohu commented 4 years ago

Also do I. Have you solved it?

YixuanLi commented 4 years ago

We are unable to update the github repo at this moment. However, we have recently built another repo which provides ODIN as well as many other OOD detection methods. Can you try this: https://github.com/jfc43/informative-outlier-mining?

huberl commented 3 years ago

I face the same problem @praveen5733 @tangbohu

YixuanLi commented 3 years ago

Can you provide details on your experimental results, so that I can take a look at the difference? It's possible that the performance will have some variations across model runs.

On Thu, Jan 7, 2021 at 2:56 PM lhuber notifications@github.com wrote:

I face the same problem @praveen5733 https://github.com/praveen5733 @tangbohu https://github.com/tangbohu

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/odin/issues/13#issuecomment-756378878, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABUGVB6C2X2TRLAHY4JGRLLSYYNV5ANCNFSM4TJ6YCBQ .

Jiejiegary commented 3 years ago

Hi, all, I have the similar problem. I trained densenet (and wideresnet) on cifar10 where models have normal test accuracy. When I test the model with odin in this task, I saw a pretty huge gap between the results and the reported ones. Maybe I miss something here.

YixuanLi commented 3 years ago

For wideresnet, you can refer to our latest paper: https://github.com/wetliu/energy_ood. It's also recommended to use energy score as it's parameter-free and gives a performance that's comparable or better than ODIN.

For ODIN, you can typically get a ballpark performance estimation by setting the temperature to be T=1000.