Closed ChristopherBrix closed 1 year ago
Hi,
I just added the license. Here are some responses to your question:
5 independent runs mean 5 different models. This repo only provides minimal code to reproduce the main results.
For places image index and other methods you are interested in, you can take a look at https://github.com/jfc43/informative-outlier-mining. We build the repo mainly based on that code base.
Thank you for the interesting paper and for making your source code available!
I tried to replicate your results on the CIFAR-100 network and noticed some issues:
loadtxt
doesn't seem to support\n
as a delimiter - do I need to install some additional packages to make this work? For now, I changed the delimiter there, and here (also, I had to remove the final trailing delimiter)python3 ood_eval.py --in-dataset CIFAR-100 [--p 90]
, I get the following results:These scores are close to those you report in Table 9, but not exactly the same. Is there some randomness involved? Especially LSUN_resize is off by quite a bit. Also, you state that you report standard deviations across 5 independent runs, but only do so for DICE, why is this the case?
Edit: Also, could you add a license to your code, so we can build upon it in future work?
Edit2: I'm unable to replicate all rows in Table 9 other than MSP, Energy and DICE. For Odin and Mahalanobis, there's a flag I can set to use this technique, but it requires some config values that I don't have. For the others, I don't know how to run them at all.