can‘t find result of SEM

[Esther-PAN]

Hello, I truly appreciate your comprehensive work. While reviewing the results of fs-ood incorporated in v1.5, I noticed that the sem method, which initially proposed this work, doesn't seem to be compared. May I ask what considerations led to this decision? If I've missed something, please kindly point it out.
Also, could you please share the training configuration and estimated duration for imgnet1k fs-ood?
Thank you.

[zjysteven]

I don't recall a specific reason for not including SEM. It should have been included, and it could be that we somehow miss it during the very large-scale evaluation.

On ImageNet-1K fs-ood: All the post-hoc methods share the same pre-trained torchvision checkpoint as the backbone (no training involved). For data augmentation training methods, we also mostly used pre-trained checkpoints. For whatever method that does need training, here is the setup quoted from the paper:

For training methods, we focus on ResNet-50 and use official checkpoints when possible. Otherwise, we fine-tune the torchvision pre-trained checkpoint for 30 epochs with a learning rate of 0.001.

The training time is highly dependent on the specific method, so it's hard to say.

[Esther-PAN]

Thank you for your prompt reply.~ Since the SEM method was the first to propose f-ood and serves as the baseline for this task, I was wondering if it would be possible to supplement the results of this method.

[zjysteven]

We would very much appreciate community's help/contribution on this~