How can we train the baseline model? OE (outlier exposure)

[hyunjunChhoi]

Hello, Thanks for your work.

How can we train the baseline model such as OE (outlier exposure)

I cannot approach the similar baseline with the OE original code

Could you help

[htwang14]

Hi, can you try the following command?

python stage1.py --gpu 0 --ds cifar10 --Lambda 0.5 --Lambda2 0 --drp <where_you_store_all_your_datasets> --srp <where_to_save_the_ckpt>

By setting Lambda2 as 0, we are removing the PASCL loss. By setting Lambda as 0.5, we are using the OE loss.

Let me know if you have further questions.

Thank you!

[hyunjunChhoi]

Thanks for your answer.

I can reproduce the OE model

However, I cannot reproduce the EnergyOE

Accuracy of the EnergyOE is far behind from the report of paper

How can I reproduce the EnergyOE ????(I just pasted the EnergyOE fine-tuning code and learn)

[shuaiNJU]

Hi, can you try the following command?

python stage1.py --gpu 0 --ds cifar10 --Lambda 0.5 --Lambda2 0 --drp <where_you_store_all_your_datasets> --srp <where_to_save_the_ckpt>

By setting Lambda2 as 0, we are removing the PASCL loss. By setting Lambda as 0.5, we are using the OE loss.

Let me know if you have further questions.

Thank you!

Hi, thanks to your great work. But I have recently failed to re-produce the results of OE in your paper (e.g., Texture, the 92.59% AUROC and 83.32% AUPR, CIFAR10-LT) by the provided train command and pretrained model. I got the results of:

Texture: 90.77 (AUROC), 73.99 (AUPR)

Any suggestion?