Jungwoo Oh

Yes, the memory peaked when creating the predicate columns. It would be great if we can expand the predicates grammar to support "any of codes" operation without creating intermediate predicate...

FYI, when I used only ~70 predicates to define the task and process the same data, it took about 14 minutes with ~15GB of RAM. This number is much less...

Hi, Could you let me know how you calculated the macro-AUROC? Did you get AUROCs for each of 83 attributes and macro-average them?

This criterion currently calculates **MICRO** AUROC by default, not **MACRO**. Since the current `multi_head_binary_cross_entropy` criterion is not supporting macro averaging option at the moment, you may need to get AUROCs...

Can you confirm if you train SE-WRN model without changing any default hyper-parameters (e.g., # of epochs, learning rate, ...)?

Did you sample 10% from the test set and use them for LLM evaluation? Could you let me know which openai model you are using now?

If I share the weights file for SE-WRN that has been used for the experiments in the original paper, could you run the LLM experiments using it? For now, I...

I've just sent an email to you. Please check it.

Hi, any updates on this issue? If not, I will close this issue.