Inquiry about the validation for EBMs

[tmp12316]

Thank you for your amazing works.

However, I found that the unknown annotations in t1_valare used to fit the distribution and the model is updated using loss. Then, there are two problems:

1. Is the pre-trained model used for t1_test from t1_train or t1_val.yaml? Due to the model update in val, I think it should come from t1_train.
2. In the process of validation, we use the unknown and known annotations to fit EBMs, which can be seen as learning a distribution using small-scaled validation with unknown labels? In this case, will the Open-set degenerate into few-shot learning?

Please correct me if I am the wrong~

[JosephKJ]

Hi @Wyman123,

Is the pre-trained model used for t1_test from t1_train or t1_val.yaml? Due to the model update in val, I think it should come from t1_train.

It is coming from t1_train. You can refer to run.sh.

Open-set degenerate into few-shot learning?

Kindly read through Salman's message here.

Kindly reopen this issue if we need to discuss more. Thank you.