Why replace the predicted labels?

[bhosalems]

Why do we have the below replacement in the evaluate.py?

make the label ids consistent

    preds['label'] = preds['label'].replace(self.activity_index) 

What happens if the validation or the test dataset does not have all the possible labels?

[tzzcl]

We follow the original ActivityNet evaluation code to replace the original string with the index to eval predictions. For your case, I think you should manually build the activity_index with the union of your all data.

[bhosalems]

thanks for the reply.