BCE
Currently, we have several BCE-related configs. Some need to be updated after #132, some have been introduced in #130.
Goal
Have exactly two configs for BCE:
- An unweighted BCE
bce_unweighted.yml
- A weighted BCE
bce_weighted.yml that uses the same beta as in the Chebifier paper
If someone wants another BCE variation, I'd say they should write the config file themselves.
@schnamo Are there other BCE configurations (for non-multilabel problems) that you would like to include as "standard configurations"?
Other loss configs
weighting_chebi100.yml is clearly outdatedfocal_loss_12.yml seems very specific - why do we need a number of classes for multilabel @schnamo? The implementation says that is only needed for multiclass problems.- The rest can stay as it is imo
BCE
Currently, we have several BCE-related configs. Some need to be updated after #132, some have been introduced in #130.
Goal
Have exactly two configs for BCE:
bce_unweighted.ymlbce_weighted.ymlthat uses the same beta as in the Chebifier paperIf someone wants another BCE variation, I'd say they should write the config file themselves.
@schnamo Are there other BCE configurations (for non-multilabel problems) that you would like to include as "standard configurations"?
Other loss configs
weighting_chebi100.ymlis clearly outdatedfocal_loss_12.ymlseems very specific - why do we need a number of classes for multilabel @schnamo? The implementation says that is only needed for multiclass problems.