For biodiv and crop yield we're planning to assign importance values (3 categories?) to the different columns.
With this, caption templates could be sampled with varying frequency based on importance.
If importance is added to the aux_classes.csv files, it is read directly by the caption builders.
Later this could be used for curriculum learning style learning (changing captioning strategy depending on accuracy of different columns).
For biodiv and crop yield we're planning to assign importance values (3 categories?) to the different columns.
With this, caption templates could be sampled with varying frequency based on importance.
If importance is added to the
aux_classes.csvfiles, it is read directly by the caption builders.Later this could be used for curriculum learning style learning (changing captioning strategy depending on accuracy of different columns).