Confusion about dynamic ray sampling

[BCL123456-BAL]

Rays are sampled separately for each semantic class.
If random non-repetitive sampling is adopted, how to ensure that the light in the small semantic area is not exhausted before the end of a epoch of training.
How to ensure that all rays of the background or large semantic regions are used in one epoch, taking random repeatable words