Hi,
This is more of a theoretical query. Is there a way to make sure all embeddings in the queue(maintained in the MoCo v2 training mechanism) are actually negative samples?
e.g., there is a possibility that one or more embeddings in the queue are actually positive examples with respect to the current key/query.
Has this been handled in MaskContrast, or could you point to any work on the same?
Thanks.