Cosine seems to work well for CLIP in the embedding space, don't see why that wouldn't work for us.
I already wrote code to compare the correlation, euclidean, and cosine distances, but only for the hierarchical code. Apply those same metrics to KMeans/KNN/etc.
Cosine seems to work well for CLIP in the embedding space, don't see why that wouldn't work for us.
I already wrote code to compare the correlation, euclidean, and cosine distances, but only for the hierarchical code. Apply those same metrics to KMeans/KNN/etc.