How to interpret SamPredictor.features

I'm interested in finding an embedding vector for each segment (for classification and tracking of segments).
Can SamPredictor.features be used for this? Its not clear to me how to interpret this. I see its a 1x256x64x64 tensor - does the 256 corresponds to 0..255 possible segments, and can the 64x64 component be flattened and sensibly treated as a point in an embedding space?

Thanks in advance.

(Could this project enable the community discussion of a discord group -- maybe this and many other equations are more like requests for help or clarity than a bug).