Shifting Semantic ID Values

[seanlaw]

Currently, the semantic IDs that are generated (4 integer values) appear to only have values between [0,255]. However, given that these are essentially a set of 4 tokens, don't we need to shift the values by column_index * 255? So, for example, if a semantic ID was identified to be [13, 14, 15, 1], would it make sense to convert this to [13 + (0 * 255), 14 + (1 * 255), 15 + (2 * 255), 1 + (3 * 255)], which results in a semantic ID of [13, 269, 525, 766] so that we can differentiate between the different codebooks? Or is this already accounted for behind the scenes?

[JacoCheung]

Are you looking for this ? Offsets are added when training the GR model.