#continuous-inputs

Question 1: Single choice

A transformer for text looks up each discrete token in an embedding table. How is it usually adapted to continuous inputs such as images or audio waveforms?

Cut it into patches or frames and project each one linearly to a vector1
Round every pixel or audio sample to an integer and look it up in an embedding table2
Feed the whole image or recording as a single token3
Sort the samples by value so that similar values become neighbouring tokens4

Timed test