Expand description
Feature standardization.
Raw φ scales vary wildly (counts 0–5, log-octave axes ~0–1, log crest
0–4); a Gaussian prior over θ only makes sense on a common scale. The
standardizer is re-fit at every posterior fit, over the union of the
observation log and the live pool, and persisted with the taste
profile — θ is only meaningful relative to the standardization that
produced it, so a profile carries both or neither.
It is a view of the data, not the data: the log stores raw φ, so a re-fit standardizer simply re-expresses the same evidence on a scale that still matches where the pool actually is.
Structs§
- Standardizer
- Per-dimension affine standardization:
(x - mean) / std.