Build a key by clustering classes and describing each split

Classes are clustered on the mean embedding of their images (Ward's method), giving a binary tree. Each split is then described by the vocabulary pair that best separates its two sides: the pair maximising |(L.yes - L.no) - (R.yes - R.no)| over the side means, not used above that split. The "Has feature" lead goes to the side the feature scores higher on. Machine answers use the same two prompts.

Usage

keyFromClusters(
  image,
  images,
  vocabulary,
  textEmb,
  template = "{x}",
  order = NULL,
  maxPerSide = 10,
  calibrate = c("none", "midpoint"),
  desc = "Key generated from image clusters",
  meta = NULL
)

Arguments

image: images x d matrix of unit-length image embeddings.
images: imageSet() data frame, one row per row of image.
vocabulary: As in vocabularyPrompts().
textEmb: Text embeddings of vocabularyPrompts(vocabulary, template), rows named by prompt (as embedTexts() returns them).
template: Must match the one used to make textEmb.
order: Optional function of a character vector of labels returning a number; at each split the child with the smaller value comes first.
maxPerSide: Most images per side used for the side means.
calibrate: "none" (threshold 0) or "midpoint" (halfway between the two sides' mean scores).
desc, meta: Title and metadata.

Value

A dimoose key with clip_text_pair features.