Vocabulary
The vocabulary maps category names to categories. Category and value names use lower-case letters, digits and single hyphens, so a label such as time-of-day:golden-hour is never ambiguous. Each value maps to a description that tells the model what it means, or null when the name says enough.
type |
Answers | Options |
|---|---|---|
single |
Exactly one value, such as people:crowd. |
values (at least two), question, minConfidence (default 0), allowUnlisted. |
multi |
Any number of values. | values, question, threshold (default 0.5) applied to each value, allowUnlisted. |
flag |
Yes or no. Labelled with the category name. | question (required), threshold (default 0.5). |
question defaults to one built from the category name. Write your own when the name alone is ambiguous.
allowUnlisted is off by default, so the model must choose from your values. Turn it on to also offer “none of these fits” (single) or “something not listed” (multi). The tag then reports unlisted, the probability that a value is missing. A single category whose likeliest answer is unlisted gets no value. Amos can’t name the missing value; use this to find images where the vocabulary falls short.
image-tagger is a Ben Hodgson projectGitHub