API reference

from zeropoint import ZeroPoint, Decision, UnsupportedLanguage, ModelFilesError, __version__

ZeroPoint(model_dir, threads=2)

Loads a model: the one shipped with the package when model_dir is not given, or a folder written by save. Every model file is checked against its sha256.

Argument	Default	What it does
model_dir	the shipped model	folder of a saved model
threads	2	CPU threads for the encoder; does not change the results

Raises ModelFilesError if a model file is missing or does not match its hash, and ValueError if threads is not a positive integer.

ZeroPoint.decide(text, options)

Chooses the option that fits text, or None. Returns a Decision.

Raises ValueError for an empty text, an empty list, an empty option or the same option twice, TypeError when the text or an option is not a string, and UnsupportedLanguage when more than 20% of the letters of the text or of an option are outside the Latin script.

ZeroPoint.decide_many(texts, options)

Decides several texts. options is one list for all texts, or one list per text. Returns a list of Decision, each exactly what decide gives for that text alone.

ZeroPoint.fit(texts, labels, *, groups=None, steps=200, frozen_layers=8, seed=0, threads=None)

Adapts the model to your labelled texts. labels holds the right option for each text, or None for texts that fit none of them. Returns (model, report); the report compares the model before and after on a held-out 20% of the examples. Needs pip install "zeropoint[fit]". See Adapting to your labels.

ZeroPoint.recalibrate(texts, labels, *, groups=None, seed=0)

Refits only the temperature and the none level. Needs no PyTorch. Returns (model, report).

ZeroPoint.save(path)

Writes the model to the folder path, with a manifest holding the sha256 of every file and the report of fit or recalibrate. Load it with ZeroPoint(path).

ZeroPoint attributes

Attribute	What it holds
languages	codes of the 25 supported languages
model_id	first 12 characters of the hash of the model's training checkpoint
threads	CPU threads used
report	the report of fit or recalibrate, for a model made by them; None otherwise

Decision

Field	Type	What it holds
choice	str or None	the chosen option, or None when none is the most probable answer
confidence	float	probability of choice (of none when choice is None)
p_none	float	probability that none of the options fits
probabilities	dict[str, float]	probability of each option, in the order given; with p_none they add up to 1
depths	dict[str, float]	depth of each option's well, 0 or more
text_truncated	bool	the text was longer than 192 word pieces and its end was ignored
options_truncated	tuple[str, ...]	options longer than 16 word pieces, whose end was ignored
blocks	int	reads of 20 options the list needed
latency_ms	float	time the decision took

Printing a Decision shows the options with their probabilities, none, the choice and the time.

Exceptions

Exception	Raised when
UnsupportedLanguage (a ValueError)	the text or an option is mostly outside the Latin script
ModelFilesError (a RuntimeError)	a model file is missing or does not match its sha256