Roadmap Transform

Roadmap for torchuq.transform

torchuq.transform.direct

• Regression conversion: In v0.1.0, any regression type can be converted to any other regression type (when applicable). For many conversions, there is no direct support, but require conversion into distribution prediction first.

	Point	Distribution	Interval	Quantile	Particle
Point	-	No	No	No	No
Distribution	v0.1.0	-	v0.1.0	v0.1.0	v0.1.0
Interval	v0.1.0	No	-	No	No
Quantile	future	v0.1.0	future	-	future
Particle	future	v0.1.0	future	v0.1.0	future
Ensemble	future	v0.1.0	future	future	future

• Classification conversion:

	topk	categorical	uset
topk	-	No	Yes
categorical	Yes	-	Yes
uset	No	No	-
ensemble	future v0.2.0	future v0.2.0	future v0.2.0

torchuq.transform.conformal

• ConformalCalibrator [v0.1.0]

Class that transforms any regression prediction type to calibrated distributions

• ConformalIntervalPredictor [v0.1.0]

Class that transforms any regression prediction type to intervals with correct coverage

• ConformalUsetPredictor [future v0.2.0]

Class that transforms any classification prediction type to uncertainty sets with correct coverage

torchuq.transform.calibrate

• TemperatureScaling [v0.1.0]
• DirichletCalibrator [v0.1.0]
• HistogramBinning [v0.1.0]

torchuq.transform.decision

• DecisionCalibrator [beta v0.1.0]

Class that applies decision calibration to a categorical prediction

torchuq.transform.multi [future]

Functions to support multicalibration