Output calibrator

The output calibration calibrates the model outputs. We explicitly support a temperature scaling output calibrator for classification, and a temperature scaling output calibrator for regression.

Alternatively, you can bring in your own output calibrator by overwriting Module.

class fortuna.output_calibrator.classification.ClassificationTemperatureScaler(parent=<flax.linen.module._Sentinel object>, name=None)

Classification temperature scaling. It scales the logits with a scalar temperature parameters. Let \(o\) be output logits and \(\phi\) be a scalar parameter. Then the scaling can be seen as \(g(\phi, o) = \exp(-\phi) o\).

name: Optional[str] = None

parent: Union[Type[Module], Scope, Type[_Sentinel], None] = None

scope: Optional[Scope] = None

class fortuna.output_calibrator.regression.RegressionTemperatureScaler(parent=<flax.linen.module._Sentinel object>, name=None)

Regression temperature scaling. It multiplies the variance with a scalar temperature parameters. Let \(v\) be the variance outputs and \(\phi\) be a scalar parameter. Then the scaling can be seen as \(g(\phi, o) = \exp(\phi) v\).

name: Optional[str] = None

parent: Union[Type[Module], Scope, Type[_Sentinel], None] = None

scope: Optional[Scope] = None

class fortuna.output_calib_model.state.OutputCalibState(step, apply_fn, params, tx, opt_state, encoded_name=(79, 117, 116, 112, 117, 116, 67, 97, 108, 105, 98, 83, 116, 97, 116, 101), frozen_params=None, dynamic_scale=None, mutable=None)

classmethod init(params, mutable=None, optimizer=None, **kwargs)

Initialize an output calibration state.

Parameters:

• params (CalibParams) – The calibration parameters.

• optimizer (Optional[OptaxOptimizer]) – An Optax optimizer associated with the calibration state.

• mutable (Optional[CalibMutable]) – The calibration mutable objects.

Returns:

A calibration state.

Return type:

Any

classmethod init_from_dict(d, optimizer=None, **kwargs)

Initialize a calibration state from a dictionary.

Parameters:

• d (Union[Dict, FrozenDict]) – A dictionary with as keys the calibrators and as values their initializations.

• optimizer (Optional[OptaxOptimizer]) – An optax optimizer to assign to the calibration state.

Returns:

A calibration state.

Return type:

OutputCalibState