bnn #

Basic building blocks for Bayesian neural networks.

Classes:

Name	Description
`TemperatureScaler`	Temperature scaling.

TemperatureScaler #

TemperatureScaler(
    loss_fn: Module = CrossEntropyLoss(),
    lr: float = 0.1,
    max_iter: int = 100,
    tolerance_grad: float = 1e-07,
    tolerance_change: float = 1e-09,
    history_size: int = 100,
)

Bases: Module

Temperature scaling.

Tunes the temperature parameter of the model in the last layer to minimize the negative log likelihood of the validation set.

Based on On Calibration of Modern Neural Networks.

Parameters:

Name	Type	Description	Default
`loss_fn`	`Module`	The loss function to be used for calibration.	`CrossEntropyLoss()`
`lr`	`float`	Learning rate for the optimizer.	`0.1`
`max_iter`	`int`	Maximum number of iterations per optimization step.	`100`
`tolerance_grad`	`float`	Tolerance for the gradient.	`1e-07`
`tolerance_change`	`float`	Tolerance for the change in the loss function / parameters.	`1e-09`
`history_size`	`int`	Size of the history for the LBFGS optimizer.	`100`

Methods:

Name	Description
`optimize`	Optimizes the temperature of the model.

Attributes:

Name	Type	Description
`history_size`
`loss_fn`
`lr`
`max_iter`
`tolerance_change`
`tolerance_grad`

history_size #

history_size = history_size

loss_fn #

loss_fn = loss_fn

lr #

lr = lr

max_iter #

max_iter = max_iter

tolerance_change #

tolerance_change = tolerance_change

tolerance_grad #

tolerance_grad = tolerance_grad

optimize #

optimize(model: BNNModule, dataloader: DataLoader) -> None

Optimizes the temperature of the model.

Parameters:

Name	Type	Description	Default
`model`	`BNNModule`	The model to be calibrated, assumed to return logits.	required
`dataloader`	`DataLoader`	The dataloader for the dataset to calibrate on (typically the validation set).	required