alibi.models.pytorch.metrics

This module contains a loss wrapper and a definition of various monitoring metrics used during training. The model to be trained inherits form :py:class:alibi.explainers.models.pytorch.model.Model and represents a simplified version of the tensorflow.keras API for training and monitoring the model. Currently it is used internally to test the functionalities for the Pytorch backend. To be discussed if the module will be exposed to the user in future versions.

AccuracyMetric

Inherits from: Metric, ABC

Accuracy monitoring metric.

Constructor

AccuracyMetric(self, name: str = 'accuracy')

Name	Type	Default	Description
name	str	'accuracy'	Name of the metric.
reduction			Metric's reduction type. Possible values mean

Methods

compute_metric

compute_metric(y_pred: Union[torch.Tensor, numpy.ndarray], y_true: Union[torch.Tensor, numpy.ndarray]) -> None

Name	Type	Default	Description
y_pred	Union[torch.Tensor, numpy.ndarray]		Predicted label.
y_true	Union[torch.Tensor, numpy.ndarray]		True label.

Returns

• Type: None

LossContainer

Loss wrapped to monitor the average loss throughout training.

Constructor

LossContainer(self, loss: Callable[[torch.Tensor, torch.Tensor], torch.Tensor], name: str)

Name	Type	Default	Description
loss	Callable[[.[<class 'torch.Tensor'>, <class 'torch.Tensor'>]], torch.Tensor]		Loss function.
name	str		Name of the loss function

Methods

reset

reset()

Resets the loss.

result

result() -> Dict[str, float]

Computes the average loss obtain by dividing the cumulated loss by the number of steps

Returns

• Type: Dict[str, float]

Metric

Inherits from: ABC

Monitoring metric object. Supports two types of reduction: mean and sum.

Constructor

Metric(self, reduction: alibi.models.pytorch.metrics.Reduction = <Reduction.MEAN: 'mean'>, name: str = 'unknown')

Name	Type	Default	Description
reduction	alibi.models.pytorch.metrics.Reduction	<Reduction.MEAN: 'mean'>	Metric's reduction type. Possible values mean
name	str	'unknown'	Name of the metric.

Methods

compute_metric

compute_metric(y_pred: Union[torch.Tensor, numpy.ndarray], y_true: Union[torch.Tensor, numpy.ndarray])

Name	Type	Default	Description
y_pred	Union[torch.Tensor, numpy.ndarray]
y_true	Union[torch.Tensor, numpy.ndarray]

reset

reset()

Resets the monitoring metric.

result

result() -> Dict[str, float]

Computes the result according to the reduction procedure.

Returns

• Type: Dict[str, float]

update_state

update_state(values: numpy.ndarray)

Update the state of the metric by summing up the metric values and updating the counts by adding the number of instances for which the metric was computed (first dimension).

Name	Type	Default	Description
values	numpy.ndarray

Reduction

Inherits from: Enum

Reduction operation supported by the monitoring metrics.