metric

Metrics to use for training or explanations.

Classes:

Name	Description
Metric	Represents a metric for the explainable model.
TorchGlobalMetric	Represent a metric to be used in the training or evaluation process for the whole model (global metric).
TorchLeafMetric	Represent a metric to be used during the training or the evaluation process, per leaf.
DictMetrics	Dictionary of metrics (torch eval etc...) used in training or explain stage.

Metric

Represents a metric for the explainable model.

TorchGlobalMetric

Represent a metric to be used in the training or evaluation process for the whole model (global metric).

Parameters:

Name	Type	Description	Default
metric	partial[Metric]	The torchmetrics metric to use, as partial.	required
on_raw_data	bool	If set, value are compute from raw data and not preprocessed data (inputs or targets).	False
reduced_dimensions	tuple[int, ...]	The dimensions to reduce in the statistics. For instance, given a batch with dimensions (batch size x num_timestamps x num_channels), and a reduced_dimensions=(0,2), it returns a per timestamp statistic, averaged over the batch and channels. Default (0,), meaning the statistic is reduced over the batch only.	(0, )

Attributes:

Name	Type	Description
metric	partial[Metric]
on_raw_data	bool
reduced_dimensions	tuple[int, ...]

metric: partial[torchmetrics.metric.Metric]

on_raw_data: bool = False

reduced_dimensions: tuple[int, ...] = (0)

TorchLeafMetric

Represent a metric to be used during the training or the evaluation process, per leaf.

Parameters:

Name	Type	Description	Default
metric	partial[Metric]	The torchmetrics metric to use, as partial.	required
on_raw_data		If set, value are compute from raw data and not preprocessed data (inputs or targets).	required
reduced_dimensions		The dimensions to reduce in the statistics. For instance, given a batch with dimensions (batch size x num_timestamps x num_channels), and a reduced_dimensions=(0,2), it returns a per timestamp statistic, averaged over the batch and channels. Default (0,), meaning the statistic is reduced over the batch only.	required

Attributes:

Name	Type	Description
metric	partial[Metric]
on_raw_data	bool
reduced_dimensions	tuple[int, ...]

metric: partial[torchmetrics.metric.Metric]

on_raw_data: bool = False

reduced_dimensions: tuple[int, ...] = (0)

DictMetrics

Dictionary of metrics (torch eval etc...) used in training or explain stage.

If a metric is: - from torchmetrics, it will be computed globally (for the whole model) and per leaf. - a TorchGlobalMetric, it will be computed globally only. - a TorchLeafMetric, it will be computed per leaf only.

In the training stage, metrics help to monitor the training. In the explain stage, metrics are displayed in XpViz along with explanations. Dictionary keys are will be used as metric names in XpViz.