Metrics#

Metrics are specified in the dataset configuration using shorthand syntax. The tracking_metric attribute specifies the metric to be used for early stopping. The metrics attribute in the dataset configuration specifies a list of metrics to be used during training.

Tip

To create custom metrics, refer to the custom metrics tutorial.

For example, a dataset for classification could specify the following metrics:

conf/dataset/ExampleDataset.yaml#
1id: ExampleDataset
2_target_: example_dataset.ExampleDataset
3...
4
5tracking_metric: autrainer.metrics.Accuracy
6metrics:
7  - autrainer.metrics.Accuracy
8  - autrainer.metrics.UAR
9  - autrainer.metrics.F1

Abstract Metric#

class autrainer.metrics.AbstractMetric(name, fn, fallback, **fn_kwargs)[source]#

Abstract class for metrics.

Parameters:

  • name (str) – The name of the metric.

  • fn (Callable) – The function to compute the metric.

  • fallback (float) – The fallback value if the metric is NaN.

  • **fn_kwargs (dict) – Additional keyword arguments to pass to the function.

__call__(*args, **kwargs)[source]#

Compute the metric.

Parameters:

  • *args – Positional arguments for the function.

  • **kwargs – Keyword arguments for the function.

Return type:

float

Returns:

The score.

abstract property starting_metric: float#

The starting metric value.

Returns:

The starting metric value.

abstract property suffix: str#

The suffix of the metric.

Returns:

The suffix of the metric.

abstract static get_best(a)[source]#

Get the best metric value from a series of scores.

Parameters:

a (Union[Series, ndarray]) – Pandas series or numpy array of scores.

Return type:

float

Returns:

Best metric value.

abstract static get_best_pos(a)[source]#

Get the position of the best metric value from a series of scores.

Parameters:

a (Union[Series, ndarray]) – Pandas series or numpy array of scores.

Return type:

int

Returns:

Position of the best metric value.

abstract static compare(a, b)[source]#

Compare two scores and return True if the first score is better.

Parameters:

  • a (Union[int, float]) – First score.

  • b (Union[int, float]) – Second score.

Return type:

bool

Returns:

True if the first score is better.

unitary(y_true, y_pred)[source]#

Unitary evaluation of metric.

Metric computed for each individual label in the case of multilabel classification or regression.

In the default case, computes the metric in the same way as globally.

Parameters:

  • y_true (ndarray) – ground truth values.

  • y_pred (ndarray) – prediction values.

Return type:

float

Returns:

The unitary score.

class autrainer.metrics.BaseAscendingMetric(name, fn, fallback=None, **fn_kwargs)[source]#

Base for ascending metrics with higher values being better.

Parameters:

  • name (str) – The name of the metric.

  • fn (Callable) – The function to compute the metric.

  • fallback (Optional[float]) – The fallback value if the metric is NaN. If None, the fallback value is set to -1e32. Defaults to None.

  • **fn_kwargs (dict) – Additional keyword arguments to pass to the function.

property starting_metric: float#

Ascending metric starting value.

Returns:

-1e32

property suffix: str#

Ascending metric suffix.

Returns:

“max”

class autrainer.metrics.BaseDescendingMetric(name, fn, fallback=None, **fn_kwargs)[source]#

Base for descending metrics with lower values being better.

Parameters:

  • name (str) – The name of the metric.

  • fn (Callable) – The function to compute the metric.

  • fallback (Optional[float]) – The fallback value if the metric is NaN. If None, the fallback value is set to 1e32. Defaults to None.

  • **fn_kwargs (dict) – Additional keyword arguments to pass to the function.

property starting_metric: float#

Descending metric starting value.

Returns:

1e32

property suffix: str#

Descending metric suffix.

Returns:

“min”

Classification Metrics#

class autrainer.metrics.Accuracy[source]#

Accuracy metric using audmetric.accuracy.

class autrainer.metrics.UAR[source]#

Unweighted average recall metric using audmetric.unweighted_average_recall.

class autrainer.metrics.F1[source]#

F1 metric using audmetric.unweighted_average_fscore.

Multi-label Classification Metrics#

class autrainer.metrics.MLAccuracy[source]#

Accuracy metric using sklearn.metrics.accuracy_score.

class autrainer.metrics.MLF1Macro[source]#

F1 macro metric using sklearn.metrics.f1_score.

class autrainer.metrics.MLF1Micro[source]#

F1 micro metric using sklearn.metrics.f1_score.

class autrainer.metrics.MLF1Weighted[source]#

F1 weighted metric using sklearn.metrics.f1_score.

(Multi-target) Regression Metrics#

class autrainer.metrics.CCC[source]#

Concordance correlation coefficient metric using audmetric.concordance_cc for (multi-target) regression. The metric is calculated for each target separately and the mean is returned.

class autrainer.metrics.MAE[source]#

Mean absolute error metric using audmetric.mean_absolute_error for (multi-target) regression. The metric is calculated for each target separately and the mean is returned.

class autrainer.metrics.MSE[source]#

Mean squared error metric using audmetric.mean_squared_error for (multi-target) regression. The metric is calculated for each target separately and the mean is returned.

class autrainer.metrics.PCC[source]#

Pearson correlation coefficient metric using audmetric.pearson_cc for (multi-target) regression. The metric is calculated for each target separately and the mean is returned.

On this page