extra-keras-metrics 1.3.1

Additional metrics integrated with the keras NN library, taken directly from `Tensorflow <https://www.tensorflow.org/api_docs/python/tf/metrics/>`_

Additional metrics integrated with the Keras NN library, taken directly from Tensorflow

How do I install this package?

As usual, just download it using pip:

pip install extra_keras_metrics

Tests Coverage

Since some software handling coverages sometimes get slightly different results, here’s three of them:

How do I use this package?

Just by importing it you will be able to access all the non-parametric metrics, such as “auprc” and “auroc”:

import extra_keras_metrics

model = my_keras_model()
model.compile(
    optimizer="sgd",
    loss="binary_crossentropy",
    metrics=["auroc", "auprc"]
)

For the parametric metrics, such as “average_precision_at_k”, you will need to import them, such as:

from extra_keras_metrics import average_precision_at_k

model = my_keras_model()
model.compile(
    optimizer="sgd",
    loss="binary_crossentropy",
    metrics=[average_precision_at_k(1), average_precision_at_k(2)]
)

This way in the history of the model you will find both the metrics indexed as “average_precision_at_k_1” and “average_precision_at_k_2” respectively.

Which metrics do I get?

You will get all the following metrics taken directly from Tensorflow. At the time of writing, the ones available are the following:

The non-parametric ones are (tested against their conterpart from sklearn):

• AUPRC (tested against sklearn’s average_precision_score).

• AUROC (tested against sklearn’s roc_auc_score).

• false_negatives (tested against false negatives from sklearn’s confusion_matrix).

• false_positives (tested against false positives from sklearn’s confusion_matrix).

• mean_absolute_error (tested against sklearn’s mean_absolute_error)

• mean_squared_error (tested against sklearn’s mean_squared_error)

• precision (tested against sklearn’s precision_score)

• recall (tested against sklearn’s recall_score)

• root_mean_squared_error (tested against squared root of sklean’s mean_squared_error)

• true_negatives (tested against true negatives from sklearn’s confusion_matrix)

• true_positives (tested against true positives from sklearn’s confusion_matrix)

The parametric ones are (only execution is tested, no baseline in sklearn was available):

• average_precision_at_k

• precision_at_k

• recall_at_k

• mean_iou

• sensitivity_at_specificity

• specificity_at_sensitivity

Extras

I’ve created also another couple packages you might enjoy: one, called extra_keras_utils that contains some commonly used code for Keras projects and plot_keras_history which automatically plots a keras training history.