TensorFlow Model Remediation is a library that provides solutions for machine learning practitioners working to create and train models in a way that reduces or eliminates user harm resulting from underlying performance biases.
You can install the package from pip
:
$ pip install tensorflow-model-remediation
Note: Make sure you are using TensorFlow 2.x.
This library contains a collection of machine learning remediation techniques for addressing potential bias in a model.
Currently TensorFlow Model Remediation contains the below techniques:
MinDiff technique: Typically used to ensure that a model predicts the preferred label equally well for all values of a sensitive attribute. Helpful when trying to achieve equality of opportunity.
Counterfactual Logit Pairing technique: Typically used to ensure that a model’s prediction does not change between “counterfactual pairs”, where the sensitive attribute referenced in a feature is different. Helpful when trying to achieve counterfactual fairness.
We recommend starting with the overview guide to get an idea of TensorFlow Model Remediation. Next try one of our interactive guides like the
Counterfactual tutorial notebook.
import tensorflow_model_remediation as tfmr
import tensorflow as tf
# Start by defining a Keras model.
original_model = ...
# Next pick the remediation technique you'd like to use. For example, a
# MinDiff implementation might look like the below:
# Set the MinDiff weight and choose a loss.
min_diff_loss = tfmr.min_diff.losses.MMDLoss()
min_diff_weight = 1.0 # Hyperparamater to be tuned.
# Create a MinDiff model.
min_diff_model = tfmr.min_diff.keras.MinDiffModel(
original_model, min_diff_loss, min_diff_weight)
# Compile the MinDiff model as you normally would do with the original model.
min_diff_model.compile(...)
# Create a MinDiff Dataset and train the min_diff_model on it.
min_diff_model.fit(min_diff_dataset, ...)
If you're interested in learning more about responsible AI practices, including
fairness, please see Google AI's Responsible AI Practices.
tensorflow/model_remediation
is Apache 2.0 licensed. See the
LICENSE
file.