What-If Tool 1.4 released - Fun new stuff

[James Wexler]

witwidget 1.4 pip package has a bunch of new features, including:

- One-click sorting of partial dependence plots by how dynamic they are.

- Ability to define custom datapoint distance functions in python for counterfactual exploration and for creating custom charts.

- Ability to slice numeric features by bucketed groups for performance and fairness analysis.

- Ability to explore counterfactuals of regression models (by defining a delta in prediction value that should be used for looking for counterfactuals).

Check out the updated web demos at https://pair-code.github.io/what-if-tool/ along with the updated text toxicity colab for counterfactuals based on tf.Hub's Universal Sentence Embeddings and the smile detection colab for counterfactuals based on mean image color.

I created some GIFs showing some of these features here: https://twitter.com/bengiswex/status/1178750489879560192

[James Wexler]

One quick update: We just pushed out version 1.4.1 with some fixes:

- Version 1.4 had a regression where in some cases, when using the threshold optimization buttons for binary classification models (in the Performance + Fairness tab), the correct thresholds were being set on the wrong slices.

- If you make use of a custom prediction function (through .set_custom_predict_fn), AND also provide the dataset to WitConfigBuilder as JSON (as opposed to tf Example protos), then the examples to infer were being passed to the custom prediction function as Example protos. This has been fixed so that they are passed as JSON objects, just as they were provided to WitConfigBuilder.