What-If Tool 1.7 release and redesigned website

[jwe...@google.com]

The What-If Tool version 1.7.0 has been released, as new versions of all available pip packages (witwidget and tensorboard-plugin-wit), and the release notes have been updated. As of this version, we only ship python3 packages. We also no longer ship witwidget-gpu as tensorflow now provides CPU and GPU capability through just the standard tensorflow package.

The whatif-tool.dev website has been redesigned to include 12 tutorials that describe in detail different features of the tool and ways to perform analysis using it. We'll be adding plenty more tutorials as well. It also contains a FAQ and links to other WIT resources.

Main new features:

- Custom prediction functions can be provided to WIT in TensorBoard, matching

the capability already in notebook mode. This allows use of WIT in TensorBoard

with any python-accessible model (e.g. sklearn, pytorch, keras etc...).

- Help links have been updated to point to the new tutorials on the redesigned

What-If Tool website.

- When attributions are provided, allows counterfactual finding by closest

attributions, as opposed to closest feature values.

- When attributions are provided, added global mean attribution tables in

the performance tab for each slice and the entire dataset.

- Added AUC display for the PR and ROC curves.

Please reach out with any questions. Thanks!

-James