This version greatly increases the number of datapoints that can be loaded into the tool, in both TensorBoard and notebook modes. WIT can now handle on the order of hundreds of thousands of datapoints for simple tabular datasets. This version also enables witwidget in JupyterLab version 2.x.
- #110 - Change file loading rules in TensorBoard.
- #120 - Allow larger number of examples in colab.
- #121 - Display actual server error text in tool.
- #144 - Jupyter notebook scaling update.
- #128 - Handle large number of examples in TensorBoard.
- #134 - Enable WitWidget in JupyterLab2.
This release has a few notable major 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, allow counterfactual finding by closest attributions.
- When attributions are provided, add global mean attribution tables in the performance tab for each slice and the entire dataset.
Also, python 2 support has been removed. Please use python 3.
- #59 - Display global attributions table in performance tab.
- #60 - Find counterfactual by attribution distance.
- #92 - Added AUC display for PR and ROC curves.
- #94 - Allow custom prediction functions in TensorBoard.
- #101 - Added links to new help page and tutorials.
This is the first release that includes the tensorboard-plugin-wit package that enables WIT to be used in TensorBoard as a dynamic plugin. It also removes the dependency on the tensorflow-serving-api package.
- #41 - Add delayed rendering option to WitWidget.
- #45 - Finish WIT as a dynamic TensorBoard plugin.
- #47 - Remove dependency on tensorflow-serving-api package.
- #43 - Update in-tool text for clearer UX.
This is the first release made from the source code since it was moved into the pair-code/what-if-tool repository. This release ensures WitWidget will work in the upcoming Chrome version 80, which removes HTML import support.
- #30 - Remove link element usage from notebook mode.
This release is the first official release with support for attribution value display in the What-If Tool. This includes support for Cloud AI Platform models that return attribution values through their explain endpoint.
- #2882 - Cloud AI Platform attribution updates.
- #2909 - Attribution display improvements.
- #2897 - Update regression demo with attributions.
- #2901 - Fix JupyterHub widget display error.
- #2916 - WitWidget small fixes including PD plot range editing.
- #2846 - Fix an issue with setting examples in Jupyter notebooks.
- #2816 - Fix an issue with TFMA slicing callback.
- #2787 - Add ml service name and version for CAIP.
- #2781 - Add callback creator for TFMA slicing usage.
- #2785 - Fix sliced intersection.
- #2744 - Allow set_example to accept JSON.
- #2762 - Fix cloud model usage with JSON input.
Note that as of 1.4.1, if you provide a custom prediction function, and provide your examples to WitConfigBuilder as JSON (not Example protos), then the examples will be passed to your custom predict function in that JSON format. This is technically a breaking API change but the existing behavior was incorrect and this specific configuration is so rarely used that we wanted to correct it immediately without any need for special code/flags.
- #2716 - Better custom predict fn when provided JSON input.
- #2716 - Fix fairness threshold setting bug introduced in 1.4.0.
- #2607 - Add ability to set custom distance function for counterfactuals and distance visualizations.
- #2461 - Add ability to sort PD plots by interestingness.
- #2660 - Add ability to consume arbitrary prediction-time information.
- #2678 - Can now slice by numeric features in Performance & Fairness tab.
- #2647 - Add counterfactual analysis for regression models.
- #2663 - Visual updates for displaying attributions.
- #2630 - Added developers guide documentation.
- #2698 - Add ability to adjust attributions.
- #2682 - Fix issue with threshold sliders not updating on fairness button presses.
- #2669 - Fix PD plots in python3.
- #2648 - Fix image handling broken from Polymer 2 update.