As part of our goal of providing more frequent releases, we are excited to announce VW 9.2.0! Some highlights include a large restructuring of VW’s library structure which better organizes related source and header files, the ability to call learn and predict directly from a string in java, access to model weights in Python, and colored output logs for errors, warnings, and general info.

Highlights

Refactoring of library structure

Historically VW has clumped together most of the source and headers files for its reductions into one directory. As more features and targets were added, it seemed necessary to better organize this structure so that new and current users could compartmentalize different parts of VW. Check out this PR for more details.

Support predict and learn directly from string in Java

You can now predict and learn directly from a string in Java with the predictFromString and learnFromString methods.

Access model weights in Python

Model weights can now be accessed directly in Python in JSON format using the json_weights function.

Experimental: Hardening of Automl and Non-Stationary Exploration Algorithms

Both reductions have gone through major revamps to improve their statistical accuracy.

Mean field implementation for logistic regression

Logistic regression can now take non-binary costs via mean field implementation. E.g. for loss p in [0,1] the update is p*update(1) + (1-p)*update(-1).

Colored logs

Logs from VW’s command line interface will now display errors in red, warnings in yellow, and info in green.

Thank you

A huge thank you and welcome to all of the new contributors since the last release:

• Byron Xu
• Bhavya Saraf
• Shaokun Zhang
• Oren Michaely
• Akshay Krishnamurthy
• Scott Inglis

And of course thank you to existing contributors:

• Ariel Faigon
• Griffin Bassman
• Olga Vrousgos
• Jack Gerrits
• Eduardo Salinas
• Wangda Zhang
• Markus Cozowicz
• Mark Rucker
• Rajan Chari
• Peter Chang
• Alexey Taymanov