https://github.com/outbrain-inc/fwumious_wabbit

Fwumious Wabbit, fast on-line machine learning toolkit written in Rust
https://github.com/outbrain-inc/fwumious_wabbit

data-stream-mining deep-learning factorization-machines incremental-learning logistic-regression machine-learning online-learning rust

Last synced: 9 days ago
JSON representation

Fwumious Wabbit, fast on-line machine learning toolkit written in Rust

• Host: GitHub
• URL: https://github.com/outbrain-inc/fwumious_wabbit
• Owner: outbrain-inc
• License: other
• Created: 2020-09-14T12:52:38.000Z (over 4 years ago)
• Default Branch: main
• Last Pushed: 2024-06-20T09:21:56.000Z (6 months ago)
• Last Synced: 2024-12-10T21:41:43.293Z (16 days ago)
• Topics: data-stream-mining, deep-learning, factorization-machines, incremental-learning, logistic-regression, machine-learning, online-learning, rust
• Language: Rust
• Homepage:
• Size: 4.4 MB
• Stars: 134
• Watchers: 14
• Forks: 19
• Open Issues: 6
• Metadata Files:
  • Readme: README.md
  • Changelog: CHANGELOG
  • License: LICENSE.md

Awesome Lists containing this project

README

        
Fwumious Wabbit is

- a very fast machine learning tool

- built with Rust

- inspired by and partially compatible with Vowpal Wabbit (much love! read more about compatibility [here](COMPATIBILITY.md))

- currently supports logistic regression and (deep) field-aware factorization machines

[![Rust-UbuntuLatest](https://github.com/outbrain/fwumious_wabbit/actions/workflows/rust.yml/badge.svg?branch=main)](https://github.com/outbrain/fwumious_wabbit/actions/workflows/rust.yml)

[![Rust-Ubuntu18](https://github.com/outbrain/fwumious_wabbit/actions/workflows/rust-Ubuntu18.yml/badge.svg)](https://github.com/outbrain/fwumious_wabbit/actions/workflows/rust-Ubuntu18.yml)

[![Gitter](https://badges.gitter.im/FwumiousWabbit/community.svg)](https://gitter.im/FwumiousWabbit/community?utm_source=badge&utm_medium=badge&utm_campaign=pr-badge)

Fwumious Wabbit is actively used in Outbrain for offline research, as well as for some production flows. It

enables "high bandwidth research" when doing feature engineering, feature

selection, hyperparameter tuning, and the like.

Data scientists can train hundreds of models over hundreds of millions of examples in

a matter of hours on a single machine.

For our tested scenarios it is almost two orders of magnitude faster than the

fastest Tensorflow implementation of Logistic Regression and FFMs that we could

come up with. It is an order of magnitude faster than Vowpal Wabbit for some specific use-cases.

Check out our [benchmark](BENCHMARK.md), here's a teaser:

![benchmark results](benchmark_results.png)

**Why is it faster?** (see [here](SPEED.md) for more details)

- Only implements Logistic Regression and (Deep) Field-aware Factorization Machines

- Uses hashing trick, lookup table for AdaGrad and a tight encoding format for the "input cache"

- Features' namespaces have to be declared up-front

- Prefetching of weights from memory (avoiding pipeline stalls)

- Written in Rust with heavy use of code specialization (via macros and traits)

- Special emphasis on efficiency of sparse operations and serving

# Weight patching

This repo also contains the patching algorithm that enables very fast weight diff computation see `weight_patcher` for more details.