v3.0.0: LLM support, Improved UI, Aggregator Refactor

Changes summary

Machine Learning and collaborative learning

• Add LLM support based on gpt-tfjs, for collaborative pre-training and fine-tuning #643 #644
• Split the collaborative training logic into the Aggregator and Client abstractions #578.
• Support text datasets and tokenization via Transformers.js #577 #651
• Benchmark gpt-tfjs #659
• Add browser support for LLM (tokenization, memory requirements, IndexedDB storage) #675
• Improve all base ML tasks #616 #660

Technical

• Update and extend documentation #614 #636, a table of content can be found in DEV.md
• Greatly simplify repository management via NPM workspaces #617
• Extend ESLint coverage #625
• Upgrade codebase to ES2022 modules and typescript #652
• Upgrade Node.js from v16 to v20 #653
• Update dependencies and fix security issues #655
• discojs is now a standalone package #664
  • discojs-node & discojs-web provide platform specific helpers
• Publish a container image to easily run the server #664
• Rolling releases for NPM packages and container image, each push to develop builds and deploys latest version #664

User Interface

• Fix bugs #608 #624
• Add user feedback during error handling #624
• Improve navigation #668
• UI design improvements #547 #686 #692

Downgrades

• The Byzantine-Robust aggregator support is temporarily inactive, due to the aggregator refactoring #578

Additional PRs merged

• v2.1.1 by @s314cy in #544
• Download and infer from remote models by @s314cy in #545
• clarify TASK docs by @morganridel in #550
• Fixing up assess(...) function by @francesco98 in #558
• more task doc update by @morganridel in #553
• adds generic training functions for trainer by @morganridel in #548
• Inference feature in the validation UI of the web client by @francesco98 in #543
• Updating GeoTags model by @francesco98 in #570
• Validation Metrics & UI Updates by @s314cy in #547
• Fix Geotags & Web Client by @s314cy in #574
• Add converter script to keras by @walidabn in #567
• Implementing end-to-end test with Byzantine-Robust Aggregator for cifar10 task by @francesco98 in #557
• aggregator and client refactor by @LucasTrg in #578
• Adding the skin_mnist task to the default tasks by @walidabn in #579
• Add support for text data & tokenization by @s314cy in #577
• Add API docs for disco by @s314cy in #580
• Reduce GCloud usage by @s314cy in #592
• Reduce GCloud usage by @s314cy in #593
• Reduce GCloud usage by @s314cy in #594
• reduce gcloud usage by @s314cy in #595
• Task creation form privacy parameters default by @walidabn in #596
• Remove stale package by @s314cy in #598
• CLI: Add placeholder for custom tasks and data by @s314cy in #599
• Refactoring created mismatched function calls by @JulienVig in #608
• Fix training related failures by @JulienVig in #616
• Update DISCO documentation by @JulienVig in #614
• Fix flaky titanic validator test by @JulienVig in #630
• *: use npm workspaces by @tharvik in #617
• Solve UI bugs and add UI error feedback by @JulienVig in #624
• full eslint by @tharvik in #625
• Update doc according npm workspace by @JulienVig in #636
• don't reexport tfjs/discojs-core by @tharvik in #640
• add fondations for LLM by @tharvik in #643
• #643 fixes by @tharvik in #645
• discojs-core/models: add gpt by @tharvik in #644
• bump node by @tharvik in #653
• *: upgrade to es2022 by @tharvik in #652
• Add tokenization and prompting API to GPT models by @JulienVig in #651
• updates dependencies by @tharvik in #655
• Fix gpt-tfjs bugs, add tests and refactor code by @JulienVig in #658
• small lus covid fixes by @tharvik in #661
• Benchmark GPT-tfjs by @JulienVig in #659
• LUS COVID demo by @JulienVig in #660
• Add benchmark results summary in cli/README.md by @JulienVig in #663
• docker: freeze node version by @tharvik in #672
• Improves and fixes the web-client UI by @JulienVig in #668
• rolling deploy by @tharvik in #664
• discojs/aggregator: rework tests by @tharvik in #676
• Create a Skin Condition classification task by @JulienVig in #678
• webapp/ModelLibrary: add some tests by @tharvik in #681
• Integrate the wikitext task in the webapp by @JulienVig in #675
• Fix DEV.md style by @JulienVig in #685
• UI polishing by @JulienVig in #686
• *: prerelease fixes by @tharvik in #688
• UI tweaks by @JulienVig in #692
• add auto-release by @tharvik in #693
• dependabot: fix for monorepo by @tharvik in #698
• webapp/TrainingInformation: fix graph jumps by @tharvik in #696
• docker: bump base image by @tharvik in #706
• Fix preprocessing memory leaks by @JulienVig in #711
• Update README.md by @JulienVig in #709
• Fix decentralized learning fail by @JulienVig in #708
• readd some privacy by @tharvik in #712
• Display landing page before server boots by @JulienVig in #720
• small UI changes by @tharvik in #725
• release v3 by @tharvik in #724

New Contributors

• @LucasTrg made their first contribution in #578
• @JulienVig made their first contribution in #608

Full Changelog: v2.1.1...v3.0.0

v2.1.1

Disco v2.1.1 Release Notes

New Features

• Disco Server
  • Exposes an API to easily start a server on a given port and add new tasks (as the server host)
  • Supports digest verification of task models being downloaded from the Internet, useful for pre-trained or custom-hosted models
• Disco.js
  • Exposes fetchTasks and pushTask functions to easily fetch tasks or add a new one to a remote Disco server (as a client)
• Web Client
  • Fixed the display of task parameters

New Server Package

The Disco Server is now available as a package on NPM, easy to use and integrate within your project.

Versioning

Both the Disco.js and Disco Server libs follow the same versioning scheme. This is why both Disco.js and Disco Server will receive a version bump even if only one of the two libs is receiving code changes.

Moreover, Disco Server being dependent on Disco.js, it means that Disco Server version 2.1.1 will require and be compatible with any Disco.js version between >=2.1.1 and <3.0.0

v2.0.0

Disco v2.0 release notes

Main new features

• Byzantine-robust training for federated learning
• Fully P2P communication for decentralized training
• Task Creation Form - Bring new ML tasks to Disco by simply filling a web form: no programming knowledge required
• More clear split of the Disco projects into discojs, web-client, cli and server
• Disco.js available as two independent NPM packages:
  • @epfml/discojs running in the browser as native JS
  • @epfml/discojs-node running on Node.js
• Web Client: Lots of UI/UX improvements
• CLI: Improved CLI for running experiments without any browser or UI. Self-contained example here
• Development & Documentation
  • Development support for Apple Silicon and Windows (in addition to Linux and Mac)
  • Much improved documentation, both for task contributors and for developers
• Miscellaneous
  • The GitHub repository's download size was reduced from 186MiB down to 3.51MiB!

Closed issues

More than 50 closed issues, bug fixes and new features:
https://github.com/epfml/disco/milestone/6?closed=1

v1.0.0

Initial Disco 🔮 release.

Supporting decentralized ✨ and federated 🌟 learning of arbitrary deep learning models with clients in browser or javascript.
Privacy: supports secure aggregation and differential privacy

See ReadMe, documentation and developer guide for more information
https://github.com/epfml/disco/