Reproductions of research papers and demos on edge-toolkit. Each one takes a published result or demo, reruns it with edge-toolkit's hub and, where applicable, its runners in place of the original infrastructure, and records what changed.
- Split learning for health: Distributed deep learning without sharing raw patient data (Vepakomma, Gupta, Swedish and Raskar, MIT Media Lab, 2018): split-learning-mit
- On-Device Training Under 256KB Memory (Lin et al., 2022): tiny-training-mit
Each reproduction is self-contained under reproductions/<name>/:
README.md: what is being reproduced, how to run it, and how it differs from the original.mise.tomlandmise.lock: the tools and tasks to set it up and run it, and the tool versions locked for it.modules/: the edge-toolkit modules written for it.
A reproduction may also contain:
upstream/: original code as a Git submodule pinned to the reproduced version.scenario.yamlanddeployment/: an et-cli cluster input and its generated deployment.screenshots/: screenshots of a run of it.
Browser-hosted modules that do not need a headless runner may use hand-written mise tasks and generate large build inputs locally.
Install mise, then clone with the submodules:
git clone --recurse-submodules https://github.com/edge-toolkit/research.gitFrom there, follow the README of the reproduction you want to run.
The repository-wide checks run from the root:
mise installmise run checkThat runs dprint (Markdown, JSON, YAML), editorconfig-checker, gitleaks, ruff (Python lint and format), taplo (TOML)
and typos, and checks every committed mise.lock is up to date with its mise.toml. mise run fmt applies the
formatters. Upstream submodules, virtualenvs and node_modules are excluded throughout, and the generated
deployment/ directories are left to et-cli rather than the formatters.
Tool versions are locked in the mise.lock beside each mise.toml, so latest installs what was last locked. After
changing a mise.toml, run mise lock in its directory; mise lock --bump moves the latest tools to their newest
releases. The generated deployment/ directories are the exception: they always install the newest hub and runner.