Lab22 CLI

lab22 is the command-line tool for downloading the datasets you are licensed for on the Lab22 platform. It needs Python 3.9 or newer and nothing beyond the standard library.

Install

python3 -m venv ~/lab22-venv           # Python 3.9 or newer; the one macOS ships is fine
source ~/lab22-venv/bin/activate
pip3 install lab22

Use

export LAB22_API_KEY=l22_sk_live_...   # from the API Keys page; the only way to authenticate
                                       # (older keys start vz_sk_live_ and work too)

lab22 datasets                         # what you can download, and where to request access
lab22 describe ego4d                   # its parts, and the values each filter takes, with file counts
lab22 download ego4d -o ~/lab22-data/ego4d --parts clips annotations --benchmarks FHO --dry-run
lab22 download ego4d -o ~/lab22-data/ego4d --parts clips annotations --benchmarks FHO -y
lab22 verify ~/lab22-data/ego4d --checksum

python3 -m lab22 works the same as lab22.

A command written for the official ego4d or egoexo tool runs unchanged with lab22 in front. lab22 ego4d and lab22 egoexo take the official tools' options, with their spellings, abbreviations and defaults, and accept the AWS options and ignore them. They add four options of their own, which work only when spelled in full: --dry-run, --shard N/M, --refresh-manifest and --ignore-disk-space.

lab22 ego4d -o ~/lab22-data/ego4d --datasets clips --benchmarks EM --dry-run
lab22 egoexo -o ~/lab22-data/egoexo4d --parts annotations metadata -y

lab22 download <dataset> takes the official tools' options, spellings and defaults as well (--parts/--datasets, --uids/--video_uids, --uid-file, --universities, --benchmarks, --splits, --views, --metadata/--no-metadata, --version/--release, --config_path, --list-datasets, --manifest, --force, --delete), with the AWS options accepted and ignored. On top: --shard N/M for splitting one download across machines, --workers, --dry-run, -y, and --refresh-manifest. A dry run fetches nothing but does create the output folder and save the manifest in it, so the real run that follows targets the same release. Files land in the official layout under -o; .lab22/ inside it holds the saved manifest, a log and a done-record per run. Rerunning the same command skips completed files; a file that was interrupted starts again from its beginning. Interrupted or failed downloads leave *.part files, which a run and verify report and never delete (shards may share a folder); delete them when no run is using the folder.

Exit codes: 0 done · 1 usage · 2 key problem · 3 not entitled · 4 incomplete/unreachable.

The full guide is on the Lab22 platform, signed in, under Documentation → CLI: installation, API keys, choosing what to download, each dataset's parts and layout, verifying, large and sharded downloads, a command reference and troubleshooting. Its source is docs/ in the source repository; working on the CLI itself is covered in DEVELOPMENT.md there.

Coming from vizo

lab22 is the new name of the vizo CLI, with the same commands and options. pip3 install --upgrade vizo installs lab22 and keeps a vizo command that runs it, so existing scripts keep working; pip3 install lab22 installs it directly. If you installed vizo with pipx, run pipx install lab22.

VIZO_API_KEY and VIZO_API_URL are still read when the LAB22_ variables are not set, and keys that start with vz_sk_live_ still work.

Keep downloading into the folders you already have, such as ~/vizo-data/ego4d, and leave their .vizo/ folder where it is. lab22 keeps using it, so a download resumes as before. Upgrade vizo on every machine and in every job that writes to your folders, because vizo 0.0.x knows only .vizo/. In a folder lab22 started, it makes a new .vizo/ beside .lab22/, which lab22 then refuses, and its --delete deletes what is in .lab22/.