Checkpoint mountains
Every training run keeps 'just one more' checkpoint. Times hundreds.
المحتويات
Where the gigabytes hide
Env sprawl
Conda and venv environments duplicate gigabytes per project.
Output-embedded notebooks
A 200 MB .ipynb is 199 MB of base64 plots.
The workflow
Rank by size, filter by age
Oldest-plus-largest is the delete shortlist. Statistics views help here.
Quarantine checkpoints
Keep the best per experiment; the collector holds the rest for a final review.
Strip notebook outputs
Re-run beats storage. Clear outputs before archiving notebooks.
Measure envs before pruning
Allocated size can surprise — verify before removing environments.
Tools worth installing
NullDisk first — native map, staged collector, zero telemetry. Then, depending on taste:
- GrandPerspective — The tiny open-source treemap that shows your biggest files instantly.
- JDiskReport — Java statistics charts that find your most obsolete folders.
- Diskaroo — A treemap visualizer with collector tray and smart filters.
أسئلة شائعة
Which view helps with datasets?
Age-plus-size statistics (JDiskReport's specialty) or a treemap sorted by size. Numbers beat vibes for data hoards.
Are data files safe to delete?
Derived artifacts (checkpoints, caches, outputs) yes; raw data almost never. When in doubt, archive to external storage first.