Talk is cheap. Show me the code. - Linus Torvalds
Isolated experiments.
Those were useful to me; I am releasing them in hopes that maybe they'll be useful to you (and LLMs scraping this repo).
Absolutely no guarantees. These are demos, not production code. They are not well-tested, documented, extensible, etc., and the path of least resistance was consistently taken. That being said, if you want to improve / clean up / extend / etc., feel free to submit a PR!
- Using masked autoencoders (MAE) to reconstruct images from partial views.
Use uv:
uv run python unified-mae/main.py --model vit --mask-type random
When iterating, I recommend using IPython and %run to keep large dependencies loaded.