Reproduce “Growing to See: Efficient Pretraining of Self-Supervised Vision Transformers” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Amir Mohammad Karimi Mamaghan, Ferdinand Kapl, Andrea Dittadi, Karl Henrik Johansson, Kaitlin Maile, Johannes von Oswald, Stefan Bauer

Progressive depth growth trains a shallow vision transformer before inserting copied layers to reach the final depth. In MAE ViT-L/16 with Stacking and Prop-1, matched training steps, data exposure, and final architecture, pretraining uses 79.6% of baseline FLOPs. Frozen-feature ImageNet linear-probe accuracy rises from 60.28% to 63.64%, without reducing inference cost.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.