Reproduce “Reversible 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
Karttikeya Mangalam, Haoqi Fan, Yanghao Li, Chao-Yuan Wu, Bo Xiong, Christoph Feichtenhofer, Jitendra Malik
Reversible Vision Transformers (Rev-ViT) and Reversible Multiscale Vision Transformers (Rev-MViT) from Meta AI and UC Berkeley address the GPU memory bottleneck in deep learning. These architectures allow for up to a 15.5x reduction in GPU memory footprint, enabling significantly larger batch sizes and higher training throughput for deeper models, while maintaining competitive accuracy across image classification, object detection, and video classification tasks.
Key insights
- Reconfigured residual connections, specifically removing internal skips within attention/MLP sub-blocks, are crucial for stable training and scaling of deeper reversible Vision Transformers.
- Reversible architectures inherently possess stronger regularization, necessitating modified training recipes (e.g., lighter augmentations, weaker stochastic depth, higher weight decay) to achieve peak performance.
- Memory savings directly enable significantly larger batch sizes, which can lead to improved training throughput and more efficient hardware utilization, particularly for deeper and higher-resolution models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.