Reproduce “Efficient ConvBN Blocks for Transfer Learning and Beyond” 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
Kaichao You, Guo Qin, Anchang Bao, Meng Cao, Ping Huang, Jiulong Shan, Mingsheng Long
Tune dynamically transforms convolution weights using tracked BatchNorm statistics during fine-tuning pretrained vision models, preserving Eval-mode behavior without permanently folding normalization into the convolution. It reduces saved feature-map memory. In a ResNet-50 experiment, Tune used about half the backbone memory and 10% less computation time than Eval. It targets Eval-mode transfer learning, not training from scratch.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.