Reproduce “Iwin Transformer: Hierarchical Vision Transformer using Interleaved Windows” 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
Simin Huo, Ning Li
The Iwin Transformer combines interleaved-window attention with depthwise convolution: it rearranges patches so local windows connect distant regions, restores their order, and uses convolution to link neighbors and supply implicit positional information. For ImageNet-22K-pretrained Iwin-L fine-tuned on ImageNet-1K, top-1 accuracy rises from 86.4% at 224² to 87.4% at 384². This is fine-tuning, not zero-shot transfer, and remains below ConvNeXt-L’s 87.5%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.