Reproduce “GroupedMixer: An Entropy Model with Group-wise Token-Mixers for Learned Image Compression” 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
Daxin Li, Yuanchao Bai, Kai Wang, Junjun Jiang, Xianming Liu, Wen Gao
A transformer-based entropy model for learned image compression, GroupedMixer, achieves leading compression performance and fast coding speeds by integrating group-wise token-mixers and context cache optimization. The model provides up to 17.84% BD-Rate savings over VVC while being orders of magnitude faster than prior transformer-based methods.
Key insights
- Decomposing global causal self-attention into specialized inner-group and cross-group token-mixers effectively balances comprehensive context modeling with computational efficiency.
- Context cache optimization is critical for achieving rapid inference speeds in group-wise autoregressive models and for enabling efficient multi-step fine-tuning during training.
- Weight sharing across groups in a single transformer architecture improves parameter efficiency and scalability while effectively modeling complex spatial-channel correlations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.