xmu-xiaoma666/external-attention-pytorch

This repository, called FightingCV, is a collection of PyTorch implementations covering dozens of attention mechanisms and related neural network building blocks drawn from recent computer vision research papers. The goal is to help researchers and developers understand how these techniques work by providing clean, standalone code for each method, separate from the larger training frameworks that paper authors usually publish. The library is organized into four main categories. The first is an attention series with over 37 different mechanisms, including self-attention, external attention, squeeze-and-excitation, CBAM, coordinate attention, and many others from published papers. The second is a backbone series covering popular architectures such as ResNet, MobileViT, ConvNeXt, and various vision transformer variants. The third is an MLP series with implementations of MLP-Mixer, ResMLP, and similar structures. The fourth covers convolution techniques including depthwise separable convolution, dynamic convolution, and involution. Installation is available via pip using the fightingcv-attention package, or you can clone the repository directly. Usage follows a consistent pattern: import the module you want, create an instance with your chosen parameters, and pass a tensor through it. The README provides a short code snippet for each method showing the expected input shape and how to initialize the module. The project is primarily aimed at researchers who want to prototype with these building blocks or study paper implementations without digging through task-specific codebases. Documentation is written in Chinese, with a companion English README in the repository. The full README is longer than what was shown.