Download
ResNeXt is a deep neural network architecture for image classification built on the idea of aggregated residual transformations. Instead of simply increasing depth or width, ResNeXt introduces a new dimension called cardinality, which refers to the number of parallel transformation paths (i.e. the number of “branches”) that are aggregated together. Each branch is a small transformation (e.g. bottleneck block) and their outputs are summed—this enables richer representation without excessive parameter blowup. The design is modular and homogeneous, making it relatively easy to scale (by tuning cardinality, width, depth) and adopt in existing residual frameworks. The official repository offers a Torch (Lua) implementation with code for training, evaluation, and pretrained models on ImageNet. In practice, ResNeXt models often outperform standard ResNet models of comparable complexity.
Features
- Aggregated residual transformations combining multiple parallel branches
- Introduces “cardinality” as a new architectural dimension
- Modular bottleneck blocks with easy scaling across width/depth/cardinality
- Torch implementation with training and evaluation scripts
- Pretrained models for ImageNet classification
- Compatibility with residual architectures and straightforward integration
Project Samples
