theFoxofSky/ddfnet

Decouples standard convolution into separate spatial and channel dynamic filters, enabling flexible filter combinations (multiplicative or additive) that unify existing weight-dynamic operations like CARAFE and depthwise filtering. Implemented as a modular CUDA operation in PyTorch that integrates with timm-based ResNet architectures, achievable by copying the ddf folder into existing projects. Pre-trained ImageNet models are provided, with the approach demonstrating competitive accuracy (79.1% top-1 on ResNet50) while maintaining parameter efficiency.

223 stars. No commits in the last 6 months.