CoreNet is a deep neural network toolkit for training standard and novel models across tasks including foundation models, object classification, object detection, and semantic segmentation.
The toolkit addresses the need for a flexible, production-ready framework to train models at various scales. It provides researchers and engineers with a unified codebase that supports both small and large-scale neural networks, enabling experimentation with novel architectures and training approaches. The library includes built-in support for multiple computer vision and language tasks, with pre-trained models and training recipes available for reference implementations.
Developers should adopt CoreNet if they are training vision models, language models, or multimodal foundation models and want access to Apple's research implementations. The toolkit suits projects requiring reproducible training pipelines with published recipes, particularly those building on architectures like MobileViT, FastVit, or OpenELM. The project includes training and evaluation recipes alongside pre-trained model weights, making it valuable for teams seeking to replicate or extend published research. CoreNet evolved from an earlier library called CVNets, which focused specifically on computer vision; the newer toolkit expands scope to include language models and multimodal systems.
Development activity shows consistent research output with multiple peer-reviewed publications demonstrating the toolkit's application to problems including efficient language models, vision-language models, mobile architectures, and data augmentation techniques. The project maintains organized documentation with a projects folder containing detailed recipes and model links for each published work. Installation requirements specify Python and PyTorch versions with optional dependencies for audio and video processing, indicating active maintenance of compatibility with current deep learning infrastructure. The codebase uses Jupyter Notebooks as a primary language, suggesting an emphasis on interactive documentation and reproducible examples alongside the core library code.
