apple/corenet

What it solves

CoreNet is a deep neural network toolkit designed to simplify the training of both standard and novel AI models. It provides a unified framework for researchers and engineers to build and scale models for a wide range of tasks, from small-scale specialized models to large-scale foundation models like LLMs and CLIP.

How it works

The library is organized around a registry-based system where models and datasets are categorized by task (e.g., classification, detection, language modeling). Users can define their training and evaluation pipelines using YAML configuration files, which reference specific model classes, datasets, and optimizers. It supports multi-node training and integrates with MLX for efficient execution on Apple Silicon.

Who it’s for

It is primarily intended for AI researchers and engineers who need a reproducible way to train deep neural networks across different modalities (text, image, audio, and video).

Highlights

• Broad Task Support: Supports object classification, object detection, semantic segmentation, and language modeling.
• Reading-Ready Recipes: Includes reproducible training recipes and pre-trained weights for numerous Apple research publications (e.g., OpenELM, CatLIP, FastVit).
• Multi-Modal Capabilities: Provides specialized data loaders and encoders for audio, video, and text.
• Apple Silicon Optimization: Includes MLX examples for running models efficiently on Mac hardware.