CoreNet is a deep neural network library designed for training a wide range of models, from small vision tasks to large-scale foundation models like LLMs and CLIP. It provides researchers and engineers with tools for standard and novel architectures, offering reproducible training recipes and pre-trained models for various applications including object classification, detection, and semantic segmentation.

How It Works

CoreNet employs a modular design, organizing models and datasets by task. Model implementations are registered via a decorator, allowing easy selection in YAML configurations. The library includes components for loss functions, metrics, optimizers, schedulers, and data handling, supporting flexible training pipelines. It also features specific examples for Apple Silicon via MLX, demonstrating efficient execution on the platform.

Quick Start & Requirements

• Installation: Requires Git LFS. Recommended Python 3.10+ (Linux) or 3.9+ (macOS) and PyTorch >= v2.1.0. Install via pip install --editable . after cloning and pulling LFS content. Optional dependencies for audio/video processing include libsox-dev and ffmpeg (Linux) or sox and ffmpeg (macOS).
• Setup: Cloning and setting up the environment is estimated to take a few minutes.
• Resources: Official documentation and tutorials are available within the repository. MLX examples for Apple Silicon are also provided.

Highlighted Details

• Supports training of foundation models like CLIP and LLMs, as well as vision tasks like object detection and segmentation.
• Includes numerous research publications from Apple that utilize CoreNet, with associated training recipes and pre-trained models.
• Features MLX examples for efficient model execution on Apple Silicon hardware.
• Evolved from CVNets, expanding its scope to include language modeling and multi-modal applications.

Maintenance & Community

Maintained by Maxwell Horton, Mohammad Sekhavat, Yanzi Jin, and Dmitry Belenko. Community contributions are welcomed via pull requests, adhering to a provided Code of Conduct.

Licensing & Compatibility

The license is available in the LICENSE file. No specific license type is mentioned in the README, but compatibility for commercial use or closed-source linking is not detailed.

Limitations & Caveats

The README does not explicitly state the license type, which could impact commercial adoption. Case sensitivity issues on macOS file systems are noted, requiring specific commands (cd $(pwd -P)) to navigate the repository correctly.