CoreNet: A library for training deep neural networks

Original Article ↗ Hacker News Discussion ↗

Apple’s Package Management Ecosystem

  • Many are amused CoreNet’s README uses Homebrew, highlighting Apple’s lack of a native dev‑oriented package manager.
  • Debate over Homebrew vs MacPorts:
    • Homebrew seen as de facto standard with more up‑to‑date packages.
    • MacPorts preferred by some for isolated, reproducible /local stacks and richer build options.
  • Some mention nix/nix-darwin as an emerging alternative, though awareness seems uneven.
  • Question whether Apple “should Sherlock” Homebrew; some argue a package manager is now core OS functionality, others note the backlash big vendors get either way.

What CoreNet Is

  • CoreNet is described as a mid‑level training/inference framework built on PyTorch tensors.
  • It reimplements training loops, schedulers, and some optimizers, while often inheriting from PyTorch/torchvision classes.
  • Config/YAML‑driven, somewhat like fairseq/NeMo; good for standard pipelines, less suited to heavily hacking architectures.

MLX, Apple Silicon, and Performance

  • CoreNet currently trains with PyTorch but provides MLX “examples” that focus mainly on inference and model conversion.
  • Claims of MLX speedups on Apple Silicon for CLIP (e.g., significant FP16 speedups vs PyTorch), but broader benchmarks are missing.
  • Some report Apple Silicon (e.g., M2 Max) can be competitive or better than high-end NVIDIA GPUs for certain small/medium fine‑tuning tasks, but this is workload-dependent.
  • Training on Apple Silicon is viewed as practical for small‑scale or on‑device–targeted models; its value for large-scale server training is questioned.

Comparison to Other Frameworks

  • Skeptics say it’s “basically PyTorch with an Apple logo” and ask why use it over Hugging Face Transformers + MPS.
  • Supporters like its cleaner, more modular implementations of common models, metrics, and blocks, and see it as better suited for building new models than for inference.

Apple’s AI Strategy and Perception

  • Some see CoreNet and related releases (MLX, OpenELM, CatLIP mention) as Apple scrambling to catch up on AI.
  • Others argue Apple has long invested in on‑device ML (Neural Engine, CV/NLP features, accessibility) and that public perception underestimates this.
  • Debate over whether Apple has lagged in AI research output relative to its size, though recent publications and open training frameworks are noted as a shift.