// the find
apple-aiml-research/ml-cvnets
CVNets: A library for training computer vision networks
CVNets is Apple's PyTorch library for training vision models - classification, detection, segmentation, and a CLIP-style multimodal setup - built around their own MobileViT line of mobile-efficient architectures. It's aimed at researchers and ML engineers who want a single codebase to train or reproduce mobile-focused vision models rather than stitching together separate repos per architecture.
Genuinely broad model coverage in one consistent config-driven framework - MobileViT v1/v2, ResNet, EfficientNet, Swin, ViT, SSD, Mask R-CNN, DeepLabv3, PSPNet, CLIP, and ByteFormer all share the same training/eval engine instead of being separate forks. It ships PyTorch-to-CoreML conversion docs, which is unusual and useful given the mobile deployment angle - most research repos stop at the PyTorch checkpoint. The architectures are backed by actual published papers (MobileViT at ICLR'22, RangeAugment, ByteFormer) rather than being reimplementations of someone else's idea with no provenance.
The changelog's last real entry is July 2023 despite a 2026 push date, so it's unclear whether recent activity is new models/features or just dependency bumps - worth checking commit history before assuming active development. Setup is the classic research-code pattern: Conda plus a pinned constraints.txt, not a pip-installable package, so expect friction integrating it into an existing project. The maintainer list is small (four people, one 'previous') for a repo this wide in scope, which usually means slow issue/PR turnaround for anyone outside Apple's own use cases. There's a conftest.py but no visible CI status or test coverage info in the README, so you're trusting the benchmark numbers in the model zoo rather than seeing verified test runs.