Summary

This repository offers a curated collection of machine learning models converted to Apple's Core ML format, targeting iOS and macOS developers. It simplifies integrating advanced AI capabilities by providing pre-packaged models for diverse tasks, saving significant conversion and setup time.

How It Works

The project converts popular open-source ML models (TensorFlow, PyTorch) into Apple's Core ML .mlmodel format. These are distributed via Google Drive links for direct bundling into Xcode projects. Integration guidance, including Swift code snippets for Vision framework and Core GAN Container, is provided.

Quick Start & Requirements

• Installation: Download .mlmodel files from Google Drive links and add to an Xcode project.
• Prerequisites: Xcode, iOS/macOS development environment.
• Dependencies: Apple's Core ML framework; Swift integration examples provided.
• Model Size: Varies widely.
• Optimization: Supports quantization (16/8/4-bit) for reduced size/faster inference, with potential accuracy trade-offs.
• Resources: Links to Core ML Helpers and Core GAN Container.

Highlighted Details

• Extensive Model Zoo: Features models for Image Classification, Object Detection (YOLOv5-v10, YOLO-World), Segmentation (U2Net, MobileSAM, SAM2-Tiny), Super Resolution (ESRGAN, GFPGAN), Image Generation (Stable Diffusion variants), and more.
• Open-Vocabulary Detection: Includes YOLO-World for real-time detection with arbitrary text queries.
• Task Diversity: Covers a broad spectrum of computer vision and generative AI tasks.
• Conversion Scripts: Python scripts available for converting specific models.

Maintenance & Community

• Maintainer: Daisuke Majima (Freelance engineer, iOS/ML/AR).
• Contact: Email (rockyshikoku@gmail.com), GitHub, Twitter, Medium.
• Activity: Primarily a curated collection; updates depend on new model releases or conversion needs.

Licensing & Compatibility

• License: Each model inherits its original project's license (e.g., Apache 2.0, MIT, GNU, AGPL-3.0, Open RAIL-M), requiring individual review.
• Compatibility: For Core ML integration in Apple's ecosystem; some models note macOS execution. Copyleft licenses may impact closed-source commercial use.

Limitations & Caveats

• Conversion Integrity: Models are converted by a third party, potentially introducing subtle differences or requiring manual adjustments.
• License Due Diligence: Users must meticulously check and comply with each model's original license.
• Quantization Accuracy Loss: Weight quantization can degrade accuracy.
• Focus: Distribution hub, not for training/fine-tuning Core ML models.
• Hosting: Relies on Google Drive.