mlboydaisuke/AdaFace-IR18-CoreML

AdaFace IR-18 — Core ML

CVPR 2022

Face recognition embeddings. A 112x112 aligned face goes in, a 512-dim L2-normalised embedding comes out; compare two faces with a cosine similarity. AdaFace weights the margin by image quality during training, so low-quality and blurry faces degrade gracefully instead of collapsing together.

iResNet-18 backbone trained on CASIA-WebFace.

Core ML conversion of mk-minchul/AdaFace for on-device inference on iPhone, iPad and Mac. Converted with coremltools; the packages are stateless, so all sequencing and buffering lives in your Swift code.

Task	image feature extraction
Upstream	mk-minchul/AdaFace
Packages	1
Download size	42 MB
Minimum iOS	17.0

Files

File	Size	Compute units	SHA-256
AdaFace_IR18.mlpackage.zip	42 MB	all	-
Total	42 MB

compute_units is not a suggestion -- it is the configuration the conversion was verified against. Moving a package to a different compute unit can silently change the numerics (FP16 attention overflow) or crash on the GPU.

Download

hf download mlboydaisuke/coreml-zoo --include "adaface/*" --local-dir ./adaface
unzip './adaface/adaface/*.zip' -d ./adaface

Use in Swift

import CoreML

let config = MLModelConfiguration()
config.computeUnits = .all   // as converted — see the table above

// Unzip the .mlpackage, drop it into your Xcode target and Xcode compiles it
// at build time:
let model = try AdaFace_IR18(configuration: config)

// ...or compile a downloaded .mlpackage at runtime:
let compiled = try await MLModel.compileModel(at: mlpackageURL)
let model = try MLModel(contentsOf: compiled, configuration: config)

Demo

• Sample app — sample_apps/AdaFaceDemo, a standalone SwiftUI project.
• Models Zoo — this model is downloadable and runnable inside the Models Zoo app on the App Store, no build required.

Conversion

• Script: convert_adaface.py
• Pitfalls hit during conversion (FP16 overflow, ANE buffer limits, stride handling): docs/coreml_conversion_notes.md
• Model index: CoreML-Models

License

The conversion inherits the upstream license: MIT.

Credits

• Upstream authors: mk-minchul/AdaFace, 2022
• Core ML conversion: john-rocky (Daisuke Majima)