Compute Embedding from Apple Photos, Face Tagging, and Embeddings

compute-embedding.swift

func computeEmbedding(
    face: VNFaceObservation,
    image: CGImage,
    model: MLModel
) async throws -> [Float]? {
    // 1. crop the face out of the original image
    guard let crop = cropFace(observation: face, from: image) else { return nil }

    // 2. resize the face to the model's input size
    let pixelBuffer = try makePixelBuffer(from: crop, size: 160)
    let input = try MLDictionaryFeatureProvider(dictionary: [
        "image": MLFeatureValue(pixelBuffer: pixelBuffer)
    ])

    // 3. run the model
    let prediction = try await model.prediction(from: input)
    guard let array = prediction.featureValue(for: "embedding")?.multiArrayValue else {
        return nil
    }

    // 4. normalize the output
    return l2Normalize(array)
}