# Vision Framework와 CoreML: 온디바이스 ML iOS 면접 질문

> Vision Framework와 CoreML의 핵심 면접 질문으로 iOS 면접을 준비할 수 있습니다. 이미지 인식, 객체 감지, 온디바이스 ML을 다룹니다.

- Published: 2026-03-07
- Updated: 2026-04-29
- Author: SharpSkill
- Tags: vision, coreml, ios, machine-learning, interview
- Reading time: 14 min

---

온디바이스 머신러닝은 현대 iOS 애플리케이션에 중요한 경쟁 우위를 제공합니다. Vision Framework와 CoreML을 통해 모델을 디바이스에서 직접 실행할 수 있어 데이터 프라이버시와 실시간 성능이 보장됩니다. 이 면접 질문들은 모든 시니어 iOS 개발자가 숙지해야 할 핵심 개념을 다룹니다.

> **가이드 구성**
>
> 질문은 주제별로 정리되어 있습니다. CoreML 기초, Vision Framework, 성능 최적화, 실전 사례입니다. 각 답변에는 최신 Swift 코드와 자세한 설명이 포함되어 있습니다.

## CoreML 기초

### 1. CoreML이란 무엇이며 어떤 장점이 있습니까?

CoreML은 iOS, macOS, watchOS, tvOS 애플리케이션에 머신러닝 모델을 통합하기 위한 Apple의 프레임워크입니다. Apple 하드웨어(CPU, GPU, Neural Engine)에 맞춰 모델을 자동으로 최적화하고, 네트워크 연결 없이 온디바이스 실행을 보장합니다.

주요 장점에는 데이터 프라이버시(데이터가 디바이스를 벗어나지 않음), 낮은 지연 시간(네트워크 왕복 없음), Apple Silicon 칩의 Neural Engine을 위한 자동 최적화가 포함됩니다.

```swift
// CoreMLBasics.swift
import CoreML

// Loading a compiled CoreML model (.mlmodelc)
class ImageClassifier {
    // Model is compiled at build time to optimize loading
    private let model: VNCoreMLModel

    init() throws {
        // Configuration to use Neural Engine if available
        let config = MLModelConfiguration()
        config.computeUnits = .all  // CPU + GPU + Neural Engine

        // Load model with custom configuration
        let mlModel = try MobileNetV2(configuration: config).model
        model = try VNCoreMLModel(for: mlModel)
    }

    // Method to classify an image
    func classify(image: CGImage) async throws -> [(String, Float)] {
        // Create Vision request with CoreML model
        let request = VNCoreMLRequest(model: model)
        request.imageCropAndScaleOption = .centerCrop

        // Handler to process the image
        let handler = VNImageRequestHandler(cgImage: image, options: [:])
        try handler.perform([request])

        // Extract results
        guard let results = request.results as? [VNClassificationObservation] else {
            return []
        }

        // Return top 5 predictions with confidence
        return results.prefix(5).map { ($0.identifier, $0.confidence) }
    }
}
```

### 2. TensorFlow나 PyTorch 모델을 CoreML로 어떻게 변환합니까?

변환에는 Apple의 공식 Python 패키지인 coremltools를 사용합니다. TensorFlow, PyTorch, ONNX 및 기타 인기 있는 형식을 지원합니다. 변환에는 모델 크기를 줄이기 위한 양자화 같은 최적화가 포함될 수 있습니다.

```python
# convert_model.py
import coremltools as ct
import torch

# Conversion from PyTorch
class MyClassifier(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = torch.nn.Conv2d(3, 64, 3)
        self.fc = torch.nn.Linear(64, 10)

    def forward(self, x):
        x = self.conv(x)
        x = x.mean([2, 3])  # Global average pooling
        return self.fc(x)

# Example input for tracing
example_input = torch.rand(1, 3, 224, 224)

# Trace the PyTorch model
traced_model = torch.jit.trace(MyClassifier(), example_input)

# Convert to CoreML with metadata
mlmodel = ct.convert(
    traced_model,
    inputs=[ct.ImageType(name="image", shape=(1, 3, 224, 224))],
    classifier_config=ct.ClassifierConfig(["cat", "dog", "bird"]),
    minimum_deployment_target=ct.target.iOS17
)

# Save model with compression
mlmodel.save("MyClassifier.mlpackage")
```

`.mlpackage` 모델은 Xcode 프로젝트에 직접 추가할 수 있으며, Xcode가 자동으로 타입이 지정된 Swift 클래스를 생성합니다.

### 3. MLModel과 VNCoreMLModel의 차이점은 무엇입니까?

`MLModel`은 ML 모델을 로드하고 실행하기 위한 CoreML의 기본 클래스입니다. `VNCoreMLModel`은 CoreML 모델을 Vision Framework와 함께 사용할 수 있도록 해주는 래퍼이며, 자동 이미지 전처리와 Vision 파이프라인 통합을 제공합니다.

```swift
// MLModelVsVNCoreML.swift
import CoreML
import Vision

// Direct MLModel usage (low level)
func predictWithMLModel(features: MLFeatureProvider) async throws -> String {
    let config = MLModelConfiguration()
    let model = try MyModel(configuration: config)

    // Direct prediction with feature provider
    let prediction = try model.prediction(from: features)

    // Manual output access
    guard let output = prediction.featureValue(for: "classLabel")?.stringValue else {
        throw PredictionError.invalidOutput
    }
    return output
}

// Usage with VNCoreMLModel (high level, recommended for images)
func predictWithVision(image: CGImage) async throws -> [VNClassificationObservation] {
    let config = MLModelConfiguration()
    let mlModel = try MyModel(configuration: config).model

    // Wrapper for use with Vision
    let visionModel = try VNCoreMLModel(for: mlModel)

    // Vision automatically handles resizing and preprocessing
    let request = VNCoreMLRequest(model: visionModel)
    request.imageCropAndScaleOption = .scaleFill

    let handler = VNImageRequestHandler(cgImage: image)
    try handler.perform([request])

    return request.results as? [VNClassificationObservation] ?? []
}
```

> **어떤 것을 언제 사용해야 합니까?**
>
> 표 형식 데이터나 이미지가 아닌 입력에는 `MLModel`을 직접 사용합니다. 이미지와 관련된 모든 작업에는 `VNCoreMLModel`을 사용합니다. Vision이 형식 변환과 전처리를 자동으로 처리하기 때문입니다.

### 4. CoreML에서 다양한 iOS 버전을 어떻게 처리합니까?

CoreML은 iOS 버전마다 발전합니다. 변환 시 최소 deployment target을 정의하고, 구버전에서 사용할 수 없는 기능을 처리하는 것이 중요합니다.

```swift
// CoreMLVersioning.swift
import CoreML

class AdaptiveMLManager {
    // Check model capabilities based on iOS version
    func loadOptimalModel() throws -> MLModel {
        let config = MLModelConfiguration()

        // iOS 17+: Optimized Neural Engine with compute budget
        if #available(iOS 17, *) {
            config.computeUnits = .cpuAndNeuralEngine
            // New in iOS 17: compute power limit
            config.allowLowPrecisionAccumulationOnGPU = true
            return try AdvancedModel(configuration: config).model
        }
        // iOS 16: Enhanced GPU support
        else if #available(iOS 16, *) {
            config.computeUnits = .all
            return try StandardModel(configuration: config).model
        }
        // iOS 15: CPU only fallback for reliability
        else {
            config.computeUnits = .cpuOnly
            return try LegacyModel(configuration: config).model
        }
    }

    // Check if Neural Engine is available
    var hasNeuralEngine: Bool {
        if #available(iOS 16, *) {
            // Devices with A11+ have Neural Engine
            var sysinfo = utsname()
            uname(&sysinfo)
            let machine = String(bytes: Data(bytes: &sysinfo.machine,
                count: Int(_SYS_NAMELEN)), encoding: .ascii)?
                .trimmingCharacters(in: .controlCharacters) ?? ""

            // iPhone X and later have Neural Engine
            return machine.contains("iPhone10") ||
                   machine.hasPrefix("iPhone1") && machine.count > 7
        }
        return false
    }
}
```

## Vision Framework

### 5. Vision Framework는 어떤 종류의 요청을 지원합니까?

Vision Framework는 이미지 분석을 위한 다양한 요청을 제공합니다. 주요 카테고리로는 얼굴 감지, 텍스트 인식(OCR), 객체 감지, 비디오에서의 객체 추적, 이미지 유사도 분석이 있습니다.

```swift
// VisionRequests.swift
import Vision

class VisionAnalyzer {
    // Face detection with landmarks
    func detectFaces(in image: CGImage) async throws -> [VNFaceObservation] {
        let request = VNDetectFaceLandmarksRequest()
        request.revision = VNDetectFaceLandmarksRequestRevision3

        let handler = VNImageRequestHandler(cgImage: image)
        try handler.perform([request])

        return request.results ?? []
    }

    // Text recognition (OCR)
    func recognizeText(in image: CGImage) async throws -> [String] {
        let request = VNRecognizeTextRequest()
        request.recognitionLevel = .accurate  // .fast for real-time
        request.recognitionLanguages = ["en-US", "fr-FR"]
        request.usesLanguageCorrection = true

        let handler = VNImageRequestHandler(cgImage: image)
        try handler.perform([request])

        return request.results?.compactMap { observation in
            observation.topCandidates(1).first?.string
        } ?? []
    }

    // Object detection and classification
    func detectObjects(in image: CGImage) async throws -> [VNRecognizedObjectObservation] {
        // Use a CoreML model for detection
        let config = MLModelConfiguration()
        let detector = try YOLOv8(configuration: config)
        let visionModel = try VNCoreMLModel(for: detector.model)

        let request = VNCoreMLRequest(model: visionModel)
        request.imageCropAndScaleOption = .scaleFill

        let handler = VNImageRequestHandler(cgImage: image)
        try handler.perform([request])

        return request.results as? [VNRecognizedObjectObservation] ?? []
    }

    // Compute similarity between images
    func computeSimilarity(image1: CGImage, image2: CGImage) async throws -> Float {
        // Generate feature prints for both images
        let request = VNGenerateImageFeaturePrintRequest()

        let handler1 = VNImageRequestHandler(cgImage: image1)
        try handler1.perform([request])
        guard let print1 = request.results?.first else { throw VisionError.noResults }

        let handler2 = VNImageRequestHandler(cgImage: image2)
        try handler2.perform([request])
        guard let print2 = request.results?.first else { throw VisionError.noResults }

        // Compute distance between embeddings
        var distance: Float = 0
        try print1.computeDistance(&distance, to: print2)

        // Convert distance to similarity score (0-1)
        return 1.0 / (1.0 + distance)
    }
}
```

### 6. Vision으로 실시간 객체 추적을 어떻게 구현합니까?

객체 추적은 감지된 객체를 비디오 프레임을 통해 따라가기 위해 `VNTrackObjectRequest`를 사용합니다. 감지 observation으로 초기화하고, 이후 프레임은 동일한 요청을 사용해 추적합니다.

```swift
// ObjectTracking.swift
import Vision
import AVFoundation

class ObjectTracker: NSObject {
    private var trackingRequest: VNTrackObjectRequest?
    private let sequenceHandler = VNSequenceRequestHandler()

    // Callback to notify position updates
    var onTrackingUpdate: ((CGRect) -> Void)?
    var onTrackingLost: (() -> Void)?

    // Initialize tracking with an initial detection
    func startTracking(observation: VNDetectedObjectObservation) {
        // Create tracking request from observation
        trackingRequest = VNTrackObjectRequest(
            detectedObjectObservation: observation
        ) { [weak self] request, error in
            self?.handleTrackingResult(request: request, error: error)
        }

        // Configure tracking
        trackingRequest?.trackingLevel = .accurate  // .fast for 60fps
    }

    // Process each new video frame
    func processFrame(_ pixelBuffer: CVPixelBuffer) {
        guard let request = trackingRequest else { return }

        do {
            // Sequence handler maintains context between frames
            try sequenceHandler.perform([request], on: pixelBuffer)
        } catch {
            onTrackingLost?()
            trackingRequest = nil
        }
    }

    private func handleTrackingResult(request: VNRequest, error: Error?) {
        guard let result = request.results?.first as? VNDetectedObjectObservation else {
            onTrackingLost?()
            return
        }

        // Check tracking confidence
        if result.confidence < 0.3 {
            onTrackingLost?()
            trackingRequest = nil
            return
        }

        // Update request for next frame
        trackingRequest = VNTrackObjectRequest(detectedObjectObservation: result) {
            [weak self] request, error in
            self?.handleTrackingResult(request: request, error: error)
        }

        // Notify new position (normalized coordinates)
        DispatchQueue.main.async { [weak self] in
            self?.onTrackingUpdate?(result.boundingBox)
        }
    }
}

// Integration with AVCaptureSession
extension ObjectTracker: AVCaptureVideoDataOutputSampleBufferDelegate {
    func captureOutput(
        _ output: AVCaptureOutput,
        didOutput sampleBuffer: CMSampleBuffer,
        from connection: AVCaptureConnection
    ) {
        guard let pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer) else {
            return
        }
        processFrame(pixelBuffer)
    }
}
```

### 7. 실시간 처리를 위해 Vision 성능을 어떻게 최적화합니까?

최적화에는 여러 가지 기법이 포함됩니다. 적절한 인식 수준 사용, 전용 큐에서의 프레임 처리, 동시 요청 수 제한입니다. 정확도와 속도 간의 선택은 사용 사례에 따라 달라집니다.

```swift
// VisionOptimization.swift
import Vision
import AVFoundation

class OptimizedVisionPipeline {
    // Dedicated queue for Vision processing (avoids main thread)
    private let processingQueue = DispatchQueue(
        label: "com.app.vision",
        qos: .userInteractive,
        attributes: .concurrent
    )

    // Limit number of simultaneously processed frames
    private let semaphore = DispatchSemaphore(value: 2)

    // Reuse requests to avoid allocations
    private lazy var textRequest: VNRecognizeTextRequest = {
        let request = VNRecognizeTextRequest()
        request.recognitionLevel = .fast  // .accurate if precision > speed
        request.usesLanguageCorrection = false  // Disable for +20% perf
        request.minimumTextHeight = 0.05  // Ignore text too small
        return request
    }()

    // Reuse sequence handler for tracking
    private let sequenceHandler = VNSequenceRequestHandler()

    // Optimized frame processing
    func processFrame(_ pixelBuffer: CVPixelBuffer) {
        // Skip if pipeline is saturated
        guard semaphore.wait(timeout: .now()) == .success else {
            return  // Drop frame rather than block
        }

        processingQueue.async { [weak self] in
            defer { self?.semaphore.signal() }

            guard let self = self else { return }

            do {
                // Use sequence handler for better performance
                try self.sequenceHandler.perform(
                    [self.textRequest],
                    on: pixelBuffer,
                    orientation: .up
                )

                // Process results
                if let results = self.textRequest.results {
                    self.handleResults(results)
                }
            } catch {
                print("Vision error: \(error)")
            }
        }
    }

    // Batch processing for static images
    func processImages(_ images: [CGImage]) async throws -> [[VNObservation]] {
        // Parallel processing with TaskGroup
        try await withThrowingTaskGroup(of: (Int, [VNObservation]).self) { group in
            for (index, image) in images.enumerated() {
                group.addTask {
                    let handler = VNImageRequestHandler(cgImage: image)
                    let request = VNDetectFaceRectanglesRequest()
                    try handler.perform([request])
                    return (index, request.results ?? [])
                }
            }

            // Collect results in original order
            var results = [[VNObservation]](repeating: [], count: images.count)
            for try await (index, observations) in group {
                results[index] = observations
            }
            return results
        }
    }

    private func handleResults(_ results: [VNRecognizedTextObservation]) {
        // Async processing of results
    }
}
```

### 8. Vision으로 사람 자세 감지를 어떻게 구현합니까?

Vision Framework iOS 14+는 신체 관절을 감지하기 위해 `VNDetectHumanBodyPoseRequest`를 제공합니다. 이 기능은 피트니스 앱, AR 게임, 동작 분석에 사용됩니다.

```swift
// PoseDetection.swift
import Vision

struct DetectedPose {
    let joints: [VNHumanBodyPoseObservation.JointName: CGPoint]
    let confidence: Float

    // Calculate angle between three joints
    func angleBetween(
        _ joint1: VNHumanBodyPoseObservation.JointName,
        _ joint2: VNHumanBodyPoseObservation.JointName,
        _ joint3: VNHumanBodyPoseObservation.JointName
    ) -> Double? {
        guard let p1 = joints[joint1],
              let p2 = joints[joint2],
              let p3 = joints[joint3] else { return nil }

        let v1 = CGVector(dx: p1.x - p2.x, dy: p1.y - p2.y)
        let v2 = CGVector(dx: p3.x - p2.x, dy: p3.y - p2.y)

        let dot = v1.dx * v2.dx + v1.dy * v2.dy
        let mag1 = sqrt(v1.dx * v1.dx + v1.dy * v1.dy)
        let mag2 = sqrt(v2.dx * v2.dx + v2.dy * v2.dy)

        return acos(dot / (mag1 * mag2)) * 180 / .pi
    }
}

class PoseDetector {
    private let request = VNDetectHumanBodyPoseRequest()

    func detectPose(in image: CGImage) async throws -> DetectedPose? {
        let handler = VNImageRequestHandler(cgImage: image)
        try handler.perform([request])

        guard let observation = request.results?.first else { return nil }

        // Extract all detected joints
        var joints: [VNHumanBodyPoseObservation.JointName: CGPoint] = [:]

        // List of main joints
        let jointNames: [VNHumanBodyPoseObservation.JointName] = [
            .nose, .neck,
            .leftShoulder, .rightShoulder,
            .leftElbow, .rightElbow,
            .leftWrist, .rightWrist,
            .leftHip, .rightHip,
            .leftKnee, .rightKnee,
            .leftAnkle, .rightAnkle
        ]

        for jointName in jointNames {
            if let point = try? observation.recognizedPoint(jointName),
               point.confidence > 0.3 {
                // Convert normalized coordinates to points
                joints[jointName] = CGPoint(x: point.x, y: point.y)
            }
        }

        return DetectedPose(
            joints: joints,
            confidence: observation.confidence
        )
    }

    // Detect if person is doing a squat
    func isSquatting(pose: DetectedPose) -> Bool {
        guard let kneeAngle = pose.angleBetween(
            .leftHip, .leftKnee, .leftAnkle
        ) else { return false }

        // A squat typically has knee angle < 100°
        return kneeAngle < 100
    }
}
```

## 최적화와 운영

### 9. CoreML 모델을 양자화하여 크기를 줄이는 방법은?

양자화는 모델 크기를 줄이고 추론 속도를 높이기 위해 가중치의 정밀도를 낮춥니다(Float32에서 Float16 또는 Int8로). 그 대가로 약간의 정확도 손실이 발생합니다.

```python
# quantize_model.py
import coremltools as ct
from coremltools.models.neural_network import quantization_utils

# Load existing model
model = ct.models.MLModel("MyModel.mlpackage")

# Float16 quantization (recommended, good size/precision balance)
model_fp16 = ct.models.neural_network.quantization_utils.quantize_weights(
    model,
    nbits=16,
    quantization_mode="linear"
)
model_fp16.save("MyModel_FP16.mlpackage")

# Int8 quantization (smallest size, possible precision loss)
# Requires calibration dataset for best results
def calibration_data():
    import numpy as np
    for _ in range(100):
        yield {"image": np.random.rand(1, 3, 224, 224).astype(np.float32)}

model_int8 = ct.compression_utils.affine_quantize_weights(
    model,
    mode="linear_symmetric",
    dtype=ct.converters.mil.mil.types.int8
)
model_int8.save("MyModel_INT8.mlpackage")
```

```swift
// QuantizationComparison.swift
import CoreML

class ModelBenchmark {
    // Compare performance of different versions
    func benchmark() async throws {
        let configs: [(String, URL)] = [
            ("Full Precision", Bundle.main.url(forResource: "Model", withExtension: "mlmodelc")!),
            ("Float16", Bundle.main.url(forResource: "Model_FP16", withExtension: "mlmodelc")!),
            ("Int8", Bundle.main.url(forResource: "Model_INT8", withExtension: "mlmodelc")!)
        ]

        for (name, url) in configs {
            let model = try MLModel(contentsOf: url)

            // Measure average inference time over 100 iterations
            let startTime = CFAbsoluteTimeGetCurrent()
            for _ in 0..<100 {
                let input = try prepareInput()
                _ = try model.prediction(from: input)
            }
            let elapsed = CFAbsoluteTimeGetCurrent() - startTime

            // Model size
            let size = try FileManager.default.attributesOfItem(atPath: url.path)[.size] as? Int ?? 0

            print("\(name): \(elapsed/100*1000)ms/inference, \(size/1024/1024)MB")
        }
    }

    private func prepareInput() throws -> MLFeatureProvider {
        // Prepare test input
        fatalError("Implement based on model requirements")
    }
}
```

### 10. 큰 이미지를 처리할 때 메모리를 어떻게 관리합니까?

고해상도 이미지를 처리하면 메모리 사용량이 급증할 수 있습니다. 기법으로는 스마트 다운샘플링, 타일 기반 처리, 능동적인 리소스 해제가 있습니다.

```swift
// MemoryOptimization.swift
import Vision
import CoreImage

class MemoryEfficientProcessor {
    // Reusable CoreImage context to avoid allocations
    private let ciContext = CIContext(options: [
        .useSoftwareRenderer: false,
        .cacheIntermediates: false  // Reduces memory usage
    ])

    // Smart downsampling of large images
    func downsampleImage(at url: URL, to maxDimension: CGFloat) -> CGImage? {
        // Options for downsampling at read time (avoids loading full image)
        let options: [CFString: Any] = [
            kCGImageSourceCreateThumbnailFromImageAlways: true,
            kCGImageSourceThumbnailMaxPixelSize: maxDimension,
            kCGImageSourceCreateThumbnailWithTransform: true,
            kCGImageSourceShouldCacheImmediately: false
        ]

        guard let source = CGImageSourceCreateWithURL(url as CFURL, nil),
              let image = CGImageSourceCreateThumbnailAtIndex(source, 0, options as CFDictionary) else {
            return nil
        }

        return image
    }

    // Tile processing for very large images
    func processByTiles(
        image: CGImage,
        tileSize: CGSize,
        processor: (CGImage) throws -> [VNObservation]
    ) throws -> [VNObservation] {
        var allObservations: [VNObservation] = []

        let imageWidth = CGFloat(image.width)
        let imageHeight = CGFloat(image.height)

        // Iterate through image by tiles
        var y: CGFloat = 0
        while y < imageHeight {
            var x: CGFloat = 0
            while x < imageWidth {
                // Calculate tile rectangle
                let tileRect = CGRect(
                    x: x, y: y,
                    width: min(tileSize.width, imageWidth - x),
                    height: min(tileSize.height, imageHeight - y)
                )

                // Extract tile
                autoreleasepool {
                    if let tile = image.cropping(to: tileRect) {
                        do {
                            let observations = try processor(tile)

                            // Adjust coordinates relative to full image
                            let adjusted = observations.compactMap { obs -> VNObservation? in
                                guard let detected = obs as? VNDetectedObjectObservation else {
                                    return obs
                                }
                                // Recalculate bounding box in global coordinates
                                var box = detected.boundingBox
                                box.origin.x = (box.origin.x * tileRect.width + x) / imageWidth
                                box.origin.y = (box.origin.y * tileRect.height + y) / imageHeight
                                box.size.width = box.size.width * tileRect.width / imageWidth
                                box.size.height = box.size.height * tileRect.height / imageHeight

                                return detected
                            }
                            allObservations.append(contentsOf: adjusted)
                        } catch {
                            print("Tile processing error: \(error)")
                        }
                    }
                }

                x += tileSize.width * 0.9  // 10% overlap to avoid cutting objects
            }
            y += tileSize.height * 0.9
        }

        return allObservations
    }
}
```

> **메모리 누수 주의**
>
> 이미지 처리 루프에서는 항상 `autoreleasepool`을 사용하고, Vision 요청 클로저의 retain cycle을 확인해야 합니다.

### 11. Create ML Components로 ML 파이프라인을 어떻게 구현합니까?

Create ML Components(iOS 16+)는 사전 정의된 트랜스포머로 모듈식 ML 파이프라인을 구축할 수 있게 해줍니다. 전통적인 단일 모델보다 더 유연합니다.

```swift
// CreateMLComponents.swift
import CreateMLComponents
import CoreImage

@available(iOS 16.0, *)
class MLPipeline {
    // Image classification pipeline with preprocessing
    func createImageClassificationPipeline() throws -> some Transformer<CGImage, String> {
        // Transformer composition
        let pipeline = ImageReader()
            .appending(ImageScaler(targetSize: .init(width: 224, height: 224)))
            .appending(ImageNormalizer(mean: [0.485, 0.456, 0.406],
                                       std: [0.229, 0.224, 0.225]))
            .appending(try ImageFeaturePrint())
            .appending(try NearestNeighborClassifier<String>
                .load(from: trainingDataURL))

        return pipeline
    }

    // Custom pipeline with custom steps
    func createCustomPipeline() -> some Transformer<CIImage, AnalysisResult> {
        // Step 1: Preprocessing
        let preprocess = CIImageTransformer { image in
            // Apply CoreImage filters
            let adjusted = image
                .applyingFilter("CIColorControls", parameters: [
                    kCIInputContrastKey: 1.2,
                    kCIInputSaturationKey: 1.1
                ])
            return adjusted
        }

        // Step 2: Detection
        let detect = VisionTransformer<CIImage, [VNFaceObservation]> { image in
            let request = VNDetectFaceRectanglesRequest()
            let handler = VNImageRequestHandler(ciImage: image)
            try handler.perform([request])
            return request.results ?? []
        }

        // Step 3: Analysis
        let analyze = ResultTransformer<[VNFaceObservation], AnalysisResult> { faces in
            AnalysisResult(
                faceCount: faces.count,
                averageConfidence: faces.map(\.confidence).reduce(0, +) / Float(faces.count)
            )
        }

        return preprocess
            .appending(detect)
            .appending(analyze)
    }
}

struct AnalysisResult {
    let faceCount: Int
    let averageConfidence: Float
}
```

### 12. CoreML 모델을 어떻게 테스트하고 검증합니까?

테스트에는 정확도 검증, 성능 테스트, 통합 테스트가 포함됩니다. 다양한 디바이스와 조건에서 테스트하는 것이 매우 중요합니다.

```swift
// MLModelTests.swift
import XCTest
import CoreML
import Vision

class CoreMLModelTests: XCTestCase {
    var model: VNCoreMLModel!

    override func setUpWithError() throws {
        let config = MLModelConfiguration()
        config.computeUnits = .cpuOnly  // Reproducible on CI
        let mlModel = try MyClassifier(configuration: config).model
        model = try VNCoreMLModel(for: mlModel)
    }

    // Accuracy test with validation dataset
    func testClassificationAccuracy() async throws {
        let testCases: [(imageName: String, expectedClass: String)] = [
            ("cat_001", "cat"),
            ("dog_001", "dog"),
            ("bird_001", "bird")
        ]

        var correct = 0
        for testCase in testCases {
            let image = try loadTestImage(named: testCase.imageName)
            let prediction = try await classify(image: image)

            if prediction == testCase.expectedClass {
                correct += 1
            }
        }

        let accuracy = Double(correct) / Double(testCases.count)
        XCTAssertGreaterThan(accuracy, 0.95, "Accuracy should be > 95%")
    }

    // Performance test (inference time)
    func testInferencePerformance() throws {
        let image = try loadTestImage(named: "test_image")

        measure(metrics: [XCTClockMetric(), XCTMemoryMetric()]) {
            let request = VNCoreMLRequest(model: model)
            let handler = VNImageRequestHandler(cgImage: image)
            try? handler.perform([request])
        }
    }

    // Transformation robustness test
    func testRobustness() async throws {
        let originalImage = try loadTestImage(named: "cat_001")
        let originalPrediction = try await classify(image: originalImage)

        // Test with rotation
        let rotated = try applyTransform(originalImage, rotation: .pi / 6)
        let rotatedPrediction = try await classify(image: rotated)
        XCTAssertEqual(originalPrediction, rotatedPrediction)

        // Test with noise
        let noisy = try addNoise(to: originalImage, intensity: 0.1)
        let noisyPrediction = try await classify(image: noisy)
        XCTAssertEqual(originalPrediction, noisyPrediction)
    }

    // Edge case handling test
    func testEdgeCases() async throws {
        // Very small image
        let smallImage = try loadTestImage(named: "tiny_10x10")
        let smallResult = try await classify(image: smallImage)
        XCTAssertNotNil(smallResult)

        // Monochrome image
        let monoImage = try loadTestImage(named: "grayscale")
        let monoResult = try await classify(image: monoImage)
        XCTAssertNotNil(monoResult)
    }

    // Helpers
    private func classify(image: CGImage) async throws -> String {
        let request = VNCoreMLRequest(model: model)
        let handler = VNImageRequestHandler(cgImage: image)
        try handler.perform([request])

        guard let results = request.results as? [VNClassificationObservation],
              let top = results.first else {
            throw TestError.noResults
        }

        return top.identifier
    }

    private func loadTestImage(named: String) throws -> CGImage {
        guard let url = Bundle(for: type(of: self))
                .url(forResource: named, withExtension: "jpg"),
              let source = CGImageSourceCreateWithURL(url as CFURL, nil),
              let image = CGImageSourceCreateImageAtIndex(source, 0, nil) else {
            throw TestError.imageNotFound
        }
        return image
    }
}
```

## 시스템 디자인 질문

### 13. 운영 환경 앱을 위한 온디바이스 ML 아키텍처를 어떻게 설계합니까?

견고한 ML 아키텍처는 책임을 분리합니다. 모델, 전처리, 후처리, 캐싱입니다. 모델 업데이트와 점진적 폴백을 관리해야 합니다.

```swift
// MLArchitecture.swift
import CoreML
import Vision

// Protocol for model abstraction
protocol MLModelProvider {
    associatedtype Input
    associatedtype Output

    func predict(_ input: Input) async throws -> Output
    var modelVersion: String { get }
}

// Model manager with OTA updates
class ModelManager {
    static let shared = ModelManager()

    private var models: [String: any MLModel] = [:]
    private let modelDirectory: URL

    private init() {
        modelDirectory = FileManager.default.urls(for: .applicationSupportDirectory, in: .userDomainMask)[0]
            .appendingPathComponent("MLModels")
        try? FileManager.default.createDirectory(at: modelDirectory, withIntermediateDirectories: true)
    }

    // Load model with fallback to bundled version
    func loadModel<T: MLModel>(
        named name: String,
        type: T.Type
    ) async throws -> T {
        // Check if downloaded version exists
        let downloadedURL = modelDirectory.appendingPathComponent("\(name).mlmodelc")

        if FileManager.default.fileExists(atPath: downloadedURL.path) {
            // Validate downloaded model integrity
            do {
                let model = try await loadAndValidate(from: downloadedURL, type: type)
                return model
            } catch {
                // Fallback to bundled version if corrupted
                print("Downloaded model corrupted, falling back to bundled version")
                try? FileManager.default.removeItem(at: downloadedURL)
            }
        }

        // Load bundled version
        guard let bundledURL = Bundle.main.url(forResource: name, withExtension: "mlmodelc") else {
            throw ModelError.modelNotFound(name)
        }

        return try await loadAndValidate(from: bundledURL, type: type)
    }

    // Download and install new model version
    func updateModel(named name: String, from url: URL) async throws {
        // Download model
        let (tempURL, _) = try await URLSession.shared.download(from: url)

        // Compile model if needed
        let compiledURL: URL
        if tempURL.pathExtension == "mlmodel" {
            compiledURL = try MLModel.compileModel(at: tempURL)
        } else {
            compiledURL = tempURL
        }

        // Validate before installation
        let config = MLModelConfiguration()
        _ = try MLModel(contentsOf: compiledURL, configuration: config)

        // Install in models directory
        let destURL = modelDirectory.appendingPathComponent("\(name).mlmodelc")
        try? FileManager.default.removeItem(at: destURL)
        try FileManager.default.moveItem(at: compiledURL, to: destURL)

        // Notify app of update
        NotificationCenter.default.post(name: .modelUpdated, object: name)
    }

    private func loadAndValidate<T: MLModel>(
        from url: URL,
        type: T.Type
    ) async throws -> T {
        let config = MLModelConfiguration()
        config.computeUnits = .all

        let model = try T(contentsOf: url, configuration: config)

        // Basic model validation
        // Verify inputs/outputs match expectations

        return model
    }
}

extension Notification.Name {
    static let modelUpdated = Notification.Name("MLModelUpdated")
}
```

### 14. 운영 환경에서 오류와 모니터링을 어떻게 처리합니까?

견고한 모니터링 시스템은 성능 지표와 오류를 수집하고 원격 디버깅을 가능하게 합니다. 분석 도구와의 통합은 필수적입니다.

```swift
// MLMonitoring.swift
import OSLog

class MLMonitor {
    static let shared = MLMonitor()

    private let logger = Logger(subsystem: "com.app.ml", category: "inference")
    private var metrics: [InferenceMetric] = []

    struct InferenceMetric: Codable {
        let modelName: String
        let inferenceTime: Double
        let inputSize: CGSize?
        let confidence: Float?
        let timestamp: Date
        let success: Bool
        let errorDescription: String?
    }

    // Record an inference
    func recordInference(
        model: String,
        duration: TimeInterval,
        inputSize: CGSize? = nil,
        confidence: Float? = nil,
        error: Error? = nil
    ) {
        let metric = InferenceMetric(
            modelName: model,
            inferenceTime: duration,
            inputSize: inputSize,
            confidence: confidence,
            timestamp: Date(),
            success: error == nil,
            errorDescription: error?.localizedDescription
        )

        metrics.append(metric)

        // Log for debugging
        if let error = error {
            logger.error("ML inference failed: \(model) - \(error.localizedDescription)")
        } else {
            logger.info("ML inference: \(model) completed in \(duration)s")
        }

        // Detect anomalies
        checkForAnomalies(metric)
    }

    // Wrapper for automatic measurement
    func measure<T>(
        model: String,
        inputSize: CGSize? = nil,
        operation: () async throws -> T
    ) async rethrows -> T {
        let start = CFAbsoluteTimeGetCurrent()

        do {
            let result = try await operation()
            let duration = CFAbsoluteTimeGetCurrent() - start

            recordInference(
                model: model,
                duration: duration,
                inputSize: inputSize
            )

            return result
        } catch {
            let duration = CFAbsoluteTimeGetCurrent() - start

            recordInference(
                model: model,
                duration: duration,
                inputSize: inputSize,
                error: error
            )

            throw error
        }
    }

    // Detect performance issues
    private func checkForAnomalies(_ metric: InferenceMetric) {
        // Alert if inference time exceeds threshold
        if metric.inferenceTime > 1.0 {
            logger.warning("Slow inference detected: \(metric.modelName) took \(metric.inferenceTime)s")

            // Send alert if available
            Task {
                await AnalyticsService.shared.reportAnomaly(
                    type: .slowInference,
                    details: metric
                )
            }
        }

        // Alert if confidence is too low
        if let confidence = metric.confidence, confidence < 0.5 {
            logger.info("Low confidence prediction: \(confidence) for \(metric.modelName)")
        }
    }

    // Generate performance report
    func generateReport() -> PerformanceReport {
        let recentMetrics = metrics.filter {
            $0.timestamp > Date().addingTimeInterval(-3600)  // Last hour
        }

        let avgInferenceTime = recentMetrics.map(\.inferenceTime).reduce(0, +) / Double(recentMetrics.count)
        let successRate = Double(recentMetrics.filter(\.success).count) / Double(recentMetrics.count)

        return PerformanceReport(
            totalInferences: recentMetrics.count,
            averageInferenceTime: avgInferenceTime,
            successRate: successRate,
            modelBreakdown: Dictionary(grouping: recentMetrics, by: \.modelName)
        )
    }
}

struct PerformanceReport {
    let totalInferences: Int
    let averageInferenceTime: Double
    let successRate: Double
    let modelBreakdown: [String: [MLMonitor.InferenceMetric]]
}
```

## 결론

Vision Framework와 CoreML은 iOS에서 온디바이스 머신러닝의 토대를 이룹니다. 이러한 기술을 익히는 것은 사용자의 프라이버시를 존중하면서 고급 ML 기능을 제공하는 현대적인 애플리케이션을 개발하는 데 필수적입니다.

### 점검 체크리스트

- ✅ CoreML과 그 장점(프라이버시, 지연 시간, 오프라인) 이해
- ✅ TensorFlow/PyTorch 모델을 CoreML로 변환할 수 있음
- ✅ Vision 요청(얼굴 감지, OCR, 분류) 숙달
- ✅ 실시간 객체 추적 구현
- ✅ 성능 최적화(양자화, 메모리 관리)
- ✅ 운영 환경을 위한 견고한 ML 아키텍처 설계
- ✅ 모니터링 및 오류 처리 구축

### 핵심 포인트

온디바이스 성능은 CPU, GPU, Neural Engine 중 어느 것을 선택하느냐에 크게 좌우됩니다. 모델 양자화는 크기와 성능 사이의 훌륭한 절충안을 제공합니다. 운영 환경에서의 모니터링은 회귀를 감지하는 데 필수적입니다.

---

Source: SharpSkill (https://sharpskill.dev), tech interview preparation for your real stack.
HTML version of this page: https://sharpskill.dev/ko/blog/ios/vision-coreml-interview-on-device-ml-ios