Swift for Machine Learning: Opportunities and Limitations for Beginners

Introduction: Why Swift for Machine Learning?

The world of machine learning (ML) has traditionally been associated with Python, Jupyter Notebooks, and libraries like TensorFlow or PyTorch. However, in recent years, Swift — a programming language from Apple known for its performance and safety — has entered the arena. If you are an iOS/macOS developer looking to integrate ML into your applications, or a beginner choosing your first language for ML, this article is for you.

Swift offers a unique approach: it combines the speed of compiled languages (like C++) with the readability of scripting languages. Thanks to the Core ML and Create ML frameworks, as well as the Swift for TensorFlow project, Swift is becoming a viable alternative for machine learning tasks on Apple devices. However, this approach also has limitations. Let's break it down.

Key Capabilities of Swift for ML

1. Core ML: On-Device Model Integration

Core ML is Apple's framework for running machine learning models directly on devices (iPhone, iPad, Mac). This is a key advantage: data is not sent to a server, ensuring privacy and speed.

Example of loading and using a pre-trained model (e.g., MobileNet for image classification):

import CoreMLimport Vision

// Load the modelguard let model = try? VNCoreMLModel(for: MobileNetV2().model) else {    fatalError("Failed to load model")}

// Create a requestlet request = VNCoreMLRequest(model: model) { request, error in    guard let results = request.results as? [VNClassificationObservation] else { return }    // Print top-3 results    for result in results.prefix(3) {        print("\\(result.identifier): \\(result.confidence * 100)%")    }}

// Perform image analysisguard let image = UIImage(named: "photo.jpg")?.cgImage else { return }let handler = VNImageRequestHandler(cgImage: image)try? handler.perform([request])

This code shows how easy it is to integrate a powerful neural network into your iOS application. All you need is to convert the model from PyTorch/TensorFlow to the .mlmodel format using the coremltools utility.

2. Create ML: (Almost) No-Code Training

For simple tasks (text classification, image classification, tabular data), Apple offers Create ML — a tool with a graphical interface inside Xcode. You can train a model simply by dragging and dropping data. But if you prefer code, Create ML is also available as a framework:

import CreateMLimport Foundation

// Load training datalet data = try MLDataTable(contentsOf: URL(fileURLWithPath: "/path/to/train.csv"))

// Create a classifierlet classifier = try MLClassifier(trainingData: data, targetColumn: "label")

// Evaluate accuracylet metrics = classifier.trainingMetricsprint("Accuracy: \\(metrics.classificationError * 100)%")

// Save the modeltry classifier.write(to: URL(fileURLWithPath: "/path/to/model.mlmodel"))

This is an ideal option for an MVP (Minimum Viable Product) or prototyping. However, for complex architectures (transformers, GANs), Create ML is not suitable.

3. Swift for TensorFlow (S4TF): A Faded Star

The Swift for TensorFlow project was an ambitious attempt by Google to create an alternative to Python for ML. It offered automatic differentiation at the language level and integration with TensorFlow. Unfortunately, in 2021, Google ceased active development of S4TF. However, the legacy lives on: many ideas from S4TF have transitioned into the Swift Numerics and swift-ml libraries.

Example (if S4TF were still active):

import TensorFlow

// Simple neural networkstruct Model: Layer {    var layer1 = Dense<Float>(i