Adding Core ML to an iOS App

This year at WWDC Apple announced Core ML, a new foundational machine learning framework that delivers easy integration with Apple products. After attending the Introducing Core ML session, I was excited to fire up the Xcode 9 beta and start seeing what Core ML could do. During the session the presenter showed how simple it is to use an existing model with Core ML, and I had to see for myself.

1. Core ML Tools - Converting the Model

Along side the Core ML release, Apple also released Core ML Tools, a Python tool for converting existing models from Keras, Caffe, SciKit-learn, or libSVM formats, to the Core ML format.

Installing the tool is as easy as using pip

pip install -U coremltools

Luckily enough, I was able to find a Keras model on Github that did what I wanted my app to do, detect clickbait headlines.

With the model in hand, converting it to the Core ML format is as easy as two lines of Python code.

# The Keras model we start with
model = ConvolutionalNet(vocabulary_size=len(vocabulary),
    embedding_dimension=EMBEDDING_DIMENSION,
    input_length=SEQUENCE_LENGTH)
model.load_weights("../models/detector.h5")

# Necessary line 1, creating the Core ML Model
coreml_model = coremltools.converters.keras.convert(model,
    ["headline"],
    "clickbaityness")

# Not necessary, but recommended descriptive text
coreml_model.author = 'Saurabh Mathur, Core ML conversion by Mike Caulley'
coreml_model.license = 'GNU GPL'
coreml_model.short_description = 'Article headline clickbait detector.'
coreml_model.input_description['headline'] = 'An article headline as a tokenized array.'
coreml_model.output_description['clickbaityness'] = 'Probability that the headline is clickbait.'

# Necessary line 2, saving the model
coreml_model.save('clickbait_model.mlmodel')

Full code for the model and the conversion script is located here.

Once the script is run, you are left with a .mlmodel file that can be brought into Xcode.

2. Bringing to Model to Xcode

Keeping with the theme of easy, bringing the model into Xcode only takes a drag and a drop of the file into the Xcode window.

When you drag and drop a couple things happen.

• The model file appears in the Project Navigator with the descriptive information you added during the conversion with Core ML Tools
• An API is generated for using the model.

Here is what a function consuming the API might look like.

static func calculateClickBaitPercentageFor(_ headline: String) -> Double? {
    let cleanedHeadlineArray = headline.clean.words
    let indices = Vocab.wordsToIndices(words: cleanedHeadlineArray)
    let input = Vocab.padSequence(sequence: indices, maxLength: 20)
    let model = clickbait_model()
    let data = input.reduce([], +)
    
    guard let mlMultiArray = try? MLMultiArray(shape:[20,1,1],
                                               dataType:MLMultiArrayDataType.int32) else {
                                                fatalError("Unexpected runtime error. MLMultiArray")
    }
    
    for (index, element) in data.enumerated() {
        mlMultiArray[index] = NSNumber(integerLiteral: element)
    }
    
    if let result = try? model.prediction(headline: mlMultiArray) {
        return result.clickbaityness[0] as? Double
    } else {
        return nil
    }
}

A few things to notice.

• The model.prediction function is one of the functions that gets generated when the Core ML file is dragged into Xcode.
• The input type for Core ML prediction function is a MLMultiArray
• The input to the Keras model is an array of 20 word mappings, so our MLMultiArray size is [20,1,1]

And that quickly we are able to have machine learning in our iOS app!

If you would like to check out the full code, it is available here.