Use a custom TensorFlow Lite model on Apple platforms

If your app uses custom TensorFlow Lite models, you can use Firebase ML to deploy your models. By deploying models with Firebase, you can reduce the initial download size of your app and update your app's ML models without releasing a new version of your app. And, with Remote Config and A/B Testing, you can dynamically serve different models to different sets of users.

Prerequisites

• The MLModelDownloader library is only available for Swift.
• TensorFlow Lite runs only on devices using iOS 9 and newer.

TensorFlow Lite models

TensorFlow Lite models are ML models that are optimized to run on mobile devices. To get a TensorFlow Lite model:

• Use a pre-built model, such as one of the official TensorFlow Lite models.
• Convert a TensorFlow model, Keras model, or concrete function to TensorFlow Lite.

Before you begin

To use TensorFlowLite with Firebase, you must use CocoaPods as TensorFlowLite currently does not support installation with Swift Package Manager. See the CocoaPods installation guide for instructions on how to install MLModelDownloader.

Once installed, import Firebase and TensorFlowLite in order to use them.

import FirebaseMLModelDownloader
import TensorFlowLite

1. Deploy your model

Deploy your custom TensorFlow models using either the Firebase console or the Firebase Admin Python and Node.js SDKs. See Deploy and manage custom models.

After you add a custom model to your Firebase project, you can reference the model in your apps using the name you specified. At any time, you can deploy a new TensorFlow Lite model and download the new model onto users' devices by calling getModel() (see below).

2. Download the model to the device and initialize a TensorFlow Lite interpreter

To start the model download, call the model downloader's getModel() method, specifying the name you assigned the model when you uploaded it, whether you want to always download the latest model, and the conditions under which you want to allow downloading.

You can choose from three download behaviors:

You should disable model-related functionality—for example, grey-out or hide part of your UI—until you confirm the model has been downloaded.

let conditions = ModelDownloadConditions(allowsCellularAccess: false)
ModelDownloader.modelDownloader()
    .getModel(name: "your_model",
              downloadType: .localModelUpdateInBackground,
              conditions: conditions) { result in
        switch (result) {
        case .success(let customModel):
            do {
                // Download complete. Depending on your app, you could enable the ML
                // feature, or switch from the local model to the remote model, etc.

                // The CustomModel object contains the local path of the model file,
                // which you can use to instantiate a TensorFlow Lite interpreter.
                let interpreter = try Interpreter(modelPath: customModel.path)
            } catch {
                // Error. Bad model file?
            }
        case .failure(let error):
            // Download was unsuccessful. Don't enable ML features.
            print(error)
        }
}

Many apps start the download task in their initialization code, but you can do so at any point before you need to use the model.

3. Perform inference on input data

Get your model's input and output shapes

The TensorFlow Lite model interpreter takes as input and produces as output one or more multidimensional arrays. These arrays contain either byte, int, long, or float values. Before you can pass data to a model or use its result, you must know the number and dimensions ("shape") of the arrays your model uses.

If you built the model yourself, or if the model's input and output format is documented, you might already have this information. If you don't know the shape and data type of your model's input and output, you can use the TensorFlow Lite interpreter to inspect your model. For example:

import tensorflow as tf

interpreter = tf.lite.Interpreter(model_path="your_model.tflite")
interpreter.allocate_tensors()

# Print input shape and type
inputs = interpreter.get_input_details()
print('{} input(s):'.format(len(inputs)))
for i in range(0, len(inputs)):
    print('{} {}'.format(inputs[i]['shape'], inputs[i]['dtype']))

# Print output shape and type
outputs = interpreter.get_output_details()
print('\n{} output(s):'.format(len(outputs)))
for i in range(0, len(outputs)):
    print('{} {}'.format(outputs[i]['shape'], outputs[i]['dtype']))

1 input(s):
[  1 224 224   3] <class 'numpy.float32'>

1 output(s):
[1 1000] <class 'numpy.float32'>

Run the interpreter

For example, if your model processes images, and your model has input dimensions of [1, 224, 224, 3] floating-point values, you might have to scale the image's color values to a floating-point range as in the following example:

let image: CGImage = // Your input image
guard let context = CGContext(
  data: nil,
  width: image.width, height: image.height,
  bitsPerComponent: 8, bytesPerRow: image.width * 4,
  space: CGColorSpaceCreateDeviceRGB(),
  bitmapInfo: CGImageAlphaInfo.noneSkipFirst.rawValue
) else {
  return false
}

context.draw(image, in: CGRect(x: 0, y: 0, width: image.width, height: image.height))
guard let imageData = context.data else { return false }

var inputData = Data()
for row in 0 ..< 224 {
  for col in 0 ..< 224 {
    let offset = 4 * (row * context.width + col)
    // (Ignore offset 0, the unused alpha channel)
    let red = imageData.load(fromByteOffset: offset+1, as: UInt8.self)
    let green = imageData.load(fromByteOffset: offset+2, as: UInt8.self)
    let blue = imageData.load(fromByteOffset: offset+3, as: UInt8.self)

    // Normalize channel values to [0.0, 1.0]. This requirement varies
    // by model. For example, some models might require values to be
    // normalized to the range [-1.0, 1.0] instead, and others might
    // require fixed-point values or the original bytes.
    var normalizedRed = Float32(red) / 255.0
    var normalizedGreen = Float32(green) / 255.0
    var normalizedBlue = Float32(blue) / 255.0

    // Append normalized values to Data object in RGB order.
    let elementSize = MemoryLayout.size(ofValue: normalizedRed)
    var bytes = [UInt8](repeating: 0, count: elementSize)
    memcpy(&bytes, &normalizedRed, elementSize)
    inputData.append(&bytes, count: elementSize)
    memcpy(&bytes, &normalizedGreen, elementSize)
    inputData.append(&bytes, count: elementSize)
    memcpy(&ammp;bytes, &normalizedBlue, elementSize)
    inputData.append(&bytes, count: elementSize)
  }
}

Then, copy your input NSData to the interpreter and run it:

try interpreter.allocateTensors()
try interpreter.copy(inputData, toInputAt: 0)
try interpreter.invoke()

You can get the model's output by calling the interpreter's output(at:) method. How you use the output depends on the model you are using.

For example, if you are performing classification, as a next step, you might map the indexes of the result to the labels they represent:

let output = try interpreter.output(at: 0)
let probabilities =
        UnsafeMutableBufferPointer<Float32>.allocate(capacity: 1000)
output.data.copyBytes(to: probabilities)

guard let labelPath = Bundle.main.path(forResource: "retrained_labels", ofType: "txt") else { return }
let fileContents = try? String(contentsOfFile: labelPath)
guard let labels = fileContents?.components(separatedBy: "\n") else { return }

for i in labels.indices {
    print("\(labels[i]): \(probabilities[i])")
}

Appendix: Model security

Regardless of how you make your TensorFlow Lite models available to Firebase ML, Firebase ML stores them in the standard serialized protobuf format in local storage.

In theory, this means that anybody can copy your model. However, in practice, most models are so application-specific and obfuscated by optimizations that the risk is similar to that of competitors disassembling and reusing your code. Nevertheless, you should be aware of this risk before you use a custom model in your app.