Use a custom TensorFlow Lite model on Android

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.

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

1. If you haven't already, add Firebase to your Android project.
2. In your module (app-level) Gradle file (usually <project>/<app-module>/build.gradle.kts or <project>/<app-module>/build.gradle), add the dependency for the Firebase ML model downloader Android library. We recommend using the Firebase Android BoM to control library versioning.

   Also, as part of setting up Firebase ML model downloader, you need to add the TensorFlow Lite SDK to your app.

   Kotlin+KTX

   dependencies {
       // Import the BoM for the Firebase platform
       implementation(platform("com.google.firebase:firebase-bom:32.3.1"))
   
       // Add the dependency for the Firebase ML model downloader library
       // When using the BoM, you don't specify versions in Firebase library dependencies
       implementation("com.google.firebase:firebase-ml-modeldownloader-ktx")
   
       // Also add the dependency for the TensorFlow Lite library and specify its version
       implementation("org.tensorflow:tensorflow-lite:2.3.0")
   }
   

   By using the Firebase Android BoM, your app will always use compatible versions of Firebase Android libraries.

   (Alternative) Add Firebase library dependencies without using the BoM

   If you choose not to use the Firebase BoM, you must specify each Firebase library version in its dependency line.

   Note that if you use multiple Firebase libraries in your app, we strongly recommend using the BoM to manage library versions, which ensures that all versions are compatible.

   dependencies {
       // Add the dependency for the Firebase ML model downloader library
       // When NOT using the BoM, you must specify versions in Firebase library dependencies
       implementation("com.google.firebase:firebase-ml-modeldownloader-ktx:24.1.3")
   
       // Also add the dependency for the TensorFlow Lite library and specify its version
       implementation("org.tensorflow:tensorflow-lite:2.3.0")
   }
   

   Java

   dependencies {
       // Import the BoM for the Firebase platform
       implementation(platform("com.google.firebase:firebase-bom:32.3.1"))
   
       // Add the dependency for the Firebase ML model downloader library
       // When using the BoM, you don't specify versions in Firebase library dependencies
       implementation("com.google.firebase:firebase-ml-modeldownloader")
   
       // Also add the dependency for the TensorFlow Lite library and specify its version
       implementation("org.tensorflow:tensorflow-lite:2.3.0")
   }
   

   By using the Firebase Android BoM, your app will always use compatible versions of Firebase Android libraries.

   (Alternative) Add Firebase library dependencies without using the BoM

   If you choose not to use the Firebase BoM, you must specify each Firebase library version in its dependency line.

   Note that if you use multiple Firebase libraries in your app, we strongly recommend using the BoM to manage library versions, which ensures that all versions are compatible.

   dependencies {
       // Add the dependency for the Firebase ML model downloader library
       // When NOT using the BoM, you must specify versions in Firebase library dependencies
       implementation("com.google.firebase:firebase-ml-modeldownloader:24.1.3")
   
       // Also add the dependency for the TensorFlow Lite library and specify its version
       implementation("org.tensorflow:tensorflow-lite:2.3.0")
   }
   

3. In your app's manifest, declare that INTERNET permission is required:

   <uses-permission android:name="android.permission.INTERNET" />

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.

val conditions = CustomModelDownloadConditions.Builder()
        .requireWifi()  // Also possible: .requireCharging() and .requireDeviceIdle()
        .build()
FirebaseModelDownloader.getInstance()
        .getModel("your_model", DownloadType.LOCAL_MODEL_UPDATE_IN_BACKGROUND,
            conditions)
        .addOnSuccessListener { model: CustomModel? ->
            // 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.
            val modelFile = model?.file
            if (modelFile != null) {
                interpreter = Interpreter(modelFile)
            }
        }

CustomModelDownloadConditions conditions = new CustomModelDownloadConditions.Builder()
    .requireWifi()  // Also possible: .requireCharging() and .requireDeviceIdle()
    .build();
FirebaseModelDownloader.getInstance()
    .getModel("your_model", DownloadType.LOCAL_MODEL_UPDATE_IN_BACKGROUND, conditions)
    .addOnSuccessListener(new OnSuccessListener<CustomModel>() {
      @Override
      public void onSuccess(CustomModel model) {
        // 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.
        File modelFile = model.getFile();
        if (modelFile != null) {
            interpreter = new Interpreter(modelFile);
        }
      }
    });

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 you have an image classification model with an input shape of [1 224 224 3] floating-point values, you could generate an input ByteBuffer from a Bitmap object as shown in the following example:

val bitmap = Bitmap.createScaledBitmap(yourInputImage, 224, 224, true)
val input = ByteBuffer.allocateDirect(224*224*3*4).order(ByteOrder.nativeOrder())
for (y in 0 until 224) {
    for (x in 0 until 224) {
        val px = bitmap.getPixel(x, y)

        // Get channel values from the pixel value.
        val r = Color.red(px)
        val g = Color.green(px)
        val b = Color.blue(px)

        // Normalize channel values to [-1.0, 1.0]. This requirement depends on the model.
        // For example, some models might require values to be normalized to the range
        // [0.0, 1.0] instead.
        val rf = (r - 127) / 255f
        val gf = (g - 127) / 255f
        val bf = (b - 127) / 255f

        input.putFloat(rf)
        input.putFloat(gf)
        input.putFloat(bf)
    }
}

Bitmap bitmap = Bitmap.createScaledBitmap(yourInputImage, 224, 224, true);
ByteBuffer input = ByteBuffer.allocateDirect(224 * 224 * 3 * 4).order(ByteOrder.nativeOrder());
for (int y = 0; y < 224; y++) {
    for (int x = 0; x < 224; x++) {
        int px = bitmap.getPixel(x, y);

        // Get channel values from the pixel value.
        int r = Color.red(px);
        int g = Color.green(px);
        int b = Color.blue(px);

        // Normalize channel values to [-1.0, 1.0]. This requirement depends
        // on the model. For example, some models might require values to be
        // normalized to the range [0.0, 1.0] instead.
        float rf = (r - 127) / 255.0f;
        float gf = (g - 127) / 255.0f;
        float bf = (b - 127) / 255.0f;

        input.putFloat(rf);
        input.putFloat(gf);
        input.putFloat(bf);
    }
}

Then, allocate a ByteBuffer large enough to contain the model's output and pass the input buffer and output buffer to the TensorFlow Lite interpreter's run() method. For example, for an output shape of [1 1000] floating-point values:

val bufferSize = 1000 * java.lang.Float.SIZE / java.lang.Byte.SIZE
val modelOutput = ByteBuffer.allocateDirect(bufferSize).order(ByteOrder.nativeOrder())
interpreter?.run(input, modelOutput)

int bufferSize = 1000 * java.lang.Float.SIZE / java.lang.Byte.SIZE;
ByteBuffer modelOutput = ByteBuffer.allocateDirect(bufferSize).order(ByteOrder.nativeOrder());
interpreter.run(input, modelOutput);

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:

modelOutput.rewind()
val probabilities = modelOutput.asFloatBuffer()
try {
    val reader = BufferedReader(
            InputStreamReader(assets.open("custom_labels.txt")))
    for (i in probabilities.capacity()) {
        val label: String = reader.readLine()
        val probability = probabilities.get(i)
        println("$label: $probability")
    }
} catch (e: IOException) {
    // File not found?
}

modelOutput.rewind();
FloatBuffer probabilities = modelOutput.asFloatBuffer();
try {
    BufferedReader reader = new BufferedReader(
            new InputStreamReader(getAssets().open("custom_labels.txt")));
    for (int i = 0; i < probabilities.capacity(); i++) {
        String label = reader.readLine();
        float probability = probabilities.get(i);
        Log.i(TAG, String.format("%s: %1.4f", label, probability));
    }
} catch (IOException e) {
    // File not found?
}

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.

On Android API level 21 (Lollipop) and newer, the model is downloaded to a directory that is excluded from automatic backup.

On Android API level 20 and older, the model is downloaded to a directory named com.google.firebase.ml.custom.models in app-private internal storage. If you enabled file backup using BackupAgent, you might choose to exclude this directory.