从旧版自定义模型 API 迁移

firebase-ml-model-interpreter 库的 22.0.2 版引入了新的 getLatestModelFile() 方法,该方法可获取设备上自定义模型的位置。您可以使用此方法直接实例化 TensorFlow Lite Interpreter 对象,该对象可用于取代 FirebaseModelInterpreter 封装容器。

今后,我们推荐您采用这种方法。由于 TensorFlow Lite 解释器版本不再与 Firebase 库版本结合使用,因此您可以根据需要更灵活地升级到新版 TensorFlow Lite,或者更轻松地使用自定义 TensorFlow Lite 构建。

本页面介绍了如何从使用 FirebaseModelInterpreter 迁移到 TensorFlow Lite Interpreter

1.更新项目依赖项

更新项目的依赖项,以包含 firebase-ml-model-interpreter 库的 22.0.2 版(或更高版本)以及 tensorflow-lite 库:

旧版

implementation("com.google.firebase:firebase-ml-model-interpreter:22.0.1")

之后

implementation("com.google.firebase:firebase-ml-model-interpreter:22.0.2")
implementation("org.tensorflow:tensorflow-lite:2.0.0")

2. 创建 TensorFlow Lite 解释器而不是 FirebaseModelInterpreter

使用 getLatestModelFile() 获取设备上模型的位置并使用它创建 TensorFlow Lite Interpreter,而不是创建 FirebaseModelInterpreter

旧版

Kotlin+KTX

val remoteModel = FirebaseCustomRemoteModel.Builder("your_model").build()
val options = FirebaseModelInterpreterOptions.Builder(remoteModel).build()
val interpreter = FirebaseModelInterpreter.getInstance(options)

Java

FirebaseCustomRemoteModel remoteModel =
        new FirebaseCustomRemoteModel.Builder("your_model").build();
FirebaseModelInterpreterOptions options =
        new FirebaseModelInterpreterOptions.Builder(remoteModel).build();
FirebaseModelInterpreter interpreter = FirebaseModelInterpreter.getInstance(options);

新版

Kotlin+KTX

val remoteModel = FirebaseCustomRemoteModel.Builder("your_model").build()
FirebaseModelManager.getInstance().getLatestModelFile(remoteModel)
    .addOnCompleteListener { task ->
        val modelFile = task.getResult()
        if (modelFile != null) {
            // Instantiate an org.tensorflow.lite.Interpreter object.
            interpreter = Interpreter(modelFile)
        }
    }

Java

FirebaseCustomRemoteModel remoteModel =
        new FirebaseCustomRemoteModel.Builder("your_model").build();
FirebaseModelManager.getInstance().getLatestModelFile(remoteModel)
        .addOnCompleteListener(new OnCompleteListener<File>() {
            @Override
            public void onComplete(@NonNull Task<File> task) {
                File modelFile = task.getResult();
                if (modelFile != null) {
                    // Instantiate an org.tensorflow.lite.Interpreter object.
                    Interpreter interpreter = new Interpreter(modelFile);
                }
            }
        });

3.更新输入和输出准备代码

通过 FirebaseModelInterpreter,您可以在运行解释器时将 FirebaseModelInputOutputOptions 对象传递给解释器,从而指定模型的输入和输出形状。

对于 TensorFlow Lite 解释器,您需要改为以适合模型输入和输出的大小分配 ByteBuffer 对象。

例如,如果模型的输入形状为 [1 224 224 3] float 值且输出形状为 [1 1000] float 值,请进行以下更改:

旧版

Kotlin+KTX

val inputOutputOptions = FirebaseModelInputOutputOptions.Builder()
    .setInputFormat(0, FirebaseModelDataType.FLOAT32, intArrayOf(1, 224, 224, 3))
    .setOutputFormat(0, FirebaseModelDataType.FLOAT32, intArrayOf(1, 1000))
    .build()

val input = ByteBuffer.allocateDirect(224*224*3*4).order(ByteOrder.nativeOrder())
// Then populate with input data.

val inputs = FirebaseModelInputs.Builder()
    .add(input)
    .build()

interpreter.run(inputs, inputOutputOptions)
    .addOnSuccessListener { outputs ->
        // ...
    }
    .addOnFailureListener {
        // Task failed with an exception.
        // ...
    }

Java

FirebaseModelInputOutputOptions inputOutputOptions =
        new FirebaseModelInputOutputOptions.Builder()
                .setInputFormat(0, FirebaseModelDataType.FLOAT32, new int[]{1, 224, 224, 3})
                .setOutputFormat(0, FirebaseModelDataType.FLOAT32, new int[]{1, 1000})
                .build();

float[][][][] input = new float[1][224][224][3];
// Then populate with input data.

FirebaseModelInputs inputs = new FirebaseModelInputs.Builder()
        .add(input)
        .build();

interpreter.run(inputs, inputOutputOptions)
        .addOnSuccessListener(
                new OnSuccessListener<FirebaseModelOutputs>() {
                    @Override
                    public void onSuccess(FirebaseModelOutputs result) {
                        // ...
                    }
                })
        .addOnFailureListener(
                new OnFailureListener() {
                    @Override
                    public void onFailure(@NonNull Exception e) {
                        // Task failed with an exception
                        // ...
                    }
                });

新版

Kotlin+KTX

val inBufferSize = 1 * 224 * 224 * 3 * java.lang.Float.SIZE / java.lang.Byte.SIZE
val inputBuffer = ByteBuffer.allocateDirect(inBufferSize).order(ByteOrder.nativeOrder())
// Then populate with input data.

val outBufferSize = 1 * 1000 * java.lang.Float.SIZE / java.lang.Byte.SIZE
val outputBuffer = ByteBuffer.allocateDirect(outBufferSize).order(ByteOrder.nativeOrder())

interpreter.run(inputBuffer, outputBuffer)

Java

int inBufferSize = 1 * 224 * 224 * 3 * java.lang.Float.SIZE / java.lang.Byte.SIZE;
ByteBuffer inputBuffer =
        ByteBuffer.allocateDirect(inBufferSize).order(ByteOrder.nativeOrder());
// Then populate with input data.

int outBufferSize = 1 * 1000 * java.lang.Float.SIZE / java.lang.Byte.SIZE;
ByteBuffer outputBuffer =
        ByteBuffer.allocateDirect(outBufferSize).order(ByteOrder.nativeOrder());

interpreter.run(inputBuffer, outputBuffer);

4.更新输出处理代码

最后,不要使用 FirebaseModelOutputs 对象的 getOutput() 方法获取模型的输出,而是将 ByteBuffer 输出转换为对您的使用场景而言很方便的任何结构。

例如,如果您要执行分类,则可以进行如下更改:

旧版

Kotlin+KTX

val output = result.getOutput(0)
val probabilities = output[0]
try {
    val reader = BufferedReader(InputStreamReader(assets.open("custom_labels.txt")))
    for (probability in probabilities) {
        val label: String = reader.readLine()
        println("$label: $probability")
    }
} catch (e: IOException) {
    // File not found?
}

Java

float[][] output = result.getOutput(0);
float[] probabilities = output[0];
try {
    BufferedReader reader = new BufferedReader(
          new InputStreamReader(getAssets().open("custom_labels.txt")));
    for (float probability : probabilities) {
        String label = reader.readLine();
        Log.i(TAG, String.format("%s: %1.4f", label, probability));
    }
} catch (IOException e) {
    // File not found?
}

更新后

Kotlin+KTX

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?
}

Java

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?
}