Process input and output data with the LiteRT Support Library

Mobile application developers typically interact with typed objects such as bitmaps or primitives such as integers. However, the LiteRT interpreter API that runs the on-device machine learning model uses tensors in the form of ByteBuffer, which can be difficult to debug and manipulate. The LiteRT Android Support Library is designed to help process the input and output of LiteRT models, and make the LiteRT interpreter easier to use.

Getting Started

Import Gradle dependency and other settings

Copy the .tflite model file to the assets directory of the Android module where the model will be run. Specify that the file should not be compressed, and add the LiteRT library to the module’s build.gradle file:

android {
    // Other settings

    // Specify tflite file should not be compressed for the app apk
    aaptOptions {
        noCompress "tflite"
    }

}

dependencies {
    // Other dependencies

    // Import tflite dependencies
    implementation 'org.tensorflow:tensorflow-lite:0.0.0-nightly-SNAPSHOT'
    // The GPU delegate library is optional. Depend on it as needed.
    implementation 'com.google.ai.edge.litert:litert-gpu:0.0.0-nightly-SNAPSHOT'
    implementation 'com.google.ai.edge.litert:litert-support:0.0.0-nightly-SNAPSHOT'
}

Explore the LiteRT Support Library AAR hosted at MavenCentral for different versions of the Support Library.

Basic image manipulation and conversion

The LiteRT Support Library has a suite of basic image manipulation methods such as crop and resize. To use it, create an ImagePreprocessor and add the required operations. To convert the image into the tensor format required by the LiteRT interpreter, create a TensorImage to be used as input:

import org.tensorflow.lite.DataType;
import org.tensorflow.lite.support.image.ImageProcessor;
import org.tensorflow.lite.support.image.TensorImage;
import org.tensorflow.lite.support.image.ops.ResizeOp;

// Initialization code
// Create an ImageProcessor with all ops required. For more ops, please
// refer to the ImageProcessor Architecture section in this README.
ImageProcessor imageProcessor =
    new ImageProcessor.Builder()
        .add(new ResizeOp(224, 224, ResizeOp.ResizeMethod.BILINEAR))
        .build();

// Create a TensorImage object. This creates the tensor of the corresponding
// tensor type (uint8 in this case) that the LiteRT interpreter needs.
TensorImage tensorImage = new TensorImage(DataType.UINT8);

// Analysis code for every frame
// Preprocess the image
tensorImage.load(bitmap);
tensorImage = imageProcessor.process(tensorImage);

DataType of a tensor can be read through the metadata extractor library as well as other model information.

Basic audio data processing

The LiteRT Support Library also defines a TensorAudio class wrapping some basic audio data processing methods. It's mostly used together with AudioRecord and captures audio samples in a ring buffer.

import android.media.AudioRecord;
import org.tensorflow.lite.support.audio.TensorAudio;

// Create an `AudioRecord` instance.
AudioRecord record = AudioRecord(...)

// Create a `TensorAudio` object from Android AudioFormat.
TensorAudio tensorAudio = new TensorAudio(record.getFormat(), size)

// Load all audio samples available in the AudioRecord without blocking.
tensorAudio.load(record)

// Get the `TensorBuffer` for inference.
TensorBuffer buffer = tensorAudio.getTensorBuffer()

Create output objects and run the model

Before running the model, we need to create the container objects that will store the result:

import org.tensorflow.lite.DataType;
import org.tensorflow.lite.support.tensorbuffer.TensorBuffer;

// Create a container for the result and specify that this is a quantized model.
// Hence, the 'DataType' is defined as UINT8 (8-bit unsigned integer)
TensorBuffer probabilityBuffer =
    TensorBuffer.createFixedSize(new int[]{1, 1001}, DataType.UINT8);

Loading the model and running inference:

import java.nio.MappedByteBuffer;
import org.tensorflow.lite.InterpreterFactory;
import org.tensorflow.lite.InterpreterApi;

// Initialise the model
try{
    MappedByteBuffer tfliteModel
        = FileUtil.loadMappedFile(activity,
            "mobilenet_v1_1.0_224_quant.tflite");
    InterpreterApi tflite = new InterpreterFactory().create(
        tfliteModel, new InterpreterApi.Options());
} catch (IOException e){
    Log.e("tfliteSupport", "Error reading model", e);
}

// Running inference
if(null != tflite) {
    tflite.run(tImage.getBuffer(), probabilityBuffer.getBuffer());
}

Accessing the result

Developers can access the output directly through probabilityBuffer.getFloatArray(). If the model produces a quantized output, remember to convert the result. For the MobileNet quantized model, the developer needs to divide each output value by 255 to obtain the probability ranging from 0 (least likely) to 1 (most likely) for each category.

Optional: Mapping results to labels

Developers can also optionally map the results to labels. First, copy the text file containing labels into the module's assets directory. Next, load the label file using the following code:

import org.tensorflow.lite.support.common.FileUtil;

final String ASSOCIATED_AXIS_LABELS = "labels.txt";
List<String> associatedAxisLabels = null;

try {
    associatedAxisLabels = FileUtil.loadLabels(this, ASSOCIATED_AXIS_LABELS);
} catch (IOException e) {
    Log.e("tfliteSupport", "Error reading label file", e);
}

The following snippet demonstrates how to associate the probabilities with category labels:

import java.util.Map;
import org.tensorflow.lite.support.common.TensorProcessor;
import org.tensorflow.lite.support.common.ops.NormalizeOp;
import org.tensorflow.lite.support.label.TensorLabel;

// Post-processor which dequantize the result
TensorProcessor probabilityProcessor =
    new TensorProcessor.Builder().add(new NormalizeOp(0, 255)).build();

if (null != associatedAxisLabels) {
    // Map of labels and their corresponding probability
    TensorLabel labels = new TensorLabel(associatedAxisLabels,
        probabilityProcessor.process(probabilityBuffer));

    // Create a map to access the result based on label
    Map<String, Float> floatMap = labels.getMapWithFloatValue();
}

Current use-case coverage

The current version of the LiteRT Support Library covers:

  • common data types (float, uint8, images, audio and array of these objects) as inputs and outputs of tflite models.
  • basic image operations (crop image, resize and rotate).
  • normalization and quantization
  • file utils

Future versions will improve support for text-related applications.

ImageProcessor Architecture

The design of the ImageProcessor allowed the image manipulation operations to be defined up front and optimised during the build process. The ImageProcessor currently supports three basic preprocessing operations, as described in the three comments in the code snippet below:

import org.tensorflow.lite.support.common.ops.NormalizeOp;
import org.tensorflow.lite.support.common.ops.QuantizeOp;
import org.tensorflow.lite.support.image.ops.ResizeOp;
import org.tensorflow.lite.support.image.ops.ResizeWithCropOrPadOp;
import org.tensorflow.lite.support.image.ops.Rot90Op;

int width = bitmap.getWidth();
int height = bitmap.getHeight();

int size = height > width ? width : height;

ImageProcessor imageProcessor =
    new ImageProcessor.Builder()
        // Center crop the image to the largest square possible
        .add(new ResizeWithCropOrPadOp(size, size))
        // Resize using Bilinear or Nearest neighbour
        .add(new ResizeOp(224, 224, ResizeOp.ResizeMethod.BILINEAR));
        // Rotation counter-clockwise in 90 degree increments
        .add(new Rot90Op(rotateDegrees / 90))
        .add(new NormalizeOp(127.5, 127.5))
        .add(new QuantizeOp(128.0, 1/128.0))
        .build();

See more details here about normalization and quantization.

The eventual goal of the support library is to support all tf.image transformations. This means the transformation will be the same as TensorFlow and the implementation will be independent of the operating system.

Developers are also welcome to create custom processors. It is important in these cases to be aligned with the training process - i.e. the same preprocessing should apply to both training and inference to increase reproducibility.

Quantization

When initiating input or output objects such as TensorImage or TensorBuffer you need to specify their types to be DataType.UINT8 or DataType.FLOAT32.

TensorImage tensorImage = new TensorImage(DataType.UINT8);
TensorBuffer probabilityBuffer =
    TensorBuffer.createFixedSize(new int[]{1, 1001}, DataType.UINT8);

The TensorProcessor can be used to quantize input tensors or dequantize output tensors. For example, when processing a quantized output TensorBuffer, the developer can use DequantizeOp to dequantize the result to a floating point probability between 0 and 1:

import org.tensorflow.lite.support.common.TensorProcessor;

// Post-processor which dequantize the result
TensorProcessor probabilityProcessor =
    new TensorProcessor.Builder().add(new DequantizeOp(0, 1/255.0)).build();
TensorBuffer dequantizedBuffer = probabilityProcessor.process(probabilityBuffer);

The quantization parameters of a tensor can be read through the metadata extractor library.