Kanji Handwriting Input on Android with TensorFlow Lite

Demo video

Requirements

• If you use Keras Convolution layer as input, data_format must be channels_last. At the time of this writing (2019-07-24), Tensorflow Lite will throw error when you try to convert CNN model with data_format is channels_first to tflite model.

Minimal instructions

1. Convert your model to tflite.

• Example with Keras model.

model = tf.keras.models.load_model('model.h5')

converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite = converter.convert()

tflite_filename = 'model.tflite'
open(tflite_filename, 'wb').write(tflite)

2. Create an android project.

3. Add 'tensorflow-lite' to app/build.gradle dependencies.

• Important:

  • Add dependencies to the build.gradle in the app folder, not the build.gradle in the root project.
  • You will also need to install Android NDK

• Example app\build.gradle

...
dependencies {
    ...
    // Tensorflow Lite library
    implementation 'org.tensorflow:tensorflow-lite:0.0.0-nightly'
}
...

4. Put your tflite model in the assets folder. It should locate at app/src/main/assets.

5. Tell Gradle not to compress our tflite model.

• Add noCompress 'tflite' to app\build.gradle.

• Example app\build.gradle

...
android {
    ...
    aaptOptions {
        noCompress 'tflite'
    }
}
...

6. Create an Interpreter in Java or Kotlin

After you add tensorflow-lite to project's dependencies and sync Gradle, you should able to import org.tensorflow.lite.Interpreter.

The Keras model equivalent in TFLite is an Interpreter.

To create and Interpreter you will need the TFLite model (in form of MappedByteBuffer) and the Interpreter.Options object.

In order to run inference (or model.predict in Keras) you will need to pre-allocated a ByteBuffer for the input and a float[] (depend on your model) for the output. The input ByteBuffer will need to be fill with your data before calling Interpreter.run(input, output).

References

• Tensorflow official example - too complicated for a simple Android app that perform image classification task. However, they do help with my implementation. You just need to find the right file name to read.
• Tensorflow Lite documentation.