template-for-tflite-on-android's Introduction

Template-for-tflite-on-android

This is a template to test the reference speed of tflite model on android.

run the notebook to generate tflite model

import tensorflow as tf
from tensorflow.keras.applications import MobileNetV2

model = MobileNetV2(weights='imagenet')
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
open("mobilenet.tflite", "wb").write(tflite_model)

put tflite model in assets folder

".\tfliteDemo\app\src\main\assets"

change model name in MainActivity.java

private static final String MODEL = "mobilenet.tflite"; // model name

change the input and output shape in MainActivity.java "ModelRunAsyncTask"

you can put tflite.run(in, out) in a loop to test average reference speed

    private class ModelRunAsyncTask extends AsyncTask<Void, Void, Integer> {
        @Override
        protected Integer doInBackground(Void... args) {
//            if output is an image
//            int width = bitmap.getWidth();
//            int height = bitmap.getHeight();
//            float[][][][] out = new float[1][height][width][3];
            
            int height = 224;  // input shape
            int width = 224;
            float[][]out = new float[1][1000];  // output array
            float[][][][] in = getScaledMatrix(height, width, bitmap);
            long startTimeForReference = SystemClock.uptimeMillis();
            tflite.run(in, out);
            long endTimeForReference = SystemClock.uptimeMillis();
            referenceTime = endTimeForReference - startTimeForReference;
            Log.i(TAG, "Reference Time: " + referenceTime);

//            outputBitmap = getBitmap(height, width, out);
            Message msg = Message.obtain();
            msg.arg1 = 1;
            mHandler.sendMessage(msg);
            return 1;
        }
    }