![Compile the model
# Get an Optimizer
adam = keras.optimizers.Adam(lr=0.001)
model.compile(loss='categorical_crossentropy',
optimizer=adam,
metrics=['accuracy'])](https://image.slidesharecdn.com/deployingaiappsusingtensorflowliteinmobiledevices-190813014747/85/NUS-ISS-Learning-Day-2019-Deploying-AI-apps-using-tensor-flow-lite-in-mobile-devices-23-320.jpg)
![Configure input Shapes
public class MnistDigitClassifier {
private static final String MODEL_NAME = "mnist.tflite";
private static final int BATCH_SIZE = 1;
public static final int IMG_HEIGHT = 28;
public static final int IMG_WIDTH = 28;
private static final int NUM_CHANNEL = 1;
private static final int NUM_CLASSES = 10;
private final Interpreter.Options options = new Interpreter.Options();
private final Interpreter mInterpreter;
private final ByteBuffer mImageData;
private final int[] mImagePixels = new int[IMG_HEIGHT * IMG_WIDTH];
private final float[][] mResult = new float[1][NUM_CLASSES];
}](https://image.slidesharecdn.com/deployingaiappsusingtensorflowliteinmobiledevices-190813014747/85/NUS-ISS-Learning-Day-2019-Deploying-AI-apps-using-tensor-flow-lite-in-mobile-devices-34-320.jpg)
![Predict : Classify
public class MnistDigitClassifier {
public Result classify(Bitmap bitmap) {
convertBitmapToByteBuffer(bitmap);
mInterpreter.run(mImageData, mResult);
Result r = new Result(mResult[0], timeCost);
r.setLabel(String.valueOf(r.getNumber()));
return r;
}](https://image.slidesharecdn.com/deployingaiappsusingtensorflowliteinmobiledevices-190813014747/85/NUS-ISS-Learning-Day-2019-Deploying-AI-apps-using-tensor-flow-lite-in-mobile-devices-36-320.jpg)
NUS-ISS Learning Day 2019-Deploying AI apps using tensor flow lite in mobile devices
- 1. Deploying AI Apps using TensorFlow Lite in Mobile Devices Colab - Keras - h5 - tflite - Android #ISSLearningDay Mr. Prasanna Veerapandi (Bala), NUS-ISS 2 Aug 2018
- 2. Agenda • Intro to Deep Learning • Build & Train CNN Models - colab • Export Model to tflite • Deploy tflite in android app • Live Demo • From Train Model ….. to Deploy in device #ISSLearningDay
- 3. About me • My name is Bala (Prasanna Veerapandi) • I teach Python for Data , Ops & Things #ISSLearningDay Google : NUS ISS PYTHON
- 4. Intro to Deep Learning SGD, NN, Back Prop, Arch, CNN, Keras, tflite #ISSLearningDay
- 5. ML vs DL
- 6. Let’s see how with a simple classifier CAT vs DOG
- 7. Get Some Data- Binary Classifier (2 folders)
- 8. Choose a function y = f(x) *Simple
- 9. Choose a function y = f(x) *Intermediate
- 10. Choose a function y = f(x) *Deep
- 11. How it learns ? TF Playground
- 13. Once Trained : You get predictions
- 14. Build & Train CNN Models - colab Data Preprocessing, Transform, Scale, Bachify, Hyper Param, Arch, Training & Evaluate #ISSLearningDay
- 15. Lets see some Keras Code
- 16. Import import tensorflow as tf from tensorflow.keras.datasets import mnist from tensorflow.keras.callbacks import LearningRateScheduler from tensorflow.keras.initializers import Constant from tensorflow.keras.layers import Input, Conv2D, BatchNormalization from tensorflow.keras.layers Activation, Dropout, Flatten, Dense from tensorflow.keras.models import Model from tensorflow.keras.utils import to_categorical
- 17. Load Data NUM_CLASSES = 10 (x_train, y_train), (x_test, y_test) = mnist.load_data(path='mnist.npz') x_train = x_train.reshape(*x_train.shape, 1) x_test = x_test.reshape(*x_test.shape, 1) x_train = x_train / 255.0 x_test = x_test / 255.0 y_train = to_categorical(y_train, NUM_CLASSES) y_test = to_categorical(y_test, NUM_CLASSES) print("x_train.shape = {}, y_train.shape = {}".format(x_train.shape, y_train.shape))
- 18. Build a Model (Arch) INPUT_SHAPE = (28, 28, 1) NUM_CLASSES = 10 model = keras.Sequential() model.add( Input(shape=INPUT_SHAPE, name='input') ) model.add( Conv2D(24, kernel_size=(6, 6), strides=1) ) model.add( MaxPooling2D( pool_size=(2, 2) ) ) model.add( BatchNormalization(scale=False, beta_initializer=Constant(0.01)) ) model.add( Activation('relu') ) model.add( tf.keras.layers.Flatten() ) model.add( Dense(200) ) model.add( BatchNormalization(scale=False, beta_initializer=Constant(0.01)) ) model.add( Activation('relu')) model.add( Dropout(rate=0.25)) model.add( Dense(NUM_CLASSES, activation='softmax', name='output') )
- 19. Model Summary
- 20. Remember the model Arch... *Intermediate
- 21. Build a Model (Arch) - Add more layers... INPUT_SHAPE = (28, 28, 1) NUM_CLASSES = 10 model = keras.Sequential() model.add( Input(shape=INPUT_SHAPE, name='input') ) model.add( Conv2D(24, kernel_size=(6, 6), strides=1) ) model.add( BatchNormalization(scale=False, beta_initializer=Constant(0.01)) ) model.add( Activation('relu') ) model.add( Conv2D(48, kernel_size=(5, 5), strides=2) ) model.add( MaxPooling2D( pool_size=(2, 2) ) ) model.add( BatchNormalization(scale=False, beta_initializer=Constant(0.01)) ) model.add( Activation('relu') ) model.add( Conv2D(64, kernel_size=(4, 4), strides=2) ) model.add( MaxPooling2D( pool_size=(2, 2) ) ) model.add( BatchNormalization(scale=False, beta_initializer=Constant(0.01)) ) model.add( Activation('relu') ) model.add( tf.keras.layers.Flatten() ) model.add( Dense(200) ) model.add( BatchNormalization(scale=False, beta_initializer=Constant(0.01)) ) model.add( Activation('relu')) model.add( Dropout(rate=0.25))
- 22. Model Summary
- 23. Compile the model # Get an Optimizer adam = keras.optimizers.Adam(lr=0.001) model.compile(loss='categorical_crossentropy', optimizer=adam, metrics=['accuracy'])
- 24. Lets Train the model history = model.fit(x_train, y_train, batch_size=128, epochs=2, verbose=1, validation_data=(x_test, y_test))
- 25. Export Model to tflite Keras, .h5, tflite converter #ISSLearningDay
- 27. Export h5 to tflite model.save('mnist.h5') converter = tf.lite.TFLiteConverter.from_keras_model_file('mnist.h5') tflite_model = converter.convert() with open('mnist.tflite', 'wb') as f: f.write(tflite_model)
- 29. Deploy tflite in android app org.tensorflow:tensorflow-lite:1.13.1 , Interpreter classifier , labels.txt, Interpreter.run( data , result ) #ISSLearningDay
- 30. Android APP
- 31. build.gradle aaptOptions { noCompress "tflite" noCompress "lite" } dependencies { implementation 'org.tensorflow:tensorflow-lite:1.13.1' }
- 32. Copy model into assets folder mnist.tflite labels.txt || labels.json File Size in 1..5 Mbs
- 33. Model file size quick-draw-tflite = 442 Kb File Size in 1..5 Mbs
- 34. Configure input Shapes public class MnistDigitClassifier { private static final String MODEL_NAME = "mnist.tflite"; private static final int BATCH_SIZE = 1; public static final int IMG_HEIGHT = 28; public static final int IMG_WIDTH = 28; private static final int NUM_CHANNEL = 1; private static final int NUM_CLASSES = 10; private final Interpreter.Options options = new Interpreter.Options(); private final Interpreter mInterpreter; private final ByteBuffer mImageData; private final int[] mImagePixels = new int[IMG_HEIGHT * IMG_WIDTH]; private final float[][] mResult = new float[1][NUM_CLASSES]; }
- 35. new Interpreter() public class MnistDigitClassifier { public MnistDigitClassifier(Activity activity) throws IOException { mInterpreter = new Interpreter(loadModelFile(activity), options); mImageData = ByteBuffer.allocateDirect( 4 * BATCH_SIZE * IMG_HEIGHT * IMG_WIDTH * NUM_CHANNEL); mImageData.order(ByteOrder.nativeOrder()); } }
- 36. Predict : Classify public class MnistDigitClassifier { public Result classify(Bitmap bitmap) { convertBitmapToByteBuffer(bitmap); mInterpreter.run(mImageData, mResult); Result r = new Result(mResult[0], timeCost); r.setLabel(String.valueOf(r.getNumber())); return r; }
- 38. Appendix // How to get the bitmap Bitmap image = mFpvPaint.exportToBitmap( QuickDrawClassifier.IMG_WIDTH, QuickDrawClassifier.IMG_HEIGHT); private MappedByteBuffer loadModelFile(Activity activity) throws IOException { AssetFileDescriptor fileDescriptor = activity.getAssets().openFd(MODEL_NAME); FileInputStream inputStream = new FileInputStream(fileDescriptor.getFileDescriptor()); FileChannel fileChannel = inputStream.getChannel(); long startOffset = fileDescriptor.getStartOffset(); long declaredLength = fileDescriptor.getDeclaredLength(); return fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength); }