TensorFlow Lite conversion

Introduction

TensorFlow Lite conversion makes your machine learning models small and fast so they can run on phones and other small devices.

You want to run a model on a smartphone without internet.

You need your model to use less battery and memory.

You want to add AI features to a small gadget like a smartwatch.

You want faster predictions on devices with limited power.

You want to share your model with others to use on mobile apps.

Syntax

TensorFlow

import tensorflow as tf

# Load your trained model
model = tf.keras.models.load_model('model.h5')

# Create a converter from the model
converter = tf.lite.TFLiteConverter.from_keras_model(model)

# Convert the model to TensorFlow Lite format
lite_model = converter.convert()

# Save the converted model
with open('model.tflite', 'wb') as f:
    f.write(lite_model)

Use from_keras_model() if you have a Keras model saved or built in memory.

The converted model is saved as a .tflite file which is smaller and optimized.

Examples

Convert a model saved in TensorFlow SavedModel format.

TensorFlow

converter = tf.lite.TFLiteConverter.from_saved_model('saved_model_dir')
lite_model = converter.convert()

Convert a Keras model with optimization to reduce size and improve speed.

TensorFlow

converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
lite_model = converter.convert()

Convert a TensorFlow ConcreteFunction to TensorFlow Lite format.

TensorFlow

converter = tf.lite.TFLiteConverter.from_concrete_functions([func])
lite_model = converter.convert()

Sample Model

This code trains a small model on random data, converts it to TensorFlow Lite format, and prints the size of the converted model.

TensorFlow

import tensorflow as tf
import numpy as np

# Create a simple model
model = tf.keras.Sequential([
    tf.keras.layers.Dense(10, activation='relu', input_shape=(5,)),
    tf.keras.layers.Dense(1)
])

# Compile and train the model on dummy data
model.compile(optimizer='adam', loss='mse')
x_train = np.random.random((100, 5))
y_train = np.random.random((100, 1))
model.fit(x_train, y_train, epochs=2, verbose=1)

# Convert the trained model to TensorFlow Lite
converter = tf.lite.TFLiteConverter.from_keras_model(model)
lite_model = converter.convert()

# Save the TensorFlow Lite model
with open('simple_model.tflite', 'wb') as f:
    f.write(lite_model)

print('TensorFlow Lite model size:', len(lite_model), 'bytes')

Important Notes

TensorFlow Lite models are much smaller and faster but may lose some accuracy.

You can apply optimizations like quantization during conversion to make models even smaller.

Always test the converted model on your device to check performance and accuracy.

Summary

TensorFlow Lite conversion makes models small and fast for mobile and embedded devices.

You convert models using TFLiteConverter from Keras, SavedModel, or ConcreteFunction.

Saving the converted model creates a .tflite file ready for deployment.