Jax Model Conversion For LiteRT (TFLite)

Licensed under the Apache License, Version 2.0 (the "License");

This CodeLab demonstrates how to build a model for MNIST recognition using Jax, and how to convert it to LiteRT (TFLite). This codelab will also demonstrate how to optimize the Jax-converted TFLite model with post-training quantiztion.

Run in Google Colab View source on GitHub Download notebook

Prerequisites

It's recommended to try this feature with the newest TensorFlow nightly pip build.

pip install tf-nightly --upgrade
pip install jax --upgrade

# Make sure your JAX version is at least 0.4.20 or above.
import jax
jax.__version__

pip install orbax-export --upgrade

from orbax.export import ExportManager
from orbax.export import JaxModule
from orbax.export import ServingConfig
from orbax.export import constants

Data Preparation

Download the MNIST data with Keras dataset and pre-process.

import numpy as np
import tensorflow as tf
import functools

import time
import itertools

import numpy.random as npr

import jax.numpy as jnp
from jax import jit, grad, random
from jax.example_libraries import optimizers
from jax.example_libraries import stax

def _one_hot(x, k, dtype=np.float32):
  """Create a one-hot encoding of x of size k."""
  return np.array(x[:, None] == np.arange(k), dtype)

(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.mnist.load_data()
train_images, test_images = train_images / 255.0, test_images / 255.0
train_images = train_images.astype(np.float32)
test_images = test_images.astype(np.float32)

train_labels = _one_hot(train_labels, 10)
test_labels = _one_hot(test_labels, 10)

Build the MNIST model with Jax

def loss(params, batch):
  inputs, targets = batch
  preds = predict(params, inputs)
  return -jnp.mean(jnp.sum(preds * targets, axis=1))

def accuracy(params, batch):
  inputs, targets = batch
  target_class = jnp.argmax(targets, axis=1)
  predicted_class = jnp.argmax(predict(params, inputs), axis=1)
  return jnp.mean(predicted_class == target_class)

init_random_params, predict = stax.serial(
    stax.Flatten,
    stax.Dense(1024), stax.Relu,
    stax.Dense(1024), stax.Relu,
    stax.Dense(10), stax.LogSoftmax)

rng = random.PRNGKey(0)

Train & Evaluate the model

step_size = 0.001
num_epochs = 10
batch_size = 128
momentum_mass = 0.9


num_train = train_images.shape[0]
num_complete_batches, leftover = divmod(num_train, batch_size)
num_batches = num_complete_batches + bool(leftover)

def data_stream():
  rng = npr.RandomState(0)
  while True:
    perm = rng.permutation(num_train)
    for i in range(num_batches):
      batch_idx = perm[i * batch_size:(i + 1) * batch_size]
      yield train_images[batch_idx], train_labels[batch_idx]
batches = data_stream()

opt_init, opt_update, get_params = optimizers.momentum(step_size, mass=momentum_mass)

@jit
def update(i, opt_state, batch):
  params = get_params(opt_state)
  return opt_update(i, grad(loss)(params, batch), opt_state)

_, init_params = init_random_params(rng, (-1, 28 * 28))
opt_state = opt_init(init_params)
itercount = itertools.count()

print("\nStarting training...")
for epoch in range(num_epochs):
  start_time = time.time()
  for _ in range(num_batches):
    opt_state = update(next(itercount), opt_state, next(batches))
  epoch_time = time.time() - start_time

  params = get_params(opt_state)
  train_acc = accuracy(params, (train_images, train_labels))
  test_acc = accuracy(params, (test_images, test_labels))
  print("Epoch {} in {:0.2f} sec".format(epoch, epoch_time))
  print("Training set accuracy {}".format(train_acc))
  print("Test set accuracy {}".format(test_acc))

Convert to TFLite model.

Note here, we

  1. Export the JAX model to TF SavedModel using orbax.
  2. Call TFLite converter API to convert the TF SavedModel to .tflite model:
jax_module = JaxModule(params, predict, input_polymorphic_shape='b, ...')
converter = tf.lite.TFLiteConverter.from_concrete_functions(
    [
        jax_module.methods[constants.DEFAULT_METHOD_KEY].get_concrete_function(
            tf.TensorSpec(shape=(1, 28, 28), dtype=tf.float32, name="input")
        )
    ]
)

tflite_model = converter.convert()
with open('jax_mnist.tflite', 'wb') as f:
  f.write(tflite_model)

Check the Converted TFLite Model

Compare the converted model's results with the Jax model.

serving_func = functools.partial(predict, params)
expected = serving_func(train_images[0:1])

# Run the model with LiteRT (TFLite)
interpreter = tf.lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
interpreter.set_tensor(input_details[0]["index"], train_images[0:1, :, :])
interpreter.invoke()
result = interpreter.get_tensor(output_details[0]["index"])

# Assert if the result of TFLite model is consistent with the JAX model.
np.testing.assert_almost_equal(expected, result, 1e-5)

Optimize the Model

We will provide a representative_dataset to do post-training quantiztion to optimize the model.

def representative_dataset():
  for i in range(1000):
    x = train_images[i:i+1]
    yield [x]
x_input = jnp.zeros((1, 28, 28))
converter = tf.lite.TFLiteConverter.experimental_from_jax(
    [serving_func], [[('x', x_input)]])
tflite_model = converter.convert()
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
tflite_quant_model = converter.convert()
with open('jax_mnist_quant.tflite', 'wb') as f:
  f.write(tflite_quant_model)

Evaluate the Optimized Model

expected = serving_func(train_images[0:1])

# Run the model with LiteRT(TFLite)
interpreter = tf.lite.Interpreter(model_content=tflite_quant_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
interpreter.set_tensor(input_details[0]["index"], train_images[0:1, :, :])
interpreter.invoke()
result = interpreter.get_tensor(output_details[0]["index"])

# Assert if the result of TFLite model is consistent with the Jax model.
np.testing.assert_almost_equal(expected, result, 1e-5)

Compare the Quantized Model size

We should be able to see the quantized model is four times smaller than the original model.

du -h jax_mnist.tflite
du -h jax_mnist_quant.tflite

Download the converted models

from google.colab import files

jax_mnist_filename = "jax_mnist.tflite"

files.download(jax_mnist_filename)

from google.colab import files

jax_mnist_quant_filename = "jax_mnist_quant.tflite"

files.download(jax_mnist_quant_filename)