jit_compile.ipynb

Kernel: Python 3

Copyright 2019 The TensorFlow Authors.

#@title Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Use XLA with tf.function

View on TensorFlow.org

Run in Google Colab

Download notebook

View source on GitHub

This tutorial trains a TensorFlow model to classify the MNIST dataset, where the training function is compiled using XLA.

First, load TensorFlow and enable eager execution.

import tensorflow as tf

Then define some necessary constants and prepare the MNIST dataset.

# Size of each input image, 28 x 28 pixels
IMAGE_SIZE = 28 * 28
# Number of distinct number labels, [0..9]
NUM_CLASSES = 10
# Number of examples in each training batch (step)
TRAIN_BATCH_SIZE = 100
# Number of training steps to run
TRAIN_STEPS = 1000

# Loads MNIST dataset.
train, test = tf.keras.datasets.mnist.load_data()
train_ds = tf.data.Dataset.from_tensor_slices(train).batch(TRAIN_BATCH_SIZE).repeat()

# Casting from raw data to the required datatypes.
def cast(images, labels):
  images = tf.cast(
      tf.reshape(images, [-1, IMAGE_SIZE]), tf.float32)
  labels = tf.cast(labels, tf.int64)
  return (images, labels)

Finally, define the model and the optimizer. The model uses a single dense layer.

layer = tf.keras.layers.Dense(NUM_CLASSES)
optimizer = tf.keras.optimizers.Adam()

Define the training function

In the training function, you get the predicted labels using the layer defined above, and then minimize the gradient of the loss using the optimizer. In order to compile the computation using XLA, place it inside tf.function with jit_compile=True.

@tf.function(jit_compile=True)
def train_mnist(images, labels):
    images, labels = cast(images, labels)

    with tf.GradientTape() as tape:
      predicted_labels = layer(images)
      loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
          logits=predicted_labels, labels=labels
      ))
    layer_variables = layer.trainable_variables
    grads = tape.gradient(loss, layer_variables)
    optimizer.apply_gradients(zip(grads, layer_variables))

Train and test the model

Once you have defined the training function, define the model.

for images, labels in train_ds:
  if optimizer.iterations > TRAIN_STEPS:
    break
  train_mnist(images, labels)

And, finally, check the accuracy:

images, labels = cast(test[0], test[1])
predicted_labels = layer(images)
correct_prediction = tf.equal(tf.argmax(predicted_labels, 1), labels)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
print("Prediction accuracy after training: %s" % accuracy)

Behind the scenes, the XLA compiler has compiled the entire TF function to HLO, which has enabled fusion optimizations. Using the introspection facilities, we can see the HLO code (other interesting possible values for "stage" are optimized_hlo for HLO after optimizations and optimized_hlo_dot for a Graphviz graph):

print(train_mnist.experimental_get_compiler_ir(images, labels)(stage='hlo'))