Distributed training in TensorFlow

View on TensorFlow.org Run in Google Colab View source on GitHub


The tf.distribute.Strategy API provides an abstraction for distributing your training across multiple processing units. The goal is to allow users to enable distributed training using existing models and training code, with minimal changes.

This tutorial uses the tf.distribute.MirroredStrategy, which does in-graph replication with synchronous training on many GPUs on one machine. Essentially, it copies all of the model's variables to each processor. Then, it uses all-reduce to combine the gradients from all processors and applies the combined value to all copies of the model.

MirroredStategy is one of several distribution strategy available in TensorFlow core. You can read about more strategies at distribution strategy guide.

Keras API

This example uses the tf.keras API to build the model and training loop. For custom training loops, see this tutorial.

Import Dependencies

from __future__ import absolute_import, division, print_function, unicode_literals
# Import TensorFlow
import tensorflow as tf  #gpu
import tensorflow_datasets as tfds

import os

Download the dataset

Download the MNIST dataset and load it from TensorFlow Datasets. This returns a dataset in tf.data format.

Setting with_info to True includes the metadata for the entire dataset, which is being saved here to ds_info. Among other things, this metadata object includes the number of train and test examples.

datasets, ds_info = tfds.load(name='mnist', with_info=True, as_supervised=True)
mnist_train, mnist_test = datasets['train'], datasets['test']

Define Distribution Strategy

Create a MirroredStrategy object. This will handle distribution, and provides a context manager (tf.distribute.MirroredStrategy.scope) to build your model inside.

strategy = tf.distribute.MirroredStrategy()
print ('Number of devices: {}'.format(strategy.num_replicas_in_sync))

Setup Input pipeline

If a model is trained on multiple GPUs, the batch size should be increased accordingly so as to make effective use of the extra computing power. Moreover, the learning rate should be tuned accordingly.

# You can also do ds_info.splits.total_num_examples to get the total
# number of examples in the dataset.

num_train_examples = ds_info.splits['train'].num_examples
num_test_examples = ds_info.splits['test'].num_examples


BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync

Pixel values, which are 0-255, have to be normalized to the 0-1 range. Define this scale in a function.

def scale(image, label):
  image = tf.cast(image, tf.float32)
  image /= 255

  return image, label

Apply this function to the training and test data, shuffle the training data, and batch it for training.

train_dataset = mnist_train.map(scale).shuffle(BUFFER_SIZE).batch(BATCH_SIZE)
eval_dataset = mnist_test.map(scale).batch(BATCH_SIZE)

Create the model

Create and compile the Keras model in the context of strategy.scope.

with strategy.scope():
  model = tf.keras.Sequential([
      tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(28, 28, 1)),
      tf.keras.layers.Dense(64, activation='relu'),
      tf.keras.layers.Dense(10, activation='softmax')


Define the callbacks.

The callbacks used here are:

  • Tensorboard: This callback writes a log for Tensorboard which allows you to visualize the graphs.
  • Model Checkpoint: This callback saves the model after every epoch.
  • Learning Rate Scheduler: Using this callback, you can schedule the learning rate to change after every epoch/batch.

For illustrative purposes, add a print callback to display the learning rate in the notebook.

# Define the checkpoint directory to store the checkpoints

checkpoint_dir = './training_checkpoints'
# Name of the checkpoint files
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt_{epoch}")
# Function for decaying the learning rate.
# You can define any decay function you need.
def decay(epoch):
  if epoch < 3:
    return 1e-3
  elif epoch >= 3 and epoch < 7:
    return 1e-4
    return 1e-5
# Callback for printing the LR at the end of each epoch.
class PrintLR(tf.keras.callbacks.Callback):
  def on_epoch_end(self, epoch, logs=None):
    print ('\nLearning rate for epoch {} is {}'.format(
        epoch + 1, tf.keras.backend.get_value(model.optimizer.lr)))
callbacks = [

Train and evaluate

Now, train the model in the usual way, calling fit on the model and passing in the dataset created at the beginning of the tutorial. This step is the same whether you are distributing the training or not.

model.fit(train_dataset, epochs=10, callbacks=callbacks)

As you can see below, the checkpoints are getting saved.

# check the checkpoint directory
!ls {checkpoint_dir}

To see how the model perform, load the latest checkpoint and call evaluate on the test data.

Call evaluate as before using appropriate datasets.


eval_loss, eval_acc = model.evaluate(eval_dataset)
print ('Eval loss: {}, Eval Accuracy: {}'.format(eval_loss, eval_acc))

To see the output, you can download and view the TensorBoard logs at the terminal.

$ tensorboard --logdir=path/to/log-directory
!ls -sh ./logs

Export to SavedModel

If you want to export the graph and the variables, SavedModel is the best way of doing this. The model can be loaded back with or without the scope. Moreover, SavedModel is platform agnostic.

path = 'saved_model/'
tf.keras.experimental.export_saved_model(model, path)

Load the model without strategy.scope.

unreplicated_model = tf.keras.experimental.load_from_saved_model(path)


eval_loss, eval_acc = unreplicated_model.evaluate(eval_dataset)
print ('Eval loss: {}, Eval Accuracy: {}'.format(eval_loss, eval_acc))

What's next?

Read the distribution strategy guide.