tf.distribute.Strategy with Training Loops

View on Run in Google Colab View source on GitHub

This tutorial demonstrates how to use tf.distribute.Strategy with custom training loops. We will train a simple CNN model on the fashion MNIST dataset. The fashion MNIST dataset contains 60000 train images of size 28 x 28 and 10000 test images of size 28 x 28.

We are using custom training loops to train our model because they give us flexibility and a greater control on training. Moreover, it is easier to debug the model and the training loop.

from __future__ import absolute_import, division, print_function, unicode_literals

# Import TensorFlow
import tensorflow as tf

# Helper libraries
import numpy as np
import os


Download the fashion mnist dataset

fashion_mnist = tf.keras.datasets.fashion_mnist

(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()

# Adding a dimension to the array -> new shape == (28, 28, 1)
# We are doing this because the first layer in our model is a convolutional
# layer and it requires a 4D input (batch_size, height, width, channels).
# batch_size dimension will be added later on.
train_images = train_images[..., None]
test_images = test_images[..., None]

# Getting the images in [0, 1] range.
train_images = train_images / np.float32(255)
test_images = test_images / np.float32(255)

train_labels = train_labels.astype('int64')
test_labels = test_labels.astype('int64')

Create a strategy to distribute the variables and the graph

How does tf.distribute.MirroredStrategy strategy work?

  • All the variables and the model graph is replicated on the replicas.
  • Input is evenly distributed across the replicas.
  • Each replica calculates the loss and gradients for the input it received.
  • The gradients are synced across all the replicas by summing them.
  • After the sync, the same update is made to the copies of the variables on each replica.
# If the list of devices is not specified in the
# <a href="../../api_docs/python/tf/distribute/MirroredStrategy"><code>tf.distribute.MirroredStrategy</code></a> constructor, it will be auto-detected.
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.

BUFFER_SIZE = len(train_images)

BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync


strategy.make_dataset_iterator creates an iterator that evenly distributes the data across all the replicas.

with strategy.scope():
  train_dataset =
  (train_images, train_labels)).shuffle(BUFFER_SIZE).batch(BATCH_SIZE)
  train_iterator = strategy.make_dataset_iterator(train_dataset)

  test_dataset =
      (test_images, test_labels)).batch(BATCH_SIZE)
  test_iterator = strategy.make_dataset_iterator(test_dataset)

Model Creation

Create a model using tf.keras.Sequential. You can also use the Model Subclassing API to do this.

with strategy.scope():
  model = tf.keras.Sequential([
      tf.keras.layers.Conv2D(32, 3, activation='relu',
                             input_shape=(28, 28, 1)),
      tf.keras.layers.Conv2D(64, 3, activation='relu'),
      tf.keras.layers.Dense(64, activation='relu'),
      tf.keras.layers.Dense(10, activation='softmax')
  optimizer = tf.train.GradientDescentOptimizer(0.001)

Define the loss function

Normally, on a single machine with 1 GPU/CPU, loss is divided by the number of examples in the batch of input.

So, how should the loss be calculated when using a tf.distribute.Strategy?

  • For an example, let's say you have 4 GPU's and a batch size of 64. One batch of input is distributed across the replicas (4 GPUs), each replica getting an input of size 16.

  • The model on each replica does a forward pass with its respective input and calculates the loss. Now, instead of dividing the loss by the number of examples in its respective input (BATCH_SIZE_PER_REPLICA = 16), the loss should be divided by the GLOBAL_BATCH_SIZE (64).

Why do this?

  • This needs to be done because after the gradients are calculated on each replica, they are synced across the replicas by summing them.

How to do this in TensorFlow? * If you're writing a custom training loop, as in this tutorial, you should sum the per example losses and divide the sum by the GLOBAL_BATCH_SIZE: scale_loss = tf.reduce_sum(loss) * (1. / GLOBAL_BATCH_SIZE)

  • Using tf.reduce_mean is not recommended. Doing so divides the loss by actual per replica batch size which may vary step to step.

  • This reduction and scaling is done automatically in keras model.compile and

  • If using tf.keras.losses classes, the loss reduction needs to be explicitly specified to be one of NONE or SUM. AUTO and SUM_OVER_BATCH_SIZE are disallowed when used with tf.distribute.Strategy. AUTO is disallowed because the user should explicitly think about what reduction they want to make sure it is correct in the distributed case. SUM_OVER_BATCH_SIZE is disallowed because currently it would only divide by per replica batch size, and leave the dividing by number of replicas to the user, which might be easy to miss. So instead we ask the user do the reduction themselves explicitly.

Training loop

with strategy.scope():
  def train_step():
    def step_fn(inputs):
      images, labels = inputs
      logits = model(images)
      cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
          logits=logits, labels=labels)
      loss = tf.reduce_sum(cross_entropy) * (1.0 / BATCH_SIZE)
      train_op = optimizer.minimize(loss)
      with tf.control_dependencies([train_op]):
        return tf.identity(loss)

    per_replica_losses = strategy.experimental_run(
        step_fn, train_iterator)
    mean_loss = strategy.reduce(
        tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None)
    return mean_loss
with strategy.scope():
  iterator_init = train_iterator.initialize()
  var_init = tf.global_variables_initializer()
  loss = train_step()
  with tf.Session() as sess:[var_init])
    for epoch in range(EPOCHS):[iterator_init])
      for step in range(10000):
        if step % 1000 == 0:
          print('Epoch {} Step {} Loss {:.4f}'.format(epoch+1,

Next Steps

Try out the new tf.distribute.Strategy API on your models.