Attend the Women in ML Symposium on December 7 Register now

Custom training with tf.distribute.Strategy

Stay organized with collections Save and categorize content based on your preferences.

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

This tutorial demonstrates how to use tf.distribute.Strategy—a TensorFlow API that provides an abstraction for distributing your training across multiple processing units (GPUs, multiple machines, or TPUs)—with custom training loops. In this example, you will train a simple convolutional neural network on the Fashion MNIST dataset containing 70,000 images of size 28 x 28.

Custom training loops provide flexibility and a greater control on training. They also make it easier to debug the model and the training loop.

# Import TensorFlow
import tensorflow as tf

# Helper libraries
import numpy as np
import os

print(tf.__version__)
2022-09-01 16:54:04.601941: E tensorflow/stream_executor/cuda/cuda_blas.cc:2981] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2022-09-01 16:54:05.309343: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvrtc.so.11.1: cannot open shared object file: No such file or directory
2022-09-01 16:54:05.309619: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvrtc.so.11.1: cannot open shared object file: No such file or directory
2022-09-01 16:54:05.309634: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.
2.10.0-rc3

Download the Fashion MNIST dataset

fashion_mnist = tf.keras.datasets.fashion_mnist

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

# Add a dimension to the array -> new shape == (28, 28, 1)
# This is done 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]

# Scale the images to the [0, 1] range.
train_images = train_images / np.float32(255)
test_images = test_images / np.float32(255)
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz
29515/29515 [==============================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz
26421880/26421880 [==============================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz
5148/5148 [==============================] - 0s 0us/step
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz
4422102/4422102 [==============================] - 0s 0us/step

Create a strategy to distribute the variables and the graph

How does tf.distribute.MirroredStrategy strategy work?

  • All the variables and the model graph are replicated across 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
# `tf.distribute.MirroredStrategy` constructor, they will be auto-detected.
strategy = tf.distribute.MirroredStrategy()
INFO:tensorflow:Using MirroredStrategy with devices ('/job:localhost/replica:0/task:0/device:GPU:0', '/job:localhost/replica:0/task:0/device:GPU:1', '/job:localhost/replica:0/task:0/device:GPU:2', '/job:localhost/replica:0/task:0/device:GPU:3')
print('Number of devices: {}'.format(strategy.num_replicas_in_sync))
Number of devices: 4

Setup input pipeline

BUFFER_SIZE = len(train_images)

BATCH_SIZE_PER_REPLICA = 64
GLOBAL_BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync

EPOCHS = 10

Create the datasets and distribute them:

train_dataset = tf.data.Dataset.from_tensor_slices((train_images, train_labels)).shuffle(BUFFER_SIZE).batch(GLOBAL_BATCH_SIZE) 
test_dataset = tf.data.Dataset.from_tensor_slices((test_images, test_labels)).batch(GLOBAL_BATCH_SIZE) 

train_dist_dataset = strategy.experimental_distribute_dataset(train_dataset)
test_dist_dataset = strategy.experimental_distribute_dataset(test_dataset)
2022-09-01 16:54:11.705735: W tensorflow/core/grappler/optimizers/data/auto_shard.cc:776] AUTO sharding policy will apply DATA sharding policy as it failed to apply FILE sharding policy because of the following reason: Found an unshardable source dataset: name: "TensorSliceDataset/_2"
op: "TensorSliceDataset"
input: "Placeholder/_0"
input: "Placeholder/_1"
attr {
  key: "Toutput_types"
  value {
    list {
      type: DT_FLOAT
      type: DT_UINT8
    }
  }
}
attr {
  key: "_cardinality"
  value {
    i: 60000
  }
}
attr {
  key: "is_files"
  value {
    b: false
  }
}
attr {
  key: "metadata"
  value {
    s: "\n\024TensorSliceDataset:0"
  }
}
attr {
  key: "output_shapes"
  value {
    list {
      shape {
        dim {
          size: 28
        }
        dim {
          size: 28
        }
        dim {
          size: 1
        }
      }
      shape {
      }
    }
  }
}
attr {
  key: "replicate_on_split"
  value {
    b: false
  }
}
experimental_type {
  type_id: TFT_PRODUCT
  args {
    type_id: TFT_DATASET
    args {
      type_id: TFT_PRODUCT
      args {
        type_id: TFT_TENSOR
        args {
          type_id: TFT_FLOAT
        }
      }
      args {
        type_id: TFT_TENSOR
        args {
          type_id: TFT_UINT8
        }
      }
    }
  }
}

2022-09-01 16:54:11.753833: W tensorflow/core/grappler/optimizers/data/auto_shard.cc:776] AUTO sharding policy will apply DATA sharding policy as it failed to apply FILE sharding policy because of the following reason: Found an unshardable source dataset: name: "TensorSliceDataset/_2"
op: "TensorSliceDataset"
input: "Placeholder/_0"
input: "Placeholder/_1"
attr {
  key: "Toutput_types"
  value {
    list {
      type: DT_FLOAT
      type: DT_UINT8
    }
  }
}
attr {
  key: "_cardinality"
  value {
    i: 10000
  }
}
attr {
  key: "is_files"
  value {
    b: false
  }
}
attr {
  key: "metadata"
  value {
    s: "\n\024TensorSliceDataset:3"
  }
}
attr {
  key: "output_shapes"
  value {
    list {
      shape {
        dim {
          size: 28
        }
        dim {
          size: 28
        }
        dim {
          size: 1
        }
      }
      shape {
      }
    }
  }
}
attr {
  key: "replicate_on_split"
  value {
    b: false
  }
}
experimental_type {
  type_id: TFT_PRODUCT
  args {
    type_id: TFT_DATASET
    args {
      type_id: TFT_PRODUCT
      args {
        type_id: TFT_TENSOR
        args {
          type_id: TFT_FLOAT
        }
      }
      args {
        type_id: TFT_TENSOR
        args {
          type_id: TFT_UINT8
        }
      }
    }
  }
}

Create the model

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

def create_model():
  model = tf.keras.Sequential([
      tf.keras.layers.Conv2D(32, 3, activation='relu'),
      tf.keras.layers.MaxPooling2D(),
      tf.keras.layers.Conv2D(64, 3, activation='relu'),
      tf.keras.layers.MaxPooling2D(),
      tf.keras.layers.Flatten(),
      tf.keras.layers.Dense(64, activation='relu'),
      tf.keras.layers.Dense(10)
    ])

  return model
# Create a checkpoint directory to store the checkpoints.
checkpoint_dir = './training_checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")

Define the loss function

Normally, on a single machine with a single GPU/CPU, the loss function 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 GPUs and a batch size of 64. One batch of input is distributed across the replicas (4 GPUs), and each replica gets 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) or you can use tf.nn.compute_average_loss which takes the per example loss, optional sample weights, and GLOBAL_BATCH_SIZE as arguments and returns the scaled loss.

  • If you are using regularization losses in your model then you need to scale the loss value by the number of replicas. You can do this by using the tf.nn.scale_regularization_loss function.

  • 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 Model.fit

  • If using tf.keras.losses classes (as in the example below), 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, you need to do the reduction yourself explicitly.

  • If labels is multi-dimensional, then average the per_example_loss across the number of elements in each sample. For example, if the shape of predictions is (batch_size, H, W, n_classes) and labels is (batch_size, H, W), you will need to update per_example_loss like: per_example_loss /= tf.cast(tf.reduce_prod(tf.shape(labels)[1:]), tf.float32)

with strategy.scope():
  # Set reduction to `NONE` so you can do the reduction afterwards and divide by
  # global batch size.
  loss_object = tf.keras.losses.SparseCategoricalCrossentropy(
      from_logits=True,
      reduction=tf.keras.losses.Reduction.NONE)
  def compute_loss(labels, predictions):
    per_example_loss = loss_object(labels, predictions)
    return tf.nn.compute_average_loss(per_example_loss, global_batch_size=GLOBAL_BATCH_SIZE)

Define the metrics to track loss and accuracy

These metrics track the test loss and training and test accuracy. You can use .result() to get the accumulated statistics at any time.

with strategy.scope():
  test_loss = tf.keras.metrics.Mean(name='test_loss')

  train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
      name='train_accuracy')
  test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
      name='test_accuracy')
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).
INFO:tensorflow:Reduce to /job:localhost/replica:0/task:0/device:CPU:0 then broadcast to ('/job:localhost/replica:0/task:0/device:CPU:0',).

Training loop

# A model, an optimizer, and a checkpoint must be created under `strategy.scope`.
with strategy.scope():
  model = create_model()

  optimizer = tf.keras.optimizers.Adam()

  checkpoint = tf.train.Checkpoint(optimizer=optimizer, model=model)
def train_step(inputs):
  images, labels = inputs

  with tf.GradientTape() as tape:
    predictions = model(images, training=True)
    loss = compute_loss(labels, predictions)

  gradients = tape.gradient(loss, model.trainable_variables)
  optimizer.apply_gradients(zip(gradients, model.trainable_variables))

  train_accuracy.update_state(labels, predictions)
  return loss 

def test_step(inputs):
  images, labels = inputs

  predictions = model(images, training=False)
  t_loss = loss_object(labels, predictions)

  test_loss.update_state(t_loss)
  test_accuracy.update_state(labels, predictions)
# `run` replicates the provided computation and runs it
# with the distributed input.
@tf.function
def distributed_train_step(dataset_inputs):
  per_replica_losses = strategy.run(train_step, args=(dataset_inputs,))
  return strategy.reduce(tf.distribute.ReduceOp.SUM, per_replica_losses,
                         axis=None)

@tf.function
def distributed_test_step(dataset_inputs):
  return strategy.run(test_step, args=(dataset_inputs,))

for epoch in range(EPOCHS):
  # TRAIN LOOP
  total_loss = 0.0
  num_batches = 0
  for x in train_dist_dataset:
    total_loss += distributed_train_step(x)
    num_batches += 1
  train_loss = total_loss / num_batches

  # TEST LOOP
  for x in test_dist_dataset:
    distributed_test_step(x)

  if epoch % 2 == 0:
    checkpoint.save(checkpoint_prefix)

  template = ("Epoch {}, Loss: {}, Accuracy: {}, Test Loss: {}, "
              "Test Accuracy: {}")
  print(template.format(epoch + 1, train_loss,
                         train_accuracy.result() * 100, test_loss.result(),
                         test_accuracy.result() * 100))

  test_loss.reset_states()
  train_accuracy.reset_states()
  test_accuracy.reset_states()
INFO:tensorflow:batch_all_reduce: 8 all-reduces with algorithm = nccl, num_packs = 1
INFO:tensorflow:batch_all_reduce: 8 all-reduces with algorithm = nccl, num_packs = 1
INFO:tensorflow:batch_all_reduce: 8 all-reduces with algorithm = nccl, num_packs = 1
Epoch 1, Loss: 0.6676574945449829, Accuracy: 76.33167266845703, Test Loss: 0.5149329304695129, Test Accuracy: 80.44999694824219
Epoch 2, Loss: 0.40435999631881714, Accuracy: 85.57333374023438, Test Loss: 0.41797366738319397, Test Accuracy: 84.44999694824219
Epoch 3, Loss: 0.35420727729797363, Accuracy: 87.32666015625, Test Loss: 0.3638404309749603, Test Accuracy: 86.73999786376953
Epoch 4, Loss: 0.31962651014328003, Accuracy: 88.55833435058594, Test Loss: 0.3434653878211975, Test Accuracy: 87.3499984741211
Epoch 5, Loss: 0.3000466823577881, Accuracy: 89.11499786376953, Test Loss: 0.32996779680252075, Test Accuracy: 88.23999786376953
Epoch 6, Loss: 0.2839100956916809, Accuracy: 89.74666595458984, Test Loss: 0.3184721767902374, Test Accuracy: 88.43000030517578
Epoch 7, Loss: 0.2712290287017822, Accuracy: 90.17833709716797, Test Loss: 0.30810120701789856, Test Accuracy: 88.86000061035156
Epoch 8, Loss: 0.2557474970817566, Accuracy: 90.69000244140625, Test Loss: 0.2876024544239044, Test Accuracy: 89.69000244140625
Epoch 9, Loss: 0.24715569615364075, Accuracy: 91.00999450683594, Test Loss: 0.33483773469924927, Test Accuracy: 88.44000244140625
Epoch 10, Loss: 0.23839063942432404, Accuracy: 91.32166290283203, Test Loss: 0.2774220407009125, Test Accuracy: 89.99000549316406

Things to note in the example above:

Restore the latest checkpoint and test

A model checkpointed with a tf.distribute.Strategy can be restored with or without a strategy.

eval_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
      name='eval_accuracy')

new_model = create_model()
new_optimizer = tf.keras.optimizers.Adam()

test_dataset = tf.data.Dataset.from_tensor_slices((test_images, test_labels)).batch(GLOBAL_BATCH_SIZE)
@tf.function
def eval_step(images, labels):
  predictions = new_model(images, training=False)
  eval_accuracy(labels, predictions)
checkpoint = tf.train.Checkpoint(optimizer=new_optimizer, model=new_model)
checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))

for images, labels in test_dataset:
  eval_step(images, labels)

print('Accuracy after restoring the saved model without strategy: {}'.format(
    eval_accuracy.result() * 100))
Accuracy after restoring the saved model without strategy: 88.44000244140625

Alternate ways of iterating over a dataset

Using iterators

If you want to iterate over a given number of steps and not through the entire dataset, you can create an iterator using the iter call and explicitly call next on the iterator. You can choose to iterate over the dataset both inside and outside the tf.function. Here is a small snippet demonstrating iteration of the dataset outside the tf.function using an iterator.

for _ in range(EPOCHS):
  total_loss = 0.0
  num_batches = 0
  train_iter = iter(train_dist_dataset)

  for _ in range(10):
    total_loss += distributed_train_step(next(train_iter))
    num_batches += 1
  average_train_loss = total_loss / num_batches

  template = ("Epoch {}, Loss: {}, Accuracy: {}")
  print(template.format(epoch + 1, average_train_loss, train_accuracy.result() * 100))
  train_accuracy.reset_states()
Epoch 10, Loss: 0.22689345479011536, Accuracy: 91.796875
Epoch 10, Loss: 0.23983070254325867, Accuracy: 91.25
Epoch 10, Loss: 0.23446457087993622, Accuracy: 91.3671875
Epoch 10, Loss: 0.22298672795295715, Accuracy: 92.03125
Epoch 10, Loss: 0.21223223209381104, Accuracy: 92.109375
Epoch 10, Loss: 0.21282605826854706, Accuracy: 92.6953125
Epoch 10, Loss: 0.21947257220745087, Accuracy: 91.5234375
Epoch 10, Loss: 0.22401729226112366, Accuracy: 91.5625
Epoch 10, Loss: 0.21480241417884827, Accuracy: 91.9140625
Epoch 10, Loss: 0.21416351199150085, Accuracy: 92.3828125

Iterating inside a tf.function

You can also iterate over the entire input train_dist_dataset inside a tf.function using the for x in ... construct or by creating iterators like you did above. The example below demonstrates wrapping one epoch of training with a @tf.function decorator and iterating over train_dist_dataset inside the function.

@tf.function
def distributed_train_epoch(dataset):
  total_loss = 0.0
  num_batches = 0
  for x in dataset:
    per_replica_losses = strategy.run(train_step, args=(x,))
    total_loss += strategy.reduce(
      tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None)
    num_batches += 1
  return total_loss / tf.cast(num_batches, dtype=tf.float32)

for epoch in range(EPOCHS):
  train_loss = distributed_train_epoch(train_dist_dataset)

  template = ("Epoch {}, Loss: {}, Accuracy: {}")
  print(template.format(epoch + 1, train_loss, train_accuracy.result() * 100))

  train_accuracy.reset_states()
/tmpfs/src/tf_docs_env/lib/python3.9/site-packages/tensorflow/python/data/ops/dataset_ops.py:461: UserWarning: To make it possible to preserve tf.data options across serialization boundaries, their implementation has moved to be part of the TensorFlow graph. As a consequence, the options value is in general no longer known at graph construction time. Invoking this method in graph mode retains the legacy behavior of the original implementation, but note that the returned value might not reflect the actual value of the options.
  warnings.warn("To make it possible to preserve tf.data options across "
INFO:tensorflow:batch_all_reduce: 8 all-reduces with algorithm = nccl, num_packs = 1
Epoch 1, Loss: 0.2222447544336319, Accuracy: 91.9433364868164
Epoch 2, Loss: 0.21512489020824432, Accuracy: 92.12999725341797
Epoch 3, Loss: 0.2052260935306549, Accuracy: 92.40333557128906
Epoch 4, Loss: 0.19451268017292023, Accuracy: 92.8499984741211
Epoch 5, Loss: 0.18623782694339752, Accuracy: 93.32500457763672
Epoch 6, Loss: 0.18053726851940155, Accuracy: 93.3499984741211
Epoch 7, Loss: 0.1721896082162857, Accuracy: 93.70333862304688
Epoch 8, Loss: 0.16532976925373077, Accuracy: 94.04000091552734
Epoch 9, Loss: 0.15882055461406708, Accuracy: 94.12999725341797
Epoch 10, Loss: 0.15295268595218658, Accuracy: 94.34666442871094

Tracking training loss across replicas

Because of the loss scaling computation that is carried out, it's not recommended to use tf.keras.metrics.Mean to track the training loss across different replicas.

For example, if you run a training job with the following characteristics:

  • Two replicas
  • Two samples are processed on each replica
  • Resulting loss values: [2, 3] and [4, 5] on each replica
  • Global batch size = 4

With loss scaling, you calculate the per-sample value of loss on each replica by adding the loss values, and then dividing by the global batch size. In this case: (2 + 3) / 4 = 1.25 and (4 + 5) / 4 = 2.25.

If you use tf.keras.metrics.Mean to track loss across the two replicas, the result is different. In this example, you end up with a total of 3.50 and count of 2, which results in total/count = 1.75 when result() is called on the metric. Loss calculated with tf.keras.Metrics is scaled by an additional factor that is equal to the number of replicas in sync.

Guide and examples

Here are some examples for using distribution strategy with custom training loops:

  1. Distributed training guide
  2. DenseNet example using MirroredStrategy.
  3. BERT example trained using MirroredStrategy and TPUStrategy. This example is particularly helpful for understanding how to load from a checkpoint and generate periodic checkpoints during distributed training etc.
  4. NCF example trained using MirroredStrategy that can be enabled using the keras_use_ctl flag.
  5. NMT example trained using MirroredStrategy.

You can find more examples listed under Examples and tutorials in the Distribution strategy guide.

Next steps