tf.distribute.Strategy

TensorFlow 1 version

View source on GitHub

A state & compute distribution policy on a list of devices.

tf.distribute.Strategy(
    extended
)

See the guide for overview and examples.

In short:

• To use it with Keras compile/fit, please read.
• You may pass descendant of tf.distribute.Strategy to tf.estimator.RunConfig to specify how a tf.estimator.Estimator should distribute its computation. See guide.
• Otherwise, use tf.distribute.Strategy.scope to specify that a strategy should be used when building an executing your model. (This puts you in the "cross-replica context" for this strategy, which means the strategy is put in control of things like variable placement.)

• If you are writing a custom training loop, you will need to call a few more methods, see the guide:

  • Start by either creating a tf.data.Dataset normally or using tf.distribute.experimental_make_numpy_dataset to make a dataset out of a numpy array.

  • Use tf.distribute.Strategy.experimental_distribute_dataset to convert a tf.data.Dataset to something that produces "per-replica" values. If you want to manually specify how the dataset should be partitioned across replicas, use tf.distribute.Strategy.experimental_distribute_datasets_from_function instead.

  • Use tf.distribute.Strategy.run to run a function once per replica, taking values that may be "per-replica" (e.g. from a distributed dataset) and returning "per-replica" values. This function is executed in "replica context", which means each operation is performed separately on each replica.

  • Finally use a method (such as tf.distribute.Strategy.reduce) to convert the resulting "per-replica" values into ordinary Tensors.

A custom training loop can be as simple as:

with my_strategy.scope():
  @tf.function
  def distribute_train_epoch(dataset):
    def replica_fn(input):
      # process input and return result
      return result

    total_result = 0
    for x in dataset:
      per_replica_result = my_strategy.run(replica_fn, args=(x,))
      total_result += my_strategy.reduce(tf.distribute.ReduceOp.SUM,
                                         per_replica_result, axis=None)
    return total_result

  dist_dataset = my_strategy.experimental_distribute_dataset(dataset)
  for _ in range(EPOCHS):
    train_result = distribute_train_epoch(dist_dataset)

This takes an ordinary dataset and replica_fn and runs it distributed using a particular tf.distribute.Strategy named my_strategy above. Any variables created in replica_fn are created using my_strategy's policy, and library functions called by replica_fn can use the get_replica_context() API to implement distributed-specific behavior.

You can use the reduce API to aggregate results across replicas and use this as a return value from one iteration over the distributed dataset. Or you can use tf.keras.metrics (such as loss, accuracy, etc.) to accumulate metrics across steps in a given epoch.

See the custom training loop tutorial for a more detailed example.

Methods

experimental_assign_to_logical_device

View source

experimental_assign_to_logical_device(
    tensor, logical_device_id
)

Adds annotation that tensor will be assigned to a logical device.

# Initializing TPU system with 2 logical devices and 4 replicas.
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
tf.config.experimental_connect_to_cluster(resolver)
topology = tf.tpu.experimental.initialize_tpu_system(resolver)
device_assignment = tf.tpu.experimental.DeviceAssignment.build(
    topology,
    computation_shape=[1, 1, 2],
    num_replicas=4)
strategy = tf.distribute.experimental.TPUStrategy(
    resolver, device_assignment=device_assignment)
iterator = iter(inputs)

@tf.function()
def step_fn(inputs):
  output = tf.add(inputs, inputs)

  // Add operation will be executed on logical device 0.
  output = strategy.experimental_assign_to_logical_device(output, 0)
  return output

strategy.run(step_fn, args=(next(iterator),))

experimental_distribute_dataset

View source

experimental_distribute_dataset(
    dataset
)

Distributes a tf.data.Dataset instance provided via dataset.

The returned distributed dataset can be iterated over similar to how regular datasets can. NOTE: Currently, the user cannot add any more transformations to a distributed dataset.

The following is an example:

strategy = tf.distribute.MirroredStrategy()

# Create a dataset
dataset = dataset_ops.Dataset.TFRecordDataset([
  "/a/1.tfr", "/a/2.tfr", "/a/3.tfr", "/a/4.tfr"])

# Distribute that dataset
dist_dataset = strategy.experimental_distribute_dataset(dataset)

# Iterate over the distributed dataset
for x in dist_dataset:
  # process dataset elements
  strategy.run(train_step, args=(x,))

We will assume that the input dataset is batched by the global batch size. With this assumption, we will make a best effort to divide each batch across all the replicas (one or more workers).

In a multi-worker setting, we will first attempt to distribute the dataset by attempting to detect whether the dataset is being created out of ReaderDatasets (e.g. TFRecordDataset, TextLineDataset, etc.) and if so, attempting to shard the input files. Note that there has to be at least one input file per worker. If you have less than one input file per worker, we suggest that you should disable distributing your dataset using the method below.

If that attempt is unsuccessful (e.g. the dataset is created from a Dataset.range), we will shard the dataset evenly at the end by appending a .shard operation to the end of the processing pipeline. This will cause the entire preprocessing pipeline for all the data to be run on every worker, and each worker will do redundant work. We will print a warning if this method of sharding is selected.

You can disable dataset sharding across workers using the auto_shard_policy option in tf.data.experimental.DistributeOptions.

Within each worker, we will also split the data among all the worker devices (if more than one a present), and this will happen even if multi-worker sharding is disabled using the method above.

If the above batch splitting and dataset sharding logic is undesirable, please use experimental_distribute_datasets_from_function instead, which does not do any automatic splitting or sharding.

You can also use the element_spec property of the distributed dataset returned by this API to query the tf.TypeSpec of the elements returned by the iterator. This can be used to set the input_signature property of a tf.function.

strategy = tf.distribute.MirroredStrategy()

# Create a dataset
dataset = dataset_ops.Dataset.TFRecordDataset([
  "/a/1.tfr", "/a/2.tfr", "/a/3.tfr", "/a/4.tfr"])

# Distribute that dataset
dist_dataset = strategy.experimental_distribute_dataset(dataset)

@tf.function(input_signature=[dist_dataset.element_spec])
def train_step(inputs):
  # train model with inputs
  return

# Iterate over the distributed dataset
for x in dist_dataset:
  # process dataset elements
  strategy.run(train_step, args=(x,))

experimental_distribute_datasets_from_function

View source

experimental_distribute_datasets_from_function(
    dataset_fn
)

Distributes tf.data.Dataset instances created by calls to dataset_fn.

dataset_fn will be called once for each worker in the strategy. Each replica on that worker will dequeue one batch of inputs from the local Dataset (i.e. if a worker has two replicas, two batches will be dequeued from the Dataset every step).

This method can be used for several purposes. For example, where experimental_distribute_dataset is unable to shard the input files, this method might be used to manually shard the dataset (avoiding the slow fallback behavior in experimental_distribute_dataset). In cases where the dataset is infinite, this sharding can be done by creating dataset replicas that differ only in their random seed. experimental_distribute_dataset may also sometimes fail to split the batch across replicas on a worker. In that case, this method can be used where that limitation does not exist.

The dataset_fn should take an tf.distribute.InputContext instance where information about batching and input replication can be accessed:

def dataset_fn(input_context):
  batch_size = input_context.get_per_replica_batch_size(global_batch_size)
  d = tf.data.Dataset.from_tensors([[1.]]).repeat().batch(batch_size)
  return d.shard(
      input_context.num_input_pipelines, input_context.input_pipeline_id)

inputs = strategy.experimental_distribute_datasets_from_function(dataset_fn)

for batch in inputs:
  replica_results = strategy.run(replica_fn, args=(batch,))

To query the tf.TypeSpec of the elements in the distributed dataset returned by this API, you need to use the element_spec property of the distributed iterator. This tf.TypeSpec can be used to set the input_signature property of a tf.function.

# If you want to specify `input_signature` for a `tf.function` you must
# first create the iterator.
iterator = iter(inputs)

@tf.function(input_signature=[iterator.element_spec])
def replica_fn_with_signature(inputs):
  # train the model with inputs
  return

for _ in range(steps):
  strategy.run(replica_fn_with_signature,
      args=(next(iterator),))

experimental_distribute_values_from_function

View source

experimental_distribute_values_from_function(
    value_fn
)

Generates tf.distribute.DistributedValues from value_fn.

This function is to generate tf.distribute.DistributedValues to pass into run, reduce, or other methods that take distributed values when not using datasets.

Example usage:

1. Return constant value per replica:

strategy = tf.distribute.MirroredStrategy()
def value_fn(ctx):
  return tf.constant(1.)
distributed_values = (
     strategy.experimental_distribute_values_from_function(
       value_fn))
local_result = strategy.experimental_local_results(distributed_values)
local_result
(<tf.Tensor: shape=(), dtype=float32, numpy=1.0>,)

1. Distribute values in array based on replica_id:

strategy = tf.distribute.MirroredStrategy()
array_value = np.array([3., 2., 1.])
def value_fn(ctx):
  return array_value[ctx.replica_id_in_sync_group]
distributed_values = (
     strategy.experimental_distribute_values_from_function(
       value_fn))
local_result = strategy.experimental_local_results(distributed_values)
local_result
(3.0,)

1. Specify values using num_replicas_in_sync:

strategy = tf.distribute.MirroredStrategy()
def value_fn(ctx):
  return ctx.num_replicas_in_sync
distributed_values = (
     strategy.experimental_distribute_values_from_function(
       value_fn))
local_result = strategy.experimental_local_results(distributed_values)
local_result
(1,)

1. Place values on devices and distribute:

strategy = tf.distribute.TPUStrategy()
worker_devices = strategy.extended.worker_devices
multiple_values = []
for i in range(strategy.num_replicas_in_sync):
  with tf.device(worker_devices[i]):
    multiple_values.append(tf.constant(1.0))

def value_fn(ctx):
  return multiple_values[ctx.replica_id]

distributed_values = strategy.
  experimental_distribute_values_from_function(
  value_fn)

experimental_local_results

View source

experimental_local_results(
    value
)

Returns the list of all local per-replica values contained in value.

experimental_make_numpy_dataset

View source

experimental_make_numpy_dataset(
    numpy_input
)

Makes a tf.data.Dataset for input provided via a numpy array.

This avoids adding numpy_input as a large constant in the graph, and copies the data to the machine or machines that will be processing the input.

Note that you will likely need to use experimental_distribute_dataset with the returned dataset to further distribute it with the strategy.

Example:

numpy_input = np.ones([10], dtype=np.float32)
dataset = strategy.experimental_make_numpy_dataset(numpy_input)
dist_dataset = strategy.experimental_distribute_dataset(dataset)

experimental_replicate_to_logical_devices

View source

experimental_replicate_to_logical_devices(
    tensor
)

Adds annotation that tensor will be replicated to all logical devices.

# Initializing TPU system with 2 logical devices and 4 replicas.
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
tf.config.experimental_connect_to_cluster(resolver)
topology = tf.tpu.experimental.initialize_tpu_system(resolver)
device_assignment = tf.tpu.experimental.DeviceAssignment.build(
    topology,
    computation_shape=[1, 1, 2],
    num_replicas=4)
strategy = tf.distribute.experimental.TPUStrategy(
    resolver, device_assignment=device_assignment)

iterator = iter(inputs)

@tf.function()
def step_fn(inputs):
  images, labels = inputs
  images = strategy.experimental_split_to_logical_devices(
    inputs, [1, 2, 4, 1])

  // model() function will be executed on 8 logical devices with `inputs`
  // split 2 * 4  ways.
  output = model(inputs)

  // For loss calculation, all logical devices share the same logits
  // and labels.
  labels = strategy.experimental_replicate_to_logical_devices(labels)
  output = strategy.experimental_replicate_to_logical_devices(output)
  loss = loss_fn(labels, output)

  return loss

strategy.run(step_fn, args=(next(iterator),))

Args: tensor: Input tensor to annotate.

experimental_split_to_logical_devices

View source

experimental_split_to_logical_devices(
    tensor, partition_dimensions
)

Adds annotation that tensor will be split across logical devices.

For example, for system with 8 logical devices, if tensor is an image tensor with shape (batch_size, width, height, channel) and partition_dimensions is [1, 2, 4, 1], then tensor will be split 2 in width dimension and 4 way in height dimension and the split tensor values will be fed into 8 logical devices.

# Initializing TPU system with 8 logical devices and 1 replica.
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
tf.config.experimental_connect_to_cluster(resolver)
topology = tf.tpu.experimental.initialize_tpu_system(resolver)
device_assignment = tf.tpu.experimental.DeviceAssignment.build(
    topology,
    computation_shape=[2, 2, 2],
    num_replicas=1)
strategy = tf.distribute.experimental.TPUStrategy(
    resolver, device_assignment=device_assignment)

iterator = iter(inputs)

@tf.function()
def step_fn(inputs):
  inputs = strategy.experimental_split_to_logical_devices(
    inputs, [1, 2, 4, 1])

  // model() function will be executed on 8 logical devices with `inputs`
  // split 2 * 4  ways.
  output = model(inputs)
  return output

strategy.run(step_fn, args=(next(iterator),))

Args: tensor: Input tensor to annotate. partition_dimensions: An unnested list of integers with the size equal to rank of tensor specifying how tensor will be partitioned. The product of all elements in partition_dimensions must be equal to the total number of logical devices per replica.

reduce

View source

reduce(
    reduce_op, value, axis
)

Reduce value across replicas.

Given a per-replica value returned by run, say a per-example loss, the batch will be divided across all the replicas. This function allows you to aggregate across replicas and optionally also across batch elements. For example, if you have a global batch size of 8 and 2 replicas, values for examples [0, 1, 2, 3] will be on replica 0 and [4, 5, 6, 7] will be on replica 1. By default, reduce will just aggregate across replicas, returning [0+4, 1+5, 2+6, 3+7]. This is useful when each replica is computing a scalar or some other value that doesn't have a "batch" dimension (like a gradient). More often you will want to aggregate across the global batch, which you can get by specifying the batch dimension as the axis, typically axis=0. In this case it would return a scalar 0+1+2+3+4+5+6+7.

If there is a last partial batch, you will need to specify an axis so that the resulting shape is consistent across replicas. So if the last batch has size 6 and it is divided into [0, 1, 2, 3] and [4, 5], you would get a shape mismatch unless you specify axis=0. If you specify tf.distribute.ReduceOp.MEAN, using axis=0 will use the correct denominator of 6. Contrast this with computing reduce_mean to get a scalar value on each replica and this function to average those means, which will weigh some values 1/8 and others 1/4.

run

View source

run(
    fn, args=(), kwargs=None, options=None
)

Run fn on each replica, with the given arguments.

Executes ops specified by fn on each replica. If args or kwargs have tf.distribute.DistributedValues, such as those produced by a "distributed Dataset" or experimental_distribute_values_from_function when fn is executed on a particular replica, it will be executed with the component of tf.distribute.DistributedValues that correspond to that replica.

fn may call tf.distribute.get_replica_context() to access members such as all_reduce.

All arguments in args or kwargs should either be nest of tensors or tf.distribute.DistributedValues containing tensors or composite tensors.

Example usage:

1. Constant tensor input.

strategy = tf.distribute.MirroredStrategy()
tensor_input = tf.constant(3.0)
@tf.function
def replica_fn(input):
  return input*2.0
result = strategy.run(replica_fn, args=(tensor_input,))
result
<tf.Tensor: shape=(), dtype=float32, numpy=6.0>

1. DistributedValues input.

strategy = tf.distribute.MirroredStrategy()
@tf.function
def run():
  def value_fn(value_context):
    return value_context.num_replicas_in_sync
  distributed_values = (
    strategy.experimental_distribute_values_from_function(
      value_fn))
  def replica_fn2(input):
    return input*2
  return strategy.run(replica_fn2, args=(distributed_values,))
result = run()
result
<tf.Tensor: shape=(), dtype=int32, numpy=2>

scope

View source

scope()

Returns a context manager selecting this Strategy as current.

Inside a with strategy.scope(): code block, this thread will use a variable creator set by strategy, and will enter its "cross-replica context".