tf.distribute.TPUStrategy

Synchronous training on TPUs and TPU Pods.

Inherits From: Strategy

tf.distribute.TPUStrategy(
    tpu_cluster_resolver=None, experimental_device_assignment=None
)

To construct a TPUStrategy object, you need to run the initialization code as below:

resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.TPUStrategy(resolver)

While using distribution strategies, the variables created within the strategy's scope will be replicated across all the replicas and can be kept in sync using all-reduce algorithms.

To run TF2 programs on TPUs, you can either use .compile and .fit APIs in tf.keras with TPUStrategy, or write your own customized training loop by calling strategy.run directly. Note that TPUStrategy doesn't support pure eager execution, so please make sure the function passed into strategy.run is a tf.function or strategy.run is called inside a tf.function if eager behavior is enabled. See more details in https://www.tensorflow.org/guide/tpu.

experimental_distribute_datasets_from_function and experimental_distribute_dataset APIs can be used to distribute the dataset across the TPU workers when writing your own training loop. If you are using fit and compile methods available in tf.keras.Model, then Keras will handle the distribution for you.

An example of writing customized training loop on TPUs:

with strategy.scope():
  model = tf.keras.Sequential([
    tf.keras.layers.Dense(2, input_shape=(5,)),
  ])
  optimizer = tf.keras.optimizers.SGD(learning_rate=0.1)

def dataset_fn(ctx):
  x = np.random.random((2, 5)).astype(np.float32)
  y = np.random.randint(2, size=(2, 1))
  dataset = tf.data.Dataset.from_tensor_slices((x, y))
  return dataset.repeat().batch(1, drop_remainder=True)
dist_dataset = strategy.experimental_distribute_datasets_from_function(
    dataset_fn)
iterator = iter(dist_dataset)

@tf.function()
def train_step(iterator):

  def step_fn(inputs):
    features, labels = inputs
    with tf.GradientTape() as tape:
      logits = model(features, training=True)
      loss = tf.keras.losses.sparse_categorical_crossentropy(
          labels, logits)

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

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

train_step(iterator)

For the advanced use cases like model parallelism, you can set experimental_device_assignment argument when creating TPUStrategy to specify number of replicas and number of logical devices. Below is an example to initialize TPU system with 2 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=[1, 1, 1, 2],
    num_replicas=1)
strategy = tf.distribute.TPUStrategy(
    resolver, experimental_device_assignment=device_assignment)

Then you can run a tf.add operation only on logical device 0.

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

  # Add operation will be executed on logical device 0.
  output = strategy.experimental_assign_to_logical_device(output, 0)
  return output
dist_dataset = strategy.experimental_distribute_datasets_from_function(
    dataset_fn)
iterator = iter(dist_dataset)
strategy.run(step_fn, args=(next(iterator),))

os.environ['TF_CONFIG'] = json.dumps({
'cluster': {
'worker': ["localhost:12345", "localhost:23456"],
'ps': ["localhost:34567"]
},
'task': {'type': 'worker', 'index': 0}
})

# This implicitly uses TF_CONFIG for the cluster and current task info.
strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()

...

if strategy.cluster_resolver.task_type == 'worker':
# Perform something that's only applicable on workers. Since we set this
# as a worker above, this block will run on this particular instance.
elif strategy.cluster_resolver.task_type == 'ps':
# Perform something that's only applicable on parameter servers. Since we
# set this as a worker above, this block will not run on this particular
# instance.

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.

Note: This API is only supported in TPUStrategy for now. This adds an annotation to tensor specifying that operations on tensor will be invoked on logical core device id logical_device_id. When model parallelism is used, the default behavior is that all ops are placed on zero-th 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, 1, 2],
    num_replicas=4)
strategy = tf.distribute.TPUStrategy(
    resolver, experimental_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, options=None
)

Creates tf.distribute.DistributedDataset from tf.data.Dataset.

The returned tf.distribute.DistributedDataset can be iterated over similar to how regular datasets can. NOTE: The user cannot add any more transformations to a tf.distribute.DistributedDataset.

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 `tf.distribute.DistributedDataset`
for x in dist_dataset:
  # process dataset elements
  strategy.run(replica_fn, args=(x,))

In the code snippet above, the tf.distribute.DistributedDataset dist_dataset is batched by GLOBAL_BATCH_SIZE, and we iterate through it using for x in dist_dataset. x a tf.distribute.DistributedValues containing data for all replicas, which aggregates to a batch of GLOBAL_BATCH_SIZE. tf.distribute.Strategy.run will take care of feeding the right per-replica data in x to the right replica_fn executed on each replica.

What's under the hood of this method, when we say the tf.data.Dataset instance - dataset - gets distributed? It depends on how you set the tf.data.experimental.AutoShardPolicy through tf.data.experimental.DistributeOptions. By default, it is set to tf.data.experimental.AutoShardPolicy.AUTO. In a multi-worker setting, we will first attempt to distribute dataset by detecting whether dataset is being created out of reader datasets (e.g. tf.data.TFRecordDataset, tf.data.TextLineDataset, etc.) and if so, try 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 disable dataset sharding across workers, by setting the tf.data.experimental.DistributeOptions.auto_shard_policy to be tf.data.experimental.AutoShardPolicy.OFF.

If the attempt to shard by file is unsuccessful (i.e. the dataset is not read from files), 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 route is selected.

As mentioned before, within each worker, we will also split the data among all the worker devices (if more than one a present). This will happen even if multi-worker sharding is disabled.

If the above batch splitting and dataset sharding logic is undesirable, please use tf.distribute.Strategy.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 tf.distribute.DistributedDataset instance 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 `tf.distribute.DistributedDataset`
for x in dist_dataset:
  # process dataset elements
  strategy.run(train_step, args=(x,))

Note: The order in which the data is processed by the workers when using tf.distribute.Strategy.experimental_distribute_dataset or tf.distribute.Strategy.experimental_distribute_datasets_from_function is not guaranteed. This is typically required if you are using tf.distribute to scale prediction. You can however insert an index for each element in the batch and order outputs accordingly. Refer to this snippet for an example of how to order outputs.

experimental_distribute_datasets_from_function

View source

experimental_distribute_datasets_from_function(
    dataset_fn, options=None
)

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.

You can also use the element_spec property of the tf.distribute.DistributedDataset 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.

global_batch_size = 8
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)

strategy = tf.distribute.MirroredStrategy()
ds = strategy.experimental_distribute_datasets_from_function(dataset_fn)

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

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

train(ds)

Note: The order in which the data is processed by the workers when using tf.distribute.Strategy.experimental_distribute_dataset or tf.distribute.Strategy.experimental_distribute_datasets_from_function is not guaranteed. This is typically required if you are using tf.distribute to scale prediction. You can however insert an index for each element in the batch and order outputs accordingly. Refer to this snippet for an example of how to order outputs.

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_in_sync_group]

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.

Note: This only returns values on the worker initiated by this client. When using a tf.distribute.Strategy like tf.distribute.experimental.MultiWorkerMirroredStrategy, each worker will be its own client, and this function will only return values computed on that worker.

experimental_make_numpy_dataset

View source

experimental_make_numpy_dataset(
    numpy_input
)

Makes a tf.data.Dataset from a numpy array. (deprecated)

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:

strategy = tf.distribute.MirroredStrategy()
numpy_input = np.ones([10], dtype=np.float32)
dataset = strategy.experimental_make_numpy_dataset(numpy_input)
dataset
<TensorSliceDataset shapes: (), types: tf.float32>
dataset = dataset.batch(2)
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.

Note: This API is only supported in TPUStrategy for now. This adds an annotation to tensor tensor specifying that operations on tensor will be invoked on 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, 1, 2],
    num_replicas=4)
strategy = tf.distribute.TPUStrategy(
    resolver, experimental_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.

Note: This API is only supported in TPUStrategy for now. This adds an annotation to tensor tensor specifying that operations on tensor will be be split among multiple logical devices. Tensor tensor will be split across dimensions specified by partition_dimensions. The dimensions of tensor must be divisible by corresponding value in partition_dimensions.

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=[1, 2, 2, 2],
    num_replicas=1)
strategy = tf.distribute.TPUStrategy(
    resolver, experimental_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 the computation defined by fn on each TPU replica.

Executes ops specified by fn on each replica. If args or kwargs have tf.distribute.DistributedValues, such as those produced by a tf.distribute.DistributedDataset from tf.distribute.Strategy.experimental_distribute_dataset or tf.distribute.Strategy.experimental_distribute_datasets_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:

resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu='')
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.TPUStrategy(resolver)
@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_fn(input):
    return input * 2
  return strategy.run(replica_fn, args=(distributed_values,))
result = run()

scope

View source

scope()

Context manager to make the strategy current and distribute variables.

This method returns a context manager, and is used as follows:

strategy = tf.distribute.MirroredStrategy()
# Variable created inside scope:
with strategy.scope():
  mirrored_variable = tf.Variable(1.)
mirrored_variable
MirroredVariable:{
  0: <tf.Variable 'Variable:0' shape=() dtype=float32, numpy=1.0>
}
# Variable created outside scope:
regular_variable = tf.Variable(1.)
regular_variable
<tf.Variable 'Variable:0' shape=() dtype=float32, numpy=1.0>

What happens when Strategy.scope is entered?

• strategy is installed in the global context as the "current" strategy. Inside this scope, tf.distribute.get_strategy() will now return this strategy. Outside this scope, it returns the default no-op strategy.
• Entering the scope also enters the "cross-replica context". See tf.distribute.StrategyExtended for an explanation on cross-replica and replica contexts.
• Variable creation inside scope is intercepted by the strategy. Each strategy defines how it wants to affect the variable creation. Sync strategies like MirroredStrategy, TPUStrategy and MultiWorkerMiroredStrategy create variables replicated on each replica, whereas ParameterServerStrategy creates variables on the parameter servers. This is done using a custom tf.variable_creator_scope.
• In some strategies, a default device scope may also be entered: in MultiWorkerMiroredStrategy, a default device scope of "/CPU:0" is entered on each worker.

Note: Entering a scope does not automatically distribute a computation, except in the case of high level training framework like keras model.fit. If you're not using model.fit, you need to use strategy.run API to explicitly distribute that computation. See an example in the custom training loop tutorial.

What should be in scope and what should be outside?

There are a number of requirements on what needs to happen inside the scope. However, in places where we have information about which strategy is in use, we often enter the scope for the user, so they don't have to do it explicitly (i.e. calling those either inside or outside the scope is OK).

• Anything that creates variables that should be distributed variables must be in strategy.scope. This can be either by directly putting it in scope, or relying on another API like strategy.run or model.fit to enter it for you. Any variable that is created outside scope will not be distributed and may have performance implications. Common things that create variables in TF: models, optimizers, metrics. These should always be created inside the scope. Another source of variable creation can be a checkpoint restore - when variables are created lazily. Note that any variable created inside a strategy captures the strategy information. So reading and writing to these variables outside the strategy.scope can also work seamlessly, without the user having to enter the scope.
• Some strategy APIs (such as strategy.run and strategy.reduce) which require to be in a strategy's scope, enter the scope for you automatically, which means when using those APIs you don't need to enter the scope yourself.
• When a tf.keras.Model is created inside a strategy.scope, we capture this information. When high level training frameworks methods such as model.compile, model.fit etc are then called on this model, we automatically enter the scope, as well as use this strategy to distribute the training etc. See detailed example in distributed keras tutorial. Note that simply calling the model(..) is not impacted - only high level training framework APIs are. model.compile, model.fit, model.evaluate, model.predict and model.save can all be called inside or outside the scope.
• The following can be either inside or outside the scope: ** Creating the input datasets ** Defining tf.functions that represent your training step ** Saving APIs such as tf.saved_model.save. Loading creates variables, so that should go inside the scope if you want to train the model in a distributed way. ** Checkpoint saving. As mentioned above - checkpoint.restore may sometimes need to be inside scope if it creates variables.