tf.distribute.experimental.ParameterServerStrategy

View source on GitHub

An multi-worker tf.distribute strategy with parameter servers.

Inherits From: Strategy

tf.distribute.experimental.ParameterServerStrategy(
    cluster_resolver: tf.distribute.cluster_resolver.ClusterResolver,
    variable_partitioner: tf.distribute.experimental.partitioners.Partitioner = None
)

Used in the notebooks

Parameter server training is a common data-parallel method to scale up a machine learning model on multiple machines. A parameter server training cluster consists of workers and parameter servers. Variables are created on parameter servers and they are read and updated by workers in each step. By default, workers read and update these variables independently without synchronizing with each other. Under this configuration, it is known as asynchronous training.

In TensorFlow 2, we recommend an architecture based on central coordination for parameter server training. Each worker and parameter server runs a tf.distribute.Server, and on top of that, a coordinator task is responsible for creating resources on workers and parameter servers, dispatching functions, and coordinating the training. The coordinator uses a tf.distribute.experimental.coordinator.ClusterCoordinator to coordinate the cluster, and a tf.distribute.experimental.ParameterServerStrategy to define variables on parameter servers and computation on workers.

For the training to work, the coordinator dispatches tf.functions to be executed on remote workers. Upon receiving requests from the coordinator, a worker executes the tf.function by reading the variables from parameter servers, executing the ops, and updating the variables on the parameter servers. Each of the worker only processes the requests from the coordinator, and communicates with parameter servers, without direct interactions with other workers in the cluster.

As a result, failures of some workers do not prevent the cluster from continuing the work, and this allows the cluster to train with instances that can be occasionally unavailable (e.g. preemptible or spot instances). The coordinator and parameter servers though, must be available at all times for the cluster to make progress.

Note that the coordinator is not one of the training workers. Instead, it creates resources such as variables and datasets, dispatches tf.functions, saves checkpoints and so on. In addition to workers, parameter servers and the coordinator, an optional evaluator can be run on the side that periodically reads the checkpoints saved by the coordinator and runs evaluations against each checkpoint.

ParameterServerStrategy is supported with two training APIs: Custom Training Loop (CTL) and Keras Training API, also known as Model.fit. CTL is recommended when users prefer to define the details of their training loop, and Model.fit is recommended when users prefer a high-level abstraction and handling of training.

When using a CTL, ParameterServerStrategy has to work in conjunction with a tf.distribute.experimental.coordinator.ClusterCoordinator object.

When using Model.fit, currently only the tf.keras.utils.experimental.DatasetCreator input type is supported.

Example code for coordinator

This section provides code snippets that are intended to be run on (the only) one task that is designated as the coordinator. Note that cluster_resolver, variable_partitioner, and dataset_fn arguments are explained in the following "Cluster setup", "Variable partitioning", and "Dataset preparation" sections.

With a CTL,

# Prepare a strategy to use with the cluster and variable partitioning info.
strategy = tf.distribute.experimental.ParameterServerStrategy(
    cluster_resolver=...,
    variable_partitioner=...)
coordinator = tf.distribute.experimental.coordinator.ClusterCoordinator(
    strategy=strategy)

# Prepare a distribute dataset that will place datasets on the workers.
distributed_dataset = coordinator.create_per_worker_dataset(dataset_fn=...)

with strategy.scope():
  model = ...
  optimizer, metrics = ...  # Keras optimizer/metrics are great choices
  checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
  checkpoint_manager = tf.train.CheckpointManager(
      checkpoint, checkpoint_dir, max_to_keep=2)
  # `load_checkpoint` infers initial epoch from `optimizer.iterations`.
  initial_epoch = load_checkpoint(checkpoint_manager) or 0

@tf.function
def worker_fn(iterator):

  def replica_fn(inputs):
    batch_data, labels = inputs
    # calculate gradient, applying gradient, metrics update etc.

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

for epoch in range(initial_epoch, num_epoch):
  distributed_iterator = iter(distributed_dataset)  # Reset iterator state.
  for step in range(steps_per_epoch):

    # Asynchronously schedule the `worker_fn` to be executed on an arbitrary
    # worker. This call returns immediately.
    coordinator.schedule(worker_fn, args=(distributed_iterator,))

  # `join` blocks until all scheduled `worker_fn`s finish execution. Once it
  # returns, we can read the metrics and save checkpoints as needed.
  coordinator.join()
  logging.info('Metric result: %r', metrics.result())
  train_accuracy.reset_states()
  checkpoint_manager.save()

With Model.fit,

# Prepare a strategy to use with the cluster and variable partitioning info.
strategy = tf.distribute.experimental.ParameterServerStrategy(
    cluster_resolver=...,
    variable_partitioner=...)

# A dataset function takes a `input_context` and returns a `Dataset`
def dataset_fn(input_context):
  dataset = tf.data.Dataset.from_tensors(...)
  return dataset.repeat().shard(...).batch(...).prefetch(...)

# With `Model.fit`, a `DatasetCreator` needs to be used.
input = tf.keras.utils.experimental.DatasetCreator(dataset_fn=...)

with strategy.scope():
  model = ...  # Make sure the `Model` is created within scope.
model.compile(optimizer="rmsprop", loss="mse", steps_per_execution=..., ...)

# Optional callbacks to checkpoint the model, back up the progress, etc.
callbacks = [tf.keras.callbacks.ModelCheckpoint(...), ...]

# `steps_per_epoch` is required with `ParameterServerStrategy`.
model.fit(input, epochs=..., steps_per_epoch=..., callbacks=callbacks)

Example code for worker and parameter servers

In addition to the coordinator, there should be tasks designated as "worker" or "ps". They should run the following code to start a TensorFlow server, waiting for coordinator's requests:

# Provide a `tf.distribute.cluster_resolver.ClusterResolver` that serves
# the cluster information. See below "Cluster setup" section.
cluster_resolver = ...

server = tf.distribute.Server(
    cluster_resolver.cluster_spec(),
    job_name=cluster_resolver.task_type,
    task_index=cluster_resolver.task_id,
    protocol="grpc")

# Blocking the process that starts a server from exiting.
server.join()

Cluster setup

In order for the tasks in the cluster to know other tasks' addresses, a tf.distribute.cluster_resolver.ClusterResolver is required to be used in coordinator, worker, and ps. The tf.distribute.cluster_resolver.ClusterResolver is responsible for providing the cluster information, as well as the task type and id of the current task. See tf.distribute.cluster_resolver.ClusterResolver for more information.

If TF_CONFIG environment variable is set, a tf.distribute.cluster_resolver.TFConfigClusterResolver should be used as well.

Since there are assumptions in tf.distribute.experimental.ParameterServerStrategy around the naming of the task types, "chief", "ps", and "worker" should be used in the tf.distribute.cluster_resolver.ClusterResolver to refer to the coordinator, parameter servers, and workers, respectively.

The following example demonstrates setting TF_CONFIG for the task designated as a parameter server (task type "ps") and index 1 (the second task), in a cluster with 1 chief, 2 parameter servers, and 3 workers. Note that it needs to be set before the use of tf.distribute.cluster_resolver.TFConfigClusterResolver.

Example code for cluster setup:

os.environ['TF_CONFIG'] = '''
{
  "cluster": {
    "chief": ["chief.example.com:2222"],
    "ps": ["ps0.example.com:2222", "ps1.example.com:2222"],
    "worker": ["worker0.example.com:2222", "worker1.example.com:2222",
               "worker2.example.com:2222"]
  },
  "task": {
    "type": "ps",
    "index": 1
  }
}
'''

If you prefer to run the same binary for all tasks, you will need to let the binary branch into different roles at the beginning of the program:

# If coordinator, create a strategy and start the training program.
if cluster_resolver.task_type == 'chief':
  strategy = tf.distribute.experimental.ParameterServerStrategy(
      cluster_resolver)
  ...

# If worker/ps, create a server
elif cluster_resolver.task_type in ("worker", "ps"):
  server = tf.distribute.Server(...)
  ...

Alternatively, you can also start a bunch of TensorFlow servers in advance and connect to them later. The coordinator can be in the same cluster or on any machine that has connectivity to workers and parameter servers. This is covered in our guide and tutorial.

Variable creation with strategy.scope()

tf.distribute.experimental.ParameterServerStrategy follows the tf.distribute API contract where variable creation is expected to be inside the context manager returned by strategy.scope(), in order to be correctly placed on parameter servers in a round-robin manner:

# In this example, we're assuming having 3 ps.
strategy = tf.distribute.experimental.ParameterServerStrategy(
    cluster_resolver=...)
coordinator = tf.distribute.experimental.coordinator.ClusterCoordinator(
    strategy=strategy)

# Variables should be created inside scope to be placed on parameter servers.
# If created outside scope such as `v1` here, it would be placed on the
# coordinator.
v1 = tf.Variable(initial_value=0.0)

with strategy.scope():
  v2 = tf.Variable(initial_value=1.0)
  v3 = tf.Variable(initial_value=2.0)
  v4 = tf.Variable(initial_value=3.0)
  v5 = tf.Variable(initial_value=4.0)

# v2 through v5 are created in scope and are distributed on parameter servers.
# Default placement is round-robin but the order should not be relied on.
assert v2.device == "/job:ps/replica:0/task:0/device:CPU:0"
assert v3.device == "/job:ps/replica:0/task:1/device:CPU:0"
assert v4.device == "/job:ps/replica:0/task:2/device:CPU:0"
assert v5.device == "/job:ps/replica:0/task:0/device:CPU:0"

See distribute.Strategy.scope for more information.

Variable partitioning

Having dedicated servers to store variables means being able to divide up, or "shard" the variables across the ps. Partitioning large variable among ps is a commonly used technique to boost training throughput and mitigate memory constraints. It enables parallel computations and updates on different shards of a variable, and often yields better load balancing across parameter servers. Without sharding, models with large variables (e.g, embeddings) that can't fit into one machine's memory would otherwise be unable to train.

With tf.distribute.experimental.ParameterServerStrategy, if a variable_partitioner is provided to __init__ and certain conditions are satisfied, the resulting variables created in scope are sharded across the parameter servers, in a round-robin fashion. The variable reference returned from tf.Variable becomes a type that serves as the container of the sharded variables. One can access variables attribute of this container for the actual variable components. If building model with tf.Module or Keras, the variable components are collected in the variables alike attributes.

It is recommended to use size-based partitioners like tf.distribute.experimental.partitioners.MinSizePartitioner to avoid partitioning small variables, which could have negative impact on model training speed.

# Partition the embedding layer into 2 shards.
variable_partitioner = (
  tf.distribute.experimental.partitioners.MinSizePartitioner(
    min_shard_bytes=(256 << 10),
    max_shards = 2))
strategy = tf.distribute.experimental.ParameterServerStrategy(
  cluster_resolver=...,
  variable_partitioner = variable_partitioner)
with strategy.scope():
  embedding = tf.keras.layers.Embedding(input_dim=1024, output_dim=1024)
assert len(embedding.variables) == 2
assert isinstance(embedding.variables[0], tf.Variable)
assert isinstance(embedding.variables[1], tf.Variable)
assert embedding.variables[0].shape == (512, 1024)
assert embedding.variables[1].shape == (512, 1024)

The sharded variable container can be converted to a Tensor via tf.convert_to_tensor. This means the container can be directly used in most Python Ops where such Tensor conversion automatically happens. For example, in the above code snippet, x * self.w would implicitly apply the said tensor conversion. Note that such conversion can be expensive, as the variable components need to be transferred from multiple parameter servers to where the value is used.

tf.nn.embedding_lookup on the other hand doesn't apply the tensor conversion, and performs parallel lookups on the variable components instead. This is crucial to scale up embedding lookups when the embedding table variable is large.

When a partitioned variable is saved to a SavedModel, it will be saved as if it is one single variable. This improves serving efficiency by eliminating a number of Ops that handle the partiton aspects.

Known limitations of variable partitioning:

• Number of partitions must not change across Checkpoint saving/loading.

• After saving partitioned variables to a SavedModel, the SavedModel can't be loaded via tf.saved_model.load.

• Partition variable doesn't directly work with tf.GradientTape, please use the variables attributes to get the actual variable components and use them in gradient APIs instead.

Dataset preparation

With tf.distribute.experimental.ParameterServerStrategy, a dataset is created in each of the workers to be used for training. This is done by creating a dataset_fn that takes no argument and returns a tf.data.Dataset, and passing the dataset_fn into tf.distribute.experimental.coordinator. ClusterCoordinator.create_per_worker_dataset. We recommend the dataset to be shuffled and repeated to have the examples run through the training as evenly as possible.

def dataset_fn():
  filenames = ...
  dataset = tf.data.Dataset.from_tensor_slices(filenames)

  # Dataset is recommended to be shuffled, and repeated.
  return dataset.shuffle(buffer_size=...).repeat().batch(batch_size=...)

coordinator =
    tf.distribute.experimental.coordinator.ClusterCoordinator(strategy=...)
distributed_dataset = coordinator.create_per_worker_dataset(dataset_fn)

Limitations

• tf.distribute.experimental.ParameterServerStrategy in TF2 is experimental, and the API is subject to further changes.

• When using Model.fit, tf.distribute.experimental.ParameterServerStrategy must be used with a tf.keras.utils.experimental.DatasetCreator, and steps_per_epoch must be specified.

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

distribute_datasets_from_function

View source

distribute_datasets_from_function(
    dataset_fn, options=None
)

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

The argument dataset_fn that users pass in is an input function that has a tf.distribute.InputContext argument and returns a tf.data.Dataset instance. It is expected that the returned dataset from dataset_fn is already batched by per-replica batch size (i.e. global batch size divided by the number of replicas in sync) and sharded. tf.distribute.Strategy.distribute_datasets_from_function does not batch or shard the tf.data.Dataset instance returned from the input function. dataset_fn will be called on the CPU device of each of the workers and each generates a dataset where every replica on that worker will dequeue one batch of inputs (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. First, it allows you to specify your own batching and sharding logic. (In contrast, tf.distribute.experimental_distribute_dataset does batching and sharding for you.) 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.

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

You can use 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. Follow tf.distribute.DistributedDataset.element_spec to see an example.

For a tutorial on more usage and properties of this method, refer to the tutorial on distributed input). If you are interested in last partial batch handling, read this section.

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 regular datasets. NOTE: The user cannot add any more transformations to a tf.distribute.DistributedDataset. You can only create an iterator or examine the tf.TypeSpec of the data generated by it. See API docs of tf.distribute.DistributedDataset to learn more.

The following is an example:

global_batch_size = 2
# Passing the devices is optional.
strategy = tf.distribute.MirroredStrategy(devices=["GPU:0", "GPU:1"])
# Create a dataset
dataset = tf.data.Dataset.range(4).batch(global_batch_size)
# Distribute that dataset
dist_dataset = strategy.experimental_distribute_dataset(dataset)
@tf.function
def replica_fn(input):
  return input*2
result = []
# Iterate over the `tf.distribute.DistributedDataset`
for x in dist_dataset:
  # process dataset elements
  result.append(strategy.run(replica_fn, args=(x,)))
print(result)
[PerReplica:{
  0: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([0])>,
  1: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([2])>
}, PerReplica:{
  0: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([4])>,
  1: <tf.Tensor: shape=(1,), dtype=int64, numpy=array([6])>
}]

Three key actions happening under the hood of this method are batching, sharding, and prefetching.

In the code snippet above, dataset is batched by global_batch_size, and calling experimental_distribute_dataset on it rebatches dataset to a new batch size that is equal to the global batch size divided by the number of replicas in sync. We iterate through it using a Pythonic for loop. x is a tf.distribute.DistributedValues containing data for all replicas, and each replica gets data of the new 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.

Sharding contains autosharding across multiple workers and within every worker. First, in multi-worker distributed training (i.e. when you use tf.distribute.experimental.MultiWorkerMirroredStrategy or tf.distribute.TPUStrategy), autosharding a dataset over a set of workers means that each worker is assigned a subset of the entire dataset (if the right tf.data.experimental.AutoShardPolicy is set). This is to ensure that at each step, a global batch size of non-overlapping dataset elements will be processed by each worker. Autosharding has a couple of different options that can be specified using tf.data.experimental.DistributeOptions. Then, sharding within each worker means the method will split the data among all the worker devices (if more than one a present). This will happen regardless of multi-worker autosharding.

By default, this method adds a prefetch transformation at the end of the user provided tf.data.Dataset instance. The argument to the prefetch transformation which is buffer_size is equal to the number of replicas in sync.

If the above batch splitting and dataset sharding logic is undesirable, please use tf.distribute.Strategy.distribute_datasets_from_function instead, which does not do any automatic batching or sharding for you.

For a tutorial on more usage and properties of this method, refer to the tutorial on distributed input. If you are interested in last partial batch handling, read this section.

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(["GPU:0", "GPU:1"])
   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>,
       <tf.Tensor: shape=(), dtype=float32, numpy=1.0>)
       
   

2. Distribute values in array based on replica_id:

   strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
   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, 2.0)
       
   

3. Specify values using num_replicas_in_sync:

   strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
   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
       (2, 2)
       
   

4. 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.

gather

View source

gather(
    value, axis
)

Gather value across replicas along axis to the current device.

Given a tf.distribute.DistributedValues or tf.Tensor-like object value, this API gathers and concatenates value across replicas along the axis-th dimension. The result is copied to the "current" device, which would typically be the CPU of the worker on which the program is running. For tf.distribute.TPUStrategy, it is the first TPU host. For multi-client tf.distribute.MultiWorkerMirroredStrategy, this is the CPU of each worker.

This API can only be called in the cross-replica context. For a counterpart in the replica context, see tf.distribute.ReplicaContext.all_gather.

strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
# A DistributedValues with component tensor of shape (2, 1) on each replica
distributed_values = strategy.experimental_distribute_values_from_function(lambda _: tf.identity(tf.constant([[1], [2]])))
@tf.function
def run():
  return strategy.gather(distributed_values, axis=0)
run()
<tf.Tensor: shape=(4, 1), dtype=int32, numpy=
array([[1],
       [2],
       [1],
       [2]], dtype=int32)>

Consider the following example for more combinations:

strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1", "GPU:2", "GPU:3"])
single_tensor = tf.reshape(tf.range(6), shape=(1,2,3))
distributed_values = strategy.experimental_distribute_values_from_function(lambda _: tf.identity(single_tensor))
@tf.function
def run(axis):
  return strategy.gather(distributed_values, axis=axis)
axis=0
run(axis)
<tf.Tensor: shape=(4, 2, 3), dtype=int32, numpy=
array([[[0, 1, 2],
        [3, 4, 5]],
       [[0, 1, 2],
        [3, 4, 5]],
       [[0, 1, 2],
        [3, 4, 5]],
       [[0, 1, 2],
        [3, 4, 5]]], dtype=int32)>
axis=1
run(axis)
<tf.Tensor: shape=(1, 8, 3), dtype=int32, numpy=
array([[[0, 1, 2],
        [3, 4, 5],
        [0, 1, 2],
        [3, 4, 5],
        [0, 1, 2],
        [3, 4, 5],
        [0, 1, 2],
        [3, 4, 5]]], dtype=int32)>
axis=2
run(axis)
<tf.Tensor: shape=(1, 2, 12), dtype=int32, numpy=
array([[[0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2],
        [3, 4, 5, 3, 4, 5, 3, 4, 5, 3, 4, 5]]], dtype=int32)>

reduce

View source

reduce(
    reduce_op, value, axis
)

Reduce value across replicas and return result on current device.

strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
def step_fn():
  i = tf.distribute.get_replica_context().replica_id_in_sync_group
  return tf.identity(i)

per_replica_result = strategy.run(step_fn)
total = strategy.reduce("SUM", per_replica_result, axis=None)
total
<tf.Tensor: shape=(), dtype=int32, numpy=1>

To see how this would look with multiple replicas, consider the same example with MirroredStrategy with 2 GPUs:

strategy = tf.distribute.MirroredStrategy(devices=["GPU:0", "GPU:1"])
def step_fn():
  i = tf.distribute.get_replica_context().replica_id_in_sync_group
  return tf.identity(i)

per_replica_result = strategy.run(step_fn)
# Check devices on which per replica result is:
strategy.experimental_local_results(per_replica_result)[0].device
# /job:localhost/replica:0/task:0/device:GPU:0
strategy.experimental_local_results(per_replica_result)[1].device
# /job:localhost/replica:0/task:0/device:GPU:1

total = strategy.reduce("SUM", per_replica_result, axis=None)
# Check device on which reduced result is:
total.device
# /job:localhost/replica:0/task:0/device:CPU:0

This API is typically used for aggregating the results returned from different replicas, for reporting etc. For example, loss computed from different replicas can be averaged using this API before printing.

There are a number of different tf.distribute APIs for reducing values across replicas:

• tf.distribute.ReplicaContext.all_reduce: This differs from Strategy.reduce in that it is for replica context and does not copy the results to the host device. all_reduce should be typically used for reductions inside the training step such as gradients.
• tf.distribute.StrategyExtended.reduce_to and tf.distribute.StrategyExtended.batch_reduce_to: These APIs are more advanced versions of Strategy.reduce as they allow customizing the destination of the result. They are also called in cross replica context.

What should axis be?

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 by specifying the axis parameter accordingly.

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. With axis=None, reduce will aggregate only 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 or loss).

strategy.reduce("sum", per_replica_result, axis=None)

Sometimes, you will want to aggregate across both the global batch and all replicas. You can get this behavior 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.

strategy.reduce("sum", per_replica_result, axis=0)

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
)

Invokes fn on each replica, with the given arguments.

This method is the primary way to distribute your computation with a tf.distribute object. It invokes 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.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 is invoked under a replica context. fn may call tf.distribute.get_replica_context() to access members such as all_reduce. Please see the module-level docstring of tf.distribute for the concept of replica context.

All arguments in args or kwargs can be a nested structure of tensors, e.g. a list of tensors, in which case args and kwargs will be passed to the fn invoked on each replica. Or args or kwargs can be tf.distribute.DistributedValues containing tensors or composite tensors, i.e. tf.compat.v1.TensorInfo.CompositeTensor, in which case each fn call will get the component of a tf.distribute.DistributedValues corresponding to its replica. Note that arbitrary Python values that are not of the types above are not supported.

Example usage:

1. Constant tensor input.

   strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
   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
       PerReplica:{
         0: <tf.Tensor: shape=(), dtype=float32, numpy=6.0>,
         1: <tf.Tensor: shape=(), dtype=float32, numpy=6.0>
       }
       
   

2. DistributedValues input.

   strategy = tf.distribute.MirroredStrategy(["GPU:0", "GPU:1"])
   @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=4>
       
   

3. Use tf.distribute.ReplicaContext to allreduce values.

   strategy = tf.distribute.MirroredStrategy(["gpu:0", "gpu:1"])
   @tf.function
   def run():
      def value_fn(value_context):
        return tf.constant(value_context.replica_id_in_sync_group)
      distributed_values = (
          strategy.experimental_distribute_values_from_function(
              value_fn))
      def replica_fn(input):
        return tf.distribute.get_replica_context().all_reduce(
            "sum", input)
      return strategy.run(replica_fn, args=(distributed_values,))
   result = run()
   result
       PerReplica:{
         0: <tf.Tensor: shape=(), dtype=int32, numpy=1>,
         1: <tf.Tensor: shape=(), dtype=int32, numpy=1>
       }
       
   

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(["GPU:0", "GPU:1"])
# 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>,
  1: <tf.Variable 'Variable/replica_1:0' 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.

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 called in a strategy.scope. This can be accomplished either by directly calling the variable creating function within the scope context, or by relying on another API like strategy.run or keras.Model.fit to automatically enter it for you. Any variable that is created outside scope will not be distributed and may have performance implications. Some common objects that create variables in TF are Models, Optimizers, Metrics. Such objects should always be initialized in the scope, and any functions that may lazily create variables (e.g., Model.call(), tracing a tf.function, etc.) should similarly be called within 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 automatically, which means when using those APIs you don't need to explicitly enter the scope yourself.
• When a tf.keras.Model is created inside a strategy.scope, the Model object captures the scope information. When high level training framework methods such as model.compile, model.fit, etc. are then called, the captured scope will be automatically entered, and the associated strategy will be used to distribute the training etc. See a detailed example in distributed keras tutorial. WARNING: Simply calling model(..) does not automatically enter the captured scope -- only high level training framework APIs support this behavior: 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.