tf.distribute.experimental.partitioners.MinSizePartitioner

Partitioner that allocates a minimum size per shard.

Inherits From: Partitioner

tf.distribute.experimental.partitioners.MinSizePartitioner(
    min_shard_bytes=(256 << 10), max_shards=1, bytes_per_string=16
)

Used in the notebooks

Used in the tutorials
Parameter server training with ParameterServerStrategy

This partitioner ensures each shard has at least min_shard_bytes, and tries to allocate as many shards as possible, i.e., keeping shard size as small as possible. The maximum number of such shards (upper bound) is given by max_shards.

Examples:

partitioner = MinSizePartitioner(min_shard_bytes=4, max_shards=2)
partitions = partitioner(tf.TensorShape([6, 1]), tf.float32)
[2, 1]
partitioner = MinSizePartitioner(min_shard_bytes=4, max_shards=10)
partitions = partitioner(tf.TensorShape([6, 1]), tf.float32)
[6, 1]

# use in ParameterServerStrategy
# strategy = tf.distribute.experimental.ParameterServerStrategy(
#   cluster_resolver=cluster_resolver, variable_partitioner=partitioner)

Args
min_shard_bytes	Minimum bytes of each shard. Defaults to 256K.
max_shards	Upper bound on the number of shards. Defaults to 1.
bytes_per_string	If the partition value is of type string, this provides an estimate of how large each string is.

Methods

__call__

View source

__call__(
    shape, dtype, axis=0
)

Partitions the given shape and returns the partition results.

Examples of a partitioner that allocates a fixed number of shards:

partitioner = FixedShardsPartitioner(num_shards=2)
partitions = partitioner(tf.TensorShape([10, 3], tf.float32), axis=0)
print(partitions) # [2, 0]

Args
shape	a tf.TensorShape, the shape to partition.
dtype	a tf.dtypes.Dtype indicating the type of the partition value.
axis	The axis to partition along. Default: outermost axis.

Returns
A list of integers representing the number of partitions on each axis, where i-th value correponds to i-th axis.