tf.contrib.distribute.AllReduceCrossTowerOps

Class AllReduceCrossTowerOps

Inherits From: CrossTowerOps

Defined in tensorflow/contrib/distribute/python/cross_tower_ops.py.

Reduction using all reduce.

__init__

__init__(
    all_reduce_alg='nccl',
    num_packs=1,
    agg_small_grads_max_bytes=0,
    agg_small_grads_max_group=10
)

All-reduce implementation of CrossTowerOps.

Before performing all-reduce, tensors will be repacked or aggregated for more efficient cross-device transportation: 1) If num_packs is non-zero, pack values into num_packs splits. 2) Otherwise, if agg_small_grads_max_bytes > 0 and agg_small_grads_max_group > 0, aggregate values smaller than agg_small_grads_max_bytes into groups with at most agg_small_grads_max_group values. 3) Otherwise, no repacking or grouping will happen.

Args:

  • all_reduce_alg: the all-reduce algorithm to use, currently only "nccl" or "hierarchical_copy" are supported.
  • num_packs: see above.
  • agg_small_grads_max_bytes: see above.
  • agg_small_grads_max_group: see above. tensors.

Methods

batch_reduce

batch_reduce(
    aggregation,
    value_destination_pairs
)

Reduce PerDevice objects in a batch.

Reduce each first element in value_destination_pairs to each second element which indicates the destinations.

Args:

  • aggregation: Indicates how a variable will be aggregated. Accepted values are tf.VariableAggregation.SUM, tf.VariableAggregation.MEAN.
  • value_destination_pairs: a list or a tuple of tuples of PerDevice objects (or tensors with device set if there is one tower) and destinations.

Returns:

a list of Mirrored objects.

Raises:

  • ValueError: if value_destination_pairs is not a list or a tuple of tuples of PerDevice objects and destinations

broadcast

broadcast(
    tensor,
    destinations
)

Broadcast the tensor to destinations.

Args:

  • tensor: the tensor to broadcast.
  • destinations: the broadcast destinations.

Returns:

a Mirrored object.

reduce

reduce(
    aggregation,
    per_device_value,
    destinations
)

Reduce per_device_value to destinations.

It runs the reduction operation defined by aggregation and put the result on destinations.

Args:

Returns:

a Mirrored object.

Raises:

  • ValueError: if per_device_value is not a PerDevice object.