org.platanios.tensorflow.api.ops.training.distribute.ops
All-reduce algorithm to use.
Performs a batched all-reduce operation.
Performs a batched all-reduce operation.
Reduction method.
Per-device values to be reduced.
Sequence of reduced mirrored values.
Reduces a batch per-device values to the corresponding provided destinations.
Reduces a batch per-device values to the corresponding provided destinations.
For each tuple in valueDestinationPairs
, this method reduces each first element to each second element which
indicates the corresponding destination.
Reduction method.
Sequence of per-device values and destinations pairs. If a destination is None
,
then the destination is set to match the devices of the corresponding per-device
value.
Sequence of reduced mirrored values.
Broadcasts value
to destination
.
Broadcasts value
to destination
.
Value to broadcast.
Broadcast destination.
Broadcasted mirrored value.
Packer that can be used to repack/aggregate values before applying the all-reduce, in order to allow for more efficient cross-device transportation.
Reduces value
to destination
.
Reduces value
to destination
.
It runs the reduction operation defined by reduction
and puts the result on destination
.
Reduction method.
Per-device value to be reduced.
Reduction destination.
Reduced mirrored value.
Cross-tower ops that perform an all-reduce operation.