Join TensorFlow at Google I/O, May 11-12 Register now


UnweightedAggregationFactory for fast Walsh-Hadamard transform.

Inherits From: UnweightedAggregationFactory

The created tff.templates.AggregationProcess takes an input structure and applies the randomized fast Walsh-Hadamard transform to each tensor in the structure, reshaped to a rank-1 tensor in O(d*log(d)) time, where d is the number of elements of the tensor.


Specifically, for each tensor, the following operations are first performed at tff.CLIENTS:

  1. Flattens the tensor into a rank-1 tensor.
  2. Pads the tensor to d_2 dimensions with zeros, where d_2 is the smallest power of 2 larger than or equal to d.
  3. Multiplies the padded tensor with random +1/-1 values (i.e. flipping the signs). This is equivalent to multiplication by a diagonal matrix with Rademacher random varaibles on diagonal.
  4. Applies the fast Walsh-Hadamard transform. Steps 3 and 4 are repeated multiple times with independent randomness, if num_repeats > 1.

The resulting tensors are passed to the inner_agg_factory. After aggregation, at tff.SEREVR, inverses of these steps are applied in reverse order.

The allowed input dtypes are integers and floats. However, the dtype passed to the inner_agg_factory will always be a float.



View source

Creates a tff.aggregators.AggregationProcess without weights.

The provided value_type is a non-federated tff.Type object, that is, value_type.is_federated() should return False.

The returned tff.aggregators.AggregationProcess will be created for aggregation of values matching value_type placed at tff.CLIENTS. That is, its next method will expect type <S@SERVER, {value_type}@CLIENTS>, where S is the unplaced return type of its initialize method.

value_type A non-federated tff.Type (value_type.is_federated() returns False).

A tff.templates.AggregationProcess.