tff.simulation.compose_dataset_computation

Builds a new iterative process which constructs datasets on clients.

Given a tff.Computation that returns a tf.data.Dataset, and a tff.templates.IterativeProcess where exactly one of the arguments is a dataset placed on clients of the same type as returned by the tff.Computation, this function will construct a new tff.templates.IterativeProcess whose next function accepts a federated set of values of the same type as the parameter of the dataset_computation, maps dataset_computation over these values, and proceeds with the body of process.next.

For example, if the type signature of dataset_computation is:

(T -> U*)

and the type signature of process.next is:

(<S, {U*}@CLIENTS> -> <S, V>

then the returned tff.templates.IterativeProcess.next type signature will be:

(<S, {T}@CLIENTS> -> <S, V>)

This functionality is useful in several settings:

  • We may want to push some dataset preprocessing to happen on the clients, as opposed to preprocessing happening on the TFF simultation controller. This may be necessary, e.g., in the case that we want to shuffle client examples.
  • We may want to construct the entire dataset on the clients, given a client id. This may be useful in order to speed up distributed simulations, in order to remove a linear cost incurred in constructing and serializign the datasets on the controller.

dataset_computation An instance of tff.Computation which accepts some parameter and returns an element of tff.SequenceType.
process An instance of tff.templates.IterativeProcess whose next function accepts exactly one federated dataset, IE, element of type {B*}@CLIENTS, where B is equivalent to the return type of dataset_computation.

A new tff.templates.IterativeProcess satisfying the specification above.

TypeError If the arguments are of the wrong types, or their TFF type signatures are incompatible with the specification of this function.