tff.program.FederatedContext

An abstract interface representing a federated context.

Inherits From: SyncContext

A federated context supports invoking a limited set of tff.Computations, making guarantees about what a tff.Computation can accept as an argument and what it returns when invoked.

Restrictions on the TensorFlow Federated Type

Arguments can be nested structures of values corresponding to the TensorFlow Federated type signature of the tff.Computation:

Server-placed values must be represented by tff.program.MaterializableStructure.
Client-placed values must be represented by structures of values returned by a tff.program.FederatedDataSourceIterator.

Return values can be structures of tff.program.MaterializableValueReferences or a single tff.program.MaterializableValueReference, where a reference corresponds to the tensor-type of the TensorFlow Federated type signature in the return value of the invoked tff.Computation.

TensorFlow Federated Type to Python Representation

In order to interact with the value returned by a tff.Computation, it is helpful to be able to reason about the Python type of this value. In some way this Python type must depend on the TensorFlow Federated type signature of the associated value. To provide uniformity of experience and ease of reasoning, we specify the Python representation of values in a manner that can be stated entirely in the TensorFlow Federated typesystem.

We have chosen to limit the TensorFlow Federated type signatures of invoked tff.Computations to disallow the returning of client-placed values, tff.SequenceTypes, and tff.FunctionTypes, in order to reduced the area which needs to be supported by federated programs. Below we describe the mapping between TensorFlow Federated type signatures and Python representations of values that can be passed as arguments to or returned as results from tff.Computations.

Python representations of values that can be accepted as an arguments to or returned as a value from a tff.Computation:

TensorFlow Federated Type	Python Representation
`tff.TensorType`	`tff.program.MaterializableValueReference`
`tff.SequenceType`	`tff.program.MaterializableValueReference`
`tff.FederatedType` (server-placed)	Python representation of the `member` of the `tff.FederatedType`
`tff.StructWithPythonType`	Python container of the `tff.StructWithPythonType`
`tff.StructType` (with no Python type, all fields named)	`collections.OrderedDict`
`tff.StructType` (with no Python type, no fields named)	`tuple`

Python representations of values that can be only be accepted as an arguments to a tff.Computation:

TFF Type	Python Representation
`tff.FederatedType` (client-placed)	Opaque object returned by `tff.program.DataSourceIterator.select`
`tff.FunctionType`	`tff.Computation`

Methods

`invoke`

View source

@abc.abstractmethod
invoke(
    comp: tff.Computation,
    arg: Optional[ComputationArgValue]
) -> structure_utils.Structure[value_reference.MaterializableValueReference]

Invokes the comp with the argument arg.

Args
`comp`	The `tff.Computation` being invoked.
`arg`	The optional argument of `comp`; server-placed values must be represented by `tff.program.MaterializableStructure`, and client-placed values must be represented by structures of values returned by a `tff.program.FederatedDataSourceIterator`.

Returns
The result of invocation; a structure of `tff.program.MaterializableValueReference`.

Raises
`ValueError`	If the result type of `comp` does not contain only structures, server-placed values, or tensors.