tff.program.FederatedDataSourceIterator

Stay organized with collections Save and categorize content based on your preferences.

An abstract interface for representing federated data source iterators.

This interface abstracts away the specifics of iterating over data in a data source.

Things one can do with a data source iterator:

  • Determine the type of the data supplied by this iterator by inspecting the federated_type property. The type returned must match that of the data source that returned this iterator.

  • Return a new selection of federated data from the iterator by invoking select.

Please see tff.program.FederatedDataSource for additional context and the high-level description of how to use data sources.

federated_type The type of the data returned by calling select.

Methods

from_bytes

View source

@classmethod
@abc.abstractmethod
from_bytes(
    buffer: bytes
) -> 'Serializable'

Deserializes the object from bytes.

select

View source

@abc.abstractmethod
select(
    num_clients: Optional[int] = None
) -> object

Returns a new selection of federated data from this iterator.

The selection contains data distributed across a cohort of logical clients. The manner in which this cohort is constructed is a function of the arguments supplied here by the caller (such as num_clients) as well as by the capabilities of the data source itself (e.g., whether it offers data selected uniformly at random, performs sampling without replacement, guarantees determinism, etc.).

Args
num_clients Optional, the number of clients to select. Must be a positive integer, or None if unspecified.

Returns
An object of type federated_type representing the selected data, and that can be supplied as an argument to a tff.Computation. See tff.program.FederatedContext for more information about these types.

to_bytes

View source

@abc.abstractmethod
to_bytes() -> bytes

Serializes the object to bytes.