|View source on GitHub|
DatasetSplitFn for Federated Reconstruction training/evaluation.
tff.learning.reconstruction.build_dataset_split_fn( recon_epochs: int = 1, recon_steps_max: Optional[int] = None, post_recon_epochs: int = 1, post_recon_steps_max: Optional[int] = None, split_dataset: bool = False ) ->
DatasetSplitFn parameterizes training and evaluation computations
and enables reconstruction for multiple local epochs, multiple epochs of
post-reconstruction training, limiting the number of steps for both stages,
and splitting client datasets into disjoint halves for each stage.
Note that the returned function is used during both training and evaluation: during training, "post-reconstruction" refers to training of global variables, and during evaluation, it refers to calculation of metrics using reconstructed local variables and fixed global variables.
||The integer number of iterations over the dataset to make during reconstruction.|
If not None, the integer maximum number of steps (batches)
to iterate through during reconstruction. This maximum number of steps is
across all reconstruction iterations, i.e. it is applied after
||The integer constant number of iterations to make over client data after reconstruction.|
If not None, the integer maximum number of steps
(batches) to iterate through after reconstruction. This maximum number of
steps is across all post-reconstruction iterations, i.e. it is applied
If True, splits