tff.learning.metrics.SumThenFinalizeFactory

Aggregation Factory that sums and then finalizes the metrics.

Inherits From: UnweightedAggregationFactory

The created tff.templates.AggregationProcess uses the inner summation process created by the inner summation factory to sum unfinalized metrics from tff.CLIENTS to tff.SERVER, accumulates the summed unfinalized metrics in the state, and then finalize the metrics for both current round and total rounds. If the inner summation factory is not specified, tff.aggregators.SumFactory is used by default. The inner summation factory can also use SecAgg.

The accumulated unfinalized metrics across rounds are initialized to be the intial value of the unfinalized metrics, if the inital value is not specified, zero is used.

The next function of the created tff.templates.AggregationProcess takes the state and local unfinalized metrics reported from tff.CLIENTS, and returns a tff.templates.MeasuredProcessOutput object with the following properties:

  • state: a tuple of the state of the inner summation process and the accumulated unfinalized metrics across rounds.
  • result: a tuple of the finalized metrics of the current round and total rounds.
  • measurements: the measurements of the inner summation process.

metric_finalizers An collections.OrderedDict of metric names to finalizers, should have same keys as the unfinalized metrics. A finalizer is a function (typically a tf.function decorated callable or a tff.tensorflow.computation decorated TFF Computation) that takes in a metric's unfinalized values, and returns the finalized metric values. This can be obtained from tff.learning.models.VariableModel.metric_finalizers().
initial_unfinalized_metrics Optional. An collections.OrderedDict of metric names to the initial values of local unfinalized metrics, its structure should match that of local_unfinalized_metrics_type. If not specified, defaults to zero.
inner_summation_factory Optional. A tff.aggregators.UnweightedAggregationFactory that creates a tff.templates.AggregationProcess to sum the metrics from clients to server. If not specified, tff.aggregators.SumFactory is used. If the metrics aggregation needs SecAgg, aggregation_factory.SecureSumFactory can be used as the inner summation factory.

TypeError If any argument type mismatches.

Methods

create

View source

Creates a tff.templates.AggregationProcess for metrics aggregation.

Args
local_unfinalized_metrics_type A tff.types.StructWithPythonType (with collections.OrderedDict as the Python container) of a client's local unfinalized metrics. For example, local_unfinalized_metrics could represent the output type of tff.learning.models.VariableModel.report_local_unfinalized_metrics().

Returns
An instance of tff.templates.AggregationProcess.

Raises
TypeError If any argument type mismatches; if the metric finalizers mismatch the type of local unfinalized metrics; if the initial unfinalized metrics mismatch the type of local unfinalized metrics.