Attend the Women in ML Symposium on December 7 Register now

tff.learning.compression_aggregator

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

Creates aggregator with compression and adaptive zeroing and clipping.

Used in the notebooks

Used in the tutorials

Zeroes out extremely large values for robustness to data corruption on clients and clips in the L2 norm to moderately high norm for robustness to outliers. After weighting in mean, the weighted values are uniformly quantized to reduce the size of the model update communicated from clients to the server. For details, see Suresh et al. (2017) http://proceedings.mlr.press/v70/suresh17a/suresh17a.pdf The default configuration is chosen such that compression does not have adverse effect on trained model quality in typical tasks.

zeroing Whether to enable adaptive zeroing for data corruption mitigation.
clipping Whether to enable adaptive clipping in the L2 norm for robustness. Note this clipping is performed prior to the per-coordinate clipping required for quantization.
weighted Whether the mean is weighted (vs. unweighted).
debug_measurements_fn A callable to add measurements suitable for debugging learning algorithms, with possible values as None, tff.learning.add_debug_measurements or tff.learning.add_debug_measurements_with_mixed_dtype.
**kwargs Keyword arguments.

A tff.aggregators.AggregationFactory.

TypeError if debug_measurement_fn yields an aggregation factory whose weight type does not match weighted.