Issue in top-K sparsification implementation

apple / pfl-research

Simulation framework for accelerating research in Private Federated Learning

http://apple.github.io/pfl-research/

Apache License 2.0

259 stars 24 forks source link

Issue in top-K sparsification implementation #65

Open sohaib-idalin opened 2 months ago

sohaib-idalin commented 2 months ago

I see that you mention that top-K sparsification must zero out 1-top_k_percent percent of the model update elements with the lowest absolute values. But in the code, I see that you forgot to use the absolute values of the vector.

grananqvist commented 1 month ago

Thanks for finding this mistake!

sohaib-idalin commented 1 month ago

your welcome