Project: Lossy Gradient Compression: How Much Accuracy Can One Bit Buy?
Acknowledgment: We used some parts of the federated learning code in the following link to implement our simulations.
sadafsk/FL_RD
This code implements federated learning when the data across different users are compressed with a rate R. We use a rate-distortion approach to define a metric for distortion of gradients which are compressed. The accuracy of a DNN setup is studied versus different compression rates.
Python