KVSAgg

Federated Learning

We implement our KVSAgg in a GPU friendly way and apply it to Federated Learning.

In cifar10.sh we provided an example of how to run the program. The meaning of some of the arguments are provided in the chart below:

Argument name	meaning
dataset_dir	directory to where you store the dataset
num_epochs	total number of epochs
num_clients	total number of clients
num_workers	the number of clients participate each turn
num_devices	the number of gpus
num_rows, num_cols	in FetchSGD, it is used to set the sketch size to num_rows * num_cols
k	sample size in local topk and our algorithm
typ	0 for HyperIBLT; 1,2,3,4 for CS(1,3,10,0.5); 5,6,7,8 for CSK(1,3,10,0.5)
num_buckets	the width of HyperIBLT (also used to calculate the width of CS and CSK)

For FEMNIST, because that we have the orignal data preprocessed into 3597 clients, so the num_clients is always equal to 3597. We run for only 1 epoch for FEMNIST.

HyperIBLT

We implement HyperIBLT in C++ for CPU platforms and provide an example of computing the encoding and decoding efficiency of HyperIBLT.

References

Codes in the Federated_learning folder are forked from open source code of MinMax Sampling: A Near-optimal Global Summary for Aggregation in the Wide Area (https://github.com/Arimase97/MinMax-Sampling) and Secure Aggregation implementation (https://github.com/ammartahir24/SecureAggregation).

Recommend Projects