Two Timescale Hybrid Learning (TT-HL)

• Paper on IEEE Explore
• ArXiv: Semi-Decentralized Federated Learning with Cooperative D2D Local Model Aggregations
• Data: We currently do not have a permanent link to the data (including model parameters, data splits, and network graphs). However, please feel free to reach out to me and we will try to figure a way out to share the data dump with you over dropbox or google drive.

Setup

• The two main prerequisites are:

  • pytorch==1.4.0
  • syft==0.2.9

• The python version used for experiments: 3.6.9

Running Experiments

• Download MNIST and FMNIST.
• Update the config file src/common/config.py for correct locations of data_path.
• Generate the initial data splits, model starting points, and graphs:

cd src

# to generate data splits, fmnist and mnist must be downloaded and data_path set.
sh sh/gen_data.sh <dataset=mnist|fmnist>_<num-nodes=25|125>

# to generate initial models
sh sh/gen_init_models.sh

• All the experiments are run using shell scripts of the format src/sh/train_*.sh.
• There are two models used: fully connected nets (fcn) and support vector machines (svm).
• The corresponding shell scripts contain the substrings fcn or svm.
• Experiments are performed for two network graphs w.r.t csi and no-csi as explained in the paper. The corresponding shell scripts contain substrings csi or nocsi.

Generating plots

• Finish training first. This generates the files in the checkpoints folder that are needed to create plots.
• Alternately reach out to us for a dump of our trained meta data files.
• All the plots can be executed using shell scripts of the format src/sh/plot_*.sh.

Citation

If you find the repository or the paper useful, please cite the following paper

@article{
  lin2021semi,
  title={Semi-decentralized federated learning with cooperative D2D local model aggregations},
  author={Lin, Frank Po-Chen and Hosseinalipour, Seyyedali and Azam, Sheikh Shams and Brinton, Christopher G and Michelusi, Nicolo},
  journal={IEEE Journal on Selected Areas in Communications},
  volume={39},
  number={12},
  pages={3851--3869},
  year={2021},
  publisher={IEEE}
}