FedCode

FedCode: Communication-Efficient Federated Learning via Transferring Codebooks

Saeed Khalilian, Vasileios Tsouvalas, Tanir Ozcelebi, Nirvana Meratnia
IEEE EDGE 2024

Paper · Project · Poster · Citation

Abstract

FedCode reduces federated communication by exchanging codebooks of weight-cluster centers. Clients and the server use these centers to update their existing weight arrangements, with periodic calibration rounds transmitting cluster assignments as well. Independent uplink and downlink schedules control calibration frequency, allowing experiments to study the communication–accuracy tradeoff with vision and audio models.

Most FedCode rounds communicate cluster centers upstream and downstream. Periodic transfers of compressed weights recalibrate the client and server models. Figure 2 in the public preprint.

Installation

Use Python 3.11 and a dedicated virtual environment.

mkdir -p "$HOME/research-code" "$HOME/.venvs"
cd "$HOME/research-code"
git clone https://github.com/Nousphera/FedCode.git
cd FedCode
python3.11 -m venv "$HOME/.venvs/fedcode"
source "$HOME/.venvs/fedcode/bin/activate"
python -m pip install -r requirements.txt
python -m pip check

Data

python scripts/prepare_data.py --dataset cifar10 --root datasets --download
python scripts/prepare_data.py --dataset cifar10 --root datasets --validate

Use cifar100 for CIFAR-100 or spcm for Speech Commands v2. The dataset guide describes NPZ imports, audio validation, and the twelve-class label mapping. Prepared files feed the training loader directly.

Training

Run three rounds on generated images to exercise codebook and calibration modes:

python train.py --dataset synthetic --clients 2 --rounds 3 --epochs 1 \
  --batch-size 8 --clusters 4 --start-round 1 --broadcast-rate 3 \
  --request-rate 2 --output runs/smoke

Train ResNet-20 on prepared CIFAR-10:

python train.py --dataset cifar10 --data-root datasets --model resnet20 \
  --clients 10 --rounds 60 --epochs 4 --batch-size 128 --lr .001 \
  --clusters 64 --split noniid --alpha 10 --participation 1 --start-round 2 \
  --broadcast-rate 5 --request-rate 2 --seed 42 --output runs/cifar10-resnet

The runner trains clients sequentially and uses stock Flower protobufs for communication accounting. Client weights persist by identity between codebook and calibration rounds. Use --model mobilenet for MobileNet, or --dataset spcm --model yamnet for audio. python train.py --help lists options.

Evaluation

The runner evaluates the global model on the test split after every round and records accuracy as a fraction in rounds[].accuracy, alongside loss and upstream/downstream protobuf byte counts.

python -m json.tool runs/cifar10-resnet/results.json

For the CIFAR-100 configuration used in the table below, change the training command to --dataset cifar100 --rounds 100 and choose a new output directory. Method settings describe the calibration intervals and communication accounting.

Configuration and outputs

--broadcast-rate controls downlink calibration and --request-rate controls uplink calibration. --split noniid --alpha 10 selects a Dirichlet partition with concentration 10; use --alpha .1 for stronger class heterogeneity. Method settings explain payloads, cold-start clients, partitioning, and the calibration schedule.

Every run uses a new output directory. results.json records configuration, serialized request/response byte counts, and each round's communication mode. model-state.pt and final-parameters.npz store final weights; training.log records execution. Temporary client caches are removed after the run; --init-state initializes a new run from a weights-only state dictionary. The testing guide covers codec, training, and data checks.

Results

Reported CIFAR-100 / ResNet-20 results from Table 6 of the paper’s public preprint: ten clients, full participation, 100 rounds, four local epochs, 64 centers, and calibration rates F1 = 0.2 / F2 = 0.5. Data transmission reduction (DTR) is relative to FedAvg.

Method	Accuracy (%)	Downstream DTR (×)	Upstream DTR (×)	Joint DTR (×)
FedAvg	61.22	—	—	—
FedAvg + weight clustering	58.78	5.3	5.3	5.3
FedZip	57.98	—	19.8	1.9
FedCode	58.07	24.2	10.4	14.6

FedCode reaches 14.6× joint communication reduction at 58.07% accuracy. Dashes retain entries unreported in the source table.

Citation

If you use this work, please cite:

@inproceedings{khalilian2024fedcode,
  title = {{FedCode}: Communication-Efficient Federated Learning via Transferring Codebooks},
  author = {Khalilian, Saeed and Tsouvalas, Vasileios and Ozcelebi, Tanir and Meratnia, Nirvana},
  booktitle = {2024 IEEE International Conference on Edge Computing and Communications (EDGE)},
  year = {2024},
  pages = {99--109},
  publisher = {IEEE},
  doi = {10.1109/EDGE62653.2024.00022},
  url = {https://doi.org/10.1109/EDGE62653.2024.00022}
}

Machine-readable metadata: CITATION.cff · BibTeX.

License

MIT, with component terms in licenses/THIRD_PARTY.txt.