fedmars 0.5.1

Reference PyTorch implementation of FedMARS, a credit-driven layer-wise federated learning framework.

FedMARS

FedMARS is a PyTorch implementation of Federated Mode-Aware Reliability Scoring for layer-wise trust control in heterogeneous federated learning. The method builds mode-aware layer evidence on each client, converts that evidence into a layer credit, robustly aggregates credits on the server, and uses the same credit for layer selection, server update strength, and proximal drift control.

Install

pip install .

For the matching package release:

pip install fedmars==0.5.1

Minimal use

from fedmars import FedMARS, FedMARSConfig

config = FedMARSConfig(num_rounds=40, default_budget_fraction=0.35)
trainer = FedMARS(model, config)
history = trainer.fit(clients, server_val_loader=val_loader, server_test_loader=test_loader)

Each client must be a fedmars.ClientDataset or compatible object with client_id, dataset, and optional weight fields. Datasets must return (x, y) pairs.

Implementation choices in this submission version

Area	Submission behavior
Layer budget	Exact parameter-bit budget: selected layer bits are constrained by floor(model_bits * budget_fraction).
Output layer	No layer is forced by default. If no layer fits a very small budget, the selected set can be empty without violating the bit budget.
No-reference credit	Final credit alignment is neutral when no usable server reference exists.
Mixture objective	Entropy is controlled by the explicit mixture_entropy field.
Server momentum	EMA update: v <- m v + (1-m) delta.
Aggregation weighting	Positive credit weighting uses robust z-normalized client-layer credit.
Communication logs	Uplink, downlink, total bits, selected bits, budget bits, model bits, and budget-violation flags are recorded each round.

The source package contains only the FedMARS architecture, data/layer utilities, and method-critical unit tests. Experiment notebooks, smoke-result CSVs, and ablation-study scripts are intentionally excluded from this submission code package.