Implement FedCPD: Personalized Federated Learning with Prototype-Enhanced Representation and Memory Distillation (IJCAI 2025)

[appleiphonedddd]

Implement the FedCPD method as described in the paper title above, focusing on prototype‑enhanced representation learning and memory distillation for personalized federated learning (FL).

Goals

• End‑to‑end server–client pipeline.
• Prototype construction/maintenance per class or task (per paper).
• Memory/distillation mechanism to stabilize personalization across rounds.
• Clear configs, scripts, and documentation to reproduce reported behaviors.

Scope

• PyTorch-based implementation (preferred).
• Modular components for server, client,
• Support for common FL benchmarks and non‑IID partitions.

Proposed structure (suggested)

• fedcpd/
  • server.py, client.py, trainer.py

Tasks

• Implement client training loop (local epochs, optimizer, prototype updates).
• Add memory/distillation loss as specified in the paper.
• Implement server orchestration (rounds, broadcast, collection).
• Implement prototype handling (update/EMA/aggregation per paper).
• Write README with setup, commands, and expected metrics.
• Add scripts to reproduce main experiments and ablations.
• Add unit tests for prototype ops, memory buffer, and aggregation.

Deliverables

• Well‑documented code with docstrings and comments.

• Configuration files and example command lines.

Acceptance criteria

• Training completes on at least one dataset with non‑IID partitions.
• Results trend aligns with the paper (within reasonable tolerance).

References

• Paper: “FedCPD: Personalized Federated Learning with Prototype‑Enhanced Representation and Memory Distillation (IJCAI 2025)”
• Link(s): [FedCPD]
• Please cite the paper in README.

Labels
enhancement, help wanted, research, federated-learning, personalization