Here’s a clean and structured README.md you can use for the PEPFedPT directory.

Official implementation of PEPFedPT — a federated learning framework with class centroid / prototype alignment and Vision Transformer (ViT) backbone with CCMP integration.

This repository supports multiple datasets (CIFAR-100, DomainNet, Tiny-ImageNet, iNat, etc.) and implements federated averaging with class prototype updates for personalized federated learning.

PEPFedPT/
│
├── centroids.py
├── centroidutils.py
├── centroidinit.py
├── constants.py
├── datasets.py
├── get_class_embeddings.py
├── train.sh
│
├── models/
│   └── vit_models_pr_centroid.py
│
├── dataloaders/
├── cifar_embeddings/
├── domain_embeddings/
├── imagenet_embeddings/
├── inat_embeddings/
├── tinyimagenet_models/
├── ...

Main training orchestration script.

• Handles argument parsing
• Initializes dataset and model
• Controls communication rounds
• Manages client sampling
• Calls local training routines
• Performs aggregation
• Updates global centroids/prototypes

All flag descriptions are defined directly inside this file.

Contains core federated learning logic:

• Local training routine
• Federated averaging (FedAvg)
• Class prototype / centroid update logic
• Prototype aggregation across clients

This file implements the main optimization and aggregation mechanisms.

Vision Transformer (ViT) backbone implementation with:

• CCMP integration
• Prototype-based classification support
• Feature extraction for centroid computation

This is the primary model used during training.

Provides example commands for running experiments on:

• CIFAR-100
• DomainNet
• Tiny-ImageNet
• iNaturalist

Use this file as a reference for running different datasets and partition strategies.

• CIFAR-100
• DomainNet
• Tiny-ImageNet
• iNaturalist

Dataset handling is implemented in:

dataloaders/

Make sure datasets are downloaded and paths are correctly set.

For dataset-specific examples, see:

bash train.sh

All arguments are defined in centroids.py. Key flags include:

Refer to centroids.py for full details and defaults.

• Ensure GPU is available for ViT training.
• Prototype memory grows with number of classes.
• Non-IID partitions are supported.
• Embeddings are optionally saved under dataset-specific folders.