---
license: mit
tags:
  - federated-learning
  - knowledge-distillation
  - point-cloud
  - 3d-classification
  - benchmarking
  - privacy-preserving
  - medical-imaging
dataset_info:
  features:
    - name: point_cloud
      dtype: float32
    - name: label
      dtype: int64
  splits:
    - name: train
      num_examples: unknown
    - name: validation
      num_examples: unknown
    - name: test
      num_examples: unknown
  supervised: true
  task_categories:
    - image-classification
  task_ids:
    - 3d-image-classification
pretty_name: FLKD 3D Point Cloud Benchmark
size_categories:
  - unknown
source_datasets:
  - craniosynostosis
  - modelnet40
---

# FLKD 3D Benchmark - Output Models and Results

This directory contains pre-trained models and comprehensive results from the **Benchmarking Federated Learning and Knowledge Distillation for Point Cloud Classification** benchmark (ECCV 2026).

## Overview

This collection includes trained PointNet++ models evaluated on the 3D point cloud datasets across two training paradigms:

- **Classical Training**: Centralized, single-machine model training
- **Federated Learning + Knowledge Distillation (FLKD)**: Distributed federated learning with knowledge distillation objectives

All experiments were conducted with multiple random seeds to ensure robust statistical reporting of performance metrics.

## Directory Structure

```
RootDir/
└── <dataset_name></dataset_name>/
    ├── classification/
    │   └── pointnet2_cls_ssg_s{seed}/
    │       ├── checkpoints/
    │       │   ├── best_model.pth
    │       │   ├── last_model.pth
    │       │   └── resume.pth
    │       ├── config.json
    │       ├── logs/
    │       │   └── train.log
    │       └── metrics.jsonl
    │
    └── flkd/federated/classification/
        └── fl_pointnet2_cls_ssg_{algorithm}_s{seed}/
            ├── checkpoints/
            │   ├── best_model.pth
            │   ├── last_model.pth
            │   └── resume.pth
            ├── config.json
            ├── logs/
            │   └── train.log
            └── metrics.jsonl
```

## File Descriptions

### Checkpoints
- **best_model.pth**: Model weights achieving the highest validation accuracy during training
- **last_model.pth**: Model weights from the final training round/epoch
- **resume.pth**: Complete training state (model, optimizer, RNG) for resuming interrupted training

### Configuration
- **config.json**: Hyperparameters, dataset splits, model architecture, and training settings used for this experiment

### Logs
- **train.log**: Detailed per-epoch/round training logs including loss values and validation metrics

### Metrics
- **metrics.jsonl**: Machine-readable results in JSONL format containing:
  - Per-epoch/round accuracy metrics
  - Loss values
  - Training time information
  - Other performance indicators

## Seeds and Reproducibility

Experiments use multiple random seeds (e.g., s7, s42, s123) to report mean ± standard deviation statistics, ensuring robust statistical conclusions. Load `best_model.pth` for production use; consult `metrics.jsonl` for full seed-wise performance breakdowns.

## Federated Learning Algorithms

The FLKD benchmark evaluates these 13 FL algorithms:
- **FedAvg**: Classical federated averaging (also called "vanilla")
- **FedProx**: Proximal term regularization for heterogeneous local objectives
- **SCAFFOLD**: Control variates to reduce client drift
- **FedDyn**: Biased aggregation with consensus optimization
- **FedAvgM**: Momentum-based federated averaging
- **FedAdam**: Server-side adaptive learning rates (Adam variant)
- **FedYogi**: Server-side adaptive learning rates (Yogi variant)
- **FedAdagrad**: Server-side adaptive learning rates (AdaGrad variant)
- **FedMedian**: Robust aggregation via median
- **FedBN**: Batch norm personalization for heterogeneous local data
- **MOON**: Contrastive learning to maintain consistency
- **Ditto**: Explicit client-local personalization
- **FedNova**: Normalized aggregation for non-IID data


## Model Architecture

**PointNet++ (Single-Scale Grouping / SSG)**:
- Multi-layer hierarchical feature learning on point clouds
- Set Abstraction (SA) layers with ball query and PointNet modules
- Feature Propagation (FP) layers for upsampling
- Designed for robust 3D shape understanding

## Usage

### Loading a Model

```python
import torch

# Load the best model for a specific configuration
model = torch.load('pointnet2_cls_ssg_s123/checkpoints/best_model.pth')

# Or load with full training state (for resuming)
checkpoint = torch.load('pointnet2_cls_ssg_s123/checkpoints/resume.pth')
model_state = checkpoint['model_state']
optimizer_state = checkpoint['optimizer_state']
```

### Accessing Results

```python
import json

# Load configuration
with open('pointnet2_cls_ssg_s123/config.json') as f:
    config = json.load(f)

# Read metrics (each line is a JSON object)
with open('pointnet2_cls_ssg_s123/metrics.jsonl') as f:
    for line in f:
        epoch_metrics = json.loads(line)
        print(epoch_metrics)
```

## Citation

If you use these models or results, please cite the original paper:

```bibtex
@inproceedings{aizierjiang26benchmark,
  title={Benchmarking Federated Learning and Knowledge Distillation for Point Cloud Classification},
  author={Aizierjiang Aiersilan},
  booktitle={European Conference on Computer Vision},
  organization={Springer},
  year={2026}
}
```

## License

The models and code follow the project's original license. Please refer to the main repository for detailed license information.

## Additional Resources

- **Project Website**: https://ezharjan.github.io/FLKD3DBenchmark/
- **Main Repository**: https://github.com/Ezharjan/FLKD3DBenchmark/
- **Paper**: Available at the project website

## Notes

- All models were trained using PyTorch with CUDA acceleration
- Mixed precision (bf16) was applied where supported; exact fp32 on high performance GPUs
- Auto-resume checkpoints enable resuming interrupted training without loss of progress
- Metrics are reported as mean ± std across multiple random seeds for robust statistical assessment

---

For questions, issues, or to contribute improvements, please visit the main repository or contact the authors through the project website.