WCCN: A Parameter and Computationally Efficient Framework for Complex-Valued Deep Learning

WCCN: A Parameter and Computationally Efficient Framework for Complex-Valued Deep Learning - Official Reproduction Package

This dataset provides the complete reproduction package for the paper "WCCN: A Parameter and Computationally Efficient Framework for Complex-Valued Deep Learning." It includes source code, experimental data, trained model checkpoints, and scripts to reproduce every table and figure in the manuscript.

RESEARCH BACKGROUND

Complex-valued neural networks (CVNNs) are powerful for processing data with inherent complex structure, such as PolSAR imagery and frequency-domain signals. However, traditional CVNNs impose strict holomorphic constraints that limit expressiveness. This research introduces Wirtinger Derivative Complete Complex Network (WCCN), which relaxes these constraints by processing both the complex signal and its conjugate. This enables capturing richer, non-circular features while maintaining computational efficiency.

KEY CONTRIBUTIONS

1. Superior Parameter Efficiency: WCCN achieves comparable or better accuracy than baselines with significantly fewer parameters.

2. Robust Generalization: WCCN maintains stable performance even with limited training data.

3. Novel Components: The package includes the proposed wcPReLU activation and lightweight heads, with ablation studies confirming their effectiveness.

DATASET CONTENTS

(1) Source Code: PyTorch implementations of WCCN and baselines in "models/". For image classification, baselines include DCN, SurReal, and CDS-I. For PolSAR classification, all models (CV-CNN baseline and WcCvCnn) are implemented in "polsar/models.py".

(2) Configuration Files: JSON files in "configs/" specifying hyperparameters for every experiment.

(3) Training Logs: The "outputs/" directory contains detailed logs recording loss and accuracy for all experiments.

(4) Model Checkpoints: Trained model weights in "checkpoints/" for direct evaluation.

(5) Raw Metrics: JSON and CSV files with exact numerical values from all tables.

(6) Visualization Scripts: Scripts in "scripts/" generate all figures from the paper.

EXPERIMENTAL SCOPE

- Image Classification: CIFAR-10/100 with multiple input encodings (RGB, LAB, Sliding, RTC6).
- PolSAR Classification: Flevoland, San Francisco, and Oberpfaffenhofen datasets, comparing CV-CNN baseline with WcCvCnn.

NOTE ON OUTPUTS.ZIP: Mendeley Data restricts folder depth to 7 levels. The "outputs/" directory exceeds this limit, so it is provided as outputs.zip containing complete training logs, process files, and experimental results. Extract to root before use.

This dataset enables full reproducibility of the research findings.