gpu-pac 0.3.0

GPU-Accelerated Phase-Amplitude Coupling calculation using PyTorch

gPAC: GPU-Accelerated Phase-Amplitude Coupling

gPAC is a PyTorch-based package for efficient computation of Phase-Amplitude Coupling (PAC) using Modulation Index (MI) with GPU acceleration. It provides:

• 341.8x speedup over TensorPAC (tested on real benchmarks)
• Smart memory management with auto/chunked/sequential strategies
• Full differentiability for deep learning integration
• Production-ready with comprehensive tests and examples
• High correlation with TensorPAC (0.81 ± 0.04 across diverse PAC configurations)

🎯 Example Applications

Static PAC Analysis Comodulogram visualization

Trainable PAC Classification Deep learning integration

🔬 PAC Values Comparison with TensorPAC

Phase: 4Hz, Amp: 40Hz Correlation: 0.826

Phase: 12Hz, Amp: 100Hz Correlation: 0.730

Overall Correlation 0.811 ± 0.042 (n=16)

Click images to view full size. Ground truth PAC locations marked with crosses.

📊 Performance Benchmarks

Parameter Scaling Comparison gPAC (blue) vs TensorPAC (red)

Performance Analysis Speed & memory efficiency

Click images to view detailed performance metrics

🚀 Quick Start

# Installation
pip install gpu-pac

Quick Start

import torch
from torch.utils.data import DataLoader
from gpac import PAC
from gpac.dataset import SyntheticDataGenerator

# Generate synthetic PAC dataset
generator = SyntheticDataGenerator(fs=512, duration_sec=2.0)
dataset = generator.dataset(n_samples=100, balanced=True)
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

# Method 1: Specify frequency range and number of bands
pac_model = PAC(
    seq_len=dataset[0][0].shape[-1],
    fs=512,
    pha_range_hz=(2, 20),    # Phase: 2-20 Hz
    pha_n_bands=10,          # 10 linearly spaced bands
    amp_range_hz=(30, 100),  # Amplitude: 30-100 Hz  
    amp_n_bands=10,          # 10 linearly spaced bands
)

# Method 2: Direct band specification (alternative)
# pac_model = PAC(
#     seq_len=dataset[0][0].shape[-1],
#     fs=512,
#     pha_bands_hz=[[4, 8], [8, 12], [12, 20]],      # Theta, Alpha, Beta
#     amp_bands_hz=[[30, 50], [50, 80], [80, 120]],  # Low, Mid, High Gamma
# )

# Move to GPU if available
device = 'cuda' if torch.cuda.is_available() else 'cpu'
pac_model = pac_model.to(device)

# Process a batch
for signals, labels, metadata in dataloader:
    signals = signals.to(device)
    
    # Calculate PAC
    results = pac_model(signals)
    pac_values = results['pac']  # Shape: (batch, channels, pha_bands, amp_bands)
    
    print(f"Batch PAC shape: {pac_values.shape}")
    print(f"Max PAC value: {pac_values.max().item():.3f}")
    
    # Access frequency band definitions
    print(f"Phase bands: {pac_model.pha_bands_hz}")  # Tensor of shape (n_pha, 2) with [low, high] Hz
    print(f"Amplitude bands: {pac_model.amp_bands_hz}")  # Tensor of shape (n_amp, 2) with [low, high] Hz
    break  # Just show first batch

For more examples, see the examples directory.

🔧 Core Features

Flexible Frequency Band Configuration

• Range-based: Specify frequency range and number of bands for automatic spacing
• Direct specification: Define custom frequency bands for precise control
• Standard bands: Compatible with theta, alpha, beta, gamma conventions
• High resolution: Support for 50+ bands for detailed analysis
• Band access: Direct access to frequency band definitions via pac.pha_bands_hz and pac.amp_bands_hz properties

GPU Optimization

• Multi-GPU support: Automatic data parallelism across GPUs
• FP16 mode: Half-precision computation for 2x memory efficiency
• Torch compilation: JIT compilation for additional speedup
• Batch processing: Efficient handling of multiple signals

Scientific Features

• Permutation testing: Statistical validation with n_perm surrogates
• Z-score normalization: Automatic statistical significance testing
• Modulation Index: Standard MI calculation with 18 phase bins
• Full differentiability: Gradient support for deep learning applications

🤝 Contributing

Contributions are welcome! Please see our contributing guidelines.

📖 Citation

If you use gPAC in your research, please cite:

@software{watanabe2025gpac,
  author = {Watanabe, Yusuke},
  title = {gPAC: GPU-Accelerated Phase-Amplitude Coupling},
  year = {2025},
  url = {https://github.com/ywatanabe1989/gPAC}
}

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• TensorPAC team for the reference implementation
• For fair comparison with TensorPAC, use identical frequency bands as demonstrated in ./benchmark/pac_values_comparison_with_tensorpac/generate_16_comparison_pairs.py