FirConv: A PyTorch-based module for building trainable FIR filter ⏳

🚀 Introduction

A PyTorch-based module for building trainable FIR filters in deep neural networks. Key features:

• 🔧 Seamless integration with existing PyTorch models
• 🎚️ Pre-defined filters or learnable coefficients
• 📊 Real-time frequency adaptation visualization
• ⚡ CUDA-accelerated computation

"Enabling neural networks to understand signal processing through differentiable filters"

📦 Installation

From Source

git clone https://github.com/FunkyFrog1/FirConv.git
cd FirConv
pip install .

From PyPI (coming soon)

pip install firconv

🛠️ Basic Usage

import torch
from firconv.firconv import FirConv

# Create learnable FIR filter with 32 taps
fir = FirConv(fres=100, fs=250)

# Process 1D signal batch: (batch_size, channels, seq_len)
x = torch.randn(1, 63, 250)  # Batch of 63 signals
y = fir(x)  # Output shape: (1, 63, 250)

📚 Documentation

Argument	Type	Default	Description
fres	int	100	Filter frequency range [0-fres]
N	int	-	Filter window size(default compute by fres and fs
fs	int	250	Sample rate

📅 Release Notes

0.1.1 - 2025-02-24

• Initial public release
• Add CUDA kernel optimization
• Support multi-channel filtering

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

📜 License

Apache 2.0 © 2025 FunkyFrog