fastica_torch

A PyTorch implementation of the FastICA algorithm for Independent Component Analysis.

This package provides a GPU-accelerated implementation that replicates sklearn.decomposition.FastICA, allowing seamless use with PyTorch tensors and CUDA acceleration.

Installation

pip install fastica_torch

Or install from source:

git clone https://github.com/RichieHakim/FastICA_torch.git
cd FastICA_torch
pip install -e .

Quick Start

import torch
from fastica_torch import FastICA

# Generate mixed signals
n_samples, n_features = 1000, 5
X = torch.randn(n_samples, n_features)

# Fit FastICA and extract sources
ica = FastICA(n_components=3, random_state=42)
sources = ica.fit_transform(X)

# Transform new data
X_new = torch.randn(100, n_features)
sources_new = ica.transform(X_new)

# Reconstruct from sources
X_reconstructed = ica.inverse_transform(sources)

Features

• sklearn-compatible API: Drop-in replacement for sklearn.decomposition.FastICA
• PyTorch tensors: Native support for PyTorch tensors
• GPU acceleration: Automatic CUDA support when tensors are on GPU
• Multiple algorithms: Both parallel and deflation extraction methods
• Contrast functions: logcosh, exp, cube for negentropy approximation

API Reference

FastICA

FastICA(
    n_components=None,      # Number of components (None = all)
    algorithm="parallel",   # "parallel" or "deflation"
    whiten="unit-variance", # "unit-variance", "arbitrary-variance", or False
    fun="logcosh",          # "logcosh", "exp", "cube", or callable
    fun_args=None,          # Dict of arguments for fun (e.g., {"alpha": 1.0})
    max_iter=200,           # Maximum iterations
    tol=1e-4,               # Convergence tolerance
    w_init=None,            # Initial unmixing matrix
    whiten_solver="svd",    # "svd" or "eigh"
    random_state=None,      # Random seed for reproducibility
)

Methods

Attributes (after fitting)

Comparison with sklearn

This implementation aims to produce numerically equivalent results to sklearn's FastICA when using the same random seed and parameters. Minor differences may occur due to:

• Floating-point precision differences between NumPy and PyTorch
• Different SVD implementations
• Sign/permutation ambiguity inherent to ICA

Requirements

• Python >= 3.9
• PyTorch >= 2.0
• NumPy >= 1.20

License

MIT License - see LICENSE for details.

References

1. A. Hyvarinen and E. Oja, "Independent Component Analysis: Algorithms and Applications", Neural Networks, 13(4-5):411-430, 2000.
2. A. Hyvarinen, "Fast and Robust Fixed-Point Algorithms for Independent Component Analysis", IEEE Trans. Neural Networks, 10(3):626-634, 1999.