pytorch-pca 0.1.2.dev0

Comprehensive PyTorch implementation of PCA.

PyTorch PCA

A comprehensive PCA implementation using PyTorch, inspired by the R package pcaMethods.

Overview

This package provides a unified interface to eight PCA algorithms, all accessible via the pca function. The main entry point is pytorch_pca.pca, and the package exposes pca, PCAResult, and AllowedMethod in its public API.

Installation

pip install pytorch_pca

Features

This package provides 8 different PCA algorithms optimized for various scenarios:

• svd: Standard SVD-based PCA (fastest, complete data only)
• nipals: NIPALS algorithm (handles missing values effectively)
• rnipals: Robust NIPALS (resistant to outliers and missing values)
• ppca: Probabilistic PCA (classical probabilistic model)
• bpca: Bayesian PCA (probabilistic approach with uncertainty quantification)
• svd_impute: SVD-based PCA with missing value imputation
• rpca: Robust PCA using iterative outlier detection
• nlpca: Non-linear PCA using autoencoder neural network architecture

Quick Start

import torch
from pytorch_pca import pca

# Generate sample data
X = torch.randn(100, 20)

# Basic PCA using the unified interface
result = pca(X, method="svd", n_components=5)
print(f"Transformed data shape: {result.transformed_data.shape}")
print(f"Components shape: {result.components.shape}")
print(f"Explained variance: {result.explained_variance_ratio}")

# Alternative: use pcaMethods-style naming
print(f"Scores shape: {result.scores.shape}")
print(f"Loadings shape: {result.loadings.shape}")

# Handle missing data with NIPALS
X_missing = X.clone()
X_missing[10:20, 5:10] = float('nan')
result = pca(X_missing, method="nipals", n_components=3)

# Robust PCA for data with outliers
result = pca(X, method="rpca", n_components=3)

# Probabilistic approaches
result = pca(X, method="ppca", n_components=3)
result = pca(X, method="bpca", n_components=3)

# Non-linear PCA with neural networks
result = pca(X, method="nlpca", n_components=3)

# Reconstruct data
X_reconstructed = result.reconstruct(n_components=3)

API Reference

Unified Interface

result = pca(data, method="svd", n_components=2, center=True, scale=False, **kwargs)

Method-Specific Parameters

NIPALS Methods

nipals(data, max_iter=1000, tol=1e-6, ...)
rnipals(data, max_iter=1000, tol=1e-6, ...)

Probabilistic Methods

ppca(data, max_iter=1000, tol=1e-6, ...)
bpca(data, max_iter=1000, tol=1e-6, ...)

Robust PCA

rpca(data, max_iter=100, tol=1e-6, ...)

Non-linear PCA

nlpca(data, hidden_dims=[10, 5], max_iter=1000, lr=0.01, ...)

Result Object

The PCAResult object provides:

• transformed_data: Data projected onto principal components (n_samples, n_components)
• components: Principal components (eigenvectors) (n_components, n_features)
• eigenvalues: Eigenvalues of the covariance matrix
• explained_variance_ratio: Proportion of variance explained by each component
• method: Name of the method used
• scores: Alias for transformed_data (pcaMethods compatibility)
• loadings: Transposed components (pcaMethods compatibility)
• reconstruct(n_components=None): Reconstruct data using selected components

Method Selection Guide

• Complete data, speed priority: svd
• Missing values: nipals or svd_impute
• Outliers present: rpca or rnipals
• Uncertainty quantification: bpca
• Probabilistic modeling: ppca
• Non-linear relationships: nlpca

Dependencies

• torch: The only required dependency (>= 2.7.1)

Development & Testing

• black: Code formatting
• flake8: Code linting
• mypy: Type checking
• pytest: Test runner
• scikit-learn: For test comparisons
• setuptools, setuptools-scm: Packaging

Comprehensive tests are provided in the tests/ directory, covering all algorithms, edge cases, and robust/non-linear PCA scenarios.

Testing

# Run all tests from root
pytest ./tests

License

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

Author

Ricardo Yanzon - ricayanzon

Acknowledgments

• Inspired by the R package pcaMethods
• Built with PyTorch