PyTorch PCA

A PyTorch implementation of Principal Component Analysis (PCA) that exactly matches scikit-learn's implementation with default settings. This library provides GPU-accelerated PCA functionality with a scikit-learn compatible interface.

Features

• PCA: Standard PCA implementation (pca.py)
• Incremental PCA: Memory-efficient version that processes data in batches (incremental_pca.py, contributed by: yry)
• GPU Acceleration: Both implementations support CUDA for faster computation
• scikit-learn Compatible API: Uses familiar fit/transform methods

Usage

import torch
from pca import PCA

# Create data
X = torch.randn(100, 20)  # 100 samples, 20 features

# Initialize and fit PCA
pca = PCA(n_components=10)
pca.fit(X)

# Transform data
X_transformed = pca.transform(X)

# Or do both in one step
X_transformed = pca.fit_transform(X)

# Reconstruct original data
X_reconstructed = pca.inverse_transform(X_transformed)

For incremental PCA (processing data in batches):

from incremental_pca import IncrementalPCA

# Initialize
ipca = IncrementalPCA(n_components=10, n_features=20)

# Process batches
for batch in data_batches:
    ipca.partial_fit(batch)

# Transform new data
transformed_data = ipca.transform(new_data)

Installation

Copy the code from pca.py or incremental_pca.py into your project.

Benchmarking

Run benchmark.py to compare performance between this implementation and scikit-learn's.

References

• scikit-learn PCA
• scikit-learn IncrementalPCA

Related Work

• valentingol's torch_pca appears to be a more full featured and faster (it chooses an appropriate PCA algorithm depending on input dimensions) alternative PCA implementation also matching scikit-learn.