PNMF: Probabilistic Non-negative Matrix Factorization

A PyTorch-based implementation of Probabilistic NMF with projected gradient optimization, compatible with scikit-learn's API.

Installation

pip install pnmf

For development installation:

git clone https://github.com/luisdiaz1997/Probabilistic-NMF.git
cd Probabilistic-NMF
pip install -e .

Quick Start

from PNMF import PNMF
import numpy as np

# Create some sample data
X = np.random.rand(100, 50)

# Initialize and fit the model
model = PNMF(n_components=5, init='random', random_state=42)
W = model.fit_transform(X)  # Transformed data: (100, 5)
H = model.components_       # Components: (5, 50)

# Reconstruct the data
X_reconstructed = model.inverse_transform(W)

Model

PNMF uses a Poisson likelihood for the observed data:

• y_ij ~ Poisson(λ_ij)
• λ_ij = Σ_l W_jl exp(F_il)
• F_il ~ N(0, σ²)

where W ≥ 0 are the loadings and F are the latent factors (variational inference with Gaussian prior).

Features

• Probabilistic NMF: Uses Poisson factorization for count data
• Projected Gradients: Three modes for enforcing non-negativity:
  • softplus: Uses softplus transformation
  • exp: Uses exponential transformation
  • projected: Uses projected gradient descent (default)
• sklearn-compatible API: Works seamlessly with scikit-learn workflows
• GPU Support: Automatic CUDA detection and utilization

API Reference

PNMF(n_components=2, init='random', init_mode='projected', max_iter=200, tol=1e-4, learning_rate=0.01, random_state=None, verbose=False, device='auto')

Parameters:

• n_components: Number of latent components (default: 2)
• init: Initialization method - 'random' or 'custom' (default: 'random')
• init_mode: Non-negativity enforcement - 'softplus', 'exp', or 'projected' (default: 'projected')
• max_iter: Maximum number of iterations (default: 200)
• tol: Tolerance for convergence (default: 1e-4)
• learning_rate: Learning rate for gradient descent (default: 0.01)
• random_state: Random seed for reproducibility (default: None)
• verbose: Whether to print progress messages (default: False)
• device: Device to use - 'cpu', 'cuda', or 'auto' (default: 'auto')

Methods:

• fit(X, y=None, W=None, H=None): Fit the model to data X
• transform(X): Transform X using the fitted model
• fit_transform(X, **kwargs): Fit the model and transform X
• inverse_transform(H): Transform data back to original space

License

GNU General Public License v2.0

Author

Luis Chumpitaz Diaz

Acknowledgments

This implementation borrows code from GPzoo.