# PyTorch PCA
A comprehensive PCA implementation using PyTorch, inspired by the R package [pcaMethods](https://github.com/bioc/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.
[](https://pypi.org/project/pytorch_pca/)
## Installation
```bash
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
```python
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
```python
result = pca(data, method="svd", n_components=2, center=True, scale=False, **kwargs)
```
### Method-Specific Parameters
#### NIPALS Methods
```python
nipals(data, max_iter=1000, tol=1e-6, ...)
rnipals(data, max_iter=1000, tol=1e-6, ...)
```
#### Probabilistic Methods
```python
ppca(data, max_iter=1000, tol=1e-6, ...)
bpca(data, max_iter=1000, tol=1e-6, ...)
```
#### Robust PCA
```python
rpca(data, max_iter=100, tol=1e-6, ...)
```
#### Non-linear PCA
```python
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
```bash
# Run all tests from root
pytest ./tests
```
## License
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.
## Author
**Ricardo Yanzon** - [ricayanzon](https://github.com/ricayanzon)
## Acknowledgments
- Inspired by the R package [pcaMethods](https://github.com/bioc/pcaMethods)
- Built with [PyTorch](https://pytorch.org/)
Raw data
{
"_id": null,
"home_page": null,
"name": "pytorch-pca",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.11",
"maintainer_email": null,
"keywords": "PCA, principal component analysis, pytorch, torch, svd pca, SVD, Singular Value Decomposition, NIPALS, Nonlinear Iterative Partial Least Squares, Non-linear Iterative Partial Least Squares, RNIPALS, robust NIPALS, bpca, Bayesian PCA, ppca, Probabilistic PCA, SVD Impute, Iterative PCA, Robust PCA, Nonlinear PCA, Non-linear PCA",
"author": null,
"author_email": "Ricardo Yanzon <ricayanzon@gmail.com>",
"download_url": "https://files.pythonhosted.org/packages/a7/28/d377b1335abd413edc48fd4d52880735857d900e5d70d1992ea1939d677a/pytorch_pca-0.1.3.tar.gz",
"platform": null,
"description": "# PyTorch PCA\n\nA comprehensive PCA implementation using PyTorch, inspired by the R package [pcaMethods](https://github.com/bioc/pcaMethods).\n\n## Overview\n\nThis 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.\n\n[](https://pypi.org/project/pytorch_pca/)\n\n\n\n## Installation\n\n```bash\npip install pytorch_pca\n```\n\n## Features\n\nThis package provides 8 different PCA algorithms optimized for various scenarios:\n\n- **`svd`**: Standard SVD-based PCA (fastest, complete data only)\n- **`nipals`**: NIPALS algorithm (handles missing values effectively)\n- **`rnipals`**: Robust NIPALS (resistant to outliers and missing values)\n- **`ppca`**: Probabilistic PCA (classical probabilistic model)\n- **`bpca`**: Bayesian PCA (probabilistic approach with uncertainty quantification)\n- **`svd_impute`**: SVD-based PCA with missing value imputation\n- **`rpca`**: Robust PCA using iterative outlier detection\n- **`nlpca`**: Non-linear PCA using autoencoder neural network architecture\n\n## Quick Start\n\n```python\nimport torch\nfrom pytorch_pca import pca\n\n# Generate sample data\nX = torch.randn(100, 20)\n\n# Basic PCA using the unified interface\nresult = pca(X, method=\"svd\", n_components=5)\nprint(f\"Transformed data shape: {result.transformed_data.shape}\")\nprint(f\"Components shape: {result.components.shape}\")\nprint(f\"Explained variance: {result.explained_variance_ratio}\")\n\n# Alternative: use pcaMethods-style naming\nprint(f\"Scores shape: {result.scores.shape}\")\nprint(f\"Loadings shape: {result.loadings.shape}\")\n\n# Handle missing data with NIPALS\nX_missing = X.clone()\nX_missing[10:20, 5:10] = float('nan')\nresult = pca(X_missing, method=\"nipals\", n_components=3)\n\n# Robust PCA for data with outliers\nresult = pca(X, method=\"rpca\", n_components=3)\n\n# Probabilistic approaches\nresult = pca(X, method=\"ppca\", n_components=3)\nresult = pca(X, method=\"bpca\", n_components=3)\n\n# Non-linear PCA with neural networks\nresult = pca(X, method=\"nlpca\", n_components=3)\n\n# Reconstruct data\nX_reconstructed = result.reconstruct(n_components=3)\n```\n\n## API Reference\n\n### Unified Interface\n\n```python\nresult = pca(data, method=\"svd\", n_components=2, center=True, scale=False, **kwargs)\n```\n\n### Method-Specific Parameters\n\n#### NIPALS Methods\n```python\nnipals(data, max_iter=1000, tol=1e-6, ...)\nrnipals(data, max_iter=1000, tol=1e-6, ...)\n```\n\n#### Probabilistic Methods\n```python\nppca(data, max_iter=1000, tol=1e-6, ...)\nbpca(data, max_iter=1000, tol=1e-6, ...)\n```\n\n#### Robust PCA\n```python\nrpca(data, max_iter=100, tol=1e-6, ...)\n```\n\n#### Non-linear PCA\n```python\nnlpca(data, hidden_dims=[10, 5], max_iter=1000, lr=0.01, ...)\n```\n\n### Result Object\n\nThe `PCAResult` object provides:\n\n- `transformed_data`: Data projected onto principal components `(n_samples, n_components)`\n- `components`: Principal components (eigenvectors) `(n_components, n_features)`\n- `eigenvalues`: Eigenvalues of the covariance matrix\n- `explained_variance_ratio`: Proportion of variance explained by each component\n- `method`: Name of the method used\n- `scores`: Alias for `transformed_data` (pcaMethods compatibility)\n- `loadings`: Transposed components (pcaMethods compatibility)\n- `reconstruct(n_components=None)`: Reconstruct data using selected components\n\n## Method Selection Guide\n\n- **Complete data, speed priority**: `svd`\n- **Missing values**: `nipals` or `svd_impute`\n- **Outliers present**: `rpca` or `rnipals`\n- **Uncertainty quantification**: `bpca`\n- **Probabilistic modeling**: `ppca`\n- **Non-linear relationships**: `nlpca`\n\n## Dependencies\n\n- **torch**: The only required dependency (>= 2.7.1)\n\n## Development & Testing\n\n- **black**: Code formatting\n- **flake8**: Code linting\n- **mypy**: Type checking\n- **pytest**: Test runner\n- **scikit-learn**: For test comparisons\n- **setuptools**, **setuptools-scm**: Packaging\n\nComprehensive tests are provided in the `tests/` directory, covering all algorithms, edge cases, and robust/non-linear PCA scenarios.\n\n## Testing\n\n```bash\n# Run all tests from root\npytest ./tests\n```\n\n## License\n\nThis project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.\n\n## Author\n\n**Ricardo Yanzon** - [ricayanzon](https://github.com/ricayanzon)\n\n## Acknowledgments\n\n- Inspired by the R package [pcaMethods](https://github.com/bioc/pcaMethods)\n- Built with [PyTorch](https://pytorch.org/)\n",
"bugtrack_url": null,
"license": null,
"summary": "Comprehensive PyTorch implementation of PCA.",
"version": "0.1.3",
"project_urls": {
"GitHub": "https://github.com/ricayanzon/pytorch_pca",
"Issues": "https://github.com/ricayanzon/pytorch_pca/issues"
},
"split_keywords": [
"pca",
" principal component analysis",
" pytorch",
" torch",
" svd pca",
" svd",
" singular value decomposition",
" nipals",
" nonlinear iterative partial least squares",
" non-linear iterative partial least squares",
" rnipals",
" robust nipals",
" bpca",
" bayesian pca",
" ppca",
" probabilistic pca",
" svd impute",
" iterative pca",
" robust pca",
" nonlinear pca",
" non-linear pca"
],
"urls": [
{
"comment_text": null,
"digests": {
"blake2b_256": "0c41c4ff30a1895ab35705072b0b7228740212903dfa149d70814a55d76e5d68",
"md5": "503bb5b4cd177a46db884145447cd035",
"sha256": "21ecb9c79e358626b337a4597578d9eb57e6532ccea0921d859737b0a5d2694d"
},
"downloads": -1,
"filename": "pytorch_pca-0.1.3-py3-none-any.whl",
"has_sig": false,
"md5_digest": "503bb5b4cd177a46db884145447cd035",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.11",
"size": 19244,
"upload_time": "2025-07-15T21:36:54",
"upload_time_iso_8601": "2025-07-15T21:36:54.299690Z",
"url": "https://files.pythonhosted.org/packages/0c/41/c4ff30a1895ab35705072b0b7228740212903dfa149d70814a55d76e5d68/pytorch_pca-0.1.3-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": null,
"digests": {
"blake2b_256": "a728d377b1335abd413edc48fd4d52880735857d900e5d70d1992ea1939d677a",
"md5": "ba33c42900f11667fb4f6ebbecb8dbb4",
"sha256": "0fbfb20a283093a8d5460bd5e932ae32c1ee47034cf38da8d588f0a7da6ba1a6"
},
"downloads": -1,
"filename": "pytorch_pca-0.1.3.tar.gz",
"has_sig": false,
"md5_digest": "ba33c42900f11667fb4f6ebbecb8dbb4",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.11",
"size": 24969,
"upload_time": "2025-07-15T21:36:55",
"upload_time_iso_8601": "2025-07-15T21:36:55.460457Z",
"url": "https://files.pythonhosted.org/packages/a7/28/d377b1335abd413edc48fd4d52880735857d900e5d70d1992ea1939d677a/pytorch_pca-0.1.3.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2025-07-15 21:36:55",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "ricayanzon",
"github_project": "pytorch_pca",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"requirements": [
{
"name": "torch",
"specs": [
[
"==",
"2.7.1"
]
]
}
],
"lcname": "pytorch-pca"
}
JSON
