Yeo-Johnson Transformed Distributions for PyTorch

This repository provides a PyTorch implementation of the Yeo-Johnson Transformation applied to distributions, allowing for direct sampling from transformed normal distributions. The main use case is to enable sampling from normal distributions transformed using a Yeo-Johnson Transform. Sampling from other base distributions can also be easily implemented by following the provided YeoJohnsonNormal class example.

Features

• Yeo-Johnson Transform: Implements the Yeo-Johnson transformation, which can handle both positive and negative inputs, as a PyTorch transform.
• Transformed Normal Distribution: Provides a YeoJohnsonNormal class for directly sampling from normal distributions transformed with the Yeo-Johnson transform.
• Extendibility: Designed to easily extend to other base distributions following the example provided.

Installation

This package requires PyTorch. You can install the latest version of PyTorch by following the instructions on the official PyTorch website.

Once PyTorch is installed, you can clone this repository to your local machine:

git clone https://github.com/lwelzel/yjt-distributions.git

Then, navigate to the repository directory and install the package using pip:

cd yjt-distributions
pip install -e . 

Usage

Here's how to use the YeoJohnsonNormal class to sample from a Yeo-Johnson transformed normal distribution:

import torch
from yjt-distributions import YeoJohnsonNormal

# Parameters for the normal distribution and the transformation
loc = 0.0
scale = 1.0
lbda = 0.5  # Lambda parameter for the Yeo-Johnson transform
tloc = 0.0  # Translation parameter
tscale = 1.0  # Scaling parameter

# Create a YeoJohnsonNormal distribution
yj_normal = YeoJohnsonNormal(loc=torch.tensor([loc]),
                             scale=torch.tensor([scale]),
                             lbda=torch.tensor([lbda]),
                             tloc=torch.tensor([tloc]),
                             tscale=torch.tensor([tscale]))

# Sample from the distribution
samples = yj_normal.sample((1000,))

Extending to Other Distributions

To extend this implementation to other base distributions, you can follow the pattern used in YeoJohnsonNormal. Essentially, you will need to create a new class that inherits from TransformedDistribution and specify your base distribution along with the Yeo-Johnson transform as the transformations to apply. Contributing

Contributions are welcome! Please feel free to submit pull requests or open issues to improve the repository or extend its functionality.

License

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

Original Use Case

The original use case for this was the need to sample from a transit depth dataset which was normalized using a YJ-transform. This is useful since the individual transit depths are normally distributed, however after transforming the spectra, the uncertainties are not normal anymore. Sampling from the transformed distributions is more convienient for downstream learning tasks.