I remember there is a bug in the implementation (something related to triangular_solve and cholesky_solve) but I do not have time to fix it now. Please use the code with caution. I will fix it later.
This repository contains a pure PyTorch implementation of an online sparse Gaussian process regression model proposed by Bui et. al., NeurIPS 2017.
# Create a virtual environment
conda create -n osgpr python=3.8
conda activate osgpr
# Install PyTorch (2.1.0) from https://pytorch.org/get-started/locally/. For example,
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
# Install other dependencies
pip install matplotlib tqdm git+https://github.com/gngdb/pytorch-minimize.git
Simply run the main.py file to see an 1D demo:
python main.py
If you would like to compare the results with the original GPflow implementation, install GPflow 2.6:
pip install gpflow==2.6
Run the sanity_check.py file to see the side-by-side comparison:
python sanity_check.py
