This is the repository for Bayesian Image-on-image Regression via Deep kernel learning based Gaussian Processes (BIRD-GP). The repository contains
- birdgp: a python module for BIRD-GP
- birdgp_example: a Fashion MNIST based example of using the BIRD_GP class in birdgp module
- MNIST: code for the sythetic data analysis based on the MNIST dataset (Section 4.1 in the manuscript)
- Fashion MNIST: code for the sythetic data analysis based on the MNIST dataset (Section 4.2 in the manuscript)
- HCP: code for the HCP fMRI analysis (Section 5 in the manusript)
For synthetic data analysis, the notebook digits_birdgp and fashion_birdgp use an earlier version of BIRD-GP implementation that separates Stage 1 and Stage 2. In the birdgp_example folder, we provide a Fashion MNIST based example of using the BIRD_GP class in birdgp module.
The python module can be loaded by
import sys
sys.path.append("birdgp")
import bird_gp
The module depends on numpy, torch, tdqm, scipy, matplotlib, pandas and sklearn, please make sure the dependencies are installed.
As part of BIRD-GP, the kernel learning neural network for basis fitting BFNN, the horseshoe-prior linear regression FastHorseshoeLM and the Bayesian neural network with Stein variational gradient descent svgd_bnn are also implemented as python class and can be used independently. We also provide a helper function for generating grids over images voxels generate_grids.
An example of applying BIRD-GP on a synthesized Fashion MNIST dataset is included in the birdgp_example folder.