GSVGD

[Paper][Slides][Poster]

Data

Covertype data is downloaded from https://archive.ics.uci.edu/ml/datasets/covertype

Other Dependencies

• Code for Sliced-SVGD is adapted from Wenbo Gong's repo
• Code for optimization on Grassmann manifold is adapted from Pymanopt

Run experiments

The code below provides an example of running the numerical experiments in the paper. - The .sh scripts assume 8 GPUs are available. You can also use CPUs by changing the --gpu argument in these scripts to --gpu=-1. - Note: These experiments can take several hours to finish, since they encompass various configurations of dimensions/sample sizes, along with multiple repetitions. If you wish to obtain results faster, you can reduce the number of configurations accordingly in the .sh scripts.

To run: 1. Install the GSVGD module pip install -e . 2. Run experiments (the full experiments can take several hours) ```

e.g.1 run multivariate gaussian experiment and generate plots

sh scripts/run_gaussian.sh

e.g.2 run conditioned diffusion and generate plots

sh scripts/run_diffusion.sh ```

Basic usage

``python ''' distribution: target distribution class withlog_prob` method (up to a constant term) kernel: instance of kernel class manifold: instance of Grassmann manifold class for projector update optimizer: instance of optimizer class for particle update

delta: stepsize for projector update T: initial temperature T0 X: initial particles A: initial projectors m: number of projectors epochs: number of iterations '''

instantiate GSVGD class

gsvgd = FullGSVGDBatch( target=distribution, kernel=kernel, manifold=manifold, optimizer=optimizer, delta=delta, T=T )

update particles

_ = gsvgd.fit(X=X, A=A, m=m, epochs=epochs, threshold=0.0001*m)

final particles: X (updates are done in-place)

```

Run tests

python python -m pytest

Code directory

.

requirements.txt: Dependencies.
setup.py: Setup script.
data: covertype data.
experiments: main scripts for the 5 numerical experiments.
plots: folder to hold plots.
scripts: Shell scripts to run the experiments and generate plots.
src: Source files for implementing each sampling method, and util functions for experiments.
   Sliced_KSD_Clean: Utils for Sliced SVGD adapted from Wenbo Gong.
   blr.py: Utils for Bayesian logistic regression.
   diffusion.py: Utils for conditioned duffition.
   gsvgd.py: GSVGD class (main).
   kernel.py: Kernel class.
   manifold.py: Class for optimisation on the Grassmann manifold, adapted from Pymanopt
   metrics.py: Metric class for evaluation of results.
   s_svgd.py: S-SVGD class, adapted from Wenbo Gong.
   svgd.py: SVGD class.
   utils.py: Other utils.
tests: unittests
thumbnail: Thumbnail fig.
README.md