Code used for the experiments in **"Wide Neural Networks with Bottlenecks are Deep Gaussian Processes"** (<https://arxiv.org/abs/2001.00921>).
# Overview
Code used for the experiments in the paper **"Wide Neural Networks with Bottlenecks are Deep Gaussian Processes"** (<https://arxiv.org/abs/2001.00921>).
All scripts relevant to the experiments in the paper are located in the [*paper_experiments*](paper_experiments/README.md) directory.
Older material not included in the paper is kept in the *other_scripts* directory.
# Dependencies
For the experiments in the [*paper_experiments*](paper_experiments/README.md) directory, we used:
- python 3.7.7
- pytorch 1.3.1
- numpy 1.17.4
- scikit-learn 0.22.1
All scripts relevant to the experiments in the paper are located in the [paper_experiments](./-/tree/master/paper_experiments) directory.
Older material not included in the paper is kept in the [other_scripts](./-/tree/master/other_scripts) directory.
The *\*.lsf* files are the jobscripts for the [OLCF Summit supercomputer](https://www.olcf.ornl.gov/summit).
Each jobscript runs the main script *main.py*;
MPI-rank-dependent arguments are set with the script *mapper.sh*.
The bottleneck neural network Gaussian process (NNGP) model is defined in *modules.py*,
and the loaders for the Rings, Iris, and Boston datasets are contained in *datasets.py*.
The script *functions.py* implements a memory-efficient custom gradient for the ReLU-NNGP kernel.
# Report generation
All results dictionaries of the experiments will be placed in a *results* subdirectory.
Assuming this subdirectory exists, the script *report/main.py* generates the figures and LaTeX scripts that compile to give "reports" or summaries of the results.
