https://arxiv.org/abs/2205.10772

The code is tested on Python 3.9, CUDA 11, CuDNN 8.2. To install the dependencies:

pip install -r requirements.txt -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

scripts contains the scripts to reproduce the experiments. The usage is

mkdir -p ~/run/liv
cd scripts
python main.py <exp_name> -ng <num_gpus> -nm <num_tasks_per_gpu> -train [exp-specific args]

where exp_name can be dgmm, gp-uq or demand. Each choice corresponds to one or more experiments in paper; see the corresponding file in the scripts/ directory for explanations.

Results will be logged into a directory of the form ~/run/liv/<date>-<time>-<expname>_/. Each of its subdirectory corresponds to a single experiment, within which stdout contains the test metrics. The experiment-specific scripts contain instructions for the extraction of the results.

This repository contains files from deepmind/ssl_hsic and google/flax, the licenses of which are appended to the corresponding file.