The script run.sh contains the exact configurations from the paper and can be run to replicate the results.
Assuming you have a reasonably up-to-date LaTeX distribution installed, it will generate the tables displaying the results as a file named tables.pdf and the plots as a file named result-plots.pdf.
Searching for values of the logical weight λ can be performed by running the script lambda-search.sh and specifying the number of epochs to train for.
Our experiments used 200 epochs, so by running bash lambda-search.sh 200 you will be able to replicate our exact results.
This script will generate the plots as a file named lambda-search-plots.pdf and output the optimal value of λ to the console.
Plots and tables generated from the data used in the paper can be found in the figures/ directory.
The experiments were run on Python 3.10.
The provided requirements.txt can be used to install the required packages using pip install -r requirements.txt, however, it will only install stable versions.
The exact nightly build versions used were torch 2.1.0.dev20230715 and torchvision 0.16.0.dev20230715.