martinspendl / ml-reproducibility-challenge-23

ML Reproducibility Challenge 23 - [Re] Easy Bayesian Transfer Learning with Informative Priors

Reproducibility report

This is the code for the reproducibility of parts of the paper: Pre-Train Your Loss: Easy Bayesian Transfer Learning with Informative Priors.

The code inside folder BayesianTransferLearning is copied from the authors GitHub, but adjusted for replicating the results. (THIS CODE BELONGS TO THE AUTHORS).

Installation

# using pip in a virtual environment
pip install -r requirements.txt

# using Conda
conda create --name <env_name> --file requirements.txt
conda activate <env_name>

Downloading the data

1. Run ./BayesianTransferLearning/Prapare Data/oxford-102-flowers.py to download the Oxford-102-Flowers data.
2. Run downsampled_data/create_downsampled_folders.py from the downsampled_data directory to create the subfolders with smaller data set sizes.
3. Download priors from the original authors and put them in to priors folder.

Run experiments

There are three experiments we ran corresponding to python scripts:

1. The influence of low-dimensional rank on performance (run_experiment_rank.py)

2. The influence of prior scaling on performance (run_experiment_scale.py)

3. Comparison of Bayesian and non-Bayesian learning (run_experiment_comparison.py)

To run experiments simply run the corresponding scipt. The results will accumulate in a text file (results_*.txt).