Paper link: https://arxiv.org/abs/2106.05459
This code is the implementation of the learning chain used in the paper. Single experiment with a particular learning chain can be started as:
python run_experiment.py N,
where N is the number starting from 1 to 300 meaning the particular configuration of the learning chain returned by the config_factory() function.
Every experiment saves the statistics from all induced distribution along the learning chain in the pickle file. Mode recovery costs are computed using functions from compute_pkls.py. compute_pkls.py parallelize computation of mode recovery costs across multiple cpus such that each process takes an independent pickle file.
The jupyter notebook plots.ipynb implements all the plots which are used in the paper. Please contact me if you want to get pkl files from our experiments.
- torch
- scipy
- numpy
- tqdm
- fire