KevinRojas1499/ZOD-MC

README

Zeroth-Order Sampling Methods for Non-Log-Concave Distributions

This is the official repo for the ZOD-MC.

Setting up:

To install the required packages simply run:

conda create --name zodmc python=3.12
pip install pip --upgrade
pip install -r requirements.txt

Sampling

We use wandb to track our experiments, if you don't have a wandb account you can use the sample.py file and process the samples as you want. If you do have wandb, you can sample using a diffusion based sampler from a 2d gaussian mixture by running:

python3 main.py --config config/general_config.yaml \
    --density_parameters_path config/density_parameters/2d_gmm.yaml \
    --density gmm --dimension 2

Table of config hyperparmeters to pick a diffusion method

Method	score_method	p0t_method
ZOD-MC	p0t	rejection
RDMC	p0t	ula
RSDMC	recursive	-

Your own distributions

You can implement your own distributions by adding them in the utils/densities.py file. You can add a distribution from a potential by passing it to DistributionFromPotential, and the gradients will be computed using pytorchs automatic differentiation.

Replicating our results:

After installing the required packages you may replicate all our results in the paper by running the following commands:

bash scripts/replicate_2d_mmd.sh
bash scripts/replicate_radius.sh
bash scripts/replicate_non_continuous.sh
bash scripts/replicate_mueller.sh
bash scripts/replicate_dimension.sh
bash scripts/replicate_score_error.sh