Probabilistic Submodular Models

Tech stack

• Python 3.7 or superior
• Hydra configuration framework

Initialize project

Install the virtual environment:

python3 -m venv ./venv

Activate the virtual environment:

source ./venv/bin/activate

Install third-party dependencies:

python3 -m pip install -r python/requirements.txt

How to run

This project is composed of 3 Python modules (one for each step of the application) that should be run in sequence. The output of the application is stored in the /out folder in a hierarchical fashion.

It's easier to explain the hierarchical structure of the output via an example. Consider an experiment with the demo_monotone probabilistic submodular model with a 3-dimensional ground set, M=10000 samples obtained with a metropolis sampler. The ground truth density map, which associates each subset to the probability of being sampled from the ground truth distribution, is saved in /out/demo_monotone/n-3/ground_truth.csv. The history of the samples obtained with the metropolis sampler is saved in /out/demo_monotone/n-3/M-10000/metropolis/history.csv. The cumulative probability distances are saved in /out/demo_monotone/n-3/M-10000/metropolis/cumulative_probability_distances.csv. The plots are saved in /out/demo_monotone/n-3/M-10000/metropolis/plots.

(1) Sample Step: python.mcmc_sample

• Compute the ground truth density map (which associates each subset to the probability of being sampled from the ground truth distribution) only once
• Run the specified samplers over the target probabilistic submodular models, saving the history of the samples obtained in CSV

python3 -u -m python.mcmc_sample -m \
  obj=demo_monotone,demo_non_monotone,delta_5 \
  sample_size=small,middle,big \
  sampler=gibbs,metropolis,lovasz_projection

(2) Metrics collection Step: python.metrics

• Compute the cumulative probability distances for all experiments run in the previous step, saving them in CSV

python3 -u -m python.metrics

(3) Plot generation Step: python.plotter

• Plot the cumulative probability distance between each sampler's outcome and the ground truth distribution

python3 -u -m python.plotter

Configuration

Please read and edit /python/conf/config.yaml