Optimal Continuous DR-Submodular Maximization and Applications to Provable Mean Field Inference

This repository collects source code for the paper:

"Optimal Continuous DR-Submodular Maximization and Applications to Provable Mean Field Inference"

ICML 2019. Yatao A. Bian, Joachim M. Buhmann and Andreas Krause

File Structure:

functions/: contains utility files
- exp_specs.py: contain specifications for the experiments
- utils.py: utility functions
- process_results/: contains notebook to calculate experimental stats
main.py: the main file to run different experiments
data/: contains FLID models

Usage:

The absl flags define the main usage of the code. Specifically,

flags.DEFINE_integer('problem_id', 1, 'Options: 1: ELBO, 2: PA-ELBO.')

For ELBO, we run different algorithms on the ELBO objective
For PA-ELBO, we run different algorithms on the PA-ELBO objective

flags.DEFINE_string('mode', 'run', 'Options: run: run algorithms; stats: get experimental statistics.')

'run': run different algorithms/solvers, dump the results into pickle file and plot figures
'stats': generate function values returned in all the experiments and dump them into a pickle file.

flags.DEFINE_boolean('debug', True, 'Whether it is in debug mode.')

In debug mode, one only run solvers on one fold of FLID model, which is much faster than the non-debug mode.
Be careful, it may takes hours to run the non-debug mode, especially for the PA-ELBO objectives.

Example:

Let us consider an example for the ELBO objective.

Firstly navigate to the folder in your command line and run:

$ python main.py --mode=run --debug=False --problem_id=1

Results will be stored in the ./results folder. To get the experimental stats file, you can then run:

$ python main.py --mode=stats --debug=False --problem_id=1

A pickle file called optf_1epoch.pkl will be generated in the same result folder. In order to see more details of experimental stats, you can go to the folder process_results/ and play with the notebook process_results.ipynb.

Dependencies:

The code has been tested on Ubuntu 17.10, 64 bits with Python 3.6. It should work with other OS with little change.

Copyright: