Code for Parameter-free Locally Accelerated Conditional Gradient and its experiments.

Setting up the conda environment to run experiments

We use Anaconda for Python package management. When you have it installed, you can create a new conda environment using environment.yaml:

$ conda env create --name pflacg --file=environment.yaml

Activate the pflacg conda environment for executing the code in the pflacg Python module:

$ conda activate pflacg

A brief overview of our codebase

The bulk of the code for running the experiments is located within pflacg which can be imported as a Python module from the root of this repository (or alternative if copied to the binary directory of your Python installation). All implemented algorithms are located under pflacg/algorithms and the implementation for Parameter-free Locally Accelerated Conditional Gradient is located inside pflacg/algorithms/pflacg.py. The code for construsting relevant objective functions and feasible regions is located under pflacg/experiemnts. The primary tool we use for running our experiments is pflacg/experiments/experiment_driver.py and instructions for using it are included in the next section.

Running the experiments

Using the experiments driver

Here we provide detailed instructions on how to run experiments using experiments_driver.py. Please find example configurations for the algorithms, feasible regions and objectives functions under examples, or alternative you can construct the necessary objects using your custom Python script.

python -m pflacg.experiments.experiments_driver --task run_algorithms \
    --save_location $SAVE_LOCATION \
    --objective_function_config $PATH_TO_OBJECTIVE_FUNCTION_CONFIG \
    --feasible_region_config $PATH_TO_FEASIBLE_REGION_CONFIG \
    --algorithms_config $PATH_TO_ALGORITHMS_CONFIG \
    --exit_criterion_type $EXIT_CRITERION_TYPE \
    --exit_criterion_value $EXIT_CRITERION_VALUE \
    --max_time $MAX_TIME

Optional arguments:

• --save_objects: Save all experiments related objects (objective function, feasible region and algorithms).
• --objective_function_use_pickled $PATH_TO_PICKLED_OBJECTIVE_FUNCTION: Use an existing pickled objective function. Do not use this with --objective_function_config.
• --feasible_region_use_pickled $PATH_TO_PICKLED_FEASIBLE_REGION: Use an existing pickled feasible region. Do not use this with --feasible_region_config.
• --num_cpu_per_process $NUM_CPU: Allowed number of CPUs per process. If 0, then no limit. Default is 1.
• --random_seed $SEED: Seed for random in NumPy.
• --save_logging: Save logs to a file in save_location.
• --clear_pycache: Clear python cache. Please include this option after any changes is made to the code.

Upon completion, the results from an experiment run will be save as text files within your specified $SAVE_LOCATION. You can then use experiment_driver.py to generate

Example run commands

We provide two pseudo-commands for using the experiment driver. The first comand runs algorithms over the supplied objective function and feasible, while the second command plots the run results.

$ python -m pflacg.experiments.experiments_driver --task run_algorithms \
    --save_location test_runs \
    --objective_function_use_pickled ./Quadratic-dim_5000.pickle \
    --feasible_region_use_pickled ./ProbabilitySimplexPolytope-dim_5000.pickle \
    --algorithms_config ./examples/algorithms/example_PFLaCG.json \
    --exit_criterion_type SWG \
    --exit_criterion_value 0.001 \
    --num_cpu_per_process 1 \
    --max_time 100 \
    --clear_pycache \
    --save_logging

$ python -m pflacg.experiments.experiments_driver --task plot_results \
    --save_location "your/save/location" \
    --plot_config "examples/plots/plot_config.json"

Using Jupyter notebook

Jupyter notebook comes installed with our pflacg conda environment. Start a Jupyter notebook server while in the root directory of this repository:

jupyter notebook

When inside a Jupyter notebook, you can import the pflacg module by simply running import pflacg. To set up your custom experiements, please first instantiate your desire feasible region and objective function objects, then instantiate algorithm objects such as ParameterFreeLaCG. To run your experiements, simply call the run method inside any algorihtm objects.

Example notebook

We provide a toy example notebook for using our pflacg module and it is available under examples/Example-Quadratic-ProbSimplex.ipynb.