ABCdiscrete: Approximate Bayesian Computation for Discrete Data

The repository contains the python code required to reproduce the experiments carried out in the following paper:

• Auzina, I. A., & Tomczak, J. M. (2021). Approximate bayesian computation for discrete spaces. Entropy, 23(3), 312., Entropy

Requirements

The code requires:

• python 3.5 or above
• numpy
• scipy
• scikit-image
• matplotlib
• nasbench software and its dependencies (https://github.com/google-research/nasbench)

Run the Experiments

1. Open the experiments directory
2. Select one of the experiments of interest
3. Check the settings and update the pythonpath and the datapath (if needed), see example below:
   • PYTHONPATH = '/home/username/location/ABCdiscrete/experiments'
   • DATA_PATH = '/home/username/location/nasbench_only108.tfrecord'
4. Run the experiment

the python code is ran in multiple parallel processes (MAX_PROCESS), thus, check how many nodes you have available

Evaluate the Experiments

1. Open the evaluate directory
2. Select the experiment you want to evaluate
3. Check the settings and update the pythonpath and the datapath
4. Run the evaluation

Overall Design

The repository is organized in 7 folders, which details are described below:

• experiments: the directory contains the main execution files for each experiment (every experiment has a separate execution file).
• testbeds : the directory contains the use-cases utilised for the experiments. The super class main_usecase.py specifies the functionalities that any use-case must posses (if you want to implement an additional use-case).
• algorithms: contains the super class, main_sampling.py that specifies the minimum required functions, and the subclasses:
  • population-based MCMC mcmc.py
  • population-based ABC abc.py
• kernels: contains the possible proposal distributions.
• results: the directory where the results will be stored.
• evaluate: contains the execution files for the evaluation.
• utils: contains additional functionalities such as plotting or creating text files to aid storing the results in a more user-friendly way.