Code accompanying paper: Direct Amortized Likelihood Ratio Estimation

To run the code in this repository, first run:

pip install .

This repository contains:

• ./notebooks/quadcopter_example.ipynb: This is an example notebook that runs all three likelihood ratio estimators (DNRE, BNRE, NRE) to both train and sample quadcopter designs.

• ./src/dnre/benchmark.py: This code runs the benchmark using the following command:

  python src/dnre/benchmark.py --model_type dnre --task two_moons --save_dir ./benchmark_results/two_moons/dnre --device 0
  

  Where we have selected DNRE as the approach to perform grid search over in the above command. The code includes comments for all additional options.

• ./src/dnre/benchmark_evaluate.py: This code evaluates the above best result from the grid search using the following command:

  python src/dnre/benchmark_evaluate.py --model_type dnre --task two_moons --path_dir ./benchmark_results/two_moons/dnre --device 0 --metric coverage
  

  Where we point to the same directory as above. This will evaluate the expected coverage, but there are also options for C2ST and the log posterior.

• ./data/data_dict_4490: Contains the data from the quadcopter experiment.

Acknowledgements

This project was supported by DARPA under the Symbiotic Design for Cyber-Physical Systems (SDCPS) with contract FA8750-20-C-0002. The views, opinions and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.