jchyb / masterthesis Goto Github PK

It primarly consists of heavily adjusted codes of https://github.com/wiktor145/QTrator and https://github.com/wiktor145/tranformers-qtranspilers.

• ML_measurement.ipynb contains implementations around the used models and measurements done for the purpose of evaluating their use.
• Qtrator contains the modified set of tools for quantum transpiler benchamarking based on https://github.com/wiktor145/QTrator , with modifications:
  • Changes to optimizers themselves (e.g. to the basis gate sets), with the aim of standardizing the options for better comparisons between the optimizers.
  • "custom" optimizers were added, where the optimization passes used can be adjusted with strings of zeroes and ones (otherwise they are kept similar to the QiskitLevel3 transpiler). They can be found in the QTrator/custom_optimizers directory.
  • Some directories holding previous results/tests were removed, for readability.
  • The main QTrator code (run_benchmark.py and evaluate_results.py) was modified to:
    • allow multi-process execution to be able to obtain the results faster (for the purposes of which I turned of time measurement used in the transpilation time metric, now ignored). The multi-process execution can be turned off directly in-code by replacing the pool.map commands with the commented-out for-loops.
    • allow separate testing of the "custom" optimizers - adjusted with the option test generative optimizers set in benchmark_configuration.yml.
    • allow recursively accessing directories in benchmark_results, and name resulting files accordingly (by including __ in place of \ in path, to flatten the tree).
    • not overwrite results when evaluating - instead, it creates numbered files in bench_res_csv directory (like bench_res_csv\benchmark_results (1).csv)
• tranformers-qtranspilers contains a set of scripts that are able to transform the QTrator output to a dataset for use when training and testing related ML models, with following adjustments:
  • dataset-generating scripts were replaced with create_avg_dataset_custom.py (for the "custom" transpilers) and create_avg_dataset.py (for the regular transpilers) scripts. The main adjustment there is the fact that the dataset is calculated based on multiple different QTrator benchmarking runs, not just one. The related (hard-coded) metrics are averaged-out, and muliple datasets based on them are produced, with filenames titled accordingly. Additional sub-procedure to the "custom" version of the script is added that aims to filter out the "rarely" most-performant optmizers, with the lower-bound limit of 50, to improve training results.
  • unrelated files used previously were removed for readability.
• ML_measurements.ipynb contains the ML training and measurements code used in google colab.
• custom_prelim and prelim contain the datasets of the preliminary tests
• final_ml_dataset contains the dataset used when training ML models

Some things ended up being changed last-minute, like removing some passes of compositive methods. E.g. in the tranformers-qtranspilers\other_files\create_avg_dataset_custom.py file there is a subprocedure that aims to remove an optimization pass from consideration when creating final dataset. cleanup_filter_custom.py was used for similar reasons on the preliminary results.