pySynthesis

This tool enables the automatic synthesis of Python methods by linearly combining blocks of existing code whose comments match a provided textual description (this is considered the synthesis specifications). The matching mechanism implements a behaviorizeable relation [1] and thus guarantees that there is an underlying logical system in which synthesis performs inference. In simple terms, you can learn to use this tool to easily combine blocks of code.

Dependencies
nltk

Contact
Emmanouil (Manios) Krasanakis, [email protected]

Contents

1. Demo
2. Command line
3. Citation
4. Internal step
5. Benchmarks

Demo

Install Python and necessary dependencies. Then run the file synthui.py to perform synthesis:

demo.mp4

⚠️ For other synthesis tasks, you need to provide example code.

Command line

Install Python and necessary dependencies, open a terminal in the project's folder and run the command: python synth.py -h to view all possible arguments. For example, to synthesize the following quoted phrase based on a file of example implementations, run:

python synth.py "train and evaluate and save an SVR model" examples/example.py

This outputs the following code:

def solution(x, y, path):
    svr = SVR()
    svr.train(x, y)
    pickle.dump(svr, path)
    y_hat = svr.predict(x, probs=True)
    return svr, scipy.metrics.auc(y, y_hat)

Adding textual predicates to the description either adds or clarifies functionality (to switch to a different implementation).

A list of interesting queries that can be synthesized from the example code file (try --vstrict 1.1 and --vstrict 1.2):

"Calculate PageRank."
"Calculate PageRank with normalized priors."
"Calculate PageRank with symmetric normalization and normalized priors."
"Calculate PageRank with asymmetric normalization and normalized priors."

"Train a model."
"Train an SVR model."
"Train and evaluate an SVR model."
"Train and evaluate a logistic regression model."
"Train, evaluate and save an SVR model."
"Train and evaluate and save an SVR model using normalization on training data."
"Train and evaluate a logistic regression model using standardization on training data."

Citation

If you reuse or derive code as part of your research, we ask that you cite the following work:

@article{krasanakis2021defining,
  title={Defining Behaviorizeable Relations to Enable Inference in Semi-Automatic Program Synthesis},
  author={Krasanakis, Emmanouil and Symeonidis, Andreas},
  journal={Journal of Logical and Algebraic Methods in Programming},
  pages={100714},
  year={2021},
  publisher={Elsevier}
}

Internal steps

To expose how synthesis works, add the --show-known option. To display imported code blocks add the --verbose option to show intermediate steps. To avoid renaming intermediate variables and thus keep compatible variable names between the two options, use --explain instead of the second one. For example:

python synth.py "pagerank with symmetric normalization" examples/example.py --explain --show-known

⚠️ The --explain option could introduce erroneous synthesis outcomes if the same code block is added more than one time to the synthesized program (this should happen only rarely and typically means that specifications were misunderstood).

Benchmarks

Benchmarks additionally depend on the tqdm,numpy,matplotlib libraries. After installing these, the running time measurements of our paper can be replicated by the scripts:

• benchmark_specification_complexity.py to analyse how synthesis time scales with user specification complexity
• benchmark_corpus_complexity.py to analyse how synthesis time scales with the total complexity of known specifications in corpora of various sizes

New Examples

python synth.py "run default personalized pagerank from a small graph's signal and assess it on a signal split" examples\pygrank_snippets.py

"run personalized pagerank"
"run default personalized pagerank"