Skip to content

ShounakRay/CodeSage

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CodeSage

Often times, there is often a more preferable and efficient implementation of the code a developer has written for their project. It is only through code reviews and extensive research that one is able to enhance their code quality, both from a functionality and style perspective. It is especially and relatively a less fluid and time-consuming process for fledgling coders to gain inspiration from the relevant and exemplary repositories online. Thus, the overarching problem we wish to address is the following: how can we provide developers with "better" alternative code snippets with respect to what they’re writing and explain why the alternative is better? Answering this question involves computationally searching numerous code samples online, implicitly representing the relationships between sub-optimal and "optimal" code, and autonomously reasoning to the user why one approach is better than the other

TODO:

[Link Google Doc here]

Immediate Reference:

"""
Format that Shounak needs (Joe sends this to him):
[code_id1, code_id2, ...]

Reference dictionary: {code_id: (documentation, actual_code, reputation_score)}

Shounak's output:
{code_id: intent_category}


Michael (seq2seq intermediate step):
{intent_category: (good_code -> List[code_id],
                   bad_code -> List[int])}

Dataset/Input to the seq2seq: completeGraph (every combo) between all the `good_code` and `bad_code`
                                for a given intent_category

"""

# Joe:
# TODO: Update documentation that we use
# (via CodeBert OR another API that provides better documentation) in _get_data
# TODO: Create an alternative way to get reputation scores through Code Reviewer

# Michael:
# TODO: Convert Jupyter notebook into something Shounak can directly pull an output from
# TODO: Make an arbitrary dataset and build a basic seq2seq (includes thresholding based on reputation score)
# TODO: (Eventually) Multiprocess the dataset loading

# Shounak:
# TODO: Find a better embedding for the documentation instead of TF-IDF vectorizer
# TODO: Building out KMeans supervised clustering and comparing it to SOM unsupervised,
#       sanity-check which one is better

## Structure of Respository

File Structure

the_glue.py

This is what calls on all the modules and runs the entire pipeline, end-to-end.

_accessories

This are any helper functions that will be shared across any of the Modules.

Modules

Note: every module has Extracontent which is code you're no longer using, but might be important for reference and _accessories which are helper functions for your stage process.

Stage 1 – DataIngestion

Hits the CodeBERT or alternative API. This ingests the remote dataset and spits out format: [code_id1, code_id2, ...] with reference dictionary {code_id: (documentation, actual_code, reputation_score)}.

Stage 2 – IntentClustering

Done via supervised KMeans or unsupervised self-organizing map. This takes all the code, and spits out format {code_id: intent_category} for all code_ids.

Stage 3 – Code2Code

This is our seq2seq model (RNN –> ... -> Transformer) that:

  1. Generates the dataset for the seq2seq
  2. Trains the seq2seq model

Stage 4 – Code2Explanation

This is able to ingest a given code snippet and output exemplar code snippets and explanations. Hits the Codex API.

About

A generative framework mapping "good" code to "bad" code

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •