Talk on 09.10.2018

• Code2Vec

  • Original implementation by authors is quite hard to understand
  • It uses javaparser
  • src-d implementation isn't complete
  • Attention can be improved
    • Check it during modification stage

• How to mine file pairs from Idea history

  • Go is cool
    • Works faster
    • There is a babelfish client
  • Python
    • Git for Python
    • Everything else is in Python

• Prediction

  • Author of methods

• Parsing

  • LSP (Microsoft)
    • No sdk for python
    • More complex than babelfish
  • TreeSitter
    • Work in progress
    • No Java
  • Javaparser
    • Original implementation of code2vec
    • Downsides are known
    • Hard to extend to other languages
  • Psi
    • You have to run idea to use it

• Data

  • File before / after
  • Author
    • Clean data to avoid multiple aliases for single author
  • Date of commit
  • Metadata

• Pipeline

  • Extract pairs of files
  • Build uast
  • Extract all methods
  • Match pairs of methods and detect changed ones
  • Pairs of methods are our dataset
  • Process it to be compatible with code2vec

• Motivation

  • Code review, knowledge transfer
  • Authorship of small snippets of code

• Another level of abstraction

  • Batches of methods to get importance of single method

• Tasks

  • Vladimir: mining in python
  • Egor: code2vec, review of implementations