RepoSim4Py

An embedding-approach-based tool for comparing semantic similarities between different Python repositories by using different information from repositories.

Model Details

RepoSim4Py is a pipeline based on the HuggingFace platform for generating embeddings according to specified Github Python repositories. For each Python repository, it generates embeddings at different levels based on the source code, code documentation, requirements, and README files within the repository. By taking the mean of these embeddings, a repository-level mean embedding is generated. These embeddings can be used to compute semantic similarities at different levels, for example, using cosine similarity to get comparison.

Model Description

The model used by RepoSim4Py is UniXcoder fine-tuned on code search task, using the AdvTest dataset.

Pipeline developed by: Henry65
Repository: RepoSim4Py
Model type: code understanding
Language(s): Python
License: MIT

Model Sources

Repository: UniXcoder
Paper: UniXcoder: Unified Cross-Modal Pre-training for Code Representation

Uses

Below is an example of how to use the RepoSim4Py pipeline to easily generate embeddings for GitHub Python repositories.

First, initialise the pipeline:

from transformers import pipeline

model = pipeline(model="Henry65/RepoSim4Py", trust_remote_code=True)

Then specify one (or multiple repositories in a tuple) as input and get the result as a list of dictionaries:

repo_infos = model("lazyhope/python-hello-world")
print(repo_infos)

Output (Long numpy array outputs are omitted):

[{'name': 'lazyhope/python-hello-world',
  'topics': [],
  'license': 'MIT',
  'stars': 0,
  'code_embeddings': array([[-2.07551336e+00,  2.81387949e+00,  2.35216689e+00, ...]], dtype=float32),
  'mean_code_embedding': array([[-2.07551336e+00,  2.81387949e+00,  2.35216689e+00, ...]], dtype=float32),
  'doc_embeddings': array([[-2.37494540e+00,  5.40957630e-01,  2.29580235e+00, ...]], dtype=float32),
  'mean_doc_embedding': array([[-2.37494540e+00,  5.40957630e-01,  2.29580235e+00, ...]], dtype=float32),
  'requirement_embeddings': array([[0., 0., 0., ...]], dtype=float32),
  'mean_requirement_embedding': array([[0., 0., 0., ...]], dtype=float32),
  'readme_embeddings': array([[-2.1671042 ,  2.8404987 ,  1.4761417 , ...]], dtype=float32),
  'mean_readme_embedding': array([[-1.91171765e+00,  1.65386486e+00,  9.49612021e-01, ...]], dtype=float32),
  'mean_repo_embedding': array([[-2.0755134,  2.8138795,  2.352167 , ...]], dtype=float32),
  'code_embeddings_shape': (1, 768)
  'mean_code_embedding_shape': (1, 768)
  'doc_embeddings_shape': (1, 768)
  'mean_doc_embedding_shape': (1, 768)
  'requirement_embeddings_shape': (1, 768)
  'mean_requirement_embedding_shape': (1, 768)
  'readme_embeddings_shape': (3, 768)
  'mean_readme_embedding_shape': (1, 768)
  'mean_repo_embedding_shape': (1, 3072)
}]

More specific information please refer to Example.py. Note that "github_token" is unnecessary.

Training Details

Please follow the original UniXcoder page for details of fine-tuning it on code search task.

Evaluation

We used the awesome-python list which contains over 400 Python repositories categorized in different topics, in order to label similar repositories. The evaluation metrics and results can be found in the RepoSim4Py repository, under the Embedding folder.

Acknowledgements

Many thanks to authors of the UniXcoder model and the AdvTest dataset, as well as the awesome python list for providing a useful baseline.

UniXcoder (https://github.com/microsoft/CodeBERT/tree/master/UniXcoder)
AdvTest (https://github.com/microsoft/CodeXGLUE/tree/main/Text-Code/NL-code-search-Adv)
awesome-python (https://github.com/vinta/awesome-python)

Authors

Honglin Zhang (https://github.com/liaomu0926)
Rosa Filgueira (https://www.rosafilgueira.com)

Henry65
/

RepoSim4Py