Home

Awesome

Tests License: MIT Python 3.6 Weights-And-Biases

The CodeSearchNet challenge has been concluded

We would like to thank all participants for their submissions and we hope that this challenge provided insights to practitioners and researchers about the challenges in semantic code search and motivated new research. We would like to encourage everyone to continue using the dataset and the human evaluations, which we now provide publicly. Please, see below for details, specifically the Evaluation section.

No new submissions to the challenge will be accepted.

Table of Contents

<!-- TOC depthFrom:1 depthTo:6 withLinks:1 updateOnSave:1 orderedList:0 --> <!-- /TOC -->

Quickstart

If this is your first time reading this, we recommend skipping this section and reading the following sections. The below commands assume you have Docker and Nvidia-Docker, as well as a GPU that supports CUDA 9.0 or greater. Note: you should only have to run script/setup once to download the data.

# clone this repository
git clone https://github.com/github/CodeSearchNet.git
cd CodeSearchNet/
# download data (~3.5GB) from S3; build and run the Docker container
script/setup
# this will drop you into the shell inside a Docker container
script/console
# optional: log in to W&B to see your training metrics,
# track your experiments, and submit your models to the benchmark
wandb login

# verify your setup by training a tiny model
python train.py --testrun
# see other command line options, try a full training run with default values,
# and explore other model variants by extending this baseline script
python train.py --help
python train.py

# generate predictions for model evaluation
python predict.py -r github/CodeSearchNet/0123456 # this is the org/project_name/run_id

Finally, you can submit your run to the community benchmark by following these instructions.

Introduction

Project Overview

CodeSearchNet is a collection of datasets and benchmarks that explore the problem of code retrieval using natural language. This research is a continuation of some ideas presented in this blog post and is a joint collaboration between GitHub and the Deep Program Understanding group at Microsoft Research - Cambridge. We aim to provide a platform for community research on semantic code search via the following:

  1. Instructions for obtaining large corpora of relevant data
  2. Open source code for a range of baseline models, along with pre-trained weights
  3. Baseline evaluation metrics and utilities
  4. Mechanisms to track progress on a shared community benchmark hosted by Weights & Biases

We hope that CodeSearchNet is a step towards engaging with the broader machine learning and NLP community regarding the relationship between source code and natural language. We describe a specific task here, but we expect and welcome other uses of our dataset.

More context regarding the motivation for this problem is in this technical report. Please, cite the dataset and the challenge as

@article{husain2019codesearchnet,
  title={{CodeSearchNet} challenge: Evaluating the state of semantic code search},
  author={Husain, Hamel and Wu, Ho-Hsiang and Gazit, Tiferet and Allamanis, Miltiadis and Brockschmidt, Marc},
  journal={arXiv preprint arXiv:1909.09436},
  year={2019}
}

Data

The primary dataset consists of 2 million (comment, code) pairs from open source libraries. Concretely, a comment is a top-level function or method comment (e.g. docstrings in Python), and code is an entire function or method. Currently, the dataset contains Python, Javascript, Ruby, Go, Java, and PHP code. Throughout this repo, we refer to the terms docstring and query interchangeably. We partition the data into train, validation, and test splits such that code from the same repository can only exist in one partition. Currently this is the only dataset on which we train our model. Summary statistics about this dataset can be found in this notebook

For more information about how to obtain the data, see this section.

Evaluation

The metric we use for evaluation is Normalized Discounted Cumulative Gain. Please reference this paper for further details regarding model evaluation. The evaluation script can be found here.

Annotations

We manually annotated retrieval results for the six languages from 99 general queries. This dataset is used as groundtruth data for evaluation only. Please refer to this paper for further details on the annotation process. These annotations were used to compute the scores in the leaderboard. Now that the competition has been concluded, you can find the annotations, along with the annotator comments here.

Setup

You should only have to perform the setup steps once to download the data and prepare the environment.

  1. Due to the complexity of installing all dependencies, we prepared Docker containers to run this code. You can find instructions on how to install Docker in the official docs. Additionally, you must install Nvidia-Docker to satisfy GPU-compute related dependencies. For those who are new to Docker, this blog post provides a gentle introduction focused on data science.

  2. After installing Docker, you need to download the pre-processed datasets, which are hosted on S3. You can do this by running script/setup.

    script/setup
    

    This will build Docker containers and download the datasets. By default, the data is downloaded into the resources/data/ folder inside this repository, with the directory structure described here.

The datasets you will download (most of them compressed) have a combined size of only ~ 3.5 GB.

  1. To start the Docker container, run script/console:
    script/console
    
    This will land you inside the Docker container, starting in the /src directory. You can detach from/attach to this container to pause/continue your work.

For more about the data, see Data Details below, as well as this notebook.

Data Details

Data Acquisition

If you have run the setup steps above you will already have the data, and nothing more needs to be done. The data will be available in the /resources/data folder of this repository, with this directory structure.

Schema & Format

Data is stored in jsonlines format. Each line in the uncompressed file represents one example (usually a function with an associated comment). A prettified example of one row is illustrated below.

Code, comments, and docstrings are extracted in a language-specific manner, removing artifacts of that language.

{
  'code': 'def get_vid_from_url(url):\n'
          '        """Extracts video ID from URL.\n'
          '        """\n'
          "        return match1(url, r'youtu\\.be/([^?/]+)') or \\\n"
          "          match1(url, r'youtube\\.com/embed/([^/?]+)') or \\\n"
          "          match1(url, r'youtube\\.com/v/([^/?]+)') or \\\n"
          "          match1(url, r'youtube\\.com/watch/([^/?]+)') or \\\n"
          "          parse_query_param(url, 'v') or \\\n"
          "          parse_query_param(parse_query_param(url, 'u'), 'v')",
  'code_tokens': ['def',
                  'get_vid_from_url',
                  '(',
                  'url',
                  ')',
                  ':',
                  'return',
                  'match1',
                  '(',
                  'url',
                  ',',
                  "r'youtu\\.be/([^?/]+)'",
                  ')',
                  'or',
                  'match1',
                  '(',
                  'url',
                  ',',
                  "r'youtube\\.com/embed/([^/?]+)'",
                  ')',
                  'or',
                  'match1',
                  '(',
                  'url',
                  ',',
                  "r'youtube\\.com/v/([^/?]+)'",
                  ')',
                  'or',
                  'match1',
                  '(',
                  'url',
                  ',',
                  "r'youtube\\.com/watch/([^/?]+)'",
                  ')',
                  'or',
                  'parse_query_param',
                  '(',
                  'url',
                  ',',
                  "'v'",
                  ')',
                  'or',
                  'parse_query_param',
                  '(',
                  'parse_query_param',
                  '(',
                  'url',
                  ',',
                  "'u'",
                  ')',
                  ',',
                  "'v'",
                  ')'],
  'docstring': 'Extracts video ID from URL.',
  'docstring_tokens': ['Extracts', 'video', 'ID', 'from', 'URL', '.'],
  'func_name': 'YouTube.get_vid_from_url',
  'language': 'python',
  'original_string': 'def get_vid_from_url(url):\n'
                      '        """Extracts video ID from URL.\n'
                      '        """\n'
                      "        return match1(url, r'youtu\\.be/([^?/]+)') or \\\n"
                      "          match1(url, r'youtube\\.com/embed/([^/?]+)') or "
                      '\\\n'
                      "          match1(url, r'youtube\\.com/v/([^/?]+)') or \\\n"
                      "          match1(url, r'youtube\\.com/watch/([^/?]+)') or "
                      '\\\n'
                      "          parse_query_param(url, 'v') or \\\n"
                      "          parse_query_param(parse_query_param(url, 'u'), "
                      "'v')",
  'partition': 'test',
  'path': 'src/you_get/extractors/youtube.py',
  'repo': 'soimort/you-get',
  'sha': 'b746ac01c9f39de94cac2d56f665285b0523b974',
  'url': 'https://github.com/soimort/you-get/blob/b746ac01c9f39de94cac2d56f665285b0523b974/src/you_get/extractors/youtube.py#L135-L143'
}

Summary statistics such as row counts and token length histograms can be found in this notebook

Downloading Data from S3

The shell script /script/setup will automatically download these files into the /resources/data directory. Here are the links to the relevant files for visibility:

The s3 links follow this pattern:

https://s3.amazonaws.com/code-search-net/CodeSearchNet/v2/{python,java,go,php,javascript,ruby}.zip

For example, the link for the java is:

https://s3.amazonaws.com/code-search-net/CodeSearchNet/v2/java.zip

The size of the dataset is approximately 20 GB. The various files and the directory structure are explained here.

Human Relevance Judgements

To train neural models with a large dataset we use the documentation comments (e.g. docstrings) as a proxy. For evaluation (and the leaderboard), we collected human relevance judgements of pairs of realistic-looking natural language queries and code snippets. Now that the challenge has been concluded, we provide the data here as a .csv, with the following fields:

Running Our Baseline Model

We encourage you to reproduce and extend these models, though most variants take several hours to train (and some take more than 24 hours on an AWS P3-V100 instance).

Model Architecture

Our baseline models ingest a parallel corpus of (comments, code) and learn to retrieve a code snippet given a natural language query. Specifically, comments are top-level function and method comments (e.g. docstrings in Python), and code is an entire function or method. Throughout this repo, we refer to the terms docstring and query interchangeably.

The query has a single encoder, whereas each programming language has its own encoder. The available encoders are Neural-Bag-Of-Words, RNN, 1D-CNN, Self-Attention (BERT), and a 1D-CNN+Self-Attention Hybrid.

The diagram below illustrates the general architecture of our baseline models:

alt text

Training

This step assumes that you have a suitable Nvidia-GPU with Cuda v9.0 installed. We used AWS P3-V100 instances (a p3.2xlarge is sufficient).

  1. Start the model run environment by running script/console:

    script/console
    

    This will drop you into the shell of a Docker container with all necessary dependencies installed, including the code in this repository, along with data that you downloaded earlier. By default, you will be placed in the src/ folder of this GitHub repository. From here you can execute commands to run the model.

  2. Set up W&B (free for open source projects) per the instructions below if you would like to share your results on the community benchmark. This is optional but highly recommended.

  3. The entry point to this model is src/train.py. You can see various options by executing the following command:

    python train.py --help
    

    To test if everything is working on a small dataset, you can run the following command:

    python train.py --testrun
    
  4. Now you are prepared for a full training run. Example commands to kick off training runs:

Additional notes:

References

Benchmark

We are using a community benchmark for this project to encourage collaboration and improve reproducibility. It is hosted by Weights & Biases (W&B), which is free for open source projects. Our entries in the benchmark link to detailed logs of our training and evaluation metrics, as well as model artifacts, and we encourage other participants to provide as much detail as possible.

We invite the community to submit their runs to this benchmark to facilitate transparency by following these instructions.

How to Contribute

We anticipate that the community will design custom architectures and use frameworks other than Tensorflow. Furthermore, we anticipate that additional datasets will be useful. It is not our intention to integrate these models, approaches, and datasets into this repository as a superset of all available ideas. Rather, we intend to maintain the baseline models and links to the data in this repository as a central place of reference. We are accepting PRs that update the documentation, link to your project(s) with improved benchmarks, fix bugs, or make minor improvements to the code. Here are more specific guidelines for contributing to this repository; note particularly our Code of Conduct. Please open an issue if you are unsure of the best course of action.

Other READMEs

W&B Setup

To initialize W&B:

  1. Navigate to the /src directory in this repository.

  2. If it's your first time using W&B on a machine, you will need to log in:

    $ wandb login
    
  3. You will be asked for your API key, which appears on your W&B profile settings page.

Licenses

The licenses for source code used as data for this project are provided with the data download for each language in _licenses.pkl files.

This code and documentation for this project are released under the MIT License.