Get access to World of Code servers.

Note: the code is written to be language-agnostic. So far we trained models for Python only. To reproduce the results, replace <lang> in the commands with PY

Note 2: WoC is versioned, with only 2..3 last versions available at any time. In this study, we used version Q, but at the moment of reproduction it might not be available. Due to this fact and some randomness involved in sampling, the intermediate data files will not be exactly the same as the ones shared with this package. We expect, however, the final results to be very similar regardless of the version used.

On da4 (one of WoC servers), run: ./preprocess.py PY -v

As a result, you should get these files ():

• <lang>_project_imports.csv - two columns (WoC project id + comma-separated imports)
• <lang>_dev_imports.csv - source of training data two columns (email + comma-separated imports), low usage imports omitted
• <lang>_dev_commit_counts.csv - will be used to filter out bots two columns (email + count)
• <lang>_namespace_counts_by_project.csv - will not be used two columns (namespace + count)
• <lang>_namespace_counts_by_dev.csv - filter out infrequent namespaces two columns (namespace + count)
• <lang>_vocab.csv one column, low usage ones are already filtered out

All files have no header.

Move these files to directory data (create if necessary)

Make sure folder ns_whitespaces exists. If it contains a set of non-empty csv files <lang>_whitelist.csv, it is safe to skip this step.

Note: This step takes few days on a pretty powerful machine. Consider using precomputed files.

• download latest libraries.io archive (1.4.0 was used for this study)
• extract and put projects_with_repository_fields*.csv file in this directory
• run:

./get_ns_whitelist.py projects_with_repository_fields-1.4.0-2018-12-22.csv \
    --save-path=ns_whitelist

As a result, you'll get a bunch of <lang>_whitelist.csv files in ns_whitelist directory.

Run:

python train_model.py <lang> <embedding_size> --mode=<mode> ,

Embedding size is self explanatory. Mode is one of {dev, proj}, to train on developer or project profiles, respectively.

As a result of this step, you'll get a trained model saved in models directory.

To evaluate a model, run:

python validate.py <model_path>

The output of this file will contain results of analogical reasoning, nearest neighbor-based prediction, and competing projects benchmarks.