This repository contains the code for the dual encoder model described in "Code Search as Multi-Link Translation Using Dual Encoders".

The required dependencies must be installed to run the source code.

pip install -r requirements.txt

The following commands must be run to build the directory structure for the experiments.

mkdir -p Data/{cbow,Texts,Trained_models}
mkdir Data/cbow/CodeSearch300
mkdir Models
mkdir Plots
mkdir Results/{CodeSearch,Tensors}

Trained FastText word embeddings for the datasets' can be found here. Download the files and place them inside the Data/cbow/CodeSearch300 directory. These embeddings have a dimension size of 300, trained on a FastText CBOW model.

It is possible to train a different FastText model or with a different dimension size. The raw text data for that purpose can be found here. The steps for training a FastText model in this data are as follows:-

1. Download the raw text data and place the files in the Data/Texts directory.
2. If none of the dependencies are installed, install the FastText and Pandas packages.

pip install fasttext
pip install pandas

3. Edit the embedding and dimensions variables' values in ProcessData.py to desired values. The default values are -

embedding = "cbow"
dimensions = 300

When the dimension size is not specified in fasttext.train_unsupervised(), the embeddings have a default size of 100.

4. After assigning desired values to these parameters, run the ProcessData.py file.

python ProcessData.py

To run experiments with separate language models, run the ProcessDataSeparate.py file.

python ProcessDataSeparate.py

Once the dependencies are installed, run the corresponding files to run the code search experiment on that dataset with the default parameters. The parameters values can be changed to conduct different experiments.

For CodeSearchNet Python (Limited), run -

python CodeSearch.py

For CodeSearchNet Python (Full), run -

python CodeSearchFull.py

For AdvTest Python, run -

python AdvTest.py

For DGMS, run -

python DGMS.py