Source Code

The NMT model is bases on the RNN Encoder-Decoder architecture. We rely on the seq2seq implementation.

We provide additional scripts and configurations based on seq2seq. The code is organized as follows:

• seq2seq/ : contains scripts for training and inference

  • train_test_small.sh : script that performs training and testing (for small-size methods)

  • train_test_medium.sh : script that performs training and testing (for medium-size methods)

  • inference.sh : script that performs inference and evaluates the performances of the model (it is invoked automatically by the two scripts above)

  • configs/ : contains configurations for the architecture

• dataset/ : contains datasets (training, validation, and test sets)

  • 50/ : small-size methods dataset

  • 50-100/ : medium-size methods dataset

Usage

Before using our code, make sure to have a working installation of seq2seq. Next, download our code and run the script in the folder seq2seq/

./train_test_<small/medium>.sh <dataset_path> <num_iterations> <model_path> <config_ID>

Parameters

• <dataset_path> : path to the dataset containing the folders: train, eval, test (see dataset folder)

• <num_iterations> : number of training iterations

• <model_path> : path where to save the model's checkpoints and predictions

• <config_ID> : numerical ID of the configuration to be used. The IDs available are from 1 to 10 and they described here.

Example

./train_test_small.sh ../dataset/50/ 50000 ../model/50/ 10

Downloads

Download Source Code (10 MB)

• AST differencing Tool - GumTreeDiff

• Java Lexer - ANTLR

• Java Parser - JavaParser