Implementation of proposed source code summarization method for CS 585 class project using an encoder-decoder method with attention for translating source code into descriptive comments. The base neural architecture for this model is adapted from the dynet example. This architecture is tested both with and without pretrained word embeddings that were created using the word2vec algorithm.
- All utilities, including the pickle files that contain the parallel corpus can be found in the
utils/directory - All data including the train/dev/test split dataset, model weights, and a sample of generated output from dev and test can be found in the
data/directory. - The sample output is named {embed | non_embed}_{dev | test}.nl and consists of one line per source data instance. These lines correspond to the lines in {dev | test}.nl that are the original source comment.
- All information that was used to create the word embeddings using word2vec is available in the
vector-maker/directory. The top-level script used for this process isw2v.sh
Run the following two commands using docker installed on a machine running Ubuntu 16.04.01
docker build -t hein-cs585-hw4 .docker run hein-cs585-hw4
This will display the BLEU score output for the Dev and Test portions of both the trained model with pretrained embeddings and the trained model without pretrained embeddings.
- Training was done using the
main.pyscript. Portions of this code are commented out depending on whether we are training the model with or without pretrained embeddings. - The validation set output is also generated during the training process in order to assist in debugging the model.
- Training was done using the UA HPC Ocelote computer. Training sessions utilized one Nvidia Tesla P100 per session to train for a total of 40 epochs over approximately 6 hours
- To train on Ocelote the script
run-comm-gen.pbswas used