acleclair / icpc2020_gnn Goto Github PK

Code for the ICPC 2020 paper Improved Source Code Summarization via a Graph Neural Network

License: MIT License

Python 100.00%

icpc2020_gnn's Introduction

ICPC2020 CodeGNN

This code is part of the reproducibility package for the ICPC 2020 paper "Improved Code Summarization via a Graph Neural Network" - arxiv

The reproducibility package has three parts:

the code found in this repository
the trained models, predictions, and tokenizer (.tok) files can be downloaded HERE
the fully processed data (as a pkl file) can be downloaded HERE

The dataset we used is from our NAACL'19 paper "Recommendations for Datasets for Source Code Summarization" where the unprocessed data can be found and downloaded HERE

This code uses Keras v2.3.1 and Tensorflow v1.15.2

Running the code and models

To run the trained models from the paper download the three parts of the reproducibility package and run predict.py. Predict.py takes the path to the model file as a positional argument and will output the prediction file to ./modelout/predictions.

python3 predict.py {path to model} --gpu 0 --modeltype {model type: codegnngru|codegnnbilstm|codegnndense} --data {path to data download}

python3 predict.py ./mymodels/codegnngru.h5 --gpu 0 --modeltype codegnngru --data ./mydata

To train a new model run train.py with the modeltype and gpu options set.

python3 train.py --gpu 0 --modeltype codegnnbilstm --data ./mydata

Processing Files

We have added our processing files to the processing directory. These files will not work without some tweaking due to hard coded paths/file names/databases but show how we processed our code, comments, and AST. To generate the base AST XML we use SrCML which can be downloaded HERE

Cite this work

@inproceedings{
leclair2020codegnn,
title={Improved Code Summarization via a Graph Neural Network},
author={Alex LeClair, Sakib Haque, Lingfei Wu, Collin McMillan},
booktitle={2020 IEEE/ACM International Conference on Program Comprehension},
year={2020},
month={Oct.},
doi={10.1145/3387904.3389268}
ISSN={978-1-4503-7958-8/20/05}
}

icpc2020_gnn's People

Contributors

Stargazers

Watchers

Forkers

avijit1258 xiao1i aakashba 759713112 zhichaoouyang susanalima nashid jeongwhanchoi njuzh zoudajia caidw tamal-mondal airnicco8 ishuangxin guwenjia marcbolliger saxonsa chiayisu

icpc2020_gnn's Issues

I cant fid tdats.tok,coms.tok and smls.tok.

I cant fid tdats.tok,coms.tok and smls.tok,could you please tell me where are they?

Data pre-processing

Hi,

Could you please let me know where nd how the .tok files are prepared from the Java source code?
I couldn't find this functionality under the processing directory

Many file paths do not understand where the data preprocessing steps come from, could you improve it?

I want to use your method as my baseline, the data set is different from yours, but I don’t understand your preprocessing steps because of many unknown paths and files.
Thank you very much!

ValueError: Unknown layer: GCNLayer

Hi thanks for sharing the code, I have been trying to reproduce the results presented in the paper.

To do this I have trained the codegnnbilstm model from scratch using the instructions and code provided in the repository. The model trains fine. However, when I try to run the predict.py file, I run into the following error.

ValueError: Unknown layer: GCNLayer

Implemetation of Bleu and Rouge-l is used?

Hi, thanks a lot for sharing the code and the dataset.

Would it be possible for you to share the implementation of Bleu and Rouge used in the paper? There are multiple implementations of this available from nltk and other Github repos. I understand there would not be much of a difference in the scores (decimal places), it would be really helpful if you could share so that there is consistency with what I comparing to and the paper.