DeepSoftwareAnalytics / CAST

Replication package for EMNLP 2021 paper: CAST: Enhancing Code Summarization with Hierarchical Splitting and Reconstruction of Abstract Syntax Trees

CASTS: Enhancing Code Summarization with AST Splitting

Environment

Create an environment and activate it:

conda create -n  CAST python=3.6 ipykernel -y
conda activate CAST
conda install pydot -y
pip install git+https://github.com/casics/spiral.git
pip install pandas==1.0.5 gensim==3.5.0 scikit-learn==0.19.1 pycparser==2.18 javalang==0.11.0 tqdm networkx==2.3 nltk==3.4.4 psutil gin-config prettytable
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

Install Java 1.8

Install pytorch according to your environment, see https://pytorch.org/ you can use nvcc --version to check the cuda version.

Model

CAST (Code summarization with AST Splitting and reconsTruction) hierarchically splits a large AST into a set of subtrees and learns the representation of the complete AST by aggregating its subtrees' representation learned using tree-based neural models. The architecture of CAST follows the Seq2Seq framework, including three main components, i.e., an AST encoder module, a code token encoder module, and a summarization decoder module. The AST encoder is responsible for representing the semantic and structural information of splitting AST. The code encoder encodes the lexical information of each method. The decoder integrates the multi-channel representations from the above encoders and incorporates the copy mechanism to generates the final summary.

Quick Start

Data Preparation

Download Processed the dataset by:

$ cd ./Data/
$ bash get_data.sh
$ cd ../

or run python pipeline.sh to generate preprocessed small data from scratch.

Dataset	train	valid	test
TL-CodeSum	66,935	8,359	8,367
Funcom	1,904,093	103,650	103,487

More details about AST splitting implementation can refer to Splitting

Training

You can download the trained model by:

$ bash get_trained_model.sh

The model is saved in ./output/DATASET_NAME/model.pth. DATASET_NAME is TL_CodeSum or Funcom.

You can also train it from scratch. batch size should be configured based on the GPU memory size by setting -bs xx

Train in Single GPU

$ cd ./source_code/
$ python casts.py  -cf_file "DATASET_NAME.json"  -ddp False

for example

$ python casts.py  -cf_file "TL_CodeSum.json"  -ddp False

Train in Multi-GPU

Take 4-GPUs for example:

$ cd ./source_code/
$ python -m torch.distributed.launch --nproc_per_node=4  --master_addr='localhost' --master_port=23456 casts.py  -cf_file "DATASET_NAME.json" 

note: you can find the output in ./output/DATASET_NAME.

Testing

Set -only_test is True for testing. The model is load from ./output/tl_codesum/model.pth and the log file and predict result are saved in ./output/tl_codesum/predict.json

 $ cd ./source_code/
 $ python casts.py  -cf_file "DATASET_NAME.json"  -ddp False -only_test True

for example

$ python casts.py  -cf_file "TL_CodeSum.json"  -ddp False  -only_test True

Evaluation

 $ cd ./source_code/
 $ python evaluation.py -dataset DATASET_NAME

for example

$ python evaluation.py -dataset TL_CodeSum

Human evaluation

Click here to see the result of human evaluation