On the Effectiveness of Code Representation in Automated Patch Correctness Assessment

project structure

├── core
│   ├── ods: run ODS on our dataset
│   ├── representation: construct code representation

├── tool: tools used to extract code representation
│   ├── add.jar
│   ├── coming_jar
│   ├── jdiff.jar
│   ├── property_graph

├── bases

├── pymodels
│   ├──classification: models used in APCA tasks
│      ├── ensemble: model for fusion-based APCA
│      ├── feature & tfidf: models for feature-based APCA
│      ├── graph: models for graph-based APCA
│      ├── sequence: models for sequence-based APCA
│      ├── tree: models for tree-based APCA

├── configs: model config & task config

├── evaluation: evaluation metrics used in APCA tasks

├── factory: factory mode

├── tasks: tasks customized for APCA tasks

├── tbcnn: run TBCNN on our dataset

├── tokenizer: tokenizers used in APCA tasks

├── trainer: trainers used in APCA tasks

├── dataset
│   ├── custom.py: dataset customized for APCA tasks

├── data: patch files, intermediate result & dataset for different code representation

├── utils

├── train.py

├── test.py

├── README.MD

prerequisite

We used several tool to extract different types of code representation, please download and install them first.

# install joern

Please refer to https://github.com/joernio/joern to  install joern

# install coming

Please refer to https://github.com/SpoonLabs/coming to install coming

# install add

add.jar is included in ./tool

# install property-graph

Please refer to https://github.com/Zanbrachrissik/PropertyGraph to install property-graph

extract code representation

go to corresponding folder

cd ./core/representation

extract feature representation

python graph_tree_feature.py --ttype="feature"

extract tree representation

python graph_tree_feature.py --ttype="ast"

extract graph representation

python graph_tree_feature.py --ttype="graph"

extract sequence representation

python sequence.py

You can find feature extracting result under ./output

dataset split

We employed a 5-fold cross validation in our experiment, please run dataset_split.py to split the dataset

train

In the yml formatted configuration file, please specify parameters such as the training set, validation set, model, etc. Please specify the configuration file when running the training script.

# feature,sequence,tree,graph representation
python train.py --rtype='single' --config_file='train_config.yml'

# fusion representation
python train.py --rtype='fusion'

test

python test.py --config_file='test_config.yml'

QuanjunZhang / APCARepresentation