Skip to content

Latest commit

 

History

History
189 lines (144 loc) · 5.6 KB

README.md

File metadata and controls

189 lines (144 loc) · 5.6 KB
Raw

PRSummarizer

The source code of the paper "Automatic Generation of Pull Request Description".

Dataset

Raw Data

Our collected 333K pull requests can be downloaded from here. Here is a PR example in the json file:

{
    "id": "elastic/elasticsearch_37980",
    "body": "'Eclipse build files were missing so .eclipse project files were not being generated.\\r\\nCloses #37973\\r\\n\\r\\n'",
    "cms": [
      "'Added missing eclipse-build.gradle files\\n\\nCloses #fix/37973'"
    ],
    "commits": {
      "'3e10ee798c932cc1cab1ea6ca679417408fc1416'": {
        "cm": "'Added missing eclipse-build.gradle files\\n\\nCloses #fix/37973'",
        "comments": []
      }
    }
  }
  • id: $user/$project_$prid
  • body: PR description
  • cms: the commit messages in this PR
  • commis: the commits in this PR
    • key is the SHA1 hash digest
      • cm: commit message
      • comments: source code comments added in this commit

Preprocessed Dataset

Our dataset can be downloaded from here, which contains:

  • the train, validation and test sets
  • a json file for building vocabulary

Regular Expressions

To preprocess the raw data, we used the following regular expressions:

email_pattern = r'(^|\s)<[\w.-]+@(?=[a-z\d][^.]*\.)[a-z\d.-]*[^.]>'
url_pattern = r'https?://[-a-zA-Z0-9@:%._+~#?=/]+(?=($|[^-a-zA-Z0-9@:%._+~#?=/]))'
reference_pattern = r'#[\d]+'
signature_pattern = r'^(signed-off-by|co-authored-by|also-by):'
at_pattern = r'@\S+'
structure_pattern = r'^#+'
version_pattern = r'(^|\s|-)[\d]+(\.[\d]+){1,}'
sha_pattern = r'(^|\s)[\dA-Fa-f-]{7,}(?=(\s|$))'
digit_pattern = r'(^|\s|-)[\d]+(?=(\s|$))'

Installation

Clone and Prepare Dataset

$ git clone https://github.com/Tbabm/PRSummarizer.git
$ cd PRSummarizer
$ mkdir data
# download our preprocessed dataset and place the four files in `data`
$ mkdir models

Install ROUGE

  • See here for instructions about installing ROUGE
  • Please make sure you have correctly set environment variable ROUGE to /absolute/path/to/ROUGE-RELEASE-1.5.5

Install Dependencies

Through conda:

$ conda env create -f environment.yml

OR through pip

$ pip install -r requirements.txt

Install pyrouge

  • install and test pyrouge if you haven't done it.
$ git clone https://github.com/bheinzerling/pyrouge
$ cd pyrouge
$ pip install .

# set rouge path for pyrouge
$ pyrouge_set_rouge_path ${ROUGE}

# test the installation of pyrouge
$ python -m pyrouge.test

Usage

Train

Train Attn+PG first:

python3 -m prsum.prsum train --param-path params_attn_pg.json

After training, suppose the models are stored in models/train_12345678/model/. Select the best Attn+PG model:

python3 -m prsum.prsum select_model \
                       --param_path params_attn_pg.json \
                       --model_pattern "models/train_12345678/model/model_{}_" \
                       --start_iter 1000 \
                       --end_iter 26000

Suppose the best model is model_12000_87654321. Train Attn+PG+RL based on the best model:

python3 -m prsum.prsum train \
                       --param_path params_attn_pg_rl.json \
                       --model_path "models/train_12345678/model/model_12000_87654321"

Validate

Select the best Attn+PG model:

# start_iter = the best iteration of `Attn+PG` (here, 12000) + save_interval (here, 1000)
START_ITER=13000
python3 -m prsum.prsum select_model
                       --param_path params_attn_pg_rl.json \
                       --model_pattern "models/train_12345678/model/rl_model_{}_" \
                       --start_iter $START_ITER \
                       --end_iter 41000

Suppose the best model is model_34000_98765432.

Test

Test the best Attn+PG+RL model:

python3 -m prsum.prsum decode \
                       --param_path params_attn_pg_rl.json \
                       --model_path "models/train_12345678/model/rl_model_34000_98765432" \
                       --ngram_filter 1

Now, you will get the test results.

NOTE: Your test results may be slightly different from those reported in our paper. Because the pointer generator uses the scatter_add function in pytorch. When using GPUs, this function is undeterministic. See here for more details.

Pre-trained Model

Our pre-trained model and test results can be downloaded here. To test with our pre-atrained model:

mkdir models
mv rl_model_34000 ./models
python3 -m prsum.prsum decode \
                       --param_path params_attn_pg_rl.json \
                       --model_path "models/rl_model_34000" \
                       --ngram_filter 1

Citation

If you use this code, please consider citing our paper:

@inproceedings{liu2019automatic,
  title={Automatic generation of pull request descriptions},
  author={Liu, Zhongxin and Xia, Xin and Treude, Christoph and Lo, David and Li, Shanping},
  booktitle={Proceedings of the 34th IEEE/ACM International Conference on Automated Software Engineering},
  pages={176--188},
  year={2019},
}

Thanks!

Reference