tc-codesum

This repository provides the code necessary to reproduce the results found in the paper "Análise Experimental de Abordagens para Sumarização de Código Fonte" (Experimental Analysis of Code Summarization Approaches).

Table of Contents

• Data
• CodeBERT
  • Preparation
  • Training & Eval
  • Testing
• NeuralCodeSum
  • Preparation
  • Training, Testing & Beam Search
    • NeuralCodeSum
    • RNN

Data

For training, we used a subset of 500K code and summarization pairs from the CoDesc dataset for code summarization, which can be downloaded here.

# Requirements: zip and gdown (Google Drive files)
pip install zip gdown

# Download the data
gdown https://drive.google.com/uc?id=1pH51Ttn440keobIvDbk2KAZ7t3he1Wyw

# Unpack & remove zip
unzip codesummarization.zip -d data/
rm codesummarization.zip

CodeBERT

Preparation

# Requirements
pip install torch==1.4.0
pip install transformers==2.5.0
pip install filelock

# Clone CodeBERT repo
git clone --depth 1 https://github.com/guoday/CodeBERT.git

Training & Eval

cd CodeBERT/CodeBERT/code2nl

# Parameters
lang="java" 
lr=5e-5
batch_size=16
beam_size=2
source_length=256
target_length=128
data_dir=$(pwd)
output_dir=$data_dir/CodeBERT/CodeBERT/code2nl/model
train_file=$data_dir/data/codesummarization_data/train_small/train_small.jsonl
dev_file=$data_dir/data/codesummarization_data/dev/dev.jsonl
eval_steps=10 
train_steps=10 
pretrained_model=microsoft/codebert-base 

# Run Model
python run.py --do_train --do_eval \
--model_type roberta \
--model_name_or_path $pretrained_model \
--train_filename $train_file \
--dev_filename $dev_file \
--output_dir $output_dir \
--max_source_length $source_length \
--max_target_length $target_length \
--beam_size $beam_size \
--train_batch_size $batch_size \
--eval_batch_size $batch_size \
--learning_rate $lr \
--train_steps $train_steps \
--eval_steps $eval_steps

Testing

cd CodeBERT/CodeBERT/code2nl

# Parameters
lang="java" 
beam_size=5
batch_size=32
source_length=256
target_length=128
data_dir=$(pwd)
output_dir=$data_dir/CodeBERT/CodeBERT/code2nl/model
dev_file=$data_dir/data/codesummarization_data/dev/dev.jsonl
test_file=$data_dir/data/codesummarization_data/test/test.jsonl
test_model=$data_dir/CodeBERT/CodeBERT/code2nl/model/pytorch_model.bin

# Run Model
python run.py --do_test \
--model_type roberta \
--model_name_or_path microsoft/codebert-base \
--load_model_path $test_model \
--dev_filename $dev_file \
--test_filename $test_file \
--output_dir $output_dir \
--max_source_length $source_length \
--max_target_length $target_length \
--beam_size $beam_size \
--eval_batch_size $batch_size

NeuralCodeSum

Preparation

# Clone the repository
git clone --depth 1 https://github.com/wasiahmad/NeuralCodeSum.git

# Install requirements
cd NeuralCodeSum
pip install -r requirements.txt
python setup.py develop

# Copy data files
bash copy_data.sh

# Copy files that we modified to run the models 
bash copy_modified.sh

Training, Testing & Beam Search

NeuralCodeSum

cd NeuralCodeSum/scripts/java

# Run Model
bash transformer.sh 0 code2jdoc

RNN

cd NeuralCodeSum/scripts/java

# Run Model
bash rnn.sh 0 rnn