Bug Localization with DualTextCNN

This project implements a deep learning-based bug localization system using a Dual TextCNN model. It takes natural language bug reports and ranks source files based on their likelihood of being related to the bug.

The system is designed to support tasks such as:

• Training a dual-path CNN classifier
• Evaluating ranked file predictions using IR metrics (Acc@1, Acc@5, MAP, MRR)

It is built using PyTorch and supports fully configurable training and evaluation via a command-line interface.

TEXTCNN_BUGLOCALIZATION/
│
├── artifacts/
│   ├── glove6B/                            # Pretrained GloVe embeddings
│   ├── test_bugs/                          # Bug report pickle files for testing
│   │   └── sphinx-doc+sphinx_bugs.pkl
│   └── trained_model/                      # Saved model checkpoints
│       └── best_model2.pt
│
├── data/
│   ├── data_builder.py
│   ├── glove_loader.py
│   ├── paired_text_dataset.py
│   └── datasets/
│       └── past_brid2commit/
│           ├── sphinx-doc+sphinx_bugs_brid2commit.pkl
│           └── ...                         # other repos' mappings
│
├── evaluation/
│   └── ranker.py                           # Ranking logic and IR metric 
│
├── models/
│   └── dual_text_cnn.py                    # DualTextCNN model
│
├── training/
│   └── trainer.py                          # Trainer class for training and validation
│
├── main.py                                 # CLI entry point
└── README.md

pip install -r requirements.txt

Place the following in the artifact/ folder:

• glove6B/ — Pretrained GloVe embeddings (link: https://nlp.stanford.edu/projects/glove/)
• past_brid2commit/ — Pickled bug-to-commit mappings
• *_indexing/ — Indexed source file folders (collected with: https://github.com/ThreeCirclesK/SWE-Bench-History_BugLoc)
• test_bugs/ — Test bug pickle files

Place the following in the dataset/ folder:

• SWEBench_Custom/ — Indexed files for bug localization. Unzip to use them

python main.py \
  --repodir /path/to/SWEBench_Custom \
  --reponame sphinx-doc+sphinx \
  --model_name artifact/trained_model/best_model.pt

python main.py \
  --evaluate_only \
  --repodir /path/to/SWEBench_Custom \
  --reponame sphinx-doc+sphinx \
  --model_name artifact/trained_model/best_model.pt

The evaluation script computes:

• Accuracy@1
• Accuracy@5
• Mean Reciprocal Rank (MRR)
• Mean Average Precision (MAP)

These metrics reflect the model’s ability to rank fixed files higher among candidates.

The DualTextCNN processes bug reports and source files in parallel using separate convolutional pathways. It applies max-pooling and concatenates both outputs before classification.

This project is distributed under the MIT License.