Replication Package for "The EarlyBIRD Catches the Bug: On Exploiting Early Layers of Encoder Models for More Efficient Code Classification"

This repository contains the replication package for the paper "The EarlyBIRD Catches the Bug: On Exploiting Early Layers of Encoder Models for More Efficient Code Classification" by Anastasiia Grishina, Max Hort and Leon Moonen, accepted for publication in the ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE 2023).

The paper is deposited on arXiv and at the publisher's site (ACM, Open Access), and a copy is included in this repository.

The replication package is archived on Zenodo with DOI: 10.5281/zenodo.7608802. The source code is distributed under the MIT license, the data is distributed under the CC BY 4.0 license.

Organization

The replication package is organized as follows:

• src - the source code

• requirements - txt files with Python packages and versions for replication

• data - all raw datasets used for training (to be found in Zenodo 10.5281/zenodo.7608802)

  • raw
    • devign - Devign
    • reveal - ReVeal
    • break_it_fix_it - BIFI dataset
    • exception - Exception Type dataset

• mlruns - results of experiments, the folder is created once the run.py is executed (see part II), empty folder at the time of distribution

• output - results of experiments

  • tables
    • mlflow_<dataset_name>.csv - we used MLflow to log metrics and parameters in our experiments and generated .csv files with mlflow experiments csv -x <experiment_number> -o mlflow_<dataset_name>.csv command
  • figures - figures reported in paper
  • runs - folder to store model checkpoints, if the corresponding argument is provided when running the code

• model-checkpoints - models with the best F1-weighted score on each of the four dataset - one model for one dataset. Note that the best model is not always the model with the best average improvement over the baseline reported in the paper, because of possible best-performing outliers.

• notebooks - one Jupyter notebook with code to generate figures and tables with aggregated results as reported in the paper

Usage

Python version: 3.7.9 (later versions should also work well); CUDA version: 11.6; Git LFS.

Commands below work well on Mac or Linux and should be adapted if you have a Windows machine.

I. Set up data, environment and code

1. Path to project directory

Update path/to/project to point at EarlyBIRD

export EarlyBIRD=~/path/to/EarlyBIRD

2. Download codebert checkpoint

Please, install Git LFS: https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage

Run the following from within $EarlyBIRD/:

cd $EarlyBIRD
mkdir -p checkpoints/reused/model
cd checkpoints/reused/model
git lfs install
git clone https://huggingface.co/microsoft/codebert-base
cd codebert-base/
git lfs pull
cd ../../..

3. Set up a virtual environment

cd $EarlyBIRD
python -m venv venv
source venv/bin/activate

3.1 No CUDA

python -m pip install -r requirements/requirements_no_cuda.txt

3.2 With CUDA (to run on GPU)

python -m pip install -r requirements/requirements_with_cuda.txt
python -m pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

4 Preprocess data

Raw data can be obtained from the data/ folder in Zenodo via the link 10.5281/zenodo.7608802.

After preprocessing, all datasets are stored in jsonlines (if in python) format. Naming convention: split is one of 'train', 'valid', 'test' in data/preprocessed-final/<dataset_name>/<split>.jsonl, with

{'src': "def function_1() ...", 'label': "Label1"}
{'src': "def function_2() ...", 'label': "Label2"}
...

4.1 Devign

Raw data is downloaded from https://drive.google.com/file/d/1x6hoF7G-tSYxg8AFybggypLZgMGDNHfF/view. Test, train, valid txt files are downloaded from https://github.com/microsoft/CodeXGLUE/tree/main/Code-Code/Defect-detection/dataset. All files are saved in data/raw/devign.

To preprocess raw data and save tokenization statistics with the specified tokenizer:

cd $EarlyBIRD
python -m src.preprocess \
    --dataset_name devign \
    --shrink_code \
    --config_path src/config.yaml \
    --tokenizer_path "checkpoints/reused/model/codebert-base"

4.2 ReVeal

Raw data is downloaded from https://github.com/VulDetProject/ReVeal under "Our Collected vulnerabilities from Chrome and Debian issue trackers (Often referred as Chrome+Debian or Verum dataset in this project)" and saved in data/raw/reveal.

To preprocess raw data:

cd $EarlyBIRD
python -m src.preprocess \
    --dataset_name reveal \
    --shrink_code \
    --config_path src/config.yaml \
    --tokenizer_path "checkpoints/reused/model/codebert-base"

4.3 Break-it-fix-it

Raw data is downloaded as data_minimal.zip from https://github.com/michiyasunaga/BIFI under p. 1, unzipped, and the folder orig_bad_code is saved in data/raw/break_it_fix_it.

To preprocess raw data:

cd $EarlyBIRD
python -m src.preprocess \
    --dataset_name break_it_fix_it \
    --shrink_code \
    --ratio_train 0.9 \
    --config_path src/config.yaml \
    --tokenizer_path "checkpoints/reused/model/codebert-base"

Note: The original paper contains only train and test split. Use --ratio_train to specify what part of the original train (orig-train) split will be used in train and the rest of orig-train will be used for validation during training.

4.4 Exception Type

Raw data is downloaded from https://github.com/google-research/google-research/tree/master/cubert under "2. Exception classification" (it points to this storage) and saved in data/raw/exception_type.

To preprocess raw data:

cd $EarlyBIRD
python -m src.preprocess \
    --dataset_name exception \
    --shrink_code \
    --config_path src/config.yaml \
    --tokenizer_path "checkpoints/reused/model/codebert-base"

II. Run code

Activate virtual environment (if not done so yet):

cd $EarlyBIRD
source venv/bin/activate

Example run

Run experiments with Devign using pruned models (cutoff_layers_one_layer_cls) to 3 layers (--hidden_layer_to_use 3), for example:

cd $EarlyBIRD
python -m src.run --help              # for help with command line args

python -m src.run \
    --config_path src/config.yaml \
    --model_name codebert \
    --model_path "checkpoints/reused/model/codebert-base" \
    --tokenizer_path "checkpoints/reused/model/codebert-base" \
    --dataset_name devign \
    --benchmark_name acc \
    --train \
    --test \
    -warmup 0 \
    --device cuda \
    --epochs 10 \
    -clf one_linear_layer \
    --combination_type cutoff_layers_one_layer_cls \
    --hidden_layer_to_use 3 \
    --experiment_no 12 \
    --seed 42

To run experiments on a small subset of data, use --debug argument and reduce the number of epochs. For example:

python -m src.run \
    --debug \
    --config_path src/config.yaml \
    --model_name codebert \
    --model_path "checkpoints/reused/model/codebert-base" \
    --tokenizer_path "checkpoints/reused/model/codebert-base" \
    --dataset_name devign \
    --benchmark_name acc \
    --train \
    --test \
    -warmup 0 \
    --device cuda \
    --epochs 2 \
    -clf one_linear_layer \
    --combination_type cutoff_layers_one_layer_cls \
    --hidden_layer_to_use 3 \
    --experiment_no 12 \
    --seed 42

Explore output

Your EarlyBIRD/ should contain mlruns/. If you started the run.py from another location, you will find mlruns/ one level below that location.

cd $EarlyBIRD
mlflow ui

Citation

If you build on this data or code, please cite this work by referring to the paper:

@inproceedings{grishina2023:earlybird,
   title = {The EarlyBIRD Catches the Bug: On Exploiting Early Layers of Encoder Models for More Efficient Code Classification},
   author = {Anastasiia Grishina and Max Hort and Leon Moonen},
   booktitle = {ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE)},
   year = {2023},
   publisher = {ACM},
   doi = {https://doi.org/10.1145/3611643.3616304},
   note = {Pre-print on arXiv at https://arxiv.org/abs/2305.04940}
}

Acknowledgement

The work included in this repository was supported by the Research Council of Norway through the secureIT project (#288787). The empirical evaluation benefitted from the Experimental Infrastructure for Exploration of Exascale Computing (eX3), which is financially supported by the Research Council of Norway through project #270053.