Code for the paper "DIRECT : A Transformer-based Model for Decompiled Variable Name Recovery"

This code is adapted from DIRE.

Setting up the data and packages

1. Run pip install -r requirements.txt to install the required packages.

2. Download the preprocessed data along with training-test splits from this link, and put them in data/preprocessed_data.

3. Create a symbolic link in the src folder by running ln -s data ./src/data.

Pretraining

To pretrain the BERT encoder and decoder from scratch, run

python bert_pretrain.py [-decoder]

Training

To train the DIRECT model from scratch, first pretrain the BERT encoder and decoder. Then run

python main.py -train

Prediction

To evaluate a trained DIRECT model, assuming it is saved at src/saved_checkpoints/direct.pth, run

python main.py -restore -name direct [-val] [-top_k 1] [-approx] [-conf_piece] [-short_only]

Running the above evaluation dumps the predictions to src/predictions/<fname>.pkl. To evaluate these predictions with other metrics like Top-5 accuracy, Jaccard distance and Character Error Rate, run

python top5_analysis.py -fname <fname>.pkl

Results

Model	Accuracy(%)	Top-5 Accuracy (%)	CER	Jaccard Dist
DIRE	35.8	41.5	.664	.537
DIRECT	42.8	49.3	.663	.501
Improvement	20%	19%	.2%	6.5%