This repository contains source code necessary to reproduce the model in the paper:
"PSCS: A Path-based Neural Model for SemanticCode Search"
- Python 3.6
- Pytorch 1.0.1
- NumPy
We use the CodeSearchNet as our dataset.
Download the Java dataset following the instruction of CodeSearchNet, and move it to the directory ./data/.
The data exists in ./data is the example data.
The whole process of our model, including data processing and model training, is in file pipeline.sh.
Before running, edit the file pipeline.sh following the instructions below.
Run the pipeline.sh file:
sudo bash pipeline.sh
Note: the vocabularies for natural language of train set and test set are not shared. However, the vocabularies for path need to be shared to fit the production environment. For a code retrieval system, the code snippets to be searched are already known. So we don't need to worry about the information leakage problem.
To evaluate the performance of a trained model:
Run the test.py file:
python test.py --test_epoch 100