SANCS

Code search model based the self-attention

Dependency

Successfully tested in Ubuntu 18.04

• Python == 3.7
• PyTorch == 1.6.0
• tqdm == 4.48.2
• numpy == 1.16.3
• tables == 3.6.1
• argparse

Code Structure

• attention: Self-attention network and code-description network.
• method: Code/desc representation and similarity measure mudule.
• train.py: Train and validate code/desc representation models.
• dataset.py: Dataset loader.
• configs: Basic configuration for the attention and method folder. Each function defines the hyper-parameters for the corresponding model.
• utils.py: Utilities for models and training.

Usage

Data

In our experiments, we use the dataset shared by @guxd. You can download this shared dataset from Google Drive and add this dataset folder to /data.

Configuration

Edit hyper-parameters and settings in config.py

Train

python train --mode train

Eval

python train --mode eval

References

Here are some things I looked at while writing this model.

• https://github.com/guxd/deep-code-search
• https://github.com/jadore801120/attention-is-all-you-need-pytorch