jayavardhanr / CodeRetrieval-Main

Main code for Code Retrieval

Main Code for Code Retrieval Model

Repository Structures

• /code: All code for model training and evaluation
• /data: Directory for data and model checkpoints

Code Strcuture

• codesearcher.py: The main for Code Retrieval Model:
• config.py: Configurations for models defined in the models.py. Each function defines the hyper-parameters for the corresponding model.
• data.py: dataset loader.
• utils.py: Utilities for models and training.

Usage

The /data folder provides all the datasets

1. Main staQC dataset
2. CodeNN dataset
3. CodeNN sanity check dataset (sanity - pad, unk)
4. staQC dataset to check generalization to Code Annotation (anno)
5. staQC sanity check dataset (sanity check - code empty, QB empty)

To test on specific dataset, make changes in the Config file. For Codenn dataset, comment eval and uncomment eval_codenn in codesearch python file

To train and test our model:

Configuration

1. Edit hyper-parameters and settings in config.py

- Don't run training without changing parameters, it will override saved best models
- The parameters provided load the best model for evaluation.

Model Using QB (Best Model parameters)

+ Both 512 and 256 batchsizes produced similar results. I have placed both the weights in the data directory.

Train

python codesearcher.py --mode train --use_qb 1 --code_enc bilstm --reload -1 --dropout 0.25 --emb_size 200 --lstm_dims 400 --batch_size 256

 python codesearcher.py --mode train --use_qb 1 --code_enc bilstm --reload -1 --dropout 0.25 --emb_size 200 --lstm_dims 400 --batch_size 512

Evaluation

python codesearcher.py --mode eval --use_qb 1 --code_enc bilstm --reload 1 --dropout 0.25 --emb_size 200 --lstm_dims 400 --batch_size 256

python codesearcher.py --mode eval --use_qb 1 --code_enc bilstm --reload 1 --dropout 0.25 --emb_size 200 --lstm_dims 400 --batch_size 512

Model without QB (Best Model parameters)

Train

python codesearcher.py --mode train --use_qb 0 --code_enc bilstm --reload -1 --dropout 0.35 --emb_size 200 --lstm_dims 400 --batch_size 1024

Evaluation

python codesearcher.py --mode eval --use_qb 0 --code_enc bilstm --reload 1 --dropout 0.35 --emb_size 200 --lstm_dims 400 --batch_size 1024