README.md

LiteCoder Experiment Reproducing package

• To run the pre-train objective use the following scripts:

  • Reproduce LiteCoder with all objectives:

    • Navigate the folder Pre-training containing the LiteCoder.py file

    • Then, run Python LiteCoder.py --train-tt --train-cs --train-pd

      • The pretrained model is released on hugging face, therefore it automatically loads.

  • To run the ablation studies:

    • Ablation 1: Python LiteCoder.py --train-tt
    • Ablation 2: Python LiteCoder.py --train-tt --train-cs
    • Ablation 3: Python LiteCoder.py --train-tt --train-cs --train-pd

• To Fine-tuning LiteCoder on downstream tasks:

  • Navigate to the Fine-tuning folder and then Downstream task folder:

    • Code Clone Detection:

      • Follow the instruction of readme.md file.

    • Code Translation:

      • Run setup.sh file.
      • Navigate to the scripts/finetune and run translate.sh file.

• To extract the programming language features (i.e., token type, code sememe, and code dependencies)

  • We used open source datasets to extract language features. we released the extracted datasets on the Hugging Face:

    • LT_Java : LiteCoder/LT_Java
    • LT_Python : LiteCoder/LT_Python
    • LT_Java_Dependency : LiteCoder/LT_Java_Dependency

  • Navigate to the utils directory:

    • Use either the Java or Python notebook file to run over your dataset.
    • Run the cells, for which, you want to extract the features.

• Dependencies:

  • Feature extraction dependencies:

    - pip install ast-comments
    - pip install ast
    - pip install javalang
    - pip install tree-sitter
    

  • Model training dependencies:

    - pip install transformers 
    - pip install datasets
    - pip install pytorch_lightning
    - pip install torch
    

  • Or pip install -r requirements.txt