Siamese Network for Code Clone Detection

This project implements a Siamese Network with a Bidirectional LSTM backbone for detecting semantic similarity in source code. The network predicts whether two code snippets are functionally identical, enabling use cases like code clone detection, plagiarism detection, and software refactoring.

---

Features

• Dataset: Utilizes the BigCloneBench dataset from Hugging Face's datasets library.
• Model:
  • Embedding layer for converting tokenized code into dense vectors.
  • Bidirectional LSTM to capture sequential dependencies from both directions.
  • Fully connected layers for feature extraction.
  • Cosine similarity to compute similarity between code snippet embeddings.
• Loss Function: Implements contrastive loss, ensuring that embeddings for similar code snippets are close while dissimilar embeddings are far apart.
• Evaluation:
  • Metrics include accuracy, F1-score, precision, and recall.
  • Provides detailed results on the test dataset for robust analysis.

---

Installation

Prerequisites

Ensure the following tools and libraries are installed:

• Python 3.8 or higher
• PyTorch 1.9 or newer
• Additional Python dependencies listed in requirements.txt

Installation Steps

1. Clone the repository:

   git clone https://github.com/shivbera18/CloneCodeDetection.git
   cd CloneCodeDetection

Install dependencies:

pip install -r requirements.txt Verify installation by running:

python --version python -c "import torch; print(torch.version)" Dataset The BigCloneBench dataset is used for training and testing. It is a large benchmark dataset for detecting functionally similar code snippets.

Dataset Details: Source: BigCloneBench on Hugging Face Classes: Binary labels (1 for similar code pairs, 0 for dissimilar). Preprocessing: Tokenized code snippets using Keras' Tokenizer. Padded sequences to ensure uniform input length.

Model Architecture:

The Siamese network consists of two identical sub-networks for feature extraction. Key components:

1.Embedding Layer: Maps tokenized code into dense vector representations. Embedding size: 128.

2.Bidirectional LSTM: Captures contextual relationships in code from both directions. Hidden size: 128. Two layers with dropout for regularization.

3.Fully Connected Layers: Reduces feature dimensionality. Includes Batch Normalization and Dropout.

4.Cosine Similarity: Measures the similarity between embeddings of code snippets.

5.Contrastive Loss: Encourages similar pairs to have high cosine similarity and dissimilar pairs to have low similarity.

Training To train the model, run:

python train.py

Testing Evaluate the model's performance on the test dataset:

python evaluate.py

Adjust hyperparameters (e.g., learning rate, batch size, number of epochs) in the respective scripts (train.py, evaluate.py).