graphsense

GraphSense is a powerful library designed to facilitate the training and inference of code suggestion models based on graph embeddings. By leveraging the Node2Vec algorithm, GraphSense minimizes the need for extensive data preprocessing while delivering efficient resource consumption. It seamlessly integrates FAISS for vector indexing and uses RocksDB to manage mappings between code lines and their corresponding indices.

Key Features

• Performance-Oriented: Optimize training and inference processes with a focus on speed and resource efficiency.
• Simple Usage: Train models and generate code suggestions with minimal setup and low overhead.

Requirements

• Python 3.8 or higher is necessary to run GraphSense.

Training Example

Training a model with GraphSense is straightforward. Below is a simple example:

from graphsense import GraphTrain

g = GraphTrain()
# train the model
g.line_completion(directory_path="code_files", language="Python")

Inference Example

To generate code suggestions, the inference process is also easy to implement:

from graphsense import GraphInfer

g = GraphInfer()

g.load_artifacts()  # load the artifacts to memory
suggestions = g.infer("def factorial(n):")
g.unload_artifacts()  # clean memory

print("top 10 suggestions: ", suggestions)

Performance Overview

GraphSense demonstrates impressive performance metrics compared to traditional models like the fine-tuned gpt2-medium:

• Artifact Size: 13.9 MB (GraphSense) vs 1.44 GB (gpt2-medium)
• Average Inference Time (CPU): 0.0079 seconds (GraphSense) vs 8 seconds (gpt2-medium)
• Average Memory Usage: 277.8 MB (GraphSense) vs 1800 MB (gpt2-medium)

GraphSense achieves high accuracy, with a top-10 accuracy of 89.58% when trained on the TheAlgorithms Python dataset.

Scalability Insights

GraphSense displays exceptional scalability, as reflected in the average memory usage and execution time across varying vocabulary sizes:

• For a vocabulary of 1,000,000, average memory usage reaches 734.58 MB with an execution time of just 0.0444 seconds.

GraphSense empowers developers and data scientists to enhance code completion systems effectively, combining efficiency, performance, and scalability.