Detect semantically similar python code using fine-tuned GraphCodeBERT model.

This modified GraphCodeBERT model was fine-tuned for 11 hours using an A40 server on the PoolC (1fold) dataset, which contains over 6M pairs of semantically similar python code snippets.

It is then used to predict the similarity of python code snippets in other folds of the PoolC dataset, as well as the C4 dataset. It achieved F1 scores of greater than 0.96 on all datasets in several experiments, where balanced sampling was applied.

• pip

  In your virtual environment, run:

  pip install -r requirements.txt

  to install the required packages.

• conda

  To create a new conda environment called PythonCloneDetection with the required packages, run:

  conda env create -f environment.yml

  (this may take a while to finish)

The above commands will install cpu-only version of the pytorch package. Please refer to PyTorch's official website for instructions on how to install other versions of pytorch on your machine.

1. Run python main.py --input <input_path> --output <output_path> to run CloneClassifier on the csv file at <input_path> and save its predictions at <output_path>. For example:

   python main.py --input examples/c4.csv --output results/res.csv

   The input of main.py is a csv file containing two columns named code1 and code2, where each row contains a pair of python code snippets to be compared. The output csv file will have three columns named code1, code2, and predictions, where predictions indicates whether the two code snippets in the corresponding row are semantically similar.

2. Use the command python main.py --help to see other optional arguments including max_token_size, fp16, and per_device_eval_batch_size.

3. You could also import CloneClassifier class from clone_classifier.py and use it in your own code, for example:

   import pandas as pd
   from clone_classifier import CloneClassifier
   
   
   classifier = CloneClassifier(
       max_token_size=512,
       fp16=False,  # set to True for faster inference if available
       per_device_eval_batch_size=8,
   )
   
   df = pd.read_csv("examples/c4.csv").head(10)
   res_df = classifier.predict(
       df[["code1", "code2"]], 
       # save_path="results/res.csv"
   )
   
   print(res_df["predictions"] == df["similar"])

Distributed under the MIT License. See LICENSE for more information.

• GraphCodeBERT
• Original work of the model by @snoop2head
• Dataset source from dacon
• Dataset shared by PoolC