README.md

Variational Prefix Tuning (VPT) for Code Summarization

Installation

After cloning/downloading the repository, run the following command to create the conda environment for running.

conda env create -f environment.yml
conda activate CSTCVAE

Environment tested

The code is tested in Ubuntu 24.04 environment, please make sure you have JAVA and curl installed in your environment.

Datasets

Before you train or test a model, follow instructions in data/README.md to download and preprocess the datasets.

Finetune CodeT5+ for your dataset.

Please use python notebook finetune_codeT5P.ipynb to finetune the CodeT5+ model for your dataset.

Apply Variational Prefix Tuning (VPT) for Code Summarization.

Change backbone model and dataset directory to your finetuned model in CodeT5+_VPT_bi-criteria.ipynb.
This file will train the VPT for your fine-tuned model and run inference and subset selection on the test set you spcify.

Generated Results

Results generated for the CodeT5+ using VPT method are stored in the T5_results directory, under file with name of CodeT5+_VPT_bicriteria_subset_selection_results_lang.txt.

Results generation of all other decoding methods

Please use the codeT5P_diverse_decoding.ipynb to generate results for all other decoding methods.
Results will be stored in diverse_decoding_results directory.

Experiment Results and Weights

Our model weights and inference results can be accessed from https://drive.google.com/file/d/1nyZds1A_cyS5bWaBy4rd7DKr9Ej2vZl2/view?usp=sharing For the weights included in CodeT5_VPT_Weight, please first load the backbone model CodeT5+ from huggingface, then load our provided weights. The textual predictions for conducting all the experiments in out paper in RQ_Textual_Predictions. The files are separated based on Research Questions (RQ) and datasets, with name indicating the infernce strategy used. For all Sampling and VPT results, please use Bi-criteria subset selection from the 100 results provided to obtain the number of summaries you need.