CKTyper: Croudsourcing Knowledge Enhanced Type Inference

Overview

The source code of CKTyper is in ~/code.

The experimental data of CKTyper is in ~/Data

CKTyper
|—— Code
|    |—— config
|    |—— Java
|    |    |—— LuceneIndexer
|    |    └—— Snr-LibraryParser
|    |—— Offline_Processing
|    |—— Online_Processing
|    └—— Evaluation_Result
|—— Data
|    |—— Dataset: StatType-SO & Short-SO
|    |—— Evaluation: RQ1 ~ RQ4
|    └—— note.md
|—— readme.md
└—— requrements.txt

Setting up CKTyper

1. Enviroment for running expriment:
   • Java：openjdk 11.0.19
   • Python: Python 3.11
   • Cuda 12.1
   • PyTorch 2.2.1+cu121
2. Run pip install -r requirements.txt.
3. Download punkt_tab from NLTK downloader:

import nltk
nltk.download('punkt_tab')

4. Get your API keys from openai to access ChatGPT.
5. Complete the settings for all files in the config folder, specific instructions can be found in ~/code/config/readme.md

Running CKTyper

1. cd into ~/code
2. Run offline operation: python main.py --mode offline
3. Run online operation: python main.py --mode online --pattern pipeline
4. Evaluate experimental results: python main.py --mode evaluation --operation precision

Other available arguments:

--log_file: save the log into a file
--log_level [LEVEL]: adjust the log level to [LEVEL] (info, debug, warning, error or critical)

BaseLines

MLMTyper

This method was proposed in Prompt-tuned Code Language Model as a Neural Knowledge Base for Type Inference in Statically-Typed Partial Code(ASE22).

The source code is from https://github.com/SE-qinghuang/ASE-22-TypeInference.

Modification

In order to enable the source code to run on GPU in a Linux environment, we have made minimal modifications(in ~/Prompt-tune/ASE_prompt_tuning/PromptInference/Practicality/predDictData_line.py), mainly correcting some errors and transferring the model and data to GPU. Modifications are done as follows:

• Change the jar file name from 'Spoon_FQN_3.jar' to 'Spoon_FQN.jar' (in ground_truth function)
• Change the cmd variable from '"java -jar "+jar_package+" "+lib_path' to '['java','-jar',jar_package,lib_path]' (in ground_truth function)
• Change the cmd variable from ' 'javac ' + file_path' ' to '['javac',file_path]' (in ground_truth function)
• Change the file splitter of json_fileName from '\\' to '/' (in file_read function )
• Change the file splitter of Spoon_FQN.jar from '\\' to '/'
• Change all Chinese characters in the code to English(in predictData_Gen,file_writer,Itera_Pred()init function)
• Move model and token tensors to GPU in (in Itera_Pred()init, model_Predfunction)
• Change the FQN_Truth_path variable from 'FQN_truth' to 'FQN'(in file_read function)
• Change the prompt_infer_path and prompt_tuning_path(global variable)

SnR

This method was proposed in SnR: Constraint-Based Type Inference for Incomplete Java Code Snippets(ICSE22).

The source code is from https://zenodo.org/record/5843327.

iJTyper

This method was proposed in iJTyper: An effective type inference framework for incomplete java codes by integrating constraint- and statistics-based methods

The source code is from https://github.com/zhxchen-se/iJTyper