The following repository contains the code needed to use MastroPaolo's T5 model outside the Google Colab environment.
The model was developed by Antonio Mastropaolo, Simone Scalabrino, Nathan Cooper, David Nader Palacio, Denys Poshyvanyk, Rocco Oliveto, Gabriele Bavota and the official repository is available at the following link :
Due to limitations imposed by GitHub, it was not possible to upload the PyTorch model used. It is therefore necessary to download it via the following Google Drive link:
Once downloaded, you need to move it within the project path :
Trainted-Models/with-pretraining-new/MT-Task-Balanced/Pytorch-Model
The following code was developed and tested using Python 3.10.9
In addition, it is recommended to create a Virtual Enviroment for downloading the necessary libraries:
- nlp in version 0.4.0
- sentencepiece in version 0.1.96
- dill in version 0.3.3
- numpy in version 1.26.4
- torch in version 2.4.0
- transformers in version 4.44.2
Once you have installed all the libraries, you can import your files in .tsv format for them to be parsed by the program.
To ensure that the code works properly, it is necessary to specify the path to the .tsv file to be parsed
within the file main.py (located in the root directory of the repository) as well as in the
assertion_dataset_script.py located in the Miscellaneous folder of the repository.
The following code has been tested in both Windows and Linux (Ubuntu) environments.
In both cases, the model finished execution and saved the results in the Result directory.
This code was used to evaluate the goodness of the T5 model in predicting parts of missing code. The model used is able to replace part of the unit test code that is hidden.
The results obtained are available here.