MwT-artifact

This is the abstract for the ICSE 2024 artifact evaluation of paper icse2024early-p197, "Modularizing while Training: A New Paradigm for Modularizing DNN Models". In this abstraction, we will introduce the basic information of the submitted artifact, including the paper title, the purpose of the artifact, the badges to claim, and the technology skills that the reviewers are assumed to evaluate the artifact.

**Paper title.** The corresponding paper title of this artifact is "Modularizing while Training: A New Paradigm for Modularizing DNN Models". The unique paper ID is icse2024early-p197.

**Artifact purpose.** The purpose of the submitted artifact is to verify and validate the results in the paper icse2024early-p197. We have prepared the trained models, resulting modules, and running logs for the experiments in the paper. The reproducible results include Tables 2 ~ 8.

This artifact provides three main functions, including *modular training*, *modularizing*, and *module reuse*. *Modular training* can train an N-class classification CNN model from scratch while optimizing its intra-module cohesion and inter-module coupling, resulting in a modular model. *Modularizing* can decompose the modular model into modules. Each module is responsible for one class and only retains the relevant convolutional kernels. *Module reuse* can reuse only the corresponding modules according to the classification of a target task rather than the entire model, thus reducing the inference cost.

**The badge to claim.** This artifact aims to claim the "Available" and "Reusable" badges.

**Required reviewer's technology skills.** To review this artifact, the reviewers should be familiar with Python, Docker, and the popular DL framework PyTorch.

**Required environments.** We provide a [Docker image](https://zenodo.org/records/10443993) with the necessary libraries. To review this artifact, the reviewers should have a server with the Docker environment and at least one GPU.