Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ART: Actually Robust Training

Sebastian Chwilczyński, Kacper Trębacz, Karol Cyganik, Mateusz Małecki, Dariusz Brzezinski

TL;DR

ART addresses the lack of a unified, practical framework for robust deep learning development by introducing a step-based training library that enforces best practices via validation checks and iterative design. The approach centers on ArtProjects composed of modular ArtModules, Steps, and Checks, with a MetricCalculator to track losses and metrics across steps; additional debugging and visualization utilities include decorators, loggers, and a dashboard, plus integrations with tools like Neptune and wandb. The paper demonstrates how Art translates Karpathy's Recipe for Training Neural Networks into a programmable workflow that supports reproducibility, debugging, and incremental complexity. While current implementation focuses on PyTorch, Art can be extended to other toolkits and aims to foster a community of step templates and standards, potentially aligning with regulatory frameworks such as the EU AI Act.

Abstract

Current interest in deep learning captures the attention of many programmers and researchers. Unfortunately, the lack of a unified schema for developing deep learning models results in methodological inconsistencies, unclear documentation, and problems with reproducibility. Some guidelines have been proposed, yet currently, they lack practical implementations. Furthermore, neural network training often takes on the form of trial and error, lacking a structured and thoughtful process. To alleviate these issues, in this paper, we introduce Art, a Python library designed to help automatically impose rules and standards while developing deep learning pipelines. Art divides model development into a series of smaller steps of increasing complexity, each concluded with a validation check improving the interpretability and robustness of the process. The current version of Art comes equipped with nine predefined steps inspired by Andrej Karpathy's Recipe for Training Neural Networks, a visualization dashboard, and integration with loggers such as Neptune. The code related to this paper is available at: https://github.com/SebChw/Actually-Robust-Training.

ART: Actually Robust Training

TL;DR

ART addresses the lack of a unified, practical framework for robust deep learning development by introducing a step-based training library that enforces best practices via validation checks and iterative design. The approach centers on ArtProjects composed of modular ArtModules, Steps, and Checks, with a MetricCalculator to track losses and metrics across steps; additional debugging and visualization utilities include decorators, loggers, and a dashboard, plus integrations with tools like Neptune and wandb. The paper demonstrates how Art translates Karpathy's Recipe for Training Neural Networks into a programmable workflow that supports reproducibility, debugging, and incremental complexity. While current implementation focuses on PyTorch, Art can be extended to other toolkits and aims to foster a community of step templates and standards, potentially aligning with regulatory frameworks such as the EU AI Act.

Abstract

Current interest in deep learning captures the attention of many programmers and researchers. Unfortunately, the lack of a unified schema for developing deep learning models results in methodological inconsistencies, unclear documentation, and problems with reproducibility. Some guidelines have been proposed, yet currently, they lack practical implementations. Furthermore, neural network training often takes on the form of trial and error, lacking a structured and thoughtful process. To alleviate these issues, in this paper, we introduce Art, a Python library designed to help automatically impose rules and standards while developing deep learning pipelines. Art divides model development into a series of smaller steps of increasing complexity, each concluded with a validation check improving the interpretability and robustness of the process. The current version of Art comes equipped with nine predefined steps inspired by Andrej Karpathy's Recipe for Training Neural Networks, a visualization dashboard, and integration with loggers such as Neptune. The code related to this paper is available at: https://github.com/SebChw/Actually-Robust-Training.
Paper Structure (4 sections, 2 figures)

This paper contains 4 sections, 2 figures.

Table of Contents

  1. Introduction
  2. The Art package
  3. Target users and applications
  4. Conclusions

Figures (2)

  • Figure 1: Schematic representation of Art's components. Deep learning projects are stored as ArtProjects, which consist of steps. Each step validates a version of the deep learning model (ArtModule) using a check function. If the check is successful, the project moves on to the next step, following the build from simple to complex principle. Art also offers utilities for measuring, logging, and visualizing the whole process.
  • Figure 2: Art's Dashboard. (A) List of steps and their statuses. (B) Results of different model versions. (C) Plot for comparing model versions.