DOI: 10.17587/prin.17.510-519

Features of Frameworks and Libraries for Working with Neural Networks using the Example of the Ultralytics YOLO Ecosystem

V. A. Pantyushin, Cand. Sc. (Eng.), Associate Professor, v.pantyushin@customs-academy.ru, A. S. Starostin, Cand. Sc. (Eng.), Head of the Department, as.starostin@customs-academy.ru, V. A. Katyshev, Master's Student, va.katyshev25@customs-academy.ru, Russian Customs Academy, Lyubertsy, Moscow region, 140015, Russian Federation

Corresponding author: Valery A. Pantyushin, Cand. Sc. (Eng.), Associate Professor, Russian Customs Academy, Lyubertsy, Moscow region, 140015, Russian Federation, E-mail: v.pantyushin@customs-academy.ru

The relevance of this work stems from the fact that modern artificial intelligence technologies, particularly deep learning, are rapidly penetrating all areas of human activity — from medicine and autonomous vehicles to security systems, data analysis, and copyright protection. This technological breakthrough is underpinned by convolutional neural networks, capable of solving computer vision problems such as object detection, classification, and segmentation with high accuracy. However, the practical implementation and training of such networks entails a number of methodological and algorithmic challenges that require specialists not only to have basic theoretical knowledge but also the ability to effectively use modern software tools and cloud services. Therefore, comparative analysis and mastery of popular frameworks and libraries that abstract low-level implementation details, allowing for greater focus on defining and solving applied problems, are particularly relevant. The theoretical significance of this article lies in stimulating research into improving the efficiency and performance of artificial intelligence algorithms using a detailed representation of the architecture and topology of a convolutional neural network and an analysis of the results of a deep machine learning implementation. The practical significance of this article lies in the implementation of an algorithm for deploying and training the YOLO model in Ultralytics using the Google Colab cloud service and the MS COCO image library. Using the Ultralytics YOLO ecosystem as an example, this article examines the network topology and describes the characteristics of the neural layers used. It also explains the specifics of mounting and configuring local machine devices to a Google Colab virtual machine for training the YOLO convolutional neural network. The analysis confirms that modern frameworks, such as the Ultralytics YOLO ecosystem, provide developers with powerful and convenient tools for solving a wide range of computer vision problems. They encapsulate complex stages of neural network development — from topology design to training and validation — by providing pre-trained models, a consistent API, and extensive helper functions. The use of cloud environments like Google Colab solves the problem of limited local computing resources, making deep learning technologies more efficient and accessible. Using the YOLOv8n model training example, the full workflow was demonstrated: from mounting cloud storage and installing dependencies to actually running the training process with specified hyperparameters. A detailed analysis of key metrics such as loss functions, precision, recall, and average accuracy demonstrated their critical importance for monitoring progress, diagnosing issues (such as overfitting), and objectively assessing the model's final performance. Thus, the successful application of neural networks today is based on a symbiosis of understanding their internal structure and the effective use of high-level computing platform tools and cloud services.

Keywords: network architecture, image libraries, weighting coefficients, pretrained model, convolutional neural network, markup files, loss function

pp. 510—519

For citation:
Pantyushin V. A., Starostin A. S., Katyshev V. А. Features of Frameworks and Libraries for Working with Neural Networks using the Example of the Ultralytics YOLO Ecosystem, Programmnaya Ingeneria, 2026, vol. 17, no. 9, pp. 510—519. DOI: 10.17587/prin.17.-510-519 (in Russian).

References:

1. Galushkin A. I., Tsypkin Ya. Z. Neural Networks: The His­tory of Theory Development: A Textbook for Universities, Moscow, Alliance, 2015, 840 p. (in Russian).
2. Burkov A. Machine Learning Without Fuss, Saint Petersburg, Piter SPb, 2020, 192 p. (in Russian).
3. Potapov A. S. Computer Vision Systems. A Textbook, Saint Petersburg, ITMO University, 2016, 161 p. (in Russian).
4. Behnke S. Hierarchical Neural Networks for Image Interpretation. Lecture Notes in Computer Science. Vol. 2766. Springer. 227 p. DOI: 10.1007/b11963.
5. Topol E. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again, Basic Books, 2019, 400 p.
6. Ford M. Architects of Intelligence: The truth about AI from the people building it, Packt Publishing, 2018, 554 p.
7. Giulli A., Pal S. Keras Library: A Deep Learning Toolkit, Moscow, DMK Press, 2017, 294 p. (in Russian).
8. Raska S. Python and Machine Learning, Moscow, DMK Press, 2017, 418 p. (in Russian).
9. Ciresan C. D., Meier U., Gambardella L. M., Schmidhuber J. Deep Big Simple Neural Nets Excel on Handwritten Digit Recognition (English), Neural Computation, 2010, vol. 22, no. 12, pp. 3207—3220. DOI: 10.1162/NECO_a_00052.
10. Fisher R. A. The Use of Multiple Measurements in Taxonomic Problems, Annals of Eugenics, 1963, vol. 7, no. 2, pp. 179—188. DOI: 10.1111/j.1469-1809.1936.tb02137.x.