sapiСистемный анализ и прикладная информатика«System analysis and applied information science»2309-49232414-0481Belarusian National Technical University10.21122/2309-4923-2025-2-sapi-741Research ArticleСистемный анализSystem analysisИспользование U-Net сети с механизмом внимания совместно с архитектурными надстройками для фреймворка PyTorch в рамках поиска гиперпараметров посредством библиотеки Weights & Biases для предсказывания области видимости по карте местностиApplying attention U-net with PyTorch architectural add-ons for extensive hyperparameter search with Weights & Biases for area of visibility prediction based on terrainРулькоЕ. В.RulkoE. V.

Рулько Евгений Викторович – кандидат технических наук, доцент. Начальник научно- исследовательской лаборатории моделирования военных действий. 

г. Минск

Eugene Rulko – РhD, associate professor in computer science. The head of the research laboratory of military operation simulation. 

Minsk

eugeni1533@gmail.comВоенная академия Республики БеларусьБеларусьMilitary Academy of the Republic of BelarusBelarus20251508202502410Copyright © Рулько Е.В., 20252025Рулько Е.В.Rulko E.V.Данная работа распространяется под лицензией Creative Commons Attribution 4.0.This work is licensed under a Creative Commons Attribution 4.0 License.https://sapi.bntu.by/jour/article/view/741

Текущий уровень развития глубокого обучения позволяет заменить нейронными сетями существующие специфические для моделирования военных действий алгоритмы. Поиск гиперпараметров даёт возможность определить структуры сетей, подходящие для решения соответствующих задач. Данная работа описывает процесс поиска структуры сети для предсказания зоны оптической видимости на основе фрагмента цифровой карты местности и предлагает архитектурные решения для комбинирования возможных составных частей сети, обеспечивая их совместимость в рамках поиска наилучшего решения. В качестве финального варианта предлагается использование U-Net архитектуры с поканальным механизмом внимания и энкодером на основе сети ResNet50.

Current level of development in the sphere of deep learning allows replacing existing domain-specific algorithms for military simulation with approximating neural networks. Hyperparameter search allows finding network’s architecture, appropriate for a task. This work describes that process for the task of predicting area of optical visibility, taking a fragment of a digital map as input and proposes ancillary architectural solutions for stitching building blocks together, assuring their conformation for performing search among their possible combinations within the architectural space. The final proposed result is a channel-wise attention U-Net with an encoder, based on ResNet50 backbone.

глубокое обучениеU-Netмеханизм вниманиясегментацияпоиск гиперпараметровW&Bшаблонный методdeep learningU-Netattentionsegmentationhyperparameter searchW&Btemplate methodReferencesMathematical model complex for military grouping efficiency assessment: [Electronic resource] // URL: http://en.belfortex.com/page/show/9. (Date of access: 15/11/2024).Mathematical model complex for military grouping efficiency assessment: [Electronic resource] // URL: http://en.belfortex.com/page/show/9. (Date of access: 15/11/2024).E. Rulko, et al. Application of a simulation system for optimizing solutions based on elements of the theory of reflexive control. Collection of scientific articles of the Military academy of the Republic of Belarus. 2017. № 32. P. 153–162.E. Rulko, et al. Application of a simulation system for optimizing solutions based on elements of the theory of reflexive control. Collection of scientific articles of the Military academy of the Republic of Belarus. 2017. № 32. P. 153–162.Olaf Ronneberger, et al. U-Net: Convolutional Networks for Biomedical Image Segmentation. 2015. arXiv: 1505.04597.Olaf Ronneberger, et al. U-Net: Convolutional Networks for Biomedical Image Segmentation. 2015. arXiv: 1505.04597.Ozan Oktay et al. Attention U-Net: Learning Where to Look for the Pancreas. 2018. arXiv: 1804.03999.Ozan Oktay et al. Attention U-Net: Learning Where to Look for the Pancreas. 2018. arXiv: 1804.03999.Weights & Biases: [Electronic resource] // URL: https://wandb.ai/site. (Date of access: 15/11/2024).Weights & Biases: [Electronic resource] // URL: https://wandb.ai/site. (Date of access: 15/11/2024).PyTorch implementation of U-Net, R2U-Net, attention U-Net, attention R2U-Net. https://github.com/LeeJunHyun/Image_Segmentation. 2018.PyTorch implementation of U-Net, R2U-Net, attention U-Net, attention R2U-Net. https://github.com/LeeJunHyun/ Image_Segmentation. 2018.E. Gamma, et al. “Design Patterns Elements of Reusable Object-Oriented Software,” Addison-Wesley, Massachusetts, 1995.E. Gamma, et al. “Design Patterns Elements of Reusable Object-Oriented Software,” Addison-Wesley, Massachusetts,1995.PyTorch documentation. Weighted random sampler: [Electronic resource] // URL: https://pytorch.org/docs/stable/data.html#torch.utils.data.WeightedRandomSampler. (Date of access: 15/11/2024).PyTorch documentation. Weighted random sampler: [Electronic resource] // URL: https://pytorch.org/docs/stable/data.html#torch.utils.data.WeightedRandomSampler. (Date of access: 15/11/2024).How to Improve Class Imbalance using Class Weights in Machine Learning?: [Electronic resource] // URL: https://www.analyticsvidhya.com/blog/2020/10/improve-class-imbalance-class-weights/. (Date of access: 15/11/2024).How to Improve Class Imbalance using Class Weights in Machine Learning?: [Electronic resource] // URL: https://www.analyticsvidhya.com/blog/2020/10/improve-class-imbalance-class-weights/. (Date of access: 15/11/2024).N. V. Chawla, et al. SMOTE: Synthetic Minority Over-sampling Technique. 2011. arXiv: 1106.1813.N. V. Chawla, et al. SMOTE: Synthetic Minority Over-sampling Technique. 2011. arXiv: 1106.1813.Long Chen, et al. SCA-CNN: Spatial and Channel-wise Attention in Convolutional Networks for Image Captioning. 2016. arXiv: 1611.05594.Long Chen, et al. SCA-CNN: Spatial and Channel-wise Attention in Convolutional Networks for Image Captioning. 2016. arXiv: 1611.05594.

The authors declare that there are no conflicts of interest present.