COMPARISON OF DIFFERENT CONFIGURATIONS OF THE GAN MODEL BASED ON THE AWS CLOUD COMPUTING SERVICE
DOI:
https://doi.org/10.32782/IT/2022-2-7Keywords:
Julia programming language, machine learning, machine art, Generative adversarial network (GAN), AWS cloud computing serviceAbstract
The aim. The aim of the work is to develop an optimal configuration of the system for effective learning of the machine learning model, which implements the concept of machine art. The configuration of the system includes such aspects as the architecture of the model, the learning algorithm and the platform on which this algorithm will run. Each of these aspects is developed and a solution is proposed that will show the best results of performance measurements. The methodology of the solution of the task set is a comparable analysis of productivity indices of different system configurations that were used taking into account previous studies of machine art models which realized the machine art concept. Scientific novelty. In the course of the work the further development in the direction of application of the Julia programming language based on the AWS cloud computing service was made. An effective version of the software and hardware configuration of the system for learning of a machine learning model that implements the concept of machine art is proposed for the first time. Conclusions. The results of this work could be used for implementation of an effective system for learning of a machine learning model that use the concept of machine art, as well as in further studies of the prospects of using the Julia programming language in combination with the AWS cloud computing service. All the results of the measurements of productivity and quality of training of different configurations are presented in tables and graphs.
References
Стаття авторів моделі GAN. [Electronic resource] – Access mode: https://arxiv.org/pdf/1406.2661.pdf
Функція активації maxout, яка була застосована в першій публікації про модель GAN [Electronic resource] – Access mode: https://arxiv.org/pdf/ 1302.4389v4.pdf
Стаття авторів моделі DCGAN [Electronic resource] – Access mode: https://arxiv.org/pdf/1511.06434.pdf
Свідчення спеціалістів, які підтвердили ефективність застосування функції активації LeakyReLU в моделі GAN [Electronic resource] – Access mode: https://www.reddit.com/r/MLQuestions/comments/c96mci/why_do_gans_only_work_with_leaky_relu/
Розділ книги Яна Гудфелоу, присвячений згортковим нейронним мережам [Electronic resource] – Access mode: https://www.deeplearningbook.org /contents/convnets.html
Стаття автора ідеї про додатковий шар BatchNorm в нейронній мережі [Electronic resource] – Access mode: https://arxiv.org/pdf/1502.03167.pdf
Стаття автора ідеї про додатковий шар Dropout в нейронній мережі [Electronic resource] – Access mode: https://www.cs.toronto.edu/~rsalakhu /papers/srivastava14a.pdf
Опис розрахунку функції помилки binary cross-entropy [Electronic resource] – Access mode: https://towardsdatascience.com/understanding-binary-cross-entropy-log-loss-a-visual-explanation-a3ac6025181a
Стаття автора алгоритму зворотного розповсюдження помилки та градієнтного спуску [Electronic resource] – Access mode: https://www.nature.com /articles/323533a0
Стаття автора методу оптимізації алгоритму навчання ADAM [Electronic resource] – Access mode: https://arxiv.org/pdf/1412.6980.pdf
