Institutional Repository of Vilnius University
Generatyviniai besivaržantys neuroniniai tinklai vaizdams iš teksto generuoti /
Abstract
dc:descriptionText to image synthesis problem seeks to provide an ability to generate images that you could describe with a text description. These kinds of tasks are still very challenging, but in the past several years' new solutions for this problem were introduced which use generative adversarial networks and provided impressive results. In this bachelor final thesis, an analysis of generative adversarial networks for the synthesis of images from the text was carried out, several architectures were introduced, which try to solve text to image synthesis problem and an investigation of StackGAN architecture was performed to evaluate the capabilities of this architecture and impact of different parameters on the results of the network. The results of the StackGAN investigation have shown that this architecture is capable of demonstrating good results, generated images have a wide variety of possible outcomes, the quality of the generated image even with a small epoch size is relatively good, but to achieve image quality that would be similar to the quality of the original dataset the network requires longer training with bigger epoch size and that requires a large amount of computer resources. Investigation of network training with different learning rates and batch size values has shown that smaller discriminator learning rate value can provide more stability for generator and discriminator losses and smaller batch size with small epoch size generate images with better quality than large batch size values. One-sided label smoothing also has shown that it can help with balancing losses of generator and discriminator.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Vainickaitė, Julita,
- Contributors dc:contributor
-
- Kurasova, Olga
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD107115194&prefLang=en_US