Regulariziranje normalizirajućih tokova suparničkim gubitkom

master's thesis

• Title (english): Regularization of normalizing flows with adversarial loss
• Author: Martin Sršen
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-13, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Generiranje slika važno je područje računalnog vida s mnogim zanimljivim primjenama. Posebno su zanimljivi modeli zasnovani na normalizirajućim tokovima jer istodobno nude i mogućnost procjene izglednosti i brzo generiranje. Ipak, kvaliteta generiranih slika nešto je lošija nego kod modela zasnovanih na suparničkom gubitku. Ovaj rad proučava hibridne algoritme pristupa koji bi pokušali objediniti dobre strane oba pristupa. U početku rada se uvodimo u osnove strojnog učenja, u odabrani okvir za automatsku diferencijaciju pyTorch te biblioteke za
rukovanje tenzorima i slikama. Nakon toga detaljnije se upuštamo u vrste generativnih modela, njihove prednosti i nedostatke, razvoj kroz vrijeme. Na kraju iznosimo eksperimente i rezultate te sugestije za budući razvoj.

Abstract (english)

Image generation is an important area of ​​computer vision with many interesting applications. Particularly interesting are models based on normalizing flows because they offer the possibility of estimating the probability and fast generation at the same time. However, the quality of the generated images is slightly worse than that of models based on adversial loss. This paper examines hybrid algorithms of approaches that would attempt to combine the strengths of both approaches. At the beginning of the paper, we are introduced to the basics of machine learning, to the selected framework for automatic differentiation pyTorch and the libraries for
handling tensors and images. After that, we go into more detail about the types of generative models, their advantages and disadvantages, and their development over time. Finally, we present the experiments and results and suggestions for further development.

• Keywords: generativni modeli; suparnički gubitak; normalizirajući tok; pyTorch; strojno učenje
• Keywords (english): generative models; adversial loss; normalizing flow; pyTorch; machine learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:137523
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 16:08:12