Kvantiziranje dubokih konvolucijskih modela

undergraduate thesis

• Title (english): Quantization of deep convolutional models
• Author: Tomislav Ćosić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Ivan Grubišić (član povjerenstva)
• Committee member: Josip Šarić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-06, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Duboki konvolucijski modeli danas su metoda izbora za mnoge zadatke računalnog vida. Međutim, njihova računska složenost isključuje mnoge zanimljive primjene u robotici i inteligentnom prometu. Taj problem možemo ublažiti zamjenom decimalnih računskih operacija odgovarajućim cjelobrojnim operacijama. U okviru rada, potrebno je upoznati se s konvolucijskim modelom ResNet-18. Istražiti postojeće pristupe za kvantiziranje težina dubokih modela. Odabrati slobodno dostupni skup slika te oblikovati podskupove za učenje, validaciju i testiranje. Oblikovati algoritme, validirati hiperparametre te uhodati učenje, validiranje i zaključivanje. Modificirati postupak učenja tako da na izlazu dobijemo kvantizirani model. Vrednovati naučene modele, prikazati postignutu točnost te provesti usporedbu s rezultatima iz literature. Komentirati učinkovitost učenja i zaključivanja. Predložiti pravce budućeg razvoja. Radu priložiti izvorni kod razvijenih postupaka uz potrebna objašnjenja i dokumentaciju. Citirati korištenu literaturu i navesti dobivenu pomoć.

Abstract (english)

Today, deep convolutional models are the method of choice for many tasks in computer vision. However, their computational complexity prevents many interesting applications in robotics and intelligent traffic. We can alleviate that problem by replacing floating-point operations with their corresponding integer counterparts. In the cope of this paper, it is necessary to introduce ourselves with the ResNet-18 convolutional model. Research existing approaches towards quantization of the weights in convolutional models. Select freely available dataset and form subsets for learning, validation and testing. Model the algorithm, validate hyperparameters and set up learning, validation and testing. Modify the learning procedure to get the quantized model as an output. Evaluate learned models, display acquired accuracy and compare the values with literature results. Comment on the efficiency of the learning and inference process. Suggest the directions of futher research. Attach the source code of developed procedure along with necessary explanations and documentation. Cite used literature and state received help.

• Keywords: računalni vid; duboko učenje; konvolucijski modeli; kvantizacija; klasifikacija; PyTorch; ResNet-18
• Keywords (english): computer vision; deep learning; convolutional models; quantization; classification; PyTorch; ResNet-18
• Language: croatian
• URN:NBN: urn:nbn:hr:168:953732
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-15 09:57:05