Kvantizacija dubokih konvolucijskih modela

master's thesis

• Title (english): Quantization of deep convolutional models
• Author: Filip Oreč
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Vladimir Čeperić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Ovaj rad se bavi kvantiziranjem dubokih konvolucijskih modela. Svrha kvantizacije je provođenje izračuna
u nižoj preciznosti kako bi se smanjilo vrijeme zaključivanja dubokog modela. Kvantiziran je model SwiftNet
na problemu semantičke segmentacije. Za implementaciju korišten je okvir TensorRT, te je model kvantiziran
za uređaj NVIDIA Jetson AGX Xavier. Pokazalo se kako kvantizirani modela ima puno manje vrijeme zaključivanja
u odnosu na FP32 model uz minimalan gubitak točnosti, te nešto manje vrijeme zaključivanja nego FP16 model
također uz minimalan gubitak točnosti.

Abstract (english)

The goal of this thesis is quantization of deep convolutional models. The goal of quantization is to
use lower precision for computation which lowers inference time of the deep model. Model which is
quantized in this thesis is SwiftNet model on the problem of semantic segmentation. For implementation
TensorRT framework was used and models was quantized for the NVIDIA Jetson AGX Xavier embedded hardware.
It has been shown that quantized model has much lower inference time than FP32 model with minimal
accuracy loss and little lower inference time than FP16 model also with minimal accuracy loss.

• Keywords: Kvantizacija; Konvolucijski modeli; Duboko učenje; TensorRT
• Keywords (english): Quantization; Convolutional models; Deep learning; TensorRT
• Language: croatian
• URN:NBN: urn:nbn:hr:168:565022
• 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 15:45:01