Einzeltreffer — DigiBib

A pure integer quantization method for a lightweight neural network (LNN) is provided. The method includes the following steps: acquiring a maximum value of each pixel in each of the channels of the feature map of a current layer; dividing a value of each pixel in each of the channels of the feature map by a t-th power of the maximum value, t∈[0,1]; multiplying a weight in each of the channels by the maximum value of each pixel in each of the channels of the corresponding feature map; and convolving the processed feature map with the processed weight to acquire the feature map of a next layer. The algorithm is verified on SkyNet and MobileNet respectively, and lossless INT8 quantization on SkyNet and maximum quantization accuracy so far on MobileNetv2 are achieved.

Titel:	Pure integer quantization method for lightweight neural network (LNN)
Autor/in / Beteiligte Person:	SHANGHAITECH, UNIVERSITY
Link:	View record in USPTO Patent Grants (Volltext)
Veröffentlichung:	2024
Medientyp:	Patent
Sonstiges:	Nachgewiesen in: USPTO Patent Grants Sprachen: English Patent Number: 11934,954 Publication Date: March 19, 2024 Appl. No: 17/799933 Application Filed: September 22, 2021 Assignees: SHANGHAITECH UNIVERSITY (Shanghai, CN) Claim: 1. A pure integer quantization method for implementing a lightweight neural network (LNN) in an application specific integrated circuit (ASIC) or a field-programmable gate array (FPGA), comprising the following steps: step 1: setting, by the ASIC or FPGA, a feature map with N channels, N≥1, and acquiring a maximum value of each pixel in each of N channels of a first feature map of a current layer: step 2: processing, by the ASIC or FPGA, each pixel in each of the N channels of the first feature map as follows: dividing, by the ASIC or FPGA, a value of each pixel in au n-th channel of the first feature map by a t-th power of a maximum value in the n-th channel acquired in step 1, t∈[0,1]; and acquiring, by the ASIC or FPGA, N groups of weights corresponding to N channels of a second feature map of a next layer, wherein each of the N groups of the weights comprises N weights corresponding to the N channels of the first feature map of the current layer, and processing each of the N groups of the weights as follows: multiplying, by the ASIC or FPGA, the N weights in an n-th group respectively by the maximum value of each pixel in the N channels acquired in step 1; step 3: convolving, by the ASIC or FPGA, the first feature map processed in step 2 with the N groups of the weights processed in step 2 to acquire the second feature map of the next layer; and step 4: obtaining by the ASIC or FPGA, a quantization accuracy based on a result of step 3, and tuning, by the ASIC or FPGA, the t value to obtain a maximum quantization accuracy. Claim: 2. The pure integer quantization method for the LNN according to claim 1 , wherein the current layer is any layer except a last layer in the LNN. Patent References Cited: 10527699 January 2020 Cheng et al. ; 20190042948 February 2019 Lee ; 20190279072 September 2019 Gao ; 20190294413 September 2019 Vantrease et al. ; 20200401884 December 2020 Guo ; 20210110236 April 2021 Shibata ; 20220086463 March 2022 Coban ; 105528589 April 2016 ; 110930320 March 2020 ; 111311538 June 2020 ; 111402143 July 2020 ; 111937010 November 2020 ; 112418397 February 2021 ; 112488070 March 2021 ; 112560355 March 2021 ; 113128116 July 2021 ; WO-0074850 December 2000 ; WO-2005048185 May 2005 ; 2018073975 April 2018 Other References: Cho et al. (Per-channel Quantization Level Allocation for Quantizing Convolutional Neural Networks, Nov. 2020, pp. 1-3) (Year: 2020). cited by examiner ; Kang et al. (Decoupling Representation and Classifier for Long-Tailed Recognition, Feb. 2020, pp. 1-16) (Year: 2020). cited by examiner ; Polino et al. (Model Compression via Distillation and Quantization, Feb. 2018, pp. 1-21) (Year: 2018). cited by examiner ; Benoit Jacob, et al., Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference, IEEE, 2018, pp. 2704-2713. cited by applicant ; Raghuraman Krishnamoorthi, Quantizing deep convolutional networks for efficient inference: A whitepaper, 2018, pp. 1-36. cited by applicant ; Markus Nagel, et al., Data-Free Quantization Through Weight Equalization and Bias Correction, Qualcomm AI Research, 2019. cited by applicant ; Liu Guanyu, et al., Design and implementation of real-time defogging hardware accelerator based on image fusion, Hefei University of Technology, Master's Dissertation, 2020, pp. 1-81. cited by applicant Primary Examiner: Giroux, George Attorney, Agent or Firm: Bayramoglu Law Offices LLC

Klicken Sie ein Format an und speichern Sie dann die Daten oder geben Sie eine Empfänger-Adresse ein und lassen Sie sich per Email zusenden.

BibTeX Citavi, JabRef, u.a.
(Literaturverwaltung)

PDF kein Volltext!
(Merkzettel, Notizen)

RIS Endnote, Citavi u.a.
(Literaturverwaltung)

MODS
(XML zur Weiterverarbeitung)

oder

Wählen Sie das für Sie passende Zitationsformat und kopieren Sie es dann in die Zwischenablage, lassen es sich per Mail zusenden oder speichern es als PDF-Datei.

Gewünschter Zitations-Stil:

oder

Bitte prüfen Sie, ob die Zitation formal korrekt ist, bevor Sie sie in einer Arbeit verwenden. Benutzen Sie gegebenenfalls den "Exportieren"-Dialog, wenn Sie ein Literaturverwaltungsprogramm verwenden und die Zitat-Angaben selbst formatieren wollen.

Pure integer quantization method for lightweight neural network (LNN)