ELECTRONIC DEVICE FOR COMPRESSING CONVOLUTIONAL ARTIFICIAL INTELLIGENCE NEURAL NETWORK MODEL, AND CONTROL METHOD THEREOF
Provided are an electronic device and a method for compressing a convolutional artificial intelligence neural network (convolutional neural network (CNN)) model including a convolution layer by using the electronic device. The method may comprise the steps of: identifying a convolution tensor of at...
Gespeichert in:
Hauptverfasser: | , , , , |
---|---|
Format: | Patent |
Sprache: | eng ; fre ; kor |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Provided are an electronic device and a method for compressing a convolutional artificial intelligence neural network (convolutional neural network (CNN)) model including a convolution layer by using the electronic device. The method may comprise the steps of: identifying a convolution tensor of at least one convolution layer; determining a direction in which tiling of the convolution tensor is performed, on the basis of a shape of the convolution tensor; generating a tile matrix from the convolution tensor according to the direction of tiling; performing low rank approximation (LRA) for the tile matrix to generate a U matrix and a V matrix; and generating a U convolution tensor through recombination of the U matrix, and generating a V convolution tensor through recombination of the V matrix.
L'invention concerne un dispositif électronique et un procédé de compression d'un modèle de réseau neuronal à intelligence artificielle convolutif (réseau neuronal convolutif (CNN)) comprenant une couche de convolution en utilisant le dispositif électronique. Le procédé peut comprendre les étapes suivantes : identification d'un tenseur de convolution d'au moins une couche de convolution ; détermination d'une direction dans laquelle est effectuée la juxtaposition du tenseur de convolution, sur la base d'une forme du tenseur de convolution ; génération d'une matrice juxtaposition à partir du tenseur de convolution selon la direction de juxtaposition ; réalisation d'une approximation de rang bas (LRA) pour la matrice de juxtaposition afin de générer une matrice U et une matrice V ; et génération d'un tenseur de convolution U par recombinaison de la matrice U, et génération d'un tenseur de convolution V par recombinaison de la matrice V.
전자 장치 및 전자 장치를 이용하여 하나의 컨볼루션 레이어 (Convolution layer)를 포함하는 컨볼루션 인공지능 신경망 (Convolutional Neural Network, CNN) 모델을 압축하는 방법이 제공된다. 상기 방법은, 상기 적어도 하나의 컨볼루션 레이어의 컨볼루션 텐서(Convolution Tensor)를 식별하는 단계, 상기 컨볼루션 텐서의 모양에 기초하여 상기 컨볼루션 텐서를 타일링(tiling)하는 방향을 결정하는 단계, 상기 타일링하는 방향에 따라서, 상기 컨볼루션 텐서로부터 타일 행렬(Tile matrix)을 생성하는 단계, 상기 타일 행렬에 대해서 행렬 근사화(Low Rank Approximation, LRA)를 수행함으로써, U행렬 및 V행렬을 생성하는 단계 및 상기 U행렬을 재결합함으로써 U컨볼루션 텐서를 생성하고, 상기 V행렬을 재결합함으로써 V컨볼루션 텐서를 생성하는 단계를 포함할 수 있다. |
---|