METHOD AND APPARATUS FOR EFFICIENT WINOGRAD CONVOLUTION

An objective of the present disclosure is to ensure accuracy and operation speed of a neural network by quantizing an input feature map with low bit precision to train a neural network using winograd convolution. The present invention relates to a method for operating a neural network model regardin...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: BYOUNGJO KIM, DALEON KIM, INGUK HWANG
Format: Patent
Sprache:eng ; kor
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:An objective of the present disclosure is to ensure accuracy and operation speed of a neural network by quantizing an input feature map with low bit precision to train a neural network using winograd convolution. The present invention relates to a method for operating a neural network model regarding learning method thereof. A method and an apparatus are capable of obtaining a winograd kernel, transforming an input feature map into a winograd region to acquire a winograd input feature map, respectively quantizing the winograd input feature map and the winograd kernel using a zero point parameter and a scaling parameter corresponding to the input feature map, acquiring output data by performing a convolution operation on quantized input feature map and kernel and quantizing the output data for training neural network by comparing quantized output data with a correct value of the training set. 본 개시는 뉴럴 네트워크 모델의 학습 방법에 있어서, 위노그래드 커널을 획득하고, 입력 특징 맵을 위노그래드 영역으로 변환하여 위노그래드 입력 특징 맵을 획득하고, 입력 특징 맵에 대응되는 영점 파라미터 및 스케일링 파라미터를 사용하여 위노그래드 입력 특징 맵 및 위노그래드 커널 각각을 양자화 시키고, 양자화된 입력 특징 맵과 커널을 콘볼루션 연산을 수행하여 출력 데이터를 획득하고, 출력 데이터를 양자화하여 학습 데이터(training set)의 정답값과 비교하여 뉴럴 네트워크를 학습시키는 뉴럴 네트워크 모델의 동작 방법을 개시한다.