27-06-2023 дата публикации
Номер: CN116343009A
Автор:
QIAN PING,
HAN RUI,
WANG WENHAO,
ZHENG YIMING,
XIE LINGDONG,
LI FUQIANG,
LI TE,
JIANG PENG,
JIANG XIONGWEI,
LI FEIRAN,
LIU SHUANG,
GONG LEI,
ZHOU XUEHAI,
LI XI,
QIAN JIAMING
Принадлежит:
The invention discloses a three-dimensional neural network processing method and system, an image processing method and system and a storage medium, and belongs to the technical field of neural networks and image processing. According to the three-dimensional neural network processing method based on the block cyclic matrix, by constructing a block cyclic matrix model, a calculation acceleration model and a full-frequency-domain model, the problems of memory access, calculation irregularity and the like existing in a previous model compression method such as pruning are solved, and the three-dimensional neural network 3D CNN is compressed by using the block cyclic matrix; and furthermore, fast Fourier transform (FFT) is used for accelerating calculation, so that remarkable storage and calculation compression effects are achieved on the premise of keeping a model structure rule. On the basis, activation, batch normalization and pooling operations in a frequency domain are introduced, and ...
Подробнее