Ping-pong storage method and device for sparse neural network

Abstract

The invention discloses a ping-pong storage method and device for a sparse neural network, which solve the problem of great consumption of on-chip storage resources due to one-time loading of a largeamount of weight data during calculation of the sparse neural network. The ping-pong storage device comprises a sparse processing unit, a weight storage unit M0, a weight storage unit M1, a calculation unit and a control unit. The invention discloses a ping-pong storage method for a sparse neural network. Configuration bits are added into weight data subjected to sparse processing, M0 and M1 circularly store the weight data added with the configuration bits in a ping-pong storage mode, and the calculation unit calculates a switching point of each group of data. When the next group of data is uploaded completely, data switching is carried out when convolution calculation is carried out to a switching point each time, and the control unit is responsible for controlling each unit to completethe above work. According to the method, the problem of determining the switching point during ping-pong storage of the sparse neural network is solved, and ping-pong storage of the sparse neural network is further realized.

Description

technical field [0001] The present invention relates to the field of hardware-accelerated convolutional neural networks, and in particular to a ping-pong storage method and device for sparse neural networks. Background technique [0002] Convolutional neural networks (CNN, or Deepconvolutional neural networks, DCNN) are quite different from most other networks. They are primarily used for image processing, but can also be used for other types of input, such as audio. The sparsity of the neural network actually refers to the sparsity of the weights, which converts the samples into a suitable sparse expression form, thereby simplifying the learning task and reducing the complexity of the model, which is usually called sparse coding. "Sparseness" is defined as: only a few non-zero elements or only a few elements that are much greater than zero. There is a reason for choosing to represent our input data using components with sparsity, because the vast majority of sensory data,...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the above-mentioned technical problems existing in the prior art, the present invention provides a method and device for ping-pong storage of sparse neural networks, which solves the problem of determining switching points during ping-pong storage of sparse neural networks, and further realizes the ping-pong storage of sparse neural networks. storage

Images