A vectorization realizing method for matrix local response normalization

Abstract

A vectorization realizing method for matrix local response normalization is disclosed. The method comprises accessing an input characteristic map to a vector processor and performing block transposition, so as to store data points in each channel direction in a column in a plurality of calculation results after convolutional calculation, simultaneously normalizing the transposed matrix by using each vector processing unit VPE, each vector processing unit VPE simultaneously completing the calculation of each output characteristic map and storing the results calculated by each vector processingunit VPE in columns. The method can realize the normalized vectorization of the matrix local response, and has the advantages of simple realization method, good parallelism and high processing efficiency.

Description

technical field [0001] The invention relates to the technical field of machine learning based on a convolutional neural network, in particular to a vectorized realization method of matrix local response normalization. Background technique [0002] With the rise of deep learning technology, object recognition technology based on convolutional neural network has been widely used in image recognition, speech recognition, natural language processing and other fields. The convolutional neural network is the most widely used neural network model in the current deep learning algorithm model, and it is also the model with the best recognition rate. The convolutional neural network model generally includes matrix convolution, activation function, and maximum pooling. Pooling or average pooling, local linear normalization operations, etc. [0003] Normalization is one of the commonly used modules in the convolutional neural network model, in which the local response normalization lay...

AI Technical Summary

Problems solved by technology

[0007] However, the local response normalization operation is not only computationally intensive but also memory-intensive. If a reasonable calculation method cannot be adopted, it will be difficult to give full play to its due computing advantages even with high-performance computing equipment.

In the existing technology, the local response normalization operation usually adopts a non-vectorization method, that is, only one pixel in a picture can be operated at a time, and the calculation efficiency is not high, so it cannot be applied to the above-mentioned vectors. The processor structure always realizes multi-core parallel operation, especially when the local response normalization operation is performed on multi-input feature maps, the calculation efficiency is low, and it cannot meet the high real-time application requirements

Images