CNN model, training method thereof, terminal and computer readable storage medium

Abstract

The invention discloses a CNN model, a training method thereof, a terminal and a computer readable storage medium, belongs to the field of CNN model training, and aims to solve the problem of how to analyze large-batch samples in batches on the premise of ensuring the training performance of the model. The method comprises the following steps: inserting a synchronization layer between each batch normalization layer and a previous network layer adjacent to the batch normalization layer, wherein all network layers between an input layer and the adjacent synchronization layer and all network layers between any two adjacent synchronization layers are independent network layer units respectively; and performing forward propagation on all the network layer units section by section in a data batch mode, wherein the CNN model is of a linear structure and is a model trained by the training method. And a processor in the terminal is configured to call the program instruction to execute the training method. In a computer readable storage medium, when a program instruction is executed by a processor, the processor executes the training method.

Description

technical field [0001] The invention relates to the field of CNN model training, in particular to a CNN model and its training method, a terminal, and a computer-readable storage medium. Background technique [0002] Batch normalization (English full name BatchNormalization, English abbreviation BN) can speed up the convergence of the deep neural network model, reduce the requirements for parameter initialization, and has scale invariance, making the model loss function surface smoother, so this normalization method It is widely used in deep models, especially for convolutional neural network layers. The batch normalization operator is generally placed between the convolutional layer (or fully connected layer) and the activation layer. The operator transforms the net activation value x output by the previous layer. The formula is as follows: [0003] [0004] Among them, μ and σ are the mean and variance of the set to which the net activation value x belongs, respectivel...

AI Technical Summary

Problems solved by technology

The former leads to a decrease in model training performance, while the latter requires a higher economic cost

Images