Multi-objective deep convolution generative adversarial network model and learning method thereof

Abstract

The invention discloses a multi-objective deep convolution generative adversarial network model and a learning method thereof so that problems that the exiting deep convolution generative adversarialnetwork model can not be converged easily and training is not stable can be solved. A target deep convolution generative adversarial network model is provided; on the basis of a group search strategy,multiple groups of deep convolution generative adversarial networks are trained simultaneously to realize coevolution of multiple individuals, thereby ensuring stability of model training; with a Pareto dominant mechanism, a potential optimal combination unit of generative networks and adversarial networks is selected for follow-up training in each iteration so as to ensure convergence of the model training; on the basis of the characteristics of the deep convolution network, a cross operator is designed to realize interaction of parameter information between different networks; and because of combination of the global search ability of the evolutionary algorithm and fast local search ability of the gradient descent algorithm, the accuracy and effectiveness of a learning frame put forwardnewly are ensured.

Description

technical field

[0001] The invention belongs to the technical field of deep learning, and in particular relates to a multi-objective deep convolution generative confrontation network model and a learning method thereof, which can be used for picture generation and picture classification tasks. Background technique

[0002] Generative Adversarial Networks (Generative Adversarial Networks) is a generative deep learning model, and the research on Generative Adversarial Networks has become a hot spot in the field of deep learning. The main idea of ​​generative confrontation network is to learn the probability distribution of training samples, and realize the representation and expansion of data according to the learned distribution.

[0003] The structure of the generative adversarial network is inspired by the two-person zero-sum game in game theory, that is, the sum of the gains of the two parties in the game is zero, and there is no cooperative relationship between the two pa...

Examples

Embodiment Construction

[0087] Below in conjunction with accompanying drawing, implementation steps and experimental effects of the present invention are described in further detail:

[0088] refer to figure 1 , the specific implementation steps of the present invention are as follows:

[0089] Step 1, population initialization.

[0090] Initialize n generators by random generation and n discriminator parameters And combined into n groups of generative confrontation networks in order:

[0091] Enter the local search stage to learn and optimize each group of generative confrontation networks.

[0092] Set the structure of the generator G in DCGAN: a 5-layer micro-step convolutional neural network composed of input layer→deconvolution layer→deconvolution layer→deconvolution layer→deconvolution layer→output layer, given The feature map of each layer, and determine the size of each deconvolution layer and initialize the weights and biases randomly. The parameters of each layer are set as follows...

technical field

[0001] The invention belongs to the technical field of deep learning, and in particular relates to a multi-objective deep convolution generative confrontation network model and a learning method thereof, which can be used for picture generation and picture classification tasks. Background technique

[0002] Generative Adversarial Networks (Generative Adversarial Networks) is a generative deep learning model, and the research on Generative Adversarial Networks has become a hot spot in the field of deep learning. The main idea of ​​generative confrontation network is to learn the probability distribution of training samples, and realize the representation and expansion of data according to the learned distribution.

[0003] The structure of the generative adversarial network is inspired by the two-person zero-sum game in game theory, that is, the sum of the gains of the two parties in the game is zero, and there is no cooperative relationship between the two pa...

Examples

Embodiment Construction

[0087] Below in conjunction with accompanying drawing, implementation steps and experimental effects of the present invention are described in further detail:

[0088] refer to figure 1 , the specific implementation steps of the present invention are as follows:

[0089] Step 1, population initialization.

[0090] Initialize n generators by random generation and n discriminator parameters And combined into n groups of generative confrontation networks in order:

[0091] Enter the local search stage to learn and optimize each group of generative confrontation networks.

[0092] Set the structure of the generator G in DCGAN: a 5-layer micro-step convolutional neural network composed of input layer→deconvolution layer→deconvolution layer→deconvolution layer→deconvolution layer→output layer, given The feature map of each layer, and determine the size of each deconvolution layer and initialize the weights and biases randomly. The parameters of each layer are set as follows...

Images