CN111340173A - Method and system for training generation countermeasure network for high-dimensional data and electronic equipment - Google Patents

Abstract

The present application relates to a generative adversarial network training method, system and electronic device for high-dimensional data. Including: step a: building a generative adversarial network backbone structure; step b: using tensor train decomposition algorithm to perform tensor train decomposition on the generative adversarial network backbone structure; step c: using real high-dimensional data to train a tensor train decomposition based The tensor auto-encoder of , outputs the spatial structure features of the real high-dimensional data through the tensor auto-encoder; Step d: Combine the output of the tensor auto-encoder and the last layer feature generated by the discriminator as The input of the last layer to train the generative adversarial network. The present application enables the generative adversarial network to directly generate high-dimensional data, while reducing the amount of parameters of the network, making the training of the network more stable, and improving the quality and diversity of the generated data.

Description

A Generative Adversarial Network Training Method, System and Electronics for High-Dimensional Data equipment

technical field

The present application belongs to the technical field of generative adversarial network training, and in particular relates to a generative adversarial network training method, system and electronic device for high-dimensional data.

Background technique

The main role of the generative adversarial network is to use a fake distribution to fit the real distribution through continuous learning. With the development of deep learning, the generative adversarial network can be used to generate things similar to the real world, such as: pictures, music, articles Wait. At present, generative adversarial networks are widely used in data set enhancement, face photo generation, medical image conversion, super-resolution and other fields, and its development potential is huge.

The invention patent with the application number "201910287274.6" provides a stable training method for a generative adversarial network, including: inputting a training image into a self-encoder for processing to obtain a first generated image; based on the difference between the training image and the first generated image The loss value of the pre-trained generator is obtained by training; based on the training image and the second generated image generated by the pre-trained generator, the pre-trained generator and the discriminator are trained, and based on the corresponding generator at the end of the training and the discriminator to get a generative adversarial network.

Another invention patent with the application number "201811461559.9" provides a generative confrontation network device and a training method thereof. The generative adversarial network device includes a generative network and a discriminant network. The generation network is configured to generate a first sample according to the input data; the discriminant network is coupled to the generation network and is configured to receive the first sample and perform training based on the first sample; the generation network includes a first memristor array as the first sample. An array of weight values. The generative confrontation network device can omit the process of adding noise to the fake samples generated by the generation network, thereby saving training time, reducing resource consumption and improving the training speed of the generative confrontation network.

To sum up, the existing generative adversarial networks still have the following defects:

1. Difficult to train and very unstable. It is difficult to achieve a good mutual promotion state between the generator and the discriminator. In actual training, the discriminator is easy to converge, but the generator is difficult to train.

2. The mode collapses. Generative adversarial networks are prone to fall into the state of mode collapse during learning, that is, the generator always generates similar samples and cannot continue learning.

3. The current generative adversarial network is difficult to directly generate complex high-dimensional data. The current generative adversarial network is mostly applied to generate low-dimensional data such as pictures, but when it is used to generate complex high-dimensional data, taking medical image data as an example, because the medical image data conversion algorithm uses the generative adversarial network when It is often generated by a single slice, which will lose the structural information of the original data and greatly reduce the integrity of the generated data.

SUMMARY OF THE INVENTION

The present application provides a generative adversarial network training method, system and electronic device for high-dimensional data, aiming to solve one of the above technical problems in the prior art at least to a certain extent.

In order to solve the above problems, the application provides the following technical solutions:

A generative adversarial network training method for high-dimensional data, comprising the following steps:

Step a: Build the backbone structure of the generative adversarial network;

Step b: use the tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network;

Step c: use real high-dimensional data to train a tensor autoencoder based on tensor train decomposition, and output spatial structure features of real high-dimensional data through the tensor autoencoder;

Step d: The output of the tensor autoencoder and the features of the last layer generated by the discriminator are combined as the input of the last layer to train the generative adversarial network.

The technical solutions adopted in the embodiments of the present application further include: in the step a, the GAN backbone structure is a GAN backbone structure based on 3D convolution and 3D deconvolution.

The technical solutions adopted in the embodiments of the present application further include: in the step b, the use of the tensor train decomposition algorithm to perform tensor train decomposition on the GAN backbone structure is specifically: introducing the tensor train decomposition algorithm into the All 3D convolution and 3D deconvolution layers in the backbone structure of the generative adversarial network are decomposed by tensor train to obtain 3D-TT-Conv layer, 3D-TT-deConv layer Floor.

The technical solution adopted in the embodiment of the present application further includes: the tensor train decomposition of the 3D convolution specifically includes:

卷积核为：

卷积后输出张量：

Assuming that the input three-dimensional data dimension is W×H×D and the number of channels is C, that is, the input tensor:

The convolution kernel is:

Output tensor after convolution:

3D卷积的公式为：Write each element of the output tensor as:

The formula for 3D convolution is:

Represent the dimension of each channel of the output tensor as:

H'=H-l+1

w'=W-l+1

D'=D-l+1

转化为W′H′D′×l³C大小的矩阵，对应元素变换为：the input tensor

Converted to a matrix of size W′H′D′×l ³ C, and the corresponding elements are transformed into:

转换为大小为l³C×S的矩阵，对应元素变换为：convolution kernel tensor

Converted to a matrix of size l ³ C×S, the corresponding elements are transformed as:

Perform matrix multiplication on the input matrix X and the convolution kernel matrix K to obtain an output matrix Y of size W'H'D'×S, and restore the output matrix Y to an output tensor

将矩阵K张量化为张量

对其进行张量火车分解：Apply the tensor train decomposition to the convolution kernel matrix K: decompose the input and output dimensions:

Tensor the matrix K into a tensor

Tensor train decomposition of it:

In the above formula:

转化为大小为W×H×D×C₁×...×C_d的张量

将该张量

与卷积核的Tensor-Train矩阵进行运算，得到大小为 (W-l+1)×(H-l+1)×S₁...×S_d输出张量：the input tensor

Converted to a tensor of size W×H×D×C ₁ ×...×C _d

the tensor

Operate with the Tensor-Train matrix of the convolution kernel to obtain an output tensor of size (W-l+1)×(H-l+1)×S ₁ ...×S _d :

The technical solutions adopted in the embodiments of the present application further include: in the step c, the use of real high-dimensional data to train a tensor autoencoder based on tensor train decomposition is specifically: The product operation is introduced into the autoencoder, and the n-mode product operation in the tensor operation is used to replace the matrix multiplication of the input vector and parameter matrix in the fully connected layer, and the dimension of the tensor data is directly enlarged and reduced to extract the real high-dimensional data. the spatial structure characteristics.

Another technical solution adopted by the embodiments of the present application is: a generative adversarial network training system for high-dimensional data, including:

Network backbone building module: used to build the backbone structure of generative adversarial network;

Tensor train decomposition module: used to perform tensor train decomposition on the backbone structure of the generative adversarial network using the tensor train decomposition algorithm;

Tensor autoencoder training module: used to train a tensor autoencoder based on tensor train decomposition using real high-dimensional data, and output the spatial structure features of real high-dimensional data through the tensor autoencoder;

Network training module: used to combine the output of the tensor autoencoder and the features of the last layer generated by the discriminator as the input of the last layer to train the generative adversarial network.

The technical solutions adopted in the embodiments of the present application further include: the backbone structure of the generative confrontation network is a backbone structure of the generative confrontation network based on 3D convolution and 3D deconvolution.

The technical solutions adopted in the embodiments of the present application further include: the tensor train decomposition module uses a tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network. Specifically, the tensor train decomposition algorithm is introduced into the generation All 3D convolution and 3D deconvolution layers in the backbone structure of the adversarial network are decomposed by tensor train to obtain 3D-TT-Conv layers and 3D-TT-deConv layers.

The technical solution adopted in the embodiment of the present application further includes: the tensor train decomposition of the 3D convolution specifically includes:

卷积核为：

卷积后输出张量：

Assuming that the input three-dimensional data dimension is W×H×D and the number of channels is C, that is, the input tensor:

The convolution kernel is:

Output tensor after convolution:

3D卷积的公式为：Write each element of the output tensor as:

The formula for 3D convolution is:

Represent the dimension of each channel of the output tensor as:

H'=H-l+1

W'=W-l+1

D'=D-l+1

转化为W′H′D′×l³C大小的矩阵，对应元素变换为：the input tensor

Converted to a matrix of size W′H′D′×l ³ C, and the corresponding elements are transformed into:

转换为大小为l³C×S的矩阵，对应元素变换为：convolution kernel tensor

Converted to a matrix of size l ³ C×S, the corresponding elements are transformed as:

Perform matrix multiplication on the input matrix X and the convolution kernel matrix K to obtain an output matrix Y of size W'H'D'×S, and restore the output matrix Y to an output tensor

将矩阵K张量化为张量

对其进行张量火车分解：Apply the tensor train decomposition to the convolution kernel matrix K: decompose the input and output dimensions:

Tensor the matrix K into a tensor

Tensor train decomposition of it:

In the above formula:

转化为大小为W×H×D×C₁×...×C_d的张量

将该张量

与卷积核的Tensor-Train矩阵进行运算，得到大小为 (W-l+1)×(H-l+1)×S₁...×S_d输出张量：the input tensor

Converted to a tensor of size W×H×D×C ₁ ×...×C _d

the tensor

Operate with the Tensor-Train matrix of the convolution kernel to obtain an output tensor of size (W-l+1)×(H-l+1)×S ₁ ...×S _d :

The technical solutions adopted in the embodiments of the present application further include: the tensor autoencoder training module uses real high-dimensional data to train a tensor train decomposition-based tensor autoencoder. Specifically: The product operation is introduced into the autoencoder, and the n-mode product operation in the tensor operation is used to replace the matrix multiplication of the input vector and parameter matrix in the fully connected layer, and the dimension of the tensor data is directly enlarged and reduced to extract the real high-dimensional data. the spatial structure characteristics.

Another technical solution adopted in the embodiment of the present application is: an electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

The memory stores instructions executable by the one processor, the instructions are executed by the at least one processor, so that the at least one processor can perform the above-mentioned generative adversarial network training method for high-dimensional data the following operations:

Step a: Build the backbone structure of the generative adversarial network;

Step b: use the tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network;

Step c: use real high-dimensional data to train a tensor autoencoder based on tensor train decomposition, and output spatial structure features of real high-dimensional data through the tensor autoencoder;

Step d: The output of the tensor autoencoder and the features of the last layer generated by the discriminator are combined as the input of the last layer to train the generative adversarial network.

Compared with the prior art, the beneficial effects of the embodiments of the present application are: the GAN training method, system and electronic device for high-dimensional data in the embodiments of the present application decompose the 3D volumes in the GAN by using tensor train decomposition. The product is tensorized, so that the generative adversarial network has the ability to directly generate high-dimensional data, and at the same time reduces the amount of parameters of the network, making the training of the network more stable. At the same time, by introducing tensor operations into the training process of the self-encoder, a tensorized encoder is obtained, the spatial structure features of the real high-dimensional data are output through the tensorized encoder, and the spatial structure features of the real high-dimensional data are introduced. In the training of generative adversarial network, the generator generates more realistic and diverse data, and improves the quality and diversity of the generated data.

Description of drawings

1 is a flowchart of a generative adversarial network training method for high-dimensional data according to an embodiment of the present application;

2 is a structural diagram of a generative adversarial network based on tensor autoencoder and tensor train decomposition according to an embodiment of the present application;

3 is a schematic structural diagram of a generative adversarial network training system for high-dimensional data according to an embodiment of the present application;

FIG. 4 is a schematic structural diagram of a hardware device of a generative adversarial network training method for high-dimensional data provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

In order to solve the insufficiency of the prior art, the generative adversarial network training method for high-dimensional data in the embodiments of the present application uses tensor train decomposition technology and tensor autoencoder to improve the original generative adversarial network. By introducing the tensor operation into the network structure of the generative adversarial network, the 3D convolution is tensorized and decomposed, and the tensor is used to preserve the structural information of the input information to increase the stability of the generative adversarial network training. By using the tensor autoencoder, the information of the true distribution is introduced into the generative adversarial network, and the KL distance between the true distribution and the generative distribution can be simultaneously optimized from the forward and reverse directions, thereby improving the generative adversarial network to directly generate high-dimensional data capabilities, and increase the variety of generated data.

Specifically, please refer to FIG. 1 , which is a flowchart of a generative adversarial network training method for high-dimensional data according to an embodiment of the present application. The generative adversarial network training method for high-dimensional data according to the embodiment of the present application includes the following steps:

Step 100: Build a generative adversarial network backbone structure based on 3D convolution and 3D deconvolution;

Step 200: Introduce the tensor train decomposition algorithm into all 3D convolution and 3D deconvolution layers in the backbone structure of the generative adversarial network, and use the tensor train decomposition algorithm to perform the 3D convolution and 3D deconvolution layers in the generative adversarial network. tensor train decomposition;

In step 200, the principle of the tensor train decomposition algorithm is to express each element in a high-dimensional tensor in the form of several matrix multiplications. which is:

A(i ₁ ,i ₂ ,...,id )=G ₁ (i ₁ )G ₂ (i ₂ )...G _d ( _{id )} (1)

In formula (1), G _k (i _k ) is a matrix of size r _k-1 ×r _k , and r _k represents the rank (TT-ranks) of the tensor train decomposition. To ensure that the final result is a scalar, r ₀ =r _k =1.

The present invention uses the tensor train decomposition algorithm to perform tensor train decomposition on 3D-Conv (3D convolution) and 3D-deConv (3D deconvolution) in the generative adversarial network, and obtains 3D-TT-Conv layer, 3D-TT- deConv layer. In the following embodiments, the 3D-TT-Conv layer is taken as an example to describe the application of tensor train decomposition in terms of formulas and examples.

The tensor train decomposition formula of the 3D-TT-Conv layer:

1. First, analyze the traditional 3D convolution formula:

卷积核为：

那么卷积后输出张量：

The input three-dimensional data dimension is W×H×D, and the number of channels is C, that is, the input tensor:

The convolution kernel is:

Then the output tensor after convolution:

3D卷积的公式可写为：To facilitate the formulation, write each element of the output tensor as:

The formula for 3D convolution can be written as:

2. The convolution operation is matrixed:

In order to tensor the 3D convolution, the 3D convolution operation needs to be matrixed first, and then the tensor train decomposition of the convolution kernel matrix is performed. Specific steps include:

(1) Express the dimension of each channel of the output tensor as:

H'=H-l+1

w'=W-l+l

D′=D-1+1 (3)

转化为W′H′D′×l³C大小的矩阵，对应元素变换为：the input tensor

Converted to a matrix of size W′H′D′×l ³ C, and the corresponding elements are transformed into:

转换为大小为l³C×S的矩阵，对应元素变换为：(2) The convolution kernel tensor

Converted to a matrix of size l ³ C×S, the corresponding elements are transformed as:

(3) Perform matrix multiplication on the input matrix X and the convolution kernel matrix K to obtain an output matrix Y of size W'H'D'×S, and restore the output matrix Y to an output tensor

将矩阵K张量化为张量

对其进行张量火车分解，则卷积核矩阵K的TT分解过程为：(4) Apply the tensor train decomposition to the convolution kernel matrix K: decompose the input and output dimensions:

Tensor the matrix K into a tensor

Perform tensor train decomposition on it, then the TT decomposition process of the convolution kernel matrix K is:

In formula (7):

转化为大小为W×H×D×C₁×...×C_d的张量

将该张量

与卷积核的Tensor- Train矩阵进行运算，得到大小为(W-l+1)×(H-l+1)×S₁...×S_d输出张量：(5) In order to match the convolution kernel after TT decomposition, the input tensor

Converted to a tensor of size W×H×D×C ₁ ×...×C _d

the tensor

Operates with the Tensor-Train matrix of the convolution kernel to obtain an output tensor of size (W-l+1)×(H-l+1)×S ₁ ...×S _d :

The above formula is the final 3D-TT-Conv layer tensor train decomposition operation formula.

The derivation process of the tensor train decomposition operation formula of the 3D-TT-deConv layer is the same as that of the 3D-TT-Conv layer, and will not be repeated here.

Step 300: use real high-dimensional data to train a tensor autoencoder (TT-Encoder) based on tensor train decomposition, and output the spatial structure features of real high-dimensional data through the tensor autoencoder;

In step 300, the present invention introduces the n-mode product operation in the tensor operation into the self-encoder, and uses the n-mode product operation in the tensor operation to replace the matrix multiplication of the input vector and the parameter matrix in the fully connected layer, which can be directly Enlarging and reducing the dimension of tensor data, that is, the dimension raising and reduction operations in the autoencoder, can encode high-dimensional real high-dimensional data to low-dimensional data. Since there is no operation to vectorize high-dimensional data in tensor operations, the spatial structure features of high-dimensional data can be extracted.

E.g:

Step 400: Combine the output of the tensor autoencoder and the features of the last layer generated by the discriminator as the input of the last layer, train the generative adversarial network, and obtain the generation based on the tensor autoencoder and the tensor train decomposition. adversarial network;

In step 400, the present invention can introduce the distribution of the real high-dimensional data into the training process of the generative confrontation network by combining the spatial structure features of the real high-dimensional data obtained after encoding with the features of the last layer of the discriminator, so as to achieve simultaneous The KL distance (Kullback-Leibler Divergence, relative entropy) between the real distribution and the generated distribution is simultaneously optimized in the forward and reverse directions, thereby guiding the generator to generate more realistic and diverse data, and improving the quality and diversity of the generated data. The training process of the generative adversarial network is the same as the training method of the original generative adversarial network, and will not be repeated here.

The structure of generative adversarial network based on tensor autoencoder and tensor train decomposition is shown in Figure 2. In the figure, z is random noise, TG is the 3D generator of tensor train decomposition, and TD is the 3D discriminator of tensor train decomposition. , TE is Tensor-Encoder (tensor self-encoder), which is used to directly extract the spatial structure feature f of real high-dimensional data, and h is the last layer feature of the discriminator.

Please refer to FIG. 3 , which is a schematic structural diagram of a generative adversarial network training system for high-dimensional data according to an embodiment of the present application. The generative adversarial network training system for high-dimensional data in the embodiment of the present application includes a network backbone building module, a tensor train decomposition module, a tensor autoencoder training module, and a network training module.

Network backbone building module: used to build the backbone structure of generative adversarial network based on 3D convolution and 3D deconvolution;

Tensor train decomposition module: used to introduce the tensor train decomposition algorithm into all 3D convolution and 3D deconvolution layers in the backbone structure of the generative adversarial network, and use the tensor train decomposition algorithm to analyze the 3D convolution and 3D The deconvolution layer performs tensor train decomposition; the principle of the tensor train decomposition algorithm is to express each element in a high-dimensional tensor in the form of several matrices.

which is:

A(i ₁ ,i ₂ ,...,id )=G ₁ (i ₁ )G ₂ (i ₂ )...G _d ( _{id )} (1)

In formula (1), G _k (i _k ) is a matrix of size r _k-1 ×r _k , and r _k represents the rank (TT-ranks) of the tensor train decomposition. To ensure that the final result is a scalar, r ₀ =r _k =1.

The invention uses the tensor train decomposition algorithm to decompose the tensor train of 3D-Conv (3D convolution) and 3D-deConv (3D deconvolution) in the generative confrontation network, and obtains 3D-TT-Conv layer, 3D-TT- deConv layer. In the following embodiments, the 3D-TT-Conv layer is taken as an example to describe the application of tensor train decomposition in terms of formulas and examples.

The tensor train decomposition formula of the 3D-TT-Conv layer:

1. First, analyze the traditional 3D convolution formula:

卷积核为：

那么卷积后输出张量：

The input three-dimensional data dimension is W×H×D, and the number of channels is C, that is, the input tensor:

The convolution kernel is:

Then the output tensor after convolution:

3D卷积的公式可写为：To facilitate the formulation, write each element of the output tensor as:

The formula for 3D convolution can be written as:

2. The convolution operation is matrixed:

In order to tensor the 3D convolution, the 3D convolution operation needs to be matrixed first, and then the tensor train decomposition of the convolution kernel matrix is performed. Specific steps include:

(1) Express the dimension of each channel of the output tensor as:

H'=H-l+1

w'=W-l+1

D′=D-1+1 (3)

转化为W′H′D′×l³C大小的矩阵，对应元素变换为：the input tensor

Converted to a matrix of size W′H′D′×l ³ C, and the corresponding elements are transformed into:

转换为大小为l³C×S的矩阵，对应元素变换为：(2) The convolution kernel tensor

Converted to a matrix of size l ³ C×S, the corresponding elements are transformed as:

(3) Perform matrix multiplication on the input matrix X and the convolution kernel matrix K to obtain an output matrix Y of size W'H'D'×S, and restore the output matrix Y to an output tensor

(4) Apply the tensor train decomposition to the convolution kernel matrix K:

将矩阵K张量化为张量

对其进行张量火车分解，则卷积核矩阵K的TT分解过程为：Decompose the input and output dimensions:

Tensor the matrix K into a tensor

Perform tensor train decomposition on it, then the TT decomposition process of the convolution kernel matrix K is:

In formula (7):

转化为大小为W×H×D×C₁×...×C_d的张量

将该张量

与卷积核的Tensor- Train矩阵进行运算，得到大小为(W-l+1)×(H-l+1)×S₁...×S_d输出张量：(5) In order to match the convolution kernel after TT decomposition, the input tensor

Converted to a tensor of size W×H×D×C ₁ ×...×C _d

the tensor

Operates with the Tensor-Train matrix of the convolution kernel to obtain an output tensor of size (W-l+1)×(H-l+1)×S ₁ ...×S _d :

The above formula is the final 3D-TT-Conv layer tensor train decomposition operation formula.

Tensor autoencoder training module: used to train a tensor autoencoder (TT-Encoder) based on tensor train decomposition using real high-dimensional data, and output the spatial structure features of real high-dimensional data through the tensor autoencoder; The invention introduces the n-mode product operation in the tensor operation into the self-encoder, uses the n-mode product operation in the tensor operation to replace the matrix multiplication of the input vector and the parameter matrix in the full connection layer, and can directly perform the tensor operation. The dimension of the data is enlarged and reduced, that is, the dimension-raising and dimension-reducing operations in the autoencoder can encode high-dimensional real high-dimensional data to low-dimensional data. Since there is no operation to vectorize high-dimensional data in tensor operations, the spatial structure features of high-dimensional data can be extracted.

Network training module: It is used to combine the output of the tensor auto-encoder and the last layer of features generated by the discriminator as the input of the last layer to train the generative adversarial network to obtain a tensor based auto-encoder and tensor train. Decomposed generative adversarial network; wherein, the present invention can introduce the distribution of real high-dimensional data into the training process of the generative adversarial network by combining the spatial structure features of the real high-dimensional data obtained after encoding with the features of the last layer of the discriminator , to simultaneously optimize the KL distance (Kullback-Leibler Divergence, relative entropy) between the real distribution and the generated distribution from the forward and reverse directions, thereby guiding the generator to generate more realistic and diverse data, and improving the quality and diversity of the generated data sex.

FIG. 4 is a schematic structural diagram of a hardware device of a generative adversarial network training method for high-dimensional data provided by an embodiment of the present application. As shown in Figure 4, the device includes one or more processors and memory. Taking a processor as an example, the device may further include: an input system and an output system.

The processor, the memory, the input system, and the output system may be connected by a bus or in other ways, and the connection by a bus is taken as an example in FIG. 4 .

As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor executes various functional applications and data processing of the electronic device by running the non-transitory software programs, instructions and modules stored in the memory, that is, the processing method of the above method embodiment is implemented.

The memory may include a stored program area and a stored data area, wherein the stored program area can store an operating system and an application program required by at least one function; the stored data area can store data and the like. Additionally, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, the memory may optionally include memory located remotely from the processor, which may be connected to the processing system via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

The input system can receive input numerical or character information and generate signal input. The output system may include a display device such as a display screen.

The one or more modules are stored in the memory, and when executed by the one or more processors, perform the following operations of any of the foregoing method embodiments:

Step a: Build the backbone structure of the generative adversarial network;

Step b: use the tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network;

Step c: use real high-dimensional data to train a tensor autoencoder based on tensor train decomposition, and output spatial structure features of real high-dimensional data through the tensor autoencoder;

Step d: The output of the tensor autoencoder and the features of the last layer generated by the discriminator are combined as the input of the last layer to train the generative adversarial network.

The above product can execute the method provided by the embodiments of the present application, and has functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, reference may be made to the method provided in this embodiment of the present application.

An embodiment of the present application provides a non-transitory (non-volatile) computer storage medium, where the computer storage medium stores computer-executable instructions, and the computer-executable instructions can perform the following operations:

Step a: Build the backbone structure of the generative adversarial network;

Step b: use the tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network;

Step c: use real high-dimensional data to train a tensor autoencoder based on tensor train decomposition, and output spatial structure features of real high-dimensional data through the tensor autoencoder;

Step d: The output of the tensor autoencoder and the features of the last layer generated by the discriminator are combined as the input of the last layer to train the generative adversarial network.

An embodiment of the present application provides a computer program product, the computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program includes program instructions, and when the program instructions are executed by a computer , which causes the computer to do the following:

Step a: Build the backbone structure of the generative adversarial network;

Step b: use the tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network;

Step c: use real high-dimensional data to train a tensor autoencoder based on tensor train decomposition, and output spatial structure features of real high-dimensional data through the tensor autoencoder;

Step d: The output of the tensor autoencoder and the features of the last layer generated by the discriminator are combined as the input of the last layer to train the generative adversarial network.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined in this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (11)

1. a method for generating adversarial network training for high-dimensional data, is characterized in that, comprises the following steps: Step a: Build the backbone structure of the generative adversarial network; Step b: use the tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network; Step c: use real high-dimensional data to train a tensor autoencoder based on tensor train decomposition, and output spatial structure features of real high-dimensional data through the tensor autoencoder; Step d: The output of the tensor autoencoder and the features of the last layer generated by the discriminator are combined as the input of the last layer to train the generative adversarial network. 2. The generative adversarial network training method for high-dimensional data according to claim 1, wherein in the step a, the generative adversarial network backbone structure is based on the generation of 3D convolution and 3D deconvolution Adversarial network backbone structure. 3. The generative adversarial network training method for high-dimensional data according to claim 2, characterized in that, in the step b, the use of a tensor train decomposition algorithm to tensor the generative adversarial network backbone structure The train decomposition is specifically: introducing the tensor train decomposition algorithm into all 3D convolution and 3D deconvolution layers in the backbone structure of the generative adversarial network, and performing tensor train decomposition on the 3D convolution and 3D deconvolution layers, Obtain 3D-TT-Conv layer, 3D-TT-deConv layer. 4. The generative adversarial network training method for high-dimensional data according to claim 3, wherein the tensor train decomposition to the 3D convolution specifically comprises: Assuming that the input three-dimensional data dimension is W×H×D and the number of channels is C, that is, the input tensor:

The convolution kernel is:

Output tensor after convolution:

Write each element of the output tensor as:

The formula for 3D convolution is:

Represent the dimension of each channel of the output tensor as:

the input tensor

transform into

The size of the matrix, the corresponding elements are transformed into:

convolution kernel tensor

Convert to size as

The matrix of , the corresponding elements are transformed into:

Perform matrix multiplication on the input matrix X and the convolution kernel matrix K to obtain an output matrix Y of size W'H'D'×S, and restore the output matrix Y to an output tensor

Apply the tensor train decomposition to the convolution kernel matrix K: decompose the input and output dimensions:

Tensor the matrix K into a tensor

Tensor train decomposition of it:

In the above formula:

the input tensor

Converted to a tensor of size W×H×D×C ₁ ×...×C _d

the tensor

Operates with the Tensor-Train matrix of the convolution kernel to obtain a size of

Output tensor:

5. The generative adversarial network training method for high-dimensional data according to any one of claims 1 to 4, characterized in that, in the step c, the use of real high-dimensional data to train a tensor train decomposition based on The tensor autoencoder is specifically: the n-mode product operation in the tensor operation is introduced into the autoencoder, and the n-mode product operation in the tensor operation is used to replace the matrix multiplication of the input vector and parameter matrix in the fully connected layer, The dimension of tensor data is directly enlarged and reduced, and the spatial structure features of real high-dimensional data are extracted. 6. A generative adversarial network training system for high-dimensional data, comprising: Network backbone building module: used to build the backbone structure of generative adversarial network; Tensor train decomposition module: used to perform tensor train decomposition on the backbone structure of the generative adversarial network using the tensor train decomposition algorithm; Tensor autoencoder training module: used to train a tensor autoencoder based on tensor train decomposition using real high-dimensional data, and output the spatial structure features of real high-dimensional data through the tensor autoencoder; Network training module: used to combine the output of the tensor autoencoder and the features of the last layer generated by the discriminator as the input of the last layer to train the generative adversarial network. 7. The generative adversarial network training system for high-dimensional data according to claim 6, wherein the generative adversarial network backbone structure is a generative adversarial network backbone structure based on 3D convolution and 3D deconvolution. 8 . The generative adversarial network training system for high-dimensional data according to claim 7 , wherein the tensor train decomposition module uses a tensor train decomposition algorithm to perform tensor train decomposition on the generative adversarial network backbone structure. 9 . Specifically, the tensor train decomposition algorithm is introduced into all 3D convolution and 3D deconvolution layers in the backbone structure of the generative adversarial network, and the 3D convolution and 3D deconvolution layers are decomposed by tensor train to obtain 3D - TT-Conv layer, 3D-TT-deConv layer. 9. The generative adversarial network training system for high-dimensional data according to claim 8, wherein the tensor train decomposition to the 3D convolution specifically comprises: Assuming that the input three-dimensional data dimension is W×H×D and the number of channels is C, that is, the input tensor:

The convolution kernel is:

Output tensor after convolution:

Write each element of the output tensor as:

The formula for 3D convolution is:

Represent the dimension of each channel of the output tensor as:

the input tensor

transform into

The size of the matrix, the corresponding elements are transformed into:

convolution kernel tensor

Convert to size as

The matrix of , the corresponding elements are transformed into:

Perform matrix multiplication on the input matrix X and the convolution kernel matrix K to obtain an output matrix Y of size W'H'D'×S, and restore the output matrix Y to an output tensor

Apply the tensor train decomposition to the convolution kernel matrix K: decompose the input and output dimensions:

Tensor the matrix K into a tensor

Tensor train decomposition of it:

In the above formula:

the input tensor

Converted to a tensor of size W×H×D×C ₁ ×...×C _d

the tensor

Operates with the Tensor-Train matrix of the convolution kernel to obtain a size of

Output tensor:

10. The generative adversarial network training system for high-dimensional data according to any one of claims 6 to 9, wherein the tensor autoencoder training module uses real high-dimensional data to train a tensor train decomposition based on The tensor autoencoder is specifically: the n-mode product operation in the tensor operation is introduced into the autoencoder, and the n-mode product operation in the tensor operation is used to replace the matrix multiplication of the input vector and parameter matrix in the fully connected layer. , directly enlarge and reduce the dimension of tensor data, and extract the spatial structure features of real high-dimensional data. 11. An electronic device comprising: at least one processor; and a memory communicatively coupled to the at least one processor; wherein, The memory stores instructions executable by the one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform any of the above 1 to 5 for high-speed operation. The following operations of the generative adversarial network training method for dimensional data: Step a: Build the backbone structure of the generative adversarial network; Step b: use the tensor train decomposition algorithm to perform tensor train decomposition on the backbone structure of the generative adversarial network; Step c: use real high-dimensional data to train a tensor autoencoder based on tensor train decomposition, and output spatial structure features of real high-dimensional data through the tensor autoencoder; Step d: The output of the tensor autoencoder and the features of the last layer generated by the discriminator are combined as the input of the last layer to train the generative adversarial network.

Images