Multi-channel neural network model training method, device and computer storage medium

Abstract

The present invention discloses a multi-channel neural network model training method, device, and computer storage medium. The multi-channel neural network model training method includes: when training a target model, inputting training data into an N-channel neural network model for learning. The model adopts a multi-channel network structure, including a residual network channel and N-1 densely connected network channels with different connection structures. The convolutional layers in these N-1 DenseNet channels are densely connected according to the corresponding number of interval layers. In this way, the ResNet channel can well reuse model parameters, improving the efficiency of the model. At the same time, the N-1 DenseNet channels can utilize the intermediate results of the previous layers, greatly improving the computational efficiency and performance of the model.

Description

Technical Field

[0001] The present invention relates to the field of electronic technology, and in particular to a multi-channel neural network model training method, device and computer storage medium. Background Art

[0002] Deep learning technology is playing an increasingly important role in diverse scenarios, including computer vision, speech recognition, and natural language processing. The core of deep learning lies in the design of deep model structures. In traditional residual networks, as the network deepens, the model's output is continuously modified to produce a final result. However, as residual networks accumulate data, subsequent layers can only use the accumulated results of previous layers as input, without accessing the intermediate results of previous layers. Furthermore, directly connecting the results of the current layer to all previous layers results in a large number of redundant parameters, making computation and optimization difficult. Summary of the Invention

[0003] In view of the above problems, the present invention is proposed to provide a multi-channel neural network model training method, device and computer storage medium that overcome the above problems or at least partially solve the above problems.

[0004] In a first aspect, this embodiment provides a multi-channel neural network model training method, characterized by comprising:

[0005] Determine the training data;

[0006] The training data is input into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, where N is an integer greater than 2.

[0007] Optionally, the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0008] For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0009] Optionally, the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0010] For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0011] Optionally, the data output by the ResNet channel and the N-1 DenseNet channels are processed by a convolution layer and then input into the next learning module, including:

[0012] For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

[0013] Optionally, the training data includes a training image set of classified images, and the target model is an image classification model.

[0014] In a second aspect, an embodiment of the present application provides a multi-channel neural network model training device, comprising:

[0015] a determination unit, configured to determine training data;

[0016] A model training unit is used to input the training data into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and input into the next learning module, where N is an integer greater than 2.

[0017] Optionally, the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0018] For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0019] Optionally, the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0020] For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0021] Optionally, the data output by the ResNet channel and the N-1 DenseNet channels are processed by a convolution layer and then input into the next learning module, including:

[0022] For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

[0023] Optionally, the training data includes a training image set of classified images, and the target model is an image classification model.

[0024] In a third aspect, an embodiment of the present invention further provides a multi-channel neural network model training device, comprising a processor and a memory: the memory is used to store a script program for executing the multi-channel neural network model training method of the first aspect; the processor is configured to execute the program stored in the memory.

[0025] In a fourth aspect, an embodiment of the present invention further provides a computer storage medium for storing computer software instructions used for the multi-channel neural network model training method of the first aspect, which includes a program designed for executing the multi-channel neural network model training method of the above aspect.

[0026] The above one or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:

[0027] In the technical solution of the embodiment of the present invention, when training the target model, the training data is input into an N-channel neural network model for learning. This model adopts a multi-channel network structure, including a residual network channel and N-1 densely connected network channels with different connection structures. The convolutional layers in these N-1 DenseNet channels are densely connected according to the corresponding number of interval layers. In this way, the ResNet channel can effectively reuse model parameters, improving the efficiency of the model. At the same time, the N-1 DenseNet channels can utilize the intermediate results of the previous layers, greatly improving the computational efficiency and performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0029] Figure 1 This is a flow chart of a multi-channel neural network model training method applied to a front-end device in the first embodiment of the present invention;

[0030] Figure 2Schematic diagram of a multi-channel neural network model training device according to a second embodiment of the present invention;

[0031] Figure 3 Schematic diagram of a multi-channel neural network model training device in the third embodiment of the present invention. DETAILED DESCRIPTION

[0032] This embodiment discloses a multi-channel neural network model training method, device and computer storage medium.

[0033] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Unless there is a conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0034] The term "and / or" in this document simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. Additionally, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0035] Example

[0036] The first embodiment of the present invention provides a multi-channel neural network model training method, which is applied to a server. The flow chart of the method is as follows: Figure 1 As shown, the following steps are included:

[0037] S101: Determine training data;

[0038] S102: Input the training data into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, where N is an integer greater than 2.

[0039] Specifically, in this embodiment, deep learning is a method in machine learning that is based on learning to represent data and can be applied to various applications such as object detection, image recognition, and image classification. The faster the network that applies deep learning, the higher its accuracy and the better its performance.

[0040] ResNet is one of the most commonly used deep learning models. Because ResNet directly adds the input of each layer to the output of the convolutional layer by element-wise adding the corresponding features, information from the previous layer can be directly transferred to the next layer without matrix multiplication or nonlinear transformations. Residual convolutional neural networks (ResNets) can be used to reuse features and extract high-purity features. ResNets focus on feature reuse but are not good at discovering new features.

[0041] DenseNet is one of the most commonly used deep learning models. DenseNet concatenates the output of each layer into the input of each subsequent layer. This allows for the discovery of new features using densely connected convolutional neural networks (DenseNets). However, in DenseNets, features extracted by convolution in subsequent layers are likely to have already been extracted by previous layers, resulting in a high level of redundant features. Each layer in the network receives the output features of all previous layers as input, thus extracting useful information from the outputs of all previous layers. However, as the network width and depth increase, the number of network parameters increases exponentially.

[0042] It can be seen that ResNet has a high reuse rate, but low redundancy. DenseNet can create new features, but the redundancy is high. If these two structures are combined, better performance can be achieved. Therefore, in the method in the embodiment of the present application, an N-channel neural network model is adopted, wherein the N channels include ResNet channels, and the ResNet channels adopt the network structure of the ResNet channels. The input of each convolutional layer includes the input and output of the previous layer, and the corresponding feature addition operation is performed to achieve feature reuse. For the N-1 DenseNet channels, the connection structures are different, and they are densely connected according to their corresponding number of interval layers.

[0043] Furthermore, in this embodiment, dense connection can be performed according to the number of interval layers, which can be specifically divided into the following two dense connection structures:

[0044] The first method is to segment the convolutional layers contained in each of the N-1 DenseNet channels according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0045] Specifically, in this embodiment, the connection structure of one of the DenseNet channels is described in detail. Assuming that the number of interval layers corresponding to the channel is 5, the channel is segmented at intervals of 5 layers. Each segment includes 5 convolutional layers. The input of the 5th layer in the segment is combined with the outputs of the previous 1st to 4th layers. The input of the 4th layer in the segment is the output of the previous 1st to 3rd layers. The input of the 3rd layer in the segment is the output of the previous 1st to 2nd layers. The input of the 2nd layer in the segment is the output of the 1st layer, and the input of the 1st layer in the segment is the output of the 5th layer in the previous segment. In this way, dense connections are performed within the segment every 5 layers, which can effectively control the size of the convolutional layer input data, thereby reducing the generation of unnecessary redundant features and better constructing some new features.

[0046] Furthermore, since the number of layers of dense connection intervals in each of the N-1 DenseNet channels is different, assuming that there are two DenseNet channels, the number of interval layers corresponding to the first DenseNet channel is 3, and the number of interval layers corresponding to the second DenseNet channel is 5. In this way, the densely connected structures with different numbers of interval layers are different, and the feature data constructed by different DenseNet channels are also different, which can enrich the feature information as much as possible and improve the performance of the model. Moreover, compared with the DenseNet structure in the prior art where the input of each layer is merged into the output of all subsequent layers, the DenseNet channel in the embodiment of the present application needs to process less data, and can achieve faster training of the target model.

[0047] The second type: for each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0048] Specifically, in this embodiment, the connection structure of one of the DenseNet channels is described in detail. Assuming that the number of interval layers corresponding to the channel is 5, it can be stipulated that within the DenseNet channel, the output data of each convolutional layer is only passed forward 5 layers, such as: the output data of the 1st layer can be merged into the 2nd to 6th layers, the output data of the 2nd layer can be merged into the 3rd to 7th layers, and for the 6th layer, its input data is the merged data of the output data of the 1st to 5th layers, and for the 7th layer, its input data is the merged data of the output data of the 2nd to 6th layers. In this way, the amount of data processed by each layer can be effectively controlled, and the input of each layer will also be merged into the output data of the previous layers. Furthermore, while reducing the generation of unnecessary redundant features, it can also better construct some new features, making the trained model performance better.

[0049] Similarly, since the number of layers of dense connection intervals in each of the N-1 DenseNet channels is different, assuming that there are two DenseNet channels, the number of interval layers corresponding to the first DenseNet channel is 3, and the number of interval layers corresponding to the second DenseNet channel is 5. In this way, the densely connected structures with different numbers of interval layers are different, and the feature data constructed by different DenseNet channels are also different, which can enrich the feature information as much as possible and improve the performance of the model. Moreover, compared with the DenseNet structure in the prior art where the input of each layer is merged into the output of all subsequent layers, the DenseNet channel in the embodiment of the present application needs to process less data, which can achieve faster training of the target model.

[0050] Furthermore, in this embodiment, the preset N-channel neural network model includes a preset number of learning modules, and the preset number is the number of convolutional layers contained in the entire model. In the specific implementation process, the preset number can be set according to actual needs, and this application does not impose any restrictions here.

[0051] For each convolutional layer in the N-channel neural network model, the input data for each learning module is the output data of the previous learning module. This data is split into N pieces and then fed into the N channels of the learning module. For the ResNet channel, the first input data is split from the output data of the previous learning module. This first input data has the same number of features as the ResNet channel. After the first input data is fed into the channel, it is summed with the input data of the channel in the previous learning module and then output.

[0052] For any one of the N-1 DenseNet channels, a second input data is split from the data output by the previous learning module. After the second input data is input into the channel, the second input data is merged according to the connection structure of the channel. Assuming that the DenseNet channel adopts a dense connection structure with an interval of 5 layers, and adopts the first connection method in the aforementioned embodiment, the convolution layer corresponding to the learning module is the 5th layer, then the output data of the channel in the previous 1st to 4th learning modules are merged and output. Assuming that the DenseNet channel adopts a dense connection structure with an interval of 5 layers, and adopts the second connection method in the aforementioned embodiment, the convolution layer corresponding to the learning module is the 7th layer, then the output data corresponding to the channel of the previous 2nd to 6th learning modules are merged and output.

[0053] Furthermore, in this embodiment, the convolution processing of the learning module can be implemented through the following steps:

[0054] For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

[0055] Specifically, in this embodiment, the N channels above perform corresponding operations on their respective input data and output N output data. The output data of each channel undergoes a 1*1 convolution operation to reduce the data dimension. The output data is then added together to obtain the added output data, which is then subjected to a 3*3 convolution using a group operation. The output data is then further increased in dimension through a 1*1 convolution operation, completing the learning of one learning block. Other learning blocks learn data in the same manner.

[0056] In practice, a larger N value means more channels and more data are processed, resulting in better model training results but slower training speed. A smaller N value means fewer channels and less data are processed, resulting in faster training speed, but this may compromise model training results. Therefore, the value of N can be set based on actual needs and is not a limitation in this application.

[0057] The embodiments of the present application can be applied to application scenarios such as image recognition, image classification, and scene classification. Therefore, for image classification scenarios, the training data can be classified image data, and the resulting model is the image classification model. For image recognition scenarios, the training data can be image data of a target object, and the resulting model is the image recognition model for the target object. During the specific implementation process, training data can be selected according to actual needs to obtain the corresponding target model, and this application does not impose any restrictions on this.

[0058] The method in this embodiment uses a preset N-channel neural network model, which includes a residual network channel and N-1 densely connected network channels with different connection structures. The convolutional layers in these N-1 DenseNet channels are densely connected according to the corresponding number of interval layers. In this way, the ResNet channel can effectively reuse model parameters and improve model efficiency. At the same time, the N-1 DenseNet channels can utilize the intermediate results of the previous layers, greatly improving the computational efficiency and performance of the model.

[0059] Please refer to Figure 2 The third embodiment of the present invention further provides a multi-channel neural network model training device, the device comprising:

[0060] A determination unit 201 is used to determine training data;

[0061] The model training unit 202 is used to input the training data into the N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and input into the next learning module, where N is an integer greater than 2.

[0062] As an optional embodiment, the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0063] For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0064] As an optional embodiment, the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0065] For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0066] As an optional embodiment, the data output by the ResNet channel and the N-1 DenseNet channels are processed by a convolution layer and then input into the next learning module, including:

[0067] For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

[0068] As an optional embodiment, the training data includes a training image set of classified images, and the target model is an image classification model.

[0069] Specifically, in this embodiment, the specific implementation process of the multi-channel neural network model training device applied to the terminal for multi-channel neural network model training has been described in detail in the aforementioned first embodiment, and will not be repeated here in this embodiment.

[0070] The third embodiment of the present invention also provides a multi-channel neural network model training device, please refer to Figure 3 For the sake of convenience, only the parts related to the embodiments of the present invention are shown. For specific technical details not disclosed, please refer to the method part of the embodiments of the present invention.

[0071] Figure 3The diagram shows a partial structure of a multi-channel neural network model training device provided by an embodiment of the present invention. The multi-channel neural network model training device includes a memory 501, which is used to store a script program for executing the multi-channel neural network model training method in the first embodiment. The multi-channel neural network model training device also includes a processor 502 connected to the memory 501, and the processor 502 is configured to execute the program stored in the memory 501.

[0072] When the processor 502 executes the computer script program, the steps of the multi-channel neural network model training method in the first embodiment or the second embodiment are implemented. Alternatively, when the processor executes the computer script program, the functions of each module / unit in the multi-channel neural network model training device in the third embodiment or the fourth embodiment are implemented.

[0073] Exemplarily, the computer script program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer script program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer script program in the computer device.

[0074] The device may include, but is not limited to, a processor and a memory. Figure 3 This is only an example diagram of the functional components of the multi-channel neural network model training device and does not constitute a limitation on the multi-channel neural network model training device. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the multi-channel neural network model training device may also include input and output devices, network access devices, buses, etc.

[0075] The processor 502 may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor. The processor is the control center of the computer device and connects various parts of the entire computer device using various interfaces and lines.

[0076] The memory 501 can be used to store the computer script program and / or module, and the processor realizes various functions of the computer device by running or executing the computer script program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a script program storage area and a data storage area, wherein the script program storage area can store an operating system, an application script program required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the multi-channel neural network model training device (such as audio data, video data, etc.). In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0077] In this embodiment of the present invention, the processor 502 has the following functions:

[0078] Determine the training data;

[0079] The training data is input into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, where N is an integer greater than 2.

[0080] In this embodiment of the present invention, the processor 502 further has the following functions:

[0081] For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0082] In this embodiment of the present invention, the processor 502 further has the following functions:

[0083] For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0084] In this embodiment of the present invention, the processor 502 further has the following functions:

[0085] For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

[0086] In an embodiment of the present invention, the training data includes a training image set of classified images, and the target model is an image classification model.

[0087] The sixth embodiment of the present invention provides a computer-readable storage medium having a computer script program stored thereon. If the functional unit integrated in the multi-channel neural network model training device in the third embodiment or the fourth embodiment of the present invention is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the multi-channel neural network model training method in the first embodiment or the second embodiment above, and can also be completed by instructing the relevant hardware through a computer script program. The computer script program can be stored in a computer-readable storage medium, and when the computer script program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer script program includes computer script program code, and the computer script program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device, medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0088] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0089] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

[0090] The present invention discloses A1, a multi-channel neural network model training method, characterized by comprising:

[0091] Determine the training data;

[0092] The training data is input into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, where N is an integer greater than 2.

[0093] A2. The method of claim A1, wherein the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0094] For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0095] A3. The method as claimed in claim A1, characterized in that the convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0096] For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0097] A4. The method according to claim A1, wherein the data output by the ResNet channel and the N-1 DenseNet channels are processed by a convolution layer and then input into the next learning module, comprising:

[0098] For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

[0099] A5. The method according to any one of claims A1 to A4, wherein the training data comprises a training image set of classified images, and the target model is an image classification model.

[0100] B1. A multi-channel neural network model training device, comprising:

[0101] a determination unit, configured to determine training data;

[0102] A model training unit is used to input the training data into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and input into the next learning module, where N is an integer greater than 2.

[0103] B2. The device of claim B1, wherein the convolutional layers of each of the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0104] For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0105] B3. The device of claim B1, wherein the convolutional layers of each of the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including:

[0106] For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

[0107] B4. The device according to claim B1, wherein the data output by the ResNet channel and the N-1 DenseNet channels are processed by a convolution layer and then input into the next learning module, comprising:

[0108] For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

[0109] B5. The device as described in any one of claims B1 to B4 is characterized in that the training data includes a training image set of classified images, and the target model is an image classification model.

[0110] C1. A multi-channel neural network model training device, comprising a processor and a memory:

[0111] The memory is used to store a script program for executing the method according to any one of claims 1 to 5;

[0112] The processor is configured to execute the program stored in the memory.

[0113] D1. A computer storage medium, characterized in that it is used to store computer software instructions used for the multi-channel neural network model training method described in any one of claims 1 to 5 above, which includes a program designed for executing the above-mentioned aspect for the multi-channel neural network model training method.

Claims

1. A multi-channel neural network model training method, characterized in that: include: Determine the training data; Input the training data into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures, and the convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules, and after the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module to perform a corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, where N is an integer greater than 2; The training data includes a training image set of classified images, and the target model is an image classification model.

2. The method according to claim 1, wherein The convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including: For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

3. The method according to claim 1, wherein The convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including: For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

4. The method according to claim 1, wherein The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, including: For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

5. A multi-channel neural network model training device, characterized in that: include: a determination unit, configured to determine training data; A model training unit is used to input the training data into an N-channel neural network model for learning to obtain a trained target model, wherein the N-channel neural network model includes a residual convolutional neural network ResNet channel and N-1 densely connected convolutional neural network DenseNet channels with different connection structures. The convolution layer of each channel in the N-1 DenseNet channels is densely connected according to the corresponding number of interval layers. The preset N-channel neural network model includes a preset number of learning modules. After the data output by the previous learning module corresponding to each learning module is divided into N data, one data enters the ResNet channel in the learning module for corresponding element addition operation, and the remaining N-1 data respectively enter the corresponding N-1 DenseNet channels for respective merging operations. The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, where N is an integer greater than 2; The training data includes a training image set of classified images, and the target model is an image classification model.

6. The device according to claim 5, characterized in that The convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including: For the convolutional layers contained in each DenseNet channel in the N-1 DenseNet channels, segment them according to the number of interval layers corresponding to the channel, and merge the input data corresponding to each convolutional layer in each segment with the output data of all convolutional layers before the convolutional layer in the segment. The number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

7. The device according to claim 5, characterized in that The convolutional layers of each channel in the N-1 DenseNet channels are densely connected according to their corresponding number of interval layers, including: For each convolutional layer contained in each DenseNet channel in the N-1 DenseNet channels, the output data of the convolutional layer is merged into the input data corresponding to each of the P convolutional layers after the convolutional layer, where P is the number of interval layers corresponding to the DenseNet channel, and the number of interval layers corresponding to the N-1 DenseNet channels is different from each other.

8. The device according to claim 5, wherein The data output by the ResNet channel and the N-1 DenseNet channels are processed by the convolution layer and then input into the next learning module, including: For the output data of each channel, a 1*1 convolution operation is performed on each of them, and then the corresponding bits are added to obtain the first output data. After performing 3*3 convolution and 1*1 convolution operations on the first output data in sequence, the output data corresponding to the learning module is obtained, and the output data is input into the next learning module.

9. A multi-channel neural network model training device, characterized in that: Including processor and memory: The memory is used to store a script program for executing the method according to any one of claims 1 to 4; The processor is configured to execute the program stored in the memory.

10. A computer storage medium, characterized in that Used to store computer software instructions used for the multi-channel neural network model training method described in any one of claims 1 to 4 above, which includes a program designed for executing the multi-channel neural network model training method described in the above aspect.

Images