Low-power-consumption multi-agent cooperative enhancement method

Abstract

The invention discloses a low-power-consumption multi-agent cooperative enhancement method. The method comprises the following steps: acquiring an initial task and an initial model for processing the initial task; dividing the initial task to obtain a plurality of divided sub-initial tasks; based on the plurality of sub-initial tasks, the initial model is divided to obtain a plurality of sub-initial models corresponding to the plurality of sub-initial tasks, and each sub-initial model comprises a plurality of calculation layers. The technical problems that in the prior art, due to the fact that data moves back and forth between a memory and a calculation core in the calculation process, the generated parameter quantity is large, the consumption of a battery is large, the use efficiency of the battery is low, and the calculable life cycle of a low-power-consumption agent is short are solved.

Description

Technical Field

[0001] The present invention relates to the field of computer data optimization, and in particular to a low-power multi-agent collaborative enhancement method. Background Art

[0002] Low-power intelligent agents are an important perception unit. They play an important role in the current diversified and complex perception environment with rich perception information content, longer perception duration, and demanding perception environment. Low-power intelligent agents are usually powered by batteries and include low-power processing chips and peripheral circuits designed specifically for energy consumption optimization. They can carry some intelligent AI applications for intelligent computing and run continuously with less energy consumption in application scenarios that require long-term continuous perception or in usage scenarios with energy consumption constraints, providing users with intelligent services and solutions.

[0003] Embedded low-power intelligent agents usually use low-power processors and components to reduce the energy consumption of system operation, but this also limits the computing power of the device and faces the problem of insufficient resources: (1) Limited processor performance: Embedded devices usually use low-power, low-cost processors, which have relatively low performance and may not be able to handle complex computing tasks. (2) Limited memory: Embedded devices usually have limited memory capacity, which may limit the number of applications they can run simultaneously or their ability to run large applications. (3) Environmental limitations: Embedded devices are often used in various environments, such as industrial control, automobiles, medical equipment, etc. These environments may have special requirements for device performance and power consumption, further limiting the availability of computing resources.

[0004] The basic idea to solve the current problems is to optimize computing and manage resources, select appropriate hardware and software components, and balance computing requirements, power consumption, and cost in the design. At the same time, the introduction of new technologies and solutions (such as deep learning accelerators, more efficient algorithms, etc.) may also help improve the computing performance of embedded devices. The core of the problem is that although the energy consumption required for computing itself can be gradually reduced through new computing methods, the mode determined by the centralized information processing method of the von Neumann architecture - moving data from memory to processor and back, and the energy consumption generated by the movement of sensor data in the system is the main reason.

[0005] Therefore, we proposed an inventive design, named a low-power multi-agent collaborative enhancement method. We use one / multiple groups of low-power intelligent hardware to form a swarm agent cluster. Through the swarm agent task coordination planning and memory reuse computing optimization algorithm, we reduce the battery consumption of a single agent during the computing process and improve the battery utilization efficiency, thereby extending the computing life cycle of the low-power agent. Summary of the invention

[0006] An embodiment of the present invention provides a low-power multi-agent collaborative enhancement method to at least solve the technical problem in the prior art that the movement of data back and forth between the memory and the computing core during the calculation process generates a large number of parameters, consumes a lot of battery, and has low battery utilization efficiency, thereby shortening the computing life cycle of the low-power agent.

[0007] According to one aspect of an embodiment of the present invention, a low-power multi-agent collaborative enhancement method is provided. The method may include: obtaining an initial task and processing an initial model of the initial task;

[0008] The initial task is divided to obtain a plurality of sub-initial tasks after the division; based on the plurality of sub-initial tasks, the initial model is divided to obtain a plurality of sub-initial models corresponding to the plurality of sub-initial tasks, wherein each sub-initial model includes a plurality of computing layers; in the process of each computing layer in the plurality of sub-initial models calculating the plurality of sub-initial tasks, a plurality of feature maps output by each computing layer are determined, wherein each feature map output by each computing layer includes a plurality of sub-feature maps, and each feature map output by each computing layer includes an abnormal situation part; after obtaining each feature map output by each computing layer, based on each feature map output by each computing layer feature map, determine the set of each feature map corresponding to each layer; calculate the position information of each feature map output by each computing layer based on the size of each feature map output by each computing layer of multiple sub-feature maps in each feature map output by each computing layer; determine the size of the abnormal part of each feature map based on the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map; determine the size of the normal part of all feature maps corresponding to each computing layer based on the set of each feature map corresponding to each layer and the size of the abnormal part of each feature map.

[0009] Optionally, the method further includes: a general expression of a process in which the initial task model processes the initial task is:

[0010]

[0011] 0≤c <N,0≤z<M,0≤x,y<E,E=(H-R+U) / U

[0012] in, Represents the output of the current computing layer when the processing initial task model processes the initial task. represents the input of the current computing layer when the initial task model processes the initial task, x, y, F and B respectively represent the matrices of the input, output, filter and bias parameters of the current computing layer data when the initial task model processes the initial task, k, i, j represent the number of layers of the current computing layer when the initial task model processes the initial task; the value ranges of k, i, j are 0~C-1, 0~R-1, 0~R-1 respectively, parameter c is the number of channels of the output data of the current computing layer when the initial task model processes the initial task; M and N are the filters and output data of the current computing layer when the initial task model processes the initial task; H, E and R respectively represent the sizes of the input, output data and filters of the current computing layer data when the initial task model processes the initial task, U represents the given step size of the current computing layer when the initial task model processes the initial task, and z is the value in the matrix of bias parameters.

[0013] Optionally, based on multiple sub-initial tasks, the initial model is divided to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks, including: based on the number of channels of the output feature map of the multiple sub-initial tasks at any layer or the number of layers of the multiple sub-initial tasks at any layer, the initial model is divided to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks.

[0014] Optionally, based on the number of channels of the output feature maps of the multiple sub-initial tasks at any layer or the number of layers of the multiple sub-initial tasks at any layer, the initial model is divided to obtain expressions of the multiple sub-initial models corresponding to the multiple sub-initial tasks:

[0015]

[0016] in, and Feature maps representing input data and output data; Indicates its jth th channels; Indicates the i th Channels; W i,j is the weight matrix; the symbol * represents the 2D convolution operation, B i Represents the i-th bias parameter, and the value range of j is 1 to e.

[0017] Optionally, in the process of each computing layer in the multiple sub-initialization models calculating the multiple sub-initialization tasks, an expression of each feature map of the multiple feature maps output by each computing layer is determined:

[0018]

[0019] in, It is the width and height of each feature map after each sub-initial model of each computing layer processes each initial task. n represents the number of sub-initial models, and the value range of n is 1 to l.

[0020] Optionally, after obtaining each feature map output by each computing layer, based on each feature map output by each computing layer, an expression for determining a set of each feature map corresponding to each layer is as follows:

[0021]

[0022] Among them, τ total Represents the set of each feature map corresponding to each layer.

[0023] Optionally, based on the position information of each feature map output by each computing layer of a plurality of sub-feature maps in each feature map, determining the size of the abnormal situation portion in each feature map includes:

[0024]

[0025] Among them, min and max are functions for taking the minimum and maximum values ​​respectively. Indicates the length of the abnormal part of the previous calculation layer, represents the width of the abnormal part of the current calculation layer, v represents the length of the abnormal part of each feature map, and h represents the width of the abnormal part of each feature map. Indicates the total size of the abnormal part of each feature map.

[0026] Optionally, based on the set of each feature map corresponding to each layer and the size of the abnormal part in each feature map, the expression for determining the size of the normal part of all feature maps corresponding to each calculation layer is:

[0027]

[0028] Among them, τ actual Represents the size of the normal part of all feature maps corresponding to each computational layer, Indicates the size of the abnormal part of each feature map removed from the set of each feature map corresponding to each layer.

[0029] Beneficial effects of the present invention:

[0030] The present invention proposes a low-power multi-agent collaborative enhancement method, which uses multiple intelligent devices to optimize computing energy consumption for the same task. The energy consumption optimization process can be divided into two stages: (1) splitting complex tasks so that the tasks can be adapted to be calculated on multiple agents; (2) optimizing the timing of loading tasks into memory, and jointly enhancing low-power group agents. By decomposing multiple tasks into fragments and executing them in a merged manner, AI models with larger sizes that were originally impossible to execute can be calculated, and a conclusion is derived, which solves the technical problem in the prior art that the amount of parameters generated by the movement of data back and forth between the memory and the computing core during the calculation process is large, the battery consumption is large, the battery utilization efficiency is low, and thus the low-power agent can have a short computing life cycle. The technical effect of using one or more groups of low-power intelligent hardware to form a group agent cluster during the calculation process, the group agent task overall planning and memory reuse computing optimization algorithm reduce the battery consumption of a single agent during the calculation process, improve the battery utilization efficiency, and thus make the low-power agent have a longer computing life cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0032] Figure 1 This is a flow chart of a low-power multi-agent collaborative enhancement method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only embodiments of a part of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0035] Example 1

[0036] According to an embodiment of the present invention, a low-power multi-agent collaborative enhancement method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system containing at least one set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0037] Figure 1 is a flow chart of a low-power multi-agent collaborative enhancement method according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps:

[0038] Step S101, obtaining an initial task and an initial model for processing the initial task.

[0039] In the technical solution provided in the above step S101 of the present invention, for example, the initial task may be an image classification task or an image segmentation task, etc., and the initial model for processing the initial task may be a neural network model for processing the initial task.

[0040] Step S102, dividing the initial task to obtain a plurality of sub-initial tasks.

[0041] In the technical solution provided in the above step S102 of the present invention, when the initial task is to classify an image, for example, an image is an animal, the head and body of the animal are separated, the head is a sub-initial task, and the body is a sub-initial task.

[0042] Step S103, based on the multiple sub-initial tasks, the initial model is divided to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks, wherein each sub-initial model includes multiple computing layers.

[0043] In the technical solution provided in the above step S103 of the present invention, for example, the head of the animal corresponds to a sub-initial model, the body of the animal corresponds to a sub-initial model, and each sub-initial model includes multiple computing layers.

[0044] Step S104, in the process of each computing layer in the multiple sub-initial models calculating the multiple sub-initial tasks, determine the multiple feature maps output by each computing layer, wherein each feature map output by each computing layer includes multiple sub-feature maps, and each feature map output by each computing layer includes an abnormal situation part.

[0045] In the technical solution provided in the above step S104 of the present invention, in the process of each computing layer in the multiple sub-initial models calculating the multiple sub-initial tasks, multiple feature maps output by each computing layer are output, and each feature map includes a normal situation part and an abnormal situation part.

[0046] Step S105, after obtaining each feature map output by each computing layer, determine a set of each feature map corresponding to each layer based on each feature map output by each computing layer.

[0047] In the technical solution provided in the above step S105 of the present invention, after obtaining each feature map output by each computing layer, each feature map output by each computing layer is put together to obtain a set of corresponding feature maps of each layer.

[0048] Step S106, based on the sizes of each feature map output by each computing layer and the multiple sub-feature maps in each feature map output by each computing layer, calculate the position information of each feature map output by each computing layer and the multiple sub-feature maps in each feature map output by each computing layer.

[0049] In the technical solution provided in the above step S106 of the present invention, according to the size of each feature map output by each computing layer of multiple sub-feature maps in each feature map output by each computing layer, the expression for calculating the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map is as follows:

[0050]

[0051] in, Represents the position information of each feature map output by any sub-feature map in each computing layer, Represents the horizontal coordinate of the position of each feature map output by any sub-feature map in each computing layer, The vertical coordinate representing the position of any sub-feature map in each feature map output by each computing layer.

[0052] Step S107, based on the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map, determine the size of the abnormal part of each feature map.

[0053] In the technical solution provided in the above step S107 of the present invention, the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map is calculated to obtain the size of the abnormal part in each feature map.

[0054] Step S108, based on the set of each feature map corresponding to each layer and the size of the abnormal part in each feature map, determine the size of the normal part of all feature maps corresponding to each calculation layer.

[0055] In the technical solution provided in the above step S108 of the present invention, the set of each feature map corresponding to each layer and the size of the abnormal part in each feature map are calculated to obtain the size of the normal part of all feature maps corresponding to each calculation layer.

[0056] The above method of this embodiment is further introduced below.

[0057] As an optional embodiment, the method further includes: the general expression of the process of the initial task model processing the initial task is:

[0058]

[0059] 0≤c <N,0≤z<M,0≤x,y<E,E=(H-R+U) / U

[0060] in, Represents the output of the current computing layer when the processing initial task model processes the initial task. represents the input of the current computing layer when the initial task model processes the initial task, x, y, F and B respectively represent the matrices of the input, output, filter and bias parameters of the current computing layer data when the initial task model processes the initial task, k, i, j represent the number of layers of the current computing layer when the initial task model processes the initial task; the value ranges of k, i, j are 0~C-1, 0~R-1, 0~R-1 respectively, parameter c is the number of channels of the output data of the current computing layer when the initial task model processes the initial task; M and N are the filters and output data of the current computing layer when the initial task model processes the initial task; H, E and R respectively represent the sizes of the input, output data and filters of the current computing layer data when the initial task model processes the initial task, U represents the given step size of the current computing layer when the initial task model processes the initial task, and z is the value in the matrix of bias parameters.

[0061] In this embodiment, the initial task is processed by the above formula to obtain the output of the initial task.

[0062] As an optional implementation method, step S103, based on multiple sub-initial tasks, divides the initial model to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks, including: based on the number of channels of the output feature map of the multiple sub-initial tasks at any layer or the number of layers of the multiple sub-initial tasks at any layer, divides the initial model to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks.

[0063] In this embodiment, the initial model can be divided according to the number of channels of the output feature maps of multiple sub-initial tasks at any layer or the number of layers of multiple sub-initial tasks at any layer to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks.

[0064] As an optional implementation method, based on the number of channels of the output feature map of multiple sub-initial tasks at any layer or the number of layers of multiple sub-initial tasks at any layer, the initial model is divided to obtain the expressions of multiple sub-initial models corresponding to the multiple sub-initial tasks:

[0065]

[0066] in, and Feature maps representing input data and output data; represents its j-th channel; represents the i-th channel; W i,j is the weight matrix; the symbol * represents the 2D convolution operation, B i Represents the i-th bias parameter, and the value range of j is 1 to e.

[0067] In this embodiment, the initial model is divided by the above expression to obtain multiple sub-initial models corresponding to multiple sub-initial tasks. For example, a neural network model is divided into 5 sub-initial models. This is only an example and is not specifically limited.

[0068] As an optional implementation manner, step S104, in the process of each computing layer in the multiple sub-initial models calculating the multiple sub-initial tasks, determines the expression of each feature map of the multiple feature maps output by each computing layer:

[0069]

[0070] in, It is the width and height of each feature map after each sub-initial model of each computing layer processes each initial task. n represents the number of sub-initial models, and the value range of n is 1 to l.

[0071] In this embodiment, the width and height of each feature map after each sub-initial model of each computing layer processes each initial task are calculated by the above expressions. It should be noted that k is the label of each feature map of multiple feature maps output by each computing layer, and k corresponds to l. For example, the first computing layer corresponds to 3 initial sub-models, and each sub-model corresponds to a feature map output by each sub-model. This is only an example and is not specifically limited.

[0072] As an optional implementation method, in step S105, after obtaining each feature map output by each computing layer, based on each feature map output by each computing layer, the expression of the set of each feature map corresponding to each layer is determined as follows: including:

[0073]

[0074] Among them, τ total Represents the set of each feature map corresponding to each layer.

[0075] In this embodiment, the set of each feature map corresponding to each layer is calculated by the above formula.

[0076] As an optional implementation manner, step S107, based on the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map, determines the size of the abnormal situation part in each feature map, including:

[0077]

[0078] Among them, min and max are functions for taking the minimum and maximum values ​​respectively. Indicates the length of the abnormal part of the previous calculation layer, represents the width of the abnormal part of the current calculation layer, v represents the length of the abnormal part of each feature map, and h represents the width of the abnormal part of each feature map. Indicates the total size of the abnormal part of each feature map.

[0079] In this embodiment, the size of the abnormal part in each feature map is obtained by calculating the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map. It should be noted that if the first sub-feature map is stored in the second sub-feature map, its coordinates meet such a range. As shown in the following formula:

[0080]

[0081] And in the convolution calculation process, the values ​​of the horizontal and vertical coordinates of the convolution kernel need to satisfy This proves that the coordinates of the current sub-feature map are abnormal and should be removed from the overall feature map. The x-coordinate of the convolution kernel must be expressed as The x-coordinate must be included in the feature map Within the range, the y coordinate of the convolution kernel Not a feature map Then the intersection is performed, that is, the convolution across two feature maps, which is the abnormal part.

[0082] As an optional implementation manner, in step S108, based on the set of each feature map corresponding to each layer and the size of the abnormal part in each feature map, the expression for determining the size of the normal part of all feature maps corresponding to each calculation layer is:

[0083]

[0084] Among them, τ actua l represents the size of the normal part of all feature maps corresponding to each computational layer, Indicates the size of the abnormal part of each feature map removed from the set of each feature map corresponding to each layer.

[0085] In this embodiment, the set of each feature map corresponding to each layer removes the size of the abnormal part in each feature map, and obtains the size of the normal part of all feature maps corresponding to each calculation layer.

[0086] In an embodiment of the present invention, an initial task and an initial model for processing the initial task are obtained; the initial task is divided to obtain a plurality of sub-initial tasks after division; based on the plurality of sub-initial tasks, the initial model is divided to obtain a plurality of sub-initial models corresponding to the plurality of sub-initial tasks, wherein each sub-initial model includes a plurality of computing layers; in the process of each computing layer in the plurality of sub-initial models calculating the plurality of sub-initial tasks, a plurality of feature maps output by each computing layer are determined, wherein each feature map output by each computing layer includes a plurality of sub-feature maps, and each feature map output by each computing layer includes an abnormal situation part; after obtaining each feature map output by each computing layer, based on each feature map output by each computing layer, a set of each feature map corresponding to each layer is determined; based on the size of each feature map output by each computing layer, the size of each feature map output by each computing layer is calculated. The position information of each feature map output by the computing layer; based on the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map, the size of the abnormal part of each feature map is determined; based on the set of each feature map corresponding to each layer and the size of the abnormal part of each feature map, the size of the normal part of all feature maps corresponding to each computing layer is determined, which solves the technical problems in the prior art that the amount of parameters generated by the movement of data back and forth between the memory and the computing core during the calculation process is large, the battery consumption is large, the battery utilization efficiency is low, and thus the computational life cycle of the low-power intelligent agent is short. The technical effect of reducing the battery consumption of a single intelligent agent during the calculation process by using one or more groups of low-power intelligent hardware to form a group intelligent agent cluster, the group intelligent agent task overall planning and memory reuse computing optimization algorithm are used to improve the battery utilization efficiency, thereby making the computational life cycle of the low-power intelligent agent longer.

[0087] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0088] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0089] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0090] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0091] In addition, each functional unit in each embodiment of the present invention may be integrated into a first processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0092] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A low-power multi-agent collaborative enhancement method, characterized in that: include: Obtaining initial tasks and initial models for processing the initial tasks; Divide the initial task into multiple sub-initial tasks; Based on the multiple sub-initial tasks, the initial model is divided to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks, wherein each sub-initial model includes multiple computing layers; In the process of each computing layer in the multiple sub-initialization models calculating the multiple sub-initialization tasks, multiple feature maps output by each computing layer are determined, wherein each feature map output by each computing layer includes multiple sub-feature maps, and each feature map output by each computing layer includes an abnormal situation part; After obtaining each feature map output by each computing layer, based on each feature map output by each computing layer, determine the set of each feature map corresponding to each layer; Based on the sizes of each feature map output by each computing layer and the multiple sub-feature maps in each feature map output by each computing layer, position information of each feature map output by each computing layer and the multiple sub-feature maps in each feature map output by each computing layer is calculated; Determine the size of the abnormal part in each feature map based on the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map; Based on the set of each feature map corresponding to each layer and the size of the abnormal part in each feature map, the size of the normal part of all feature maps corresponding to each calculation layer is determined.

2. The method according to claim 1, characterized in that The method further includes: the general expression of the process of the initial task model processing the initial task is: 0≤c <N,0≤z<M,0≤x,y<E,E=(H-R+U) / U in, Represents the output of the current computing layer when the processing initial task model processes the initial task. represents the input of the current computing layer when the initial task model processes the initial task, x, y, F and B respectively represent the matrices of the input, output, filter and bias parameters of the current computing layer data when the initial task model processes the initial task, k, i, j represent the number of layers of the current computing layer when the initial task model processes the initial task; the value ranges of k, i, j are 0~C-1, 0~R-1, 0~R-1 respectively, parameter c is the number of channels of the output data of the current computing layer when the initial task model processes the initial task; M and N are the filters and output data of the current computing layer when the initial task model processes the initial task; H, E and R respectively represent the sizes of the input, output data and filters of the current computing layer data when the initial task model processes the initial task, U represents the given step size of the current computing layer when the initial task model processes the initial task, and z is the value in the matrix of bias parameters.

3. The method according to claim 1, characterized in that Based on multiple sub-initial tasks, the initial model is divided to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks, including: Based on the number of channels of the output feature maps of the multiple sub-initial tasks at any layer or the number of layers of the multiple sub-initial tasks at any layer, the initial model is divided to obtain multiple sub-initial models corresponding to the multiple sub-initial tasks.

4. The method according to claim 3, characterized in that Based on the number of channels of the output feature maps of multiple sub-initial tasks at any layer or the number of layers of multiple sub-initial tasks at any layer, the initial model is divided to obtain the expressions of multiple sub-initial models corresponding to the multiple sub-initial tasks: in, and Feature maps representing input data and output data; Indicates its jth th channels; Indicates the i th Channels; W i,j is the weight matrix; the symbol * represents the 2D convolution operation, B i Represents the i-th bias parameter, and the value range of j is 1 to e.

5. The method according to claim 1, characterized in that In the process of each computing layer in the multiple sub-initialization models calculating the multiple sub-initialization tasks, the expression of each feature map of the multiple feature maps output by each computing layer is determined: in, It is the width and height of each feature map after each sub-initial model of each computing layer processes each initial task. n represents the number of sub-initial models, and the value range of n is 1 to l.

6. The method according to claim 1, characterized in that After obtaining each feature map output by each computing layer, based on each feature map output by each computing layer, the expression for determining the set of each feature map corresponding to each layer is: including: Among them, τ total Represents the set of feature maps corresponding to each layer.

7. The method according to claim 1, characterized in that Based on the position information of each feature map output by each computing layer of multiple sub-feature maps in each feature map, the size of the abnormal part in each feature map is determined, including: Among them, min and max are functions for taking the minimum and maximum values ​​respectively. Indicates the length of the abnormal part of the previous calculation layer, represents the width of the abnormal part of the current calculation layer, v represents the length of the abnormal part of each feature map, and h represents the width of the abnormal part of each feature map. Indicates the total size of the abnormal part of each feature map.

8. The method according to claim 1, characterized in that: Based on the set of each feature map corresponding to each layer and the size of the abnormal part in each feature map, the expression for determining the size of the normal part of all feature maps corresponding to each calculation layer is: Among them, τ actual Represents the size of the normal part of all feature maps corresponding to each computational layer, Indicates the size of the abnormal part of each feature map removed from the set of each feature map corresponding to each layer.

9. A processor, characterized in that: The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 8 when running.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 8.

Images