A lightweight method and system for mobile terminal models of artificial intelligence training platform

Abstract

The present invention discloses a method and system for lightweighting a model on a mobile terminal of an artificial intelligence training platform, the method comprising: obtaining a first business model on the web terminal of the artificial intelligence training platform; performing static lightweight processing based on the first business model to obtain a second business model, the static lightweight processing representing a lightweight operation of a fixed model structure that does not change with input data; performing dynamic lightweight processing based on the second business model to obtain a third business model, the dynamic lightweight processing representing a lightweight operation of selectively shielding a model forward transfer parameter that changes with input data. The present invention develops a combination of static lightweight and dynamic lightweight technologies for an electric power artificial intelligence technology training platform, realizes the operation of multiple business models on the mobile terminal of the electric power artificial intelligence technology training platform, reduces the pressure on the cloud side of the training platform, and enables users to train and learn anytime and anywhere.

Description

Technical Field

[0001] The present invention relates to the technical field of model lightweighting, and in particular to a method and system for lightweighting a mobile terminal model of an artificial intelligence training platform. Background Art

[0002] The development of the artificial intelligence industry has entered a core period, posing great challenges to the application of the power artificial intelligence industry. The actual application of artificial intelligence in the power industry system and industrial chain is still very limited. It is necessary to start from the problem and create value oriented to problem solving. Therefore, quickly finding more artificial intelligence application scenarios in the power industry to meet the growing expectations of the market and capital for it is the main driving force for the industrial upgrading and economic transformation of our power industry. At present, the power artificial intelligence capabilities lack effective support for industrial promotion and large-scale application. In order to better serve the emerging artificial intelligence industry in the power industry, the present invention develops the WEB and mobile terminals of the artificial intelligence training platform to provide power employees with artificial intelligence technology training services anytime and anywhere. Summary of the invention

[0003] The purpose of the embodiment of the present invention is to provide a method and system for lightweighting the mobile terminal model of an artificial intelligence training platform, effectively reducing the cloud pressure of the training platform, and enabling users to train and learn anytime and anywhere. The technical solution includes:

[0004] In a first aspect, a method for lightweighting a mobile terminal model of an artificial intelligence training platform is provided, the method comprising the following steps:

[0005] Obtain the first business model of the artificial intelligence training platform web terminal;

[0006] Performing static lightweight processing based on the first business model to obtain a second business model, wherein the static lightweight processing represents a lightweight operation with a fixed model structure that does not change with input data;

[0007] Dynamic lightweight processing is performed based on the second business model to obtain a third business model, wherein the dynamic lightweight processing represents a lightweight operation of selectively shielding a model forward transfer parameter that changes with input data.

[0008] In some implementations, performing static lightweight processing based on the first service model to obtain a second service model includes:

[0009] Based on the first business model, identifying and classifying the convolution kernel in each network layer in the first business model under the first classification index and the second classification index;

[0010] Based on the conditions that the convolution kernel satisfies both the first classification index and the second classification index, the convolution kernel is divided into a definite unimportant convolution kernel, a fuzzy unimportant convolution kernel and a definite important convolution kernel;

[0011] Based on the determined unimportant convolution kernels in each network layer, a determined unimportant structure is formed, and the determined unimportant structure in the first service model is removed to obtain a fourth service model;

[0012] For the fuzzy unimportant convolution kernels and the determined important convolution kernels in each network layer in the fourth business model, the pruning strategy for the fourth business model is determined according to a preset pruning strategy detection method, and the second business model is obtained based on the pruning of the fourth business model.

[0013] In some embodiments, the method of dividing the convolution kernel into a determined unimportant convolution kernel, a fuzzy unimportant convolution kernel, and a determined important convolution kernel based on the conditions that the convolution kernel satisfies both the first classification index and the second classification index comprises:

[0014] According to the value of the first classification index from large to small, the convolution kernels in each network layer are sorted in a first manner, and according to the value of the second classification index from large to small, the convolution kernels in each network layer are sorted in a second manner;

[0015] The convolution kernels in the intersection of the first preset number of convolution kernels in the first sort and the first preset number of convolution kernels in the second sort are used to determine the important convolution kernels; the convolution kernels in the intersection of the second preset number of convolution kernels in the first sort and the second preset number of convolution kernels in the second sort are used to determine the unimportant convolution kernels; the remaining convolution kernels in the convolution kernels in each network layer are used as blurred unimportant convolution kernels.

[0016] In some implementations, determining the pruning strategy for the fourth service model according to a preset pruning strategy detection method includes:

[0017] (1) selecting at least one channel as an invalid channel for each network layer in the fourth service model using a preset method, and determining the invalid channels and valid channels in the network layer;

[0018] (2) inputting the preset model training samples into the fourth business model, transmitting information through all channels of each network layer, obtaining the first output of the fourth business model, transmitting information through the effective channels of each network layer, and obtaining the second output of the fifth business model;

[0019] (3) determining whether to reselect the invalid channel selected by the preset method based on the deviation between the accuracy of the first output and the accuracy of the second output;

[0020] (4) When it is determined that the invalid channel needs to be reselected, the parameters of the preset method are adjusted based on the deviation and the process goes to step (1); when it is determined that the invalid channel does not need to be reselected, the invalid channels of each network layer in the fourth business model are deleted to obtain the second business model.

[0021] In some implementations, selecting at least one channel as an invalid channel for each network layer in the fourth service model using a preset method includes:

[0022] For each network layer, a ratio is selected from a plurality of preset ratio parameters as a network layer invalid channel ratio parameter, and channels of a corresponding ratio in the network layer are determined as invalid channels based on the invalid channel ratio parameter.

[0023] In some embodiments, adjusting the parameters of the preset method based on the deviation when determining that an invalid channel needs to be reselected includes:

[0024] Based on the optimization goal that the preset ratio parameter value of invalid channels of each network layer in the fourth service model is larger and the deviation value is smaller, the parameters of the preset method are adjusted; the optimization objective function max(F) is:

[0025] max(F)=max[(1+α(Y 5 -Y 4 ))*βR+Y 4 ], where α, β are preset adjustment parameters, Y 5 is the accuracy of the second output of the fifth business model, Y 4 is the accuracy of the first output of the fourth business model, and R is the sum of the preset proportion parameter values ​​of the invalid channels of each network layer in the fourth business model.

[0026] In some implementations, the performing dynamic lightweight processing based on the second business model to obtain a third business model includes:

[0027] Receiving a feature map output by a convolution kernel in a current network layer of a second service model;

[0028] Selecting a pixel position feature value to be masked in the feature map according to a preset feature value masking threshold;

[0029] The feature map carrying the pixel position shielding mark is input into the next network layer for calculation, wherein the feature value of the pixel position of the feature map carrying the shielding mark does not participate in the calculation of the next network layer.

[0030] In the second aspect, a lightweight system for mobile terminal models of an artificial intelligence training platform is provided, and the system includes:

[0031] A web-side model receiving unit, used to obtain a first business model of the artificial intelligence training platform web-side;

[0032] A first lightweight processing unit, configured to perform static lightweight processing based on the first business model to obtain a second business model, wherein the static lightweight processing represents a lightweight operation with a fixed model structure that does not change with input data;

[0033] The second lightweight processing unit is used to perform dynamic lightweight processing based on the second business model to obtain a third business model, wherein the dynamic lightweight processing represents a lightweight operation of selectively shielding a model forward transfer parameter that changes with input data.

[0034] According to a third aspect, an electronic device is provided, the electronic device comprising:

[0035] processor;

[0036] a memory for storing processor-executable instructions;

[0037] Among them, the processor runs the executable instructions to implement the lightweight method of the mobile terminal model of the artificial intelligence training platform as described in the first aspect.

[0038] In a fourth aspect, a computer-readable storage medium is provided, on which computer instructions are stored, characterized in that when the instructions are executed by a processor, the steps of the method for lightweighting the mobile terminal model of the artificial intelligence training platform as described in the first aspect are implemented.

[0039] The lightweight method and system of the mobile terminal model of an artificial intelligence training platform of the present invention have the following beneficial effects: through the combination of static lightweight and dynamic lightweight technology, the mobile terminal operation of business models such as power transmission, substation, and safety supervision of the electric power artificial intelligence technology training platform is realized, the cloud pressure of the training platform is reduced, and mobile AI training course learning, drills, and scene building are realized, so that users can practice and learn anytime and anywhere. The lightweighting of the model adopts a combination of static lightweight and dynamic lightweight. Based on static lightweight, the unimportant structural components of the first business model are removed, and the second business model is further operated by dynamic lightweight according to the model input data, so as to achieve the accuracy and effectiveness of the lightweight processing of the first business model, and at the same time, the reliability of business models such as power transmission, substation, and safety supervision of the mobile terminal of the artificial intelligence training platform is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flow chart of a lightweight method for a mobile terminal model of an artificial intelligence training platform in an embodiment of the present application;

[0041] Figure 2This is a schematic diagram of an implementation process of static lightweight processing of the first business model in an embodiment of the present application;

[0042] Figure 3 This is another implementation flow diagram of static lightweight processing of the first business model in an embodiment of the present application;

[0043] Figure 4 It is a schematic diagram of the lightweight system structure of the mobile terminal model of the artificial intelligence training platform in the embodiment of the present application. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] See also Figure 1 , a method for lightweighting a mobile terminal model of an artificial intelligence training platform provided in an embodiment of the present application includes:

[0046] Step 1, obtain the first business model of the artificial intelligence training platform web terminal;

[0047] Step 2: Perform static lightweight processing based on the first business model to obtain a second business model, wherein the static lightweight processing represents a lightweight operation with a fixed model structure that does not change with input data;

[0048] Step 3: Perform dynamic lightweight processing based on the second business model to obtain a third business model, wherein the dynamic lightweight processing represents a lightweight operation of selectively shielding model forward transfer parameters that change with input data.

[0049] In the embodiments of the present application, a static lightweight and dynamic lightweight combination technology is developed for the electric power artificial intelligence technology training platform to realize the mobile operation of business models such as power transmission, substation, and safety supervision of the electric power artificial intelligence technology training platform, reduce the cloud pressure of the training platform, and realize mobile AI training course learning, drills, and scenario building, so that users can train and learn anytime and anywhere.

[0050] For the lightweighting of the model, a combination of static lightweighting and dynamic lightweighting is adopted. Based on static lightweighting, unimportant structural components of the determinism of the first business model are removed, and the second business model is further run by dynamic lightweighting according to the model input data. The accuracy and effectiveness of the lightweight processing of the first business model are achieved, and at the same time, the reliability of business models such as power transmission, substation, and safety supervision on the mobile terminal of the artificial intelligence training platform is achieved.

[0051] In addition, in the present application, static lightweight processing is performed on the first business model. Based on the preset static lightweight strategy, a global lightweight operation is directly performed on the first business model, rather than performing a lightweight operation on a certain layer of the model and then training and fine-tuning the model before performing a lightweight operation on the next layer of the model. This avoids multiple iterations of lightweight processing and training fine-tuning to complete the lightweight processing of the entire model, shortens the time consumption of lightweight processing in this step, and improves the efficiency of lightweight processing.

[0052] It can be understood that the model structure of the second business model obtained after static lightweight processing is more lightweight than the first business model. In dynamic lightweight processing, the model structure of the third business model is consistent with that of the second business model. Dynamic lightweight processing is to shield the transmission of certain information during the model operation process after the input data, and does not modify the model structure. That is to say, in dynamic lightweight processing, under different input data, the second business model selects different information transmission parameters to be shielded, that is, a third business model with different information transmission process in the model is obtained.

[0053] For further information, see Figure 2 In the above step 2, static lightweight processing is performed based on the first business model to obtain a second business model, including:

[0054] Step 21, based on the first business model, identifying and classifying the convolution kernels in each network layer in the first business model under the first classification index and the second classification index;

[0055] Step 22, based on the conditions that the convolution kernel satisfies both the first classification index and the second classification index, the convolution kernel is divided into a determined unimportant convolution kernel, a fuzzy unimportant convolution kernel and a determined important convolution kernel;

[0056] Step 23, forming a determined unimportant structure based on the determined unimportant convolution kernels in each network layer, and removing the determined unimportant structure in the first service model to obtain a fourth service model;

[0057] Step 24, for the fuzzy unimportant convolution kernels and the determined important convolution kernels in each network layer in the fourth business model, determine the pruning strategy for the fourth business model according to a preset pruning strategy detection method, and obtain the second business model based on the pruning of the fourth business model.

[0058] In an embodiment of the present application, the first classification index may be the importance of information transmitted by the convolution kernel channel during the forward information transmission process of the first business model, and the second classification index may be the importance of feedback information of the convolution kernel channel during the reverse information feedback process of the first business model. In an embodiment of the present application, the importance of the convolution kernel is not only analyzed from the forward transmission process of the business model, but also the importance of the feedback information in the reverse feedback information process is combined to measure whether to prune the convolution kernel channel. The importance of the convolution kernel in the first business model is analyzed from the two classification indicators, and the two analysis results are combined to obtain a deterministic unimportant convolution kernel as the convolution kernel to be pruned, thereby obtaining a fourth business model. Further, the second business model is obtained after determining the pruning strategy for the fourth business model. It can be understood that the pruning strategy for the fourth business model is a further pruning process after the previous pruning, which achieves the maximum static lightweight processing while ensuring the performance of the mobile terminal model, thereby improving the lightweight degree of the mobile terminal business model.

[0059] Specifically, in the above step 22, the convolution kernel is divided into a determined unimportant convolution kernel, a fuzzy unimportant convolution kernel and a determined important convolution kernel based on the conditions that the convolution kernel satisfies both the first classification index and the second classification index, including:

[0060] According to the value of the first classification index from large to small, the convolution kernels in each network layer are sorted in a first manner, and according to the value of the second classification index from large to small, the convolution kernels in each network layer are sorted in a second manner;

[0061] The convolution kernels in the intersection of the first preset number of convolution kernels in the first sort and the first preset number of convolution kernels in the second sort are used to determine the important convolution kernels; the convolution kernels in the intersection of the second preset number of convolution kernels in the first sort and the second preset number of convolution kernels in the second sort are used to determine the unimportant convolution kernels; the remaining convolution kernels in the convolution kernels in each network layer are used as blurred unimportant convolution kernels.

[0062] In the embodiment of the present application, for the fusion of the two classification methods under the first classification index and the second classification index, the different classification index values ​​are sorted separately, then sorted into segments, and finally the corresponding segmented convolution kernels in the two sortings are intersected to obtain the determined unimportant convolution kernels in each network layer. It can be understood that the second preset number of convolution kernels in the first sorting is used as the unimportant convolution kernel under the first classification index, and the second preset number of convolution kernels in the second sorting is used as the unimportant convolution kernel under the second classification index, and the intersection of the two is used as the deterministic unimportant convolution kernel under the two classification indicators.

[0063] For further information, see Figure 3 In the above step 24, determining the pruning strategy for the fourth business model according to the preset pruning strategy detection method includes:

[0064] (1) selecting at least one channel as an invalid channel for each network layer in the fourth service model using a preset method, and determining the invalid channels and valid channels in the network layer;

[0065] (2) inputting the preset model training samples into the fourth business model, transmitting information through all channels of each network layer, obtaining the first output of the fourth business model, transmitting information through the effective channels of each network layer, and obtaining the second output of the fifth business model;

[0066] (3) determining whether to reselect the invalid channel selected by the preset method based on the deviation between the accuracy of the first output and the accuracy of the second output;

[0067] (4) When it is determined that the invalid channel needs to be reselected, the parameters of the preset method are adjusted based on the deviation and the process goes to step (1); when it is determined that the invalid channel does not need to be reselected, the invalid channels of each network layer in the fourth business model are deleted to obtain the second business model.

[0068] In the embodiment of the present application, the invalid channels and valid channels in the network layer are determined through step (1). At this time, it is not necessary to directly cut off the invalid channels in the fourth business model. Instead, step (2) is performed to determine the performance of the fourth business model, that is, the output accuracy, through a preset data set while retaining the entire structure of the original fourth business model. At the same time, based on the original fourth business model, the information transmission of the invalid channels is shielded, and the performance of the fifth business model is determined through a preset data set. The optimal fifth business model is explored based on the model performance loss in the two cases.

[0069] Specifically, in the above step (1), at least one channel is selected as an invalid channel for each network layer in the fourth service model by a preset method, including:

[0070] For each network layer, a ratio is selected from a plurality of preset ratio parameters as a network layer invalid channel ratio parameter, and channels of a corresponding ratio in the network layer are determined as invalid channels based on the invalid channel ratio parameter.

[0071] It can be understood that the preset method includes a method for allocating invalid channel ratio parameters for each network layer, and the preset ratio parameters can be 10%, 20%, 25%, etc. The preset method can be an intelligent optimization algorithm, such as a particle swarm optimization algorithm, etc., to implement iterative search and update of the preset method, thereby realizing the selection and update of invalid channels in step (1).

[0072] Specifically, in the above step (4), when it is determined that an invalid channel needs to be reselected, adjusting the parameters of the preset method based on the deviation includes:

[0073] Based on the optimization goal that the preset ratio parameter value of invalid channels of each network layer in the fourth service model is larger and the deviation value is smaller, the parameters of the preset method are adjusted; the optimization objective function max(F) is:

[0074] max(F)=max[(1+α(Y 5 -Y 4 ))*βR+Y 4 ], where α, β are preset adjustment parameters, Y 5 is the accuracy of the second output of the fifth business model, Y 4 is the accuracy of the first output of the fourth business model, and R is the sum of the preset proportion parameter values ​​of invalid channels of each network layer in the fourth business model.

[0075] It can be understood that the larger the preset proportion parameter value of the invalid channel of each network layer in the fourth business model, the greater the degree of lightweighting of the fourth business model. In the present application, not only the model performance loss of the fourth business model and the fifth business model is used to feedback and adjust the selection of invalid channels for each network layer in the fourth business model, but also the size of the preset proportion parameter value of the invalid channel of the fourth business model is combined to achieve the maximum lightweighting of the fourth business model while minimizing the model performance loss.

[0076] Optimize the objective function max(F) with the accuracy Y of the first output of the fourth business model 4 is the basic goal, and on this basis, α(Y 5 -Y 4 ) The model performance loss of the fourth business model and the fifth business model is smaller, and the number of invalid channels of the fourth business model driven by βR is greater.

[0077] Furthermore, in the above step 4, dynamic lightweight processing is performed based on the second business model to obtain a third business model, including:

[0078] Receiving a feature map output by a convolution kernel in a current network layer of a second service model;

[0079] Selecting a pixel position feature value to be masked in the feature map according to a preset feature value masking threshold;

[0080] The feature map carrying the pixel position shielding mark is input into the next network layer for calculation, wherein the feature value of the pixel position of the feature map carrying the shielding mark does not participate in the calculation of the next network layer.

[0081] In an embodiment of the present application, redundant feature information contained in a feature map is masked according to a preset feature value masking threshold, thereby reducing the computational processing of redundant feature information when further extracting features from the feature map in the next convolutional layer. Based on the operation of the third business model on the mobile terminal, the computational complexity and the memory usage can be effectively reduced, thereby improving the computational efficiency and operational stability of the third business model on the mobile terminal.

[0082] Specifically, the preset characteristic value shielding threshold can be pre-set based on a specific target value of the running memory occupied space. Further, the preset characteristic value shielding threshold can be dynamically changed based on a preset proportion parameter of the remaining running memory space during real-time runtime.

[0083] See also Figure 4 Based on the above-mentioned artificial intelligence training platform mobile terminal model lightweight method, the embodiment of the present application also provides an artificial intelligence training platform mobile terminal model lightweight system, the system includes:

[0084] A web-side model receiving unit, used to obtain a first business model of the artificial intelligence training platform web-side;

[0085] A first lightweight processing unit, configured to perform static lightweight processing based on the first business model to obtain a second business model, wherein the static lightweight processing represents a lightweight operation with a fixed model structure that does not change with input data;

[0086] The second lightweight processing unit is used to perform dynamic lightweight processing based on the second business model to obtain a third business model, wherein the dynamic lightweight processing represents a lightweight operation of selectively shielding a model forward transfer parameter that changes with input data.

[0087] It should be noted that the artificial intelligence training platform mobile terminal model lightweight system provided in this embodiment and the artificial intelligence training platform mobile terminal model lightweight method embodiment provided in the above embodiment belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0088] The present application also provides an electronic device, which includes:

[0089] processor;

[0090] a memory for storing processor-executable instructions;

[0091] Among them, the processor implements the above-mentioned artificial intelligence training platform mobile terminal model lightweight method by running the executable instructions.

[0092] It is understood that the memory in the above-mentioned electronic device can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. The memory in the embodiment of the present application can store data to support the operation of the terminal. In addition, the electronic device may also include a network interface, which is used to communicate with external mobile terminals, service terminals and other devices through a network connection.

[0093] The embodiment of the present application also provides a computer-readable storage medium on which computer instructions are stored, characterized in that when the instructions are executed by the processor, the steps of the above-mentioned artificial intelligence training platform mobile terminal model lightweight method are implemented. It can be understood that the above-mentioned computer-readable storage medium is a tangible device that can retain and store instructions for use by the instruction execution device, which can be a read-only memory, random access memory, read-only optical disk, magnetic tape, etc.

[0094] The present invention is not limited to the above-mentioned specific implementation modes. Various changes made by ordinary technicians in this field based on the above-mentioned concepts without creative work are all within the protection scope of the present invention.

Claims

1. A lightweight method for a mobile terminal model of an artificial intelligence training platform, characterized in that: include: Obtain the first business model of the artificial intelligence training platform web terminal; Performing static lightweight processing based on the first business model to obtain a second business model, wherein the static lightweight processing represents a lightweight operation with a fixed model structure that does not change with input data; Performing dynamic lightweight processing based on the second business model to obtain a third business model, wherein the dynamic lightweight processing represents a lightweight operation of selectively shielding a model forward transfer parameter that changes with input data; The performing static lightweight processing based on the first business model to obtain a second business model includes: Based on the first business model, identifying and classifying the convolution kernel in each network layer in the first business model under the first classification index and the second classification index; Based on the conditions that the convolution kernel satisfies both the first classification index and the second classification index, the convolution kernel is divided into a definite unimportant convolution kernel, a fuzzy unimportant convolution kernel and a definite important convolution kernel; Based on the determined unimportant convolution kernels in each network layer, a determined unimportant structure is formed, and the determined unimportant structure in the first service model is removed to obtain a fourth service model; For the fuzzy unimportant convolution kernels and the convolution kernels of the determined important convolution kernels in each network layer in the fourth business model, determine the pruning strategy for the fourth business model according to a preset pruning strategy detection method, and obtain the second business model based on the pruning of the fourth business model; The method of dividing the convolution kernel into a determined unimportant convolution kernel, a fuzzy unimportant convolution kernel and a determined important convolution kernel based on the condition that the convolution kernel satisfies the first classification index and the second classification index at the same time includes: According to the value of the first classification index from large to small, the convolution kernels in each network layer are sorted in a first manner, and according to the value of the second classification index from large to small, the convolution kernels in each network layer are sorted in a second manner; Determine the important convolution kernels based on the convolution kernels in the intersection of the first preset number of convolution kernels in the first sort and the first preset number of convolution kernels in the second sort, determine the unimportant convolution kernels based on the convolution kernels in the intersection of the second preset number of convolution kernels in the first sort and the second preset number of convolution kernels in the second sort, and use the remaining convolution kernels in the convolution kernels in each network layer as fuzzy unimportant convolution kernels; The performing dynamic lightweight processing based on the second business model to obtain a third business model includes: Receiving a feature map output by a convolution kernel in a current network layer of a second service model; Selecting a pixel position feature value to be masked in the feature map according to a preset feature value masking threshold; Inputting the feature map carrying the pixel position shielding mark into the next network layer for calculation, wherein the feature value of the pixel position of the feature map carrying the shielding mark does not participate in the calculation of the next network layer; The preset characteristic value shielding threshold is pre-set based on the target value of the running memory occupied space, and the preset characteristic value shielding threshold changes dynamically based on a preset proportion parameter of the running memory remaining space during real-time operation.

2. According to claim 1, a lightweight method for a mobile terminal model of an artificial intelligence training platform is characterized in that: The determining the pruning strategy for the fourth service model according to the preset pruning strategy detection method includes: (1) For each network layer in the fourth service model, at least one channel is selected as an invalid channel by a preset method, and the invalid channels and valid channels in the network layer are determined; (2) Inputting the preset model training samples into the fourth business model, transmitting information through all channels of each network layer, obtaining the first output of the fourth business model, transmitting information through the effective channels of each network layer, and obtaining the second output of the fifth business model; (3) determining whether to reselect the invalid channel selected by the preset method based on the deviation between the accuracy of the first output and the accuracy of the second output; (4) When it is determined that the invalid channel needs to be reselected, the parameters of the preset method are adjusted based on the deviation and the process goes to step (1). When it is determined that the invalid channel does not need to be reselected, the invalid channels of each network layer in the fourth business model are deleted to obtain the second business model.

3. According to claim 2, a lightweight method for a mobile terminal model of an artificial intelligence training platform is characterized in that: The step of selecting at least one channel as an invalid channel for each network layer in the fourth service model by a preset method includes: For each network layer, a ratio is selected from a plurality of preset ratio parameters as a network layer invalid channel ratio parameter, and channels of a corresponding ratio in the network layer are determined as invalid channels based on the invalid channel ratio parameter.

4. According to claim 2, a lightweight method for a mobile terminal model of an artificial intelligence training platform is characterized in that: The step of adjusting the parameters of the preset method based on the deviation when it is determined that the invalid channel needs to be reselected includes: Based on the fourth business model, the larger the preset ratio parameter value of the invalid channel of each network layer is, the smaller the deviation value is, and the optimization target is adjusted; the optimization target function for: ,in , , are preset adjustment parameters, is the accuracy of the second output of the fifth business model, is the accuracy of the first output of the fourth business model, It is the sum of the preset proportion parameter values ​​of the invalid channels of each network layer in the fourth service model.

5. A lightweight system for mobile terminal model of artificial intelligence training platform, characterized in that: include: A web-side model receiving unit, used to obtain a first business model of the artificial intelligence training platform web-side; A first lightweight processing unit, configured to perform static lightweight processing based on the first business model to obtain a second business model, wherein the static lightweight processing represents a lightweight operation with a fixed model structure that does not change with input data; A second lightweight processing unit is used to perform dynamic lightweight processing based on the second business model to obtain a third business model, wherein the dynamic lightweight processing represents a lightweight operation of selectively shielding a model forward transfer parameter that changes with input data; The first lightweight processing unit performs static lightweight processing based on the first service model to obtain a second service model, including: Based on the first business model, identifying and classifying the convolution kernel in each network layer in the first business model under the first classification index and the second classification index; Based on the conditions that the convolution kernel satisfies both the first classification index and the second classification index, the convolution kernel is divided into a definite unimportant convolution kernel, a fuzzy unimportant convolution kernel and a definite important convolution kernel; Based on the determined unimportant convolution kernels in each network layer, a determined unimportant structure is formed, and the determined unimportant structure in the first service model is removed to obtain a fourth service model; For the fuzzy unimportant convolution kernels and the convolution kernels of the determined important convolution kernels in each network layer in the fourth business model, determine the pruning strategy for the fourth business model according to a preset pruning strategy detection method, and obtain the second business model based on the pruning of the fourth business model; The method of dividing the convolution kernel into a determined unimportant convolution kernel, a fuzzy unimportant convolution kernel and a determined important convolution kernel based on the condition that the convolution kernel satisfies the first classification index and the second classification index at the same time includes: According to the value of the first classification index from large to small, the convolution kernels in each network layer are sorted in a first manner, and according to the value of the second classification index from large to small, the convolution kernels in each network layer are sorted in a second manner; Determine the important convolution kernels based on the convolution kernels in the intersection of the first preset number of convolution kernels in the first sort and the first preset number of convolution kernels in the second sort, determine the unimportant convolution kernels based on the convolution kernels in the intersection of the second preset number of convolution kernels in the first sort and the second preset number of convolution kernels in the second sort, and use the remaining convolution kernels in the convolution kernels in each network layer as fuzzy unimportant convolution kernels; The second lightweight processing unit performs dynamic lightweight processing based on the second service model to obtain a third service model, including: Receiving a feature map output by a convolution kernel in a current network layer of a second service model; Selecting a pixel position feature value to be masked in the feature map according to a preset feature value masking threshold; Inputting the feature map carrying the pixel position shielding mark into the next network layer for calculation, wherein the feature value of the pixel position of the feature map carrying the shielding mark does not participate in the calculation of the next network layer; The preset characteristic value shielding threshold is pre-set based on the target value of the running memory occupied space, and the preset characteristic value shielding threshold changes dynamically based on a preset proportion parameter of the running memory remaining space during real-time operation.

6. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing processor-executable instructions; The processor executes the executable instructions to implement the artificial intelligence training platform mobile terminal model lightweight method as described in any one of claims 1-4.

7. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instructions are executed by the processor, the steps of the method for lightweighting the mobile terminal model of the artificial intelligence training platform as described in any one of claims 1 to 4 are implemented.

Images