Empirical compression pruning method and system for AIGC neural network model

Abstract

The invention provides an empirical compression pruning method and system for an AIGC neural network model, and the method comprises the steps: carrying out the work logs of different AIGC services based on the AIGC neural network model, obtaining the working state information of a neural network channel, determining all unnecessary neural network channels of the AIGC neural network model, and carrying out the compression pruning of the AIGC neural network model. The judgment result of whether the AIGC neural network model is in an effective working state or not in different AIGC service scenes is judged, and a reliable experience basis is provided for a neural network channel aimed at subsequent real-time pruning optimization of the AIGC neural network model; all unnecessary neural network channels are positioned, so that the accuracy of pruning and removing the neural network channels subsequently is improved, and the neural network channels are prevented from being removed by mistake; the operation records of the AIGC neural network model in different AIGC service scenes are used as experience, model pruning optimization is effectively and comprehensively executed, the cost of model pruning optimization is reduced, and the efficiency and reliability of pruning optimization are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to an empirical compression pruning method and system for an AIGC neural network model. Background Art

[0002] Neural network models have the characteristics of high consumption of computing and memory resources. In order to improve the operating efficiency and stability of neural network models in embedded devices, it is necessary to prune and optimize the neural network models. AIGC (Artificial Intelligence Generated Content) neural network models are widely used in AI content generation. They can perform AI synthesis of different types of content such as text, images, videos and voices, greatly improving the efficiency and accuracy of AI content synthesis. The pruning optimization operations of existing AIGC neural network models are only for specific models and specific tasks. They are highly specialized, so that each time the neural network model is pruned and optimized, it needs to be redeployed, which increases the manpower and material costs of pruning and optimizing the neural network model. It is impossible to perform efficient and accurate pruning optimization operations based on historical pruning optimization experience of the neural network model. Summary of the invention

[0003] The purpose of the present invention is to provide an empirical compression pruning method and system for an AIGC neural network model, which obtains neural network channel working status information based on the work log of the AIGC neural network model executing different AIGC services, thereby determining all non-essential neural network channels existing in the AIGC neural network model, and judging whether they are in an effective working state in different AIGC service scenarios, thereby providing a reliable empirical basis for the subsequent real-time pruning optimization of the AIGC neural network model for the neural network channels; it also performs positioning processing on all non-essential neural network channels to improve the accuracy of subsequent pruning and removal of neural network channels, and avoids erroneous removal of neural network channels; it uses the operation records of the AIGC neural network model in different AIGC service scenarios as experience to effectively and comprehensively perform model pruning optimization, reduce the cost of model pruning optimization, and improve the efficiency and reliability of pruning optimization.

[0004] The present invention is achieved through the following technical solutions:

[0005] The empirical compression pruning method of AIGC neural network model includes:

[0006] Obtaining work logs of the AIGC neural network model executing different AIGC services, analyzing the work logs to obtain neural network channel work status information of the AIGC neural network model; based on the neural network channel work status information, determining all non-essential neural network channels existing in the AIGC neural network model;

[0007] Acquire data stream feature information of all non-essential neural network channels in different AIGC service scenarios, and obtain judgment results of whether all non-essential neural network channels are in effective working states in different AIGC service scenarios based on the data stream feature information; and perform positioning processing on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model;

[0008] Based on the current working status information of the AIGC neural network model, the AIGC service scenario type information currently corresponding to the AIGC neural network model is identified to determine all non-essential neural network channels in the AIGC neural network model that are not in a valid working state; then based on the position attribute information of all non-essential neural network channels that are not in a valid working state, the AIGC neural network model is compressed and pruned to obtain a pruned AIGC neural network model.

[0009] Optionally, obtaining work logs of the AIGC neural network model executing different AIGC services, analyzing the work logs to obtain neural network channel work status information of the AIGC neural network model; based on the neural network channel work status information, determining all non-essential neural network channels existing in the AIGC neural network model, including:

[0010] Obtaining a work log of the AIGC neural network model executing AIGC of different modality types; wherein the different modality types include image modality, video modality, and voice modality;

[0011] Analyze the work log to obtain the data transmission speed of all neural network channels under the AIGC neural network model, and use this as the working status information of the neural network channel;

[0012] The data transmission speed is compared with a preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, the corresponding neural network channel is determined as a necessary neural network channel; otherwise, the corresponding neural network channel is determined as a non-essential neural network channel.

[0013] Optionally, real-time monitoring of the data transmission speed of each neural network channel and determining whether the overall operation state of the neural network channel is abnormal according to the data transmission speed of each neural network channel include:

[0014] Real-time monitoring of the data transmission speed of each neural network channel;

[0015] Extract the data transmission speed corresponding to each neural network channel;

[0016] The overall operation coefficient of the neural network channel is obtained by using the data transmission speed corresponding to each neural network channel;

[0017] The overall operation coefficient of the neural network channel is obtained by the following formula:

[0018]

[0019] Where K represents the overall operation coefficient of the neural network channel; n represents the number of neural network channels; G b represents the standard deviation of the channel operation coefficients corresponding to n neural network channels; G z G represents the middle value of the channel operation coefficient corresponding to n neural network channels; p G represents the average value of the channel operation coefficients corresponding to n neural network channels; smax G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; smin G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; i represents the channel operation coefficient corresponding to the i-th neural network channel, and the channel operation coefficient is obtained by the following formula:

[0020]

[0021] Among them, G represents the channel operation coefficient corresponding to each neural network channel; m represents the number of neurons contained in each neural network channel; w i represents the weight value of the i-th neuron; P v represents the rate of change of data transmission speed corresponding to each neural network channel; P vb represents the standard deviation of the data transmission speed corresponding to each neural network channel; k represents the number of other neurons connected to the i-th neuron; w j represents the weight value corresponding to the jth neuron among the other neurons connected to the i-th neuron; P vj represents the rate of change of data processing speed of the jth other neuron among the other neurons connected to the i-th neuron; w kmaxrepresents the maximum weight value corresponding to the k other neurons among the other neurons connected to the i-th neuron; P vkb represents the standard deviation of the rate of change of data processing speed of k other neurons among the other neurons connected to the i-th neuron;

[0022] Comparing the overall operating coefficient of the neural network channel with a preset coefficient threshold;

[0023] When the overall operating coefficient of the neural network channel exceeds a preset coefficient threshold, it is determined that the overall operating state of the neural network channel is abnormal, and an abnormal risk warning is issued.

[0024] Optionally, data stream feature information of all non-essential neural network channels in different AIGC service scenarios is obtained, and based on the data stream feature information, a judgment result of whether all non-essential neural network channels are in an effective working state in different AIGC service scenarios is obtained; and positioning processing is performed on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model, including:

[0025] Obtaining data stream transmission frame loss rates of all non-essential neural network channels in different AIGC service scenarios as the data stream feature information;

[0026] Comparing the data stream transmission frame loss rate with a preset frame loss rate threshold, if the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, determining the corresponding non-essential neural network channel as not being in a valid working state in the corresponding AIGC service scenario; otherwise, determining the corresponding non-essential neural network channel as being in a valid working state in the corresponding AIGC service scenario;

[0027] All non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels within the AIGC neural network model and the connection node position information with the necessary neural network channels, which is used as the position attribute information.

[0028] Optionally, based on the current working state information of the AIGC neural network model, the AIGC service scenario type information currently corresponding to the AIGC neural network model is identified, so as to determine all non-essential neural network channels in the AIGC neural network model that are not in a valid working state; and then based on the position attribute information of all non-essential neural network channels that are not in a valid working state, the AIGC neural network model is compressed and pruned to obtain a pruned AIGC neural network model, including:

[0029] Acquire the current input working data type information of the AIGC neural network model, identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the input working data type information, and compare it with the judgment result, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state;

[0030] Based on the neural network layer position information of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the connection node position information with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, so as to obtain a pruned AIGC neural network model.

[0031] The empirical compression pruning system of AIGC neural network model includes:

[0032] A neural network channel working status analysis module is used to obtain the work logs of the AIGC neural network model executing different AIGC services, analyze the work logs, and obtain the neural network channel working status information of the AIGC neural network model;

[0033] A neural network channel identification module, used to determine all unnecessary neural network channels existing in the AIGC neural network model based on the neural network channel working state information;

[0034] A neural network channel judgment module is used to obtain data flow characteristic information of all non-essential neural network channels in different AIGC service scenarios, and based on the data flow characteristic information, obtain a judgment result of whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios;

[0035] A neural network channel positioning module, used to perform positioning processing on all non-essential neural network channels, and obtain position attribute information of all non-essential neural network channels in the AIGC neural network model;

[0036] A neural network model empirical identification module, used to identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the current working state information of the AIGC neural network model, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state;

[0037] The neural network model compression and pruning module is used to compress and prune the AIGC neural network model based on the position attribute information of all non-essential neural network channels that are not in an effective working state, so as to obtain a pruned AIGC neural network model.

[0038] Optionally, the neural network channel working state analysis module is used to obtain work logs of the AIGC neural network model executing different AIGC services, analyze the work logs, and obtain the neural network channel working state information of the AIGC neural network model, including:

[0039] Obtaining a work log of the AIGC neural network model executing AIGC of different modality types; wherein the different modality types include image modality, video modality, and voice modality;

[0040] Analyze the work log to obtain the data transmission speed of all neural network channels under the AIGC neural network model, and use this as the working status information of the neural network channel;

[0041] The neural network channel identification module is used to determine all non-essential neural network channels existing in the AIGC neural network model based on the neural network channel working state information, including:

[0042] The data transmission speed is compared with a preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, the corresponding neural network channel is determined as a necessary neural network channel; otherwise, the corresponding neural network channel is determined as a non-essential neural network channel.

[0043] Optionally, real-time monitoring of the data transmission speed of each neural network channel and determining whether the overall operation state of the neural network channel is abnormal according to the data transmission speed of each neural network channel include:

[0044] Real-time monitoring of the data transmission speed of each neural network channel;

[0045] Extract the data transmission speed corresponding to each neural network channel;

[0046] The overall operation coefficient of the neural network channel is obtained by using the data transmission speed corresponding to each neural network channel;

[0047] The overall operation coefficient of the neural network channel is obtained by the following formula:

[0048]

[0049] Where K represents the overall operation coefficient of the neural network channel; n represents the number of neural network channels; Gb represents the standard deviation of the channel operation coefficients corresponding to n neural network channels; G z G represents the middle value of the channel operation coefficient corresponding to n neural network channels; p G represents the average value of the channel operation coefficients corresponding to n neural network channels; smax G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; smin G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; i represents the channel operation coefficient corresponding to the i-th neural network channel, and the channel operation coefficient is obtained by the following formula:

[0050]

[0051] Among them, G represents the channel operation coefficient corresponding to each neural network channel; m represents the number of neurons contained in each neural network channel; w i represents the weight value of the i-th neuron; P v represents the rate of change of data transmission speed corresponding to each neural network channel; P vb represents the standard deviation of the data transmission speed corresponding to each neural network channel; k represents the number of other neurons connected to the i-th neuron; w j represents the weight value corresponding to the jth neuron among the other neurons connected to the i-th neuron; P vj represents the rate of change of data processing speed of the jth other neuron among the other neurons connected to the i-th neuron; w kmax represents the maximum weight value corresponding to the k other neurons among the other neurons connected to the i-th neuron; P vkb represents the standard deviation of the rate of change of data processing speed of k other neurons among the other neurons connected to the i-th neuron;

[0052] Comparing the overall operating coefficient of the neural network channel with a preset coefficient threshold;

[0053] When the overall operating coefficient of the neural network channel exceeds a preset coefficient threshold, it is determined that the overall operating state of the neural network channel is abnormal, and an abnormal risk warning is issued.

[0054] Optionally, the neural network channel judgment module is used to obtain data flow feature information of all non-essential neural network channels in different AIGC service scenarios, and based on the data flow feature information, obtain a judgment result of whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios, including:

[0055] Obtaining data stream transmission frame loss rates of all non-essential neural network channels in different AIGC service scenarios as the data stream feature information;

[0056] Comparing the data stream transmission frame loss rate with a preset frame loss rate threshold, if the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, determining the corresponding non-essential neural network channel as not being in a valid working state in the corresponding AIGC service scenario; otherwise, determining the corresponding non-essential neural network channel as being in a valid working state in the corresponding AIGC service scenario;

[0057] The neural network channel positioning module is used to perform positioning processing on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model, including:

[0058] All non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels within the AIGC neural network model and the connection node position information with the necessary neural network channels, which is used as the position attribute information.

[0059] Optionally, the neural network model empirical identification module is used to identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the current working state information of the AIGC neural network model, so as to determine all non-essential neural network channels in the AIGC neural network model that are not in a valid working state, including:

[0060] Acquire the current input working data type information of the AIGC neural network model, identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the input working data type information, and compare it with the judgment result, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state;

[0061] The neural network model compression and pruning module is used to compress and prune the AIGC neural network model based on the position attribute information of all non-essential neural network channels that are not in an effective working state, so as to obtain a pruned AIGC neural network model, including:

[0062] Based on the neural network layer position information of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the connection node position information with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, so as to obtain a pruned AIGC neural network model.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] The empirical compression pruning method and system of the AIGC neural network model provided in the present application obtain the working status information of the neural network channel based on the work log of the AIGC neural network model executing different AIGC services, thereby determining all non-essential neural network channels existing in the AIGC neural network model and judging whether they are in an effective working state in different AIGC service scenarios, providing a reliable empirical basis for the subsequent real-time pruning optimization of the AIGC neural network model for the neural network channels; all non-essential neural network channels are located and processed to improve the accuracy of subsequent pruning and removal of neural network channels, and avoid erroneous removal of neural network channels; the operation records of the AIGC neural network model in different AIGC service scenarios are used as experience to effectively and comprehensively perform model pruning optimization, reduce the cost of model pruning optimization, and improve the efficiency and reliability of pruning optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0066] Figure 1 A schematic diagram of the flow chart of the empirical compression pruning method for the AIGC neural network model provided by the present invention.

[0067] Figure 2 A schematic diagram of the structure of the empirical compression pruning system of the AIGC neural network model provided by the present invention. DETAILED DESCRIPTION

[0068] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. It is to be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only some structures related to the present application are shown in the accompanying drawings, rather than all structures. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0069] The terms "including" and "having" and any variations thereof in this application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices.

[0070] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0071] See also Figure 1 As shown, the empirical compression pruning method of the AIGC neural network model provided in one embodiment of the present application includes:

[0072] Obtaining work logs of the AIGC neural network model executing different AIGC services, analyzing the work logs to obtain neural network channel work status information of the AIGC neural network model; based on the neural network channel work status information, determining all non-essential neural network channels existing in the AIGC neural network model;

[0073] Acquire data stream feature information of all non-essential neural network channels in different AIGC service scenarios, and obtain judgment results of whether all non-essential neural network channels are in effective working states in different AIGC service scenarios based on the data stream feature information; and perform positioning processing on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model;

[0074] Based on the current working status information of the AIGC neural network model, the AIGC service scenario type information currently corresponding to the AIGC neural network model is identified to determine all non-essential neural network channels in the AIGC neural network model that are not in a valid working state; then based on the position attribute information of all non-essential neural network channels that are not in a valid working state, the AIGC neural network model is compressed and pruned to obtain a pruned AIGC neural network model.

[0075] The beneficial effects of the above embodiments are as follows: the empirical compression pruning method of the AIGC neural network model obtains the working status information of the neural network channel based on the work log of the AIGC neural network model executing different AIGC services, thereby determining all unnecessary neural network channels existing in the AIGC neural network model, and judging whether they are in an effective working state in different AIGC service scenarios, providing a reliable empirical basis for the subsequent real-time pruning optimization of the AIGC neural network model for the neural network channels; all unnecessary neural network channels are positioned to improve the accuracy of subsequent pruning and removal of neural network channels, and avoid erroneous removal of neural network channels; the operation records of the AIGC neural network model in different AIGC service scenarios are used as experience to effectively and comprehensively perform model pruning optimization, reduce the cost of model pruning optimization, and improve the efficiency and reliability of pruning optimization.

[0076] In another embodiment, a work log of an AIGC neural network model executing different AIGC services is obtained, and the work log is analyzed to obtain the neural network channel work status information of the AIGC neural network model; based on the neural network channel work status information, all non-essential neural network channels existing in the AIGC neural network model are determined, including:

[0077] Obtaining a work log of the AIGC neural network model executing AIGC of different modality types; wherein the different modality types include image modality, video modality, and voice modality;

[0078] The work log is analyzed to obtain the data transmission speed of all neural network channels under the AIGC neural network model, which is used as the working status information of the neural network channel;

[0079] The data transmission speed is compared with a preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, the corresponding neural network channel is determined as a necessary neural network channel; otherwise, the corresponding neural network channel is determined as a non-essential neural network channel.

[0080] The beneficial effects of the above embodiments are that the AIGC neural network model can generate AI content in different modes such as images, videos and voices, and when the AIGC neural network model generates corresponding AI content, a corresponding AIGC work log will be formed, which comprehensively records the data transmission status of all subordinate neural network channels in the process of generating AI content by the AIGC neural network model. The data transmission speed of each neural network channel is compared with the preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, it indicates that the neural network channel plays a key role in data transmission during the AIGC neural network model generating AI content, and the neural network channel is determined to be a necessary neural network channel; if the data transmission speed is less than the preset transmission speed threshold, it indicates that the high neural network channel is basically not responsible for data transmission during the AIGC neural network model generating AI content, and the neural network channel is determined to be a non-essential neural network channel. In this way, all neural network channels under the AIGC neural network model can be effectively distinguished, providing an accurate basis for subsequent neural network channel pruning optimization, and avoiding the occurrence of neural network channel mispruning.

[0081] In another embodiment, the data transmission speed of each neural network channel is monitored in real time, and whether the overall operation state of the neural network channel is abnormal is determined according to the data transmission speed of each neural network channel, including:

[0082] Real-time monitoring of the data transmission speed of each neural network channel;

[0083] Extract the data transmission speed corresponding to each neural network channel;

[0084] The overall operation coefficient of the neural network channel is obtained by using the data transmission speed corresponding to each neural network channel;

[0085] The overall operation coefficient of the neural network channel is obtained by the following formula:

[0086]

[0087] Where K represents the overall operation coefficient of the neural network channel; n represents the number of neural network channels; G b represents the standard deviation of the channel operation coefficients corresponding to n neural network channels; G z G represents the middle value of the channel operation coefficient corresponding to n neural network channels; p G represents the average value of the channel operation coefficients corresponding to n neural network channels; smax G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; smin G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons;i represents the channel operation coefficient corresponding to the i-th neural network channel, and the channel operation coefficient is obtained by the following formula:

[0088]

[0089] Among them, G represents the channel operation coefficient corresponding to each neural network channel; m represents the number of neurons contained in each neural network channel; w i represents the weight value of the i-th neuron; P v represents the rate of change of data transmission speed corresponding to each neural network channel; P vb represents the standard deviation of the data transmission speed corresponding to each neural network channel; k represents the number of other neurons connected to the i-th neuron; w j represents the weight value corresponding to the jth neuron among the other neurons connected to the i-th neuron; P vj represents the rate of change of data processing speed of the jth other neuron among the other neurons connected to the i-th neuron; w kmax represents the maximum value of the weights of the k other neurons connected to the i-th neuron; vkb represents the standard deviation of the rate of change of data processing speed of k other neurons among the other neurons connected to the i-th neuron;

[0090] Comparing the overall operating coefficient of the neural network channel with a preset coefficient threshold;

[0091] When the overall operating coefficient of the neural network channel exceeds a preset coefficient threshold, it is determined that the overall operating state of the neural network channel is abnormal, and an abnormal risk warning is issued.

[0092] The beneficial effects of the above embodiments are that by real-time monitoring of the data transmission speed of each neural network channel, the technology can instantly capture any slight changes in the network state. This real-time monitoring capability enables the system to respond quickly to any potential performance degradation or abnormal state, thereby improving the stability and reliability of the system. The scheme uses complex mathematical models and multiple parameters (such as channel operation coefficient standard deviation, median, average, neuron weight, data processing speed change rate, etc.) to comprehensively evaluate the overall operation state of the neural network channel. This multi-dimensional evaluation method can more accurately reflect the actual performance of the network and reduce the possibility of false alarms and missed reports. By comparing the overall operation coefficient of the neural network channel with the preset coefficient threshold, the technology can warn before the abnormal state occurs. This early warning mechanism helps administrators take measures in advance to avoid potential performance degradation or failure, thereby improving the availability and maintenance efficiency of the system. The mathematical model and parameter settings in the scheme can be adjusted according to actual application scenarios and requirements. This adaptability and flexibility enable the technology to adapt to neural networks of different sizes and types, improving its versatility and practicality. By real-time monitoring and evaluating the performance of the neural network channel, the technology can help administrators better understand the utilization of network resources. This helps to optimize resource allocation and improve resource utilization, thereby reducing costs and improving the overall performance of the system.

[0093] In summary, the technical effects of the above technical solutions in terms of performance indicators are mainly reflected in real-time monitoring and dynamic response, accurate evaluation of network status, early warning and prevention mechanisms, adaptability and flexibility, performance optimization and resource utilization, etc. These effects jointly improve the stability and reliability of the neural network, reduce abnormal risks, and optimize the overall performance of the system.

[0094] In another embodiment, data stream feature information of all non-essential neural network channels in different AIGC service scenarios is obtained, and based on the data stream feature information, a judgment result of whether all non-essential neural network channels are in an effective working state in different AIGC service scenarios is obtained; and positioning processing is performed on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model, including:

[0095] Obtain the data stream transmission frame loss rate of all non-essential neural network channels in different AIGC service scenarios as the characteristic information of the data stream;

[0096] The data stream transmission frame loss rate is compared with a preset frame loss rate threshold. If the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, the corresponding non-essential neural network channel is determined to be not in a valid working state in the corresponding AIGC service scenario; otherwise, the corresponding non-essential neural network channel is determined to be in a valid working state in the corresponding AIGC service scenario;

[0097] All non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels within the AIGC neural network model and the connection node position information with the necessary neural network channels, which is used as the position attribute information.

[0098] The beneficial effect of the above embodiment is to obtain the data stream transmission frame loss rate of all non-essential neural network channels in different AIGC service scenarios, so that the data transmission performance of each non-essential neural network channel in different AIGC service scenarios can be fully recorded. When the data stream transmission frame loss rate of the non-essential neural network channel in a certain AIGC service scenario is larger, it indicates that the transmission performance of the non-essential neural network channel in the corresponding AIGC service scenario is poor, and it does not have a positive effect on the AI ​​content generation of the AIGC neural network model. At this time, if the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, the corresponding non-essential neural network channel is determined to be not in an effective working state in the corresponding AIGC service scenario, which indicates that the non-essential neural network channel is a redundant neural network channel in the AIGC service scenario. Then all non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels in the AIGC neural network model and the connection node position information with the necessary neural network channel, so as to provide accurate pruning position information for subsequent accurate neural network channel pruning optimization.

[0099] In another embodiment, based on the current working state information of the AIGC neural network model, the AIGC service scenario type information currently corresponding to the AIGC neural network model is identified, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state; and then based on the position attribute information of all unnecessary neural network channels that are not in a valid working state, the AIGC neural network model is compressed and pruned, so as to obtain a pruned AIGC neural network model, including:

[0100] Obtaining the current input working data type information of the AIGC neural network model, identifying the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the input working data type information, and comparing it with the judgment result, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state;

[0101] Based on the neural network layer position information of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the connection node position information with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, so as to obtain a pruned AIGC neural network model.

[0102] The beneficial effect of the above embodiment is to obtain the current input working data type information of the AIGC neural network model, where the input working data type information may be but is not limited to whether the data to be processed in the AIGC neural network model belongs to image data, video data or voice data, and compare the input data type information with all AIGC service scenario type information of the work log of the AIGC neural network model, identify and determine the AIGC service scenario type information currently corresponding to the AIGC neural network model, and compare the aforementioned judgment results on whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios, determine all non-essential neural network channels of the AIGC neural network model that are not in a valid working state in the current AIGC service scenario, and thereby determine the neural network channels that the AIGC neural network model currently allows for pruning optimization. Then, based on the location information of the neural network layers of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the location information of the connection nodes with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, thereby achieving compressed pruning of the AIGC neural network model, reducing the cost of model pruning optimization and improving the efficiency and reliability of pruning optimization.

[0103] See also Figure 2 As shown, the empirical compression pruning system of the AIGC neural network model provided by an embodiment of the present application includes:

[0104] A neural network channel working status analysis module is used to obtain the work log of the AIGC neural network model executing different AIGC services, analyze the work log, and obtain the neural network channel working status information of the AIGC neural network model;

[0105] A neural network channel identification module, used to determine all non-essential neural network channels existing in the AIGC neural network model based on the working status information of the neural network channel;

[0106] A neural network channel judgment module is used to obtain data flow characteristic information of all non-essential neural network channels in different AIGC service scenarios, and based on the data flow characteristic information, obtain a judgment result of whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios;

[0107] A neural network channel positioning module is used to position all non-essential neural network channels and obtain the position attribute information of all non-essential neural network channels in the AIGC neural network model;

[0108] A neural network model empirical identification module is used to identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the current working state information of the AIGC neural network model, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state;

[0109] The neural network model compression and pruning module is used to compress and prune the AIGC neural network model based on the position attribute information of all unnecessary neural network channels that are not in an effective working state, so as to obtain a pruned AIGC neural network model.

[0110] The beneficial effects of the above embodiments are as follows: the empirical compression pruning system of the AIGC neural network model obtains the working status information of the neural network channel based on the work log of the AIGC neural network model executing different AIGC services, thereby determining all non-essential neural network channels existing in the AIGC neural network model, and judging whether they are in an effective working state in different AIGC service scenarios, providing a reliable empirical basis for the subsequent real-time pruning optimization of the AIGC neural network model for the neural network channels; all non-essential neural network channels are positioned to improve the accuracy of subsequent pruning and removal of neural network channels, and avoid erroneous removal of neural network channels; the operation records of the AIGC neural network model in different AIGC service scenarios are used as experience to effectively and comprehensively perform model pruning optimization, reduce the cost of model pruning optimization, and improve the efficiency and reliability of pruning optimization.

[0111] In another embodiment, the neural network channel working state analysis module is used to obtain the work log of the AIGC neural network model executing different AIGC services, analyze the work log, and obtain the neural network channel working state information of the AIGC neural network model, including:

[0112] Obtaining a work log of the AIGC neural network model executing AIGC of different modality types; wherein the different modality types include image modality, video modality, and voice modality;

[0113] The work log is analyzed to obtain the data transmission speed of all neural network channels under the AIGC neural network model, which is used as the working status information of the neural network channel;

[0114] The neural network channel identification module is used to determine all non-essential neural network channels existing in the AIGC neural network model based on the working status information of the neural network channel, including:

[0115] The data transmission speed is compared with a preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, the corresponding neural network channel is determined as a necessary neural network channel; otherwise, the corresponding neural network channel is determined as a non-essential neural network channel.

[0116] The beneficial effects of the above embodiments are that the AIGC neural network model can generate AI content in different modes such as images, videos and voices, and when the AIGC neural network model generates corresponding AI content, a corresponding AIGC work log will be formed, which comprehensively records the data transmission status of all subordinate neural network channels in the process of generating AI content by the AIGC neural network model. The data transmission speed of each neural network channel is compared with the preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, it indicates that the neural network channel plays a key role in data transmission during the AIGC neural network model generating AI content, and the neural network channel is determined to be a necessary neural network channel; if the data transmission speed is less than the preset transmission speed threshold, it indicates that the high neural network channel is basically not responsible for data transmission during the AIGC neural network model generating AI content, and the neural network channel is determined to be a non-essential neural network channel. In this way, all neural network channels under the AIGC neural network model can be effectively distinguished, providing an accurate basis for subsequent neural network channel pruning optimization, and avoiding the occurrence of neural network channel mispruning.

[0117] In another embodiment, the data transmission speed of each neural network channel is monitored in real time, and whether the overall operation state of the neural network channel is abnormal is determined according to the data transmission speed of each neural network channel, including:

[0118] Real-time monitoring of the data transmission speed of each neural network channel;

[0119] Extract the data transmission speed corresponding to each neural network channel;

[0120] The overall operation coefficient of the neural network channel is obtained by using the data transmission speed corresponding to each neural network channel;

[0121] The overall operation coefficient of the neural network channel is obtained by the following formula:

[0122]

[0123] Where K represents the overall operation coefficient of the neural network channel; n represents the number of neural network channels; G b represents the standard deviation of the channel operation coefficients corresponding to n neural network channels; G z G represents the middle value of the channel operation coefficient corresponding to n neural network channels; p G represents the average value of the channel operation coefficients corresponding to n neural network channels; smax G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; smin G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; i represents the channel operation coefficient corresponding to the i-th neural network channel, and the channel operation coefficient is obtained by the following formula:

[0124]

[0125] Among them, G represents the channel operation coefficient corresponding to each neural network channel; m represents the number of neurons contained in each neural network channel; w i represents the weight value of the i-th neuron; P v represents the rate of change of data transmission speed corresponding to each neural network channel; P vb represents the standard deviation of the data transmission speed corresponding to each neural network channel; k represents the number of other neurons connected to the i-th neuron; w j represents the weight value corresponding to the jth neuron among the other neurons connected to the i-th neuron; P vj represents the rate of change of data processing speed of the jth other neuron among the other neurons connected to the i-th neuron; w kmax represents the maximum weight value corresponding to the k other neurons among the other neurons connected to the i-th neuron; P vkb represents the standard deviation of the rate of change of data processing speed of k other neurons among the other neurons connected to the i-th neuron;

[0126] Comparing the overall operating coefficient of the neural network channel with a preset coefficient threshold;

[0127] When the overall operating coefficient of the neural network channel exceeds a preset coefficient threshold, it is determined that the overall operating state of the neural network channel is abnormal, and an abnormal risk warning is issued.

[0128] The beneficial effects of the above embodiments are that by real-time monitoring of the data transmission speed of each neural network channel, the technology can instantly capture any slight changes in the network state. This real-time monitoring capability enables the system to respond quickly to any potential performance degradation or abnormal state, thereby improving the stability and reliability of the system. The scheme uses complex mathematical models and multiple parameters (such as channel operation coefficient standard deviation, median, average, neuron weight, data processing speed change rate, etc.) to comprehensively evaluate the overall operation state of the neural network channel. This multi-dimensional evaluation method can more accurately reflect the actual performance of the network and reduce the possibility of false alarms and missed reports. By comparing the overall operation coefficient of the neural network channel with the preset coefficient threshold, the technology can warn before the abnormal state occurs. This early warning mechanism helps administrators take measures in advance to avoid potential performance degradation or failure, thereby improving the availability and maintenance efficiency of the system. The mathematical model and parameter settings in the scheme can be adjusted according to actual application scenarios and requirements. This adaptability and flexibility enable the technology to adapt to neural networks of different sizes and types, improving its versatility and practicality. By real-time monitoring and evaluating the performance of the neural network channel, the technology can help administrators better understand the utilization of network resources. This helps to optimize resource allocation and improve resource utilization, thereby reducing costs and improving the overall performance of the system.

[0129] In summary, the technical effects of the above technical solutions in terms of performance indicators are mainly reflected in real-time monitoring and dynamic response, accurate evaluation of network status, early warning and prevention mechanisms, adaptability and flexibility, performance optimization and resource utilization, etc. These effects jointly improve the stability and reliability of the neural network, reduce abnormal risks, and optimize the overall performance of the system.

[0130] In another embodiment, the neural network channel judgment module is used to obtain data flow characteristic information of all non-essential neural network channels in different AIGC service scenarios, and based on the data flow characteristic information, obtain a judgment result of whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios, including:

[0131] Obtain the data stream transmission frame loss rate of all non-essential neural network channels in different AIGC service scenarios as the characteristic information of the data stream;

[0132] The data stream transmission frame loss rate is compared with a preset frame loss rate threshold. If the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, the corresponding non-essential neural network channel is determined to be not in a valid working state in the corresponding AIGC service scenario; otherwise, the corresponding non-essential neural network channel is determined to be in a valid working state in the corresponding AIGC service scenario;

[0133] The neural network channel positioning module is used to position all non-essential neural network channels and obtain the position attribute information of all non-essential neural network channels in the AIGC neural network model, including:

[0134] All non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels within the AIGC neural network model and the connection node position information with the necessary neural network channels, which is used as the position attribute information.

[0135] The beneficial effect of the above embodiment is to obtain the data stream transmission frame loss rate of all non-essential neural network channels in different AIGC service scenarios, so that the data transmission performance of each non-essential neural network channel in different AIGC service scenarios can be fully recorded. When the data stream transmission frame loss rate of the non-essential neural network channel in a certain AIGC service scenario is larger, it indicates that the transmission performance of the non-essential neural network channel in the corresponding AIGC service scenario is poor, and it does not have a positive effect on the AI ​​content generation of the AIGC neural network model. At this time, if the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, the corresponding non-essential neural network channel is determined to be not in an effective working state in the corresponding AIGC service scenario, which indicates that the non-essential neural network channel is a redundant neural network channel in the AIGC service scenario. Then all non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels in the AIGC neural network model and the connection node position information with the necessary neural network channel, so as to provide accurate pruning position information for subsequent accurate neural network channel pruning optimization.

[0136] In another embodiment, the neural network model empirical identification module is used to identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the current working state information of the AIGC neural network model, so as to determine all non-essential neural network channels in the AIGC neural network model that are not in a valid working state, including:

[0137] Obtaining the current input working data type information of the AIGC neural network model, identifying the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the input working data type information, and comparing it with the judgment result, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state;

[0138] The neural network model compression and pruning module is used to compress and prune the AIGC neural network model based on the position attribute information of all non-essential neural network channels that are not in an effective working state, so as to obtain a pruned AIGC neural network model, including:

[0139] Based on the neural network layer position information of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the connection node position information with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, so as to obtain a pruned AIGC neural network model.

[0140] The beneficial effect of the above embodiment is to obtain the current input working data type information of the AIGC neural network model, where the input working data type information may be but is not limited to whether the data to be processed in the AIGC neural network model belongs to image data, video data or voice data, and compare the input data type information with all AIGC service scenario type information of the work log of the AIGC neural network model, identify and determine the AIGC service scenario type information currently corresponding to the AIGC neural network model, and compare the aforementioned judgment results on whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios, determine all non-essential neural network channels of the AIGC neural network model that are not in a valid working state in the current AIGC service scenario, and thereby determine the neural network channels that the AIGC neural network model currently allows for pruning optimization. Then, based on the location information of the neural network layers of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the location information of the connection nodes with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, thereby achieving compressed pruning of the AIGC neural network model, reducing the cost of model pruning optimization and improving the efficiency and reliability of pruning optimization.

[0141] In general, the empirical compression pruning method and system of the AIGC neural network model obtains the working status information of the neural network channel based on the work log of the AIGC neural network model executing different AIGC services, thereby determining all non-essential neural network channels of the AIGC neural network model and judging whether they are in an effective working state under different AIGC service scenarios, providing a reliable empirical basis for the subsequent real-time pruning optimization of the AIGC neural network model for the neural network channels; all non-essential neural network channels are located and processed to improve the accuracy of subsequent pruning and removal of neural network channels, and avoid erroneous removal of neural network channels; the operation records of the AIGC neural network model in different AIGC service scenarios are used as experience to effectively and comprehensively perform model pruning optimization, reduce the cost of model pruning optimization, and improve the efficiency and reliability of pruning optimization.

[0142] The above is only a specific implementation of the present invention, and any other improvements made based on the concept of the present invention are considered to be within the protection scope of the present invention.

Claims

1. The empirical compression pruning method of AIGC neural network model is characterized by: include: Obtaining work logs of the AIGC neural network model executing different AIGC services, analyzing the work logs to obtain neural network channel work status information of the AIGC neural network model; based on the neural network channel work status information, determining all non-essential neural network channels existing in the AIGC neural network model; Acquire data stream feature information of all non-essential neural network channels in different AIGC service scenarios, and obtain judgment results of whether all non-essential neural network channels are in effective working states in different AIGC service scenarios based on the data stream feature information; and perform positioning processing on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model; Based on the current working status information of the AIGC neural network model, the AIGC service scenario type information currently corresponding to the AIGC neural network model is identified to determine all non-essential neural network channels in the AIGC neural network model that are not in a valid working state; then based on the position attribute information of all non-essential neural network channels that are not in a valid working state, the AIGC neural network model is compressed and pruned to obtain a pruned AIGC neural network model.

2. The empirical compression pruning method of the AIGC neural network model according to claim 1, characterized in that: Obtaining work logs of the AIGC neural network model executing different AIGC services, analyzing the work logs, and obtaining neural network channel work status information of the AIGC neural network model; Based on the neural network channel working state information, all non-essential neural network channels existing in the AIGC neural network model are determined, including: Obtaining a work log of the AIGC neural network model executing AIGC of different modality types; wherein the different modality types include image modality, video modality, and voice modality; Analyze the work log to obtain the data transmission speed of all neural network channels under the AIGC neural network model, and use this as the working status information of the neural network channel; The data transmission speed is compared with a preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, the corresponding neural network channel is determined as a necessary neural network channel; otherwise, the corresponding neural network channel is determined as a non-essential neural network channel.

3. The empirical compression pruning method of the AIGC neural network model according to claim 2, characterized in that: Real-time monitoring of the data transmission speed of each neural network channel, and determining whether the overall operation state of the neural network channel is abnormal according to the data transmission speed of each neural network channel, including: Real-time monitoring of the data transmission speed of each neural network channel; Extract the data transmission speed corresponding to each neural network channel; The overall operation coefficient of the neural network channel is obtained by using the data transmission speed corresponding to each neural network channel; The overall operation coefficient of the neural network channel is obtained by the following formula: Where K represents the overall operation coefficient of the neural network channel; n represents the number of neural network channels; G b represents the standard deviation of the channel operation coefficients corresponding to n neural network channels; G z G represents the middle value of the channel operation coefficient corresponding to n neural network channels; p G represents the average value of the channel operation coefficients corresponding to n neural network channels; smax G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; smin G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; i represents the channel operation coefficient corresponding to the i-th neural network channel, and the channel operation coefficient is obtained by the following formula: Among them, G represents the channel operation coefficient corresponding to each neural network channel; m represents the number of neurons contained in each neural network channel; w i represents the weight value of the i-th neuron; P v represents the rate of change of data transmission speed corresponding to each neural network channel; P vb represents the standard deviation of the data transmission speed corresponding to each neural network channel; k represents the number of other neurons connected to the i-th neuron; w j represents the weight value corresponding to the jth neuron among the other neurons connected to the i-th neuron; P vj represents the rate of change of data processing speed of the jth other neuron among the other neurons connected to the i-th neuron; w kmax represents the maximum weight value corresponding to the k other neurons among the other neurons connected to the i-th neuron; P vkb represents the standard deviation of the rate of change of data processing speed of k other neurons among the other neurons connected to the i-th neuron; Comparing the overall operating coefficient of the neural network channel with a preset coefficient threshold; When the overall operating coefficient of the neural network channel exceeds a preset coefficient threshold, it is determined that the overall operating state of the neural network channel is abnormal, and an abnormal risk warning is issued.

4. The empirical compression pruning method of the AIGC neural network model according to claim 1, characterized in that: Acquire data stream feature information of all non-essential neural network channels in different AIGC service scenarios, and obtain judgment results of whether all non-essential neural network channels are in effective working states in different AIGC service scenarios based on the data stream feature information; and perform positioning processing on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model, including: Obtaining data stream transmission frame loss rates of all non-essential neural network channels in different AIGC service scenarios as the data stream feature information; The data stream transmission frame loss rate is compared with a preset frame loss rate threshold. If the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, the corresponding non-essential neural network channel is determined to be not in a valid working state in the corresponding AIGC service scenario; otherwise, the corresponding non-essential neural network channel is determined to be in a valid working state in the corresponding AIGC service scenario; all non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels in the AIGC neural network model and the connection node position information with the necessary neural network channels, which is used as the position attribute information.

5. The empirical compression pruning method of the AIGC neural network model according to claim 1, characterized in that: Based on the current working state information of the AIGC neural network model, the AIGC service scenario type information currently corresponding to the AIGC neural network model is identified, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state; and based on the position attribute information of all unnecessary neural network channels that are not in a valid working state, the AIGC neural network model is compressed and pruned, so as to obtain a pruned AIGC neural network model, including: Acquire the current input working data type information of the AIGC neural network model, identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the input working data type information, and compare it with the judgment result, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state; Based on the neural network layer position information of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the connection node position information with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, so as to obtain a pruned AIGC neural network model.

6. The empirical compression pruning system of AIGC neural network model is characterized by: include: A neural network channel working status analysis module is used to obtain the work logs of the AIGC neural network model executing different AIGC services, analyze the work logs, and obtain the neural network channel working status information of the AIGC neural network model; A neural network channel identification module, used to determine all unnecessary neural network channels existing in the AIGC neural network model based on the neural network channel working state information; A neural network channel judgment module is used to obtain data flow characteristic information of all non-essential neural network channels in different AIGC service scenarios, and based on the data flow characteristic information, obtain a judgment result of whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios; A neural network channel positioning module, used to perform positioning processing on all non-essential neural network channels, and obtain position attribute information of all non-essential neural network channels in the AIGC neural network model; A neural network model empirical identification module, used to identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the current working state information of the AIGC neural network model, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state; The neural network model compression and pruning module is used to compress and prune the AIGC neural network model based on the position attribute information of all non-essential neural network channels that are not in an effective working state, so as to obtain a pruned AIGC neural network model.

7. The empirical compression pruning system of the AIGC neural network model according to claim 6, characterized in that: The neural network channel working state analysis module is used to obtain the work logs of the AIGC neural network model executing different AIGC services, analyze the work logs, and obtain the neural network channel working state information of the AIGC neural network model, including: Obtaining a work log of the AIGC neural network model executing AIGC of different modality types; wherein the different modality types include image modality, video modality, and voice modality; Analyze the work log to obtain the data transmission speed of all neural network channels under the AIGC neural network model, and use this as the working status information of the neural network channel; The neural network channel identification module is used to determine all non-essential neural network channels existing in the AIGC neural network model based on the neural network channel working state information, including: The data transmission speed is compared with a preset transmission speed threshold. If the data transmission speed is greater than or equal to the preset transmission speed threshold, the corresponding neural network channel is determined as a necessary neural network channel; otherwise, the corresponding neural network channel is determined as a non-essential neural network channel.

8. The empirical compression pruning system of the AIGC neural network model according to claim 7, characterized in that: Real-time monitoring of the data transmission speed of each neural network channel, and determining whether the overall operation state of the neural network channel is abnormal according to the data transmission speed of each neural network channel, including: Real-time monitoring of the data transmission speed of each neural network channel; Extract the data transmission speed corresponding to each neural network channel; The overall operation coefficient of the neural network channel is obtained by using the data transmission speed corresponding to each neural network channel; The overall operation coefficient of the neural network channel is obtained by the following formula: Where K represents the overall operation coefficient of the neural network channel; n represents the number of neural network channels; G b represents the standard deviation of the channel operation coefficients corresponding to n neural network channels; G z G represents the middle value of the channel operation coefficient corresponding to n neural network channels; p G represents the average value of the channel operation coefficients corresponding to n neural network channels; smax G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; smin G represents the channel operation coefficient corresponding to the neural network channel with the largest number of neurons; i represents the channel operation coefficient corresponding to the i-th neural network channel, and the channel operation coefficient is obtained by the following formula: Among them, G represents the channel operation coefficient corresponding to each neural network channel; m represents the number of neurons contained in each neural network channel; w i represents the weight value of the i-th neuron; P v represents the rate of change of data transmission speed corresponding to each neural network channel; P vb represents the standard deviation of the data transmission speed corresponding to each neural network channel; k represents the number of other neurons connected to the i-th neuron; w j represents the weight value corresponding to the jth neuron among the other neurons connected to the i-th neuron; P vj represents the rate of change of data processing speed of the jth other neuron among the other neurons connected to the i-th neuron; w kmax represents the maximum weight value corresponding to the k other neurons among the other neurons connected to the i-th neuron; P vkb represents the standard deviation of the rate of change of data processing speed of k other neurons among the other neurons connected to the i-th neuron; Comparing the overall operating coefficient of the neural network channel with a preset coefficient threshold; When the overall operating coefficient of the neural network channel exceeds a preset coefficient threshold, it is determined that the overall operating state of the neural network channel is abnormal, and an abnormal risk warning is issued.

9. The empirical compression pruning system of the AIGC neural network model according to claim 6, characterized in that: The neural network channel judgment module is used to obtain data flow feature information of all non-essential neural network channels in different AIGC service scenarios, and based on the data flow feature information, obtain a judgment result of whether all non-essential neural network channels are in a valid working state in different AIGC service scenarios, including: Obtaining data stream transmission frame loss rates of all non-essential neural network channels in different AIGC service scenarios as the data stream feature information; The data stream transmission frame loss rate is compared with a preset frame loss rate threshold. If the data stream transmission frame loss rate is greater than or equal to the preset frame loss rate threshold, the corresponding non-essential neural network channel is determined to be not in a valid working state in the corresponding AIGC service scenario; otherwise, the corresponding non-essential neural network channel is determined to be in a valid working state in the corresponding AIGC service scenario; the neural network channel positioning module is used to perform positioning processing on all non-essential neural network channels to obtain position attribute information of all non-essential neural network channels in the AIGC neural network model, including: All non-essential neural network channels are positioned to obtain the neural network layer position information of all non-essential neural network channels within the AIGC neural network model and the connection node position information with the necessary neural network channels, which is used as the position attribute information.

10. The empirical compression pruning system of the AIGC neural network model according to claim 6, characterized in that: The neural network model empirical identification module is used to identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the current working state information of the AIGC neural network model, so as to determine all non-essential neural network channels in the AIGC neural network model that are not in a valid working state, including: Acquire the current input working data type information of the AIGC neural network model, identify the AIGC service scenario type information currently corresponding to the AIGC neural network model based on the input working data type information, and compare it with the judgment result, so as to determine all unnecessary neural network channels in the AIGC neural network model that are not in a valid working state; The neural network model compression and pruning module is used to compress and prune the AIGC neural network model based on the position attribute information of all non-essential neural network channels that are not in an effective working state, so as to obtain a pruned AIGC neural network model, including: Based on the neural network layer position information of all non-essential neural network channels that are not in a valid working state within the AIGC neural network model and the connection node position information with the necessary neural network channels, all non-essential neural network channels that are not in a valid working state are pruned and removed, and all remaining neural network channels of the AIGC neural network model are reconnected and integrated, so as to obtain a pruned AIGC neural network model.

Images