Semantic segmentation-based large model block transmission method and device

Abstract

The invention relates to the technical field of data processing, and discloses a large model block transmission method and device based on semantic segmentation. The method comprises the following steps: analyzing an overall structure of a large model, and identifying a key function area and a data flow path of the large model; performing semantic segmentation on the large model based on the key function region to form a plurality of semantic blocks; performing dependency definition on each semantic block by utilizing a data flow path to form a plurality of model blocks; evaluating the transmission priority of each model block by dynamically balancing the importance of each model block and the real-time network condition; and transmitting each model block through a network based on the transmission priority of each model block. The method improves the accuracy and efficiency of model segmentation, reduces the network load, optimizes the transmission efficiency and reliability of the large model in the network, and greatly improves the operation stability and efficiency of the model.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a large model block transmission method and device based on semantic segmentation. Background Art

[0002] Large machine learning models often require a lot of network resources during transmission. Currently, large data models are often divided into multiple smaller, logically related blocks or fragments according to their semantic content, and then these blocks are transmitted independently according to network conditions and the needs of the target system.

[0003] However, the current cutting of large models has the problem of low logical correlation, which leads to errors in the functions of the large models during transmission. Moreover, when the network environment is unstable, it is difficult to ensure the integrity and functionality of the large model. Summary of the invention

[0004] The present invention provides a large model block transmission method and device based on semantic segmentation, which improves the accuracy and efficiency of model segmentation, reduces network load, optimizes the transmission efficiency and reliability of large models in the network, and greatly improves the operation stability and efficiency of the model.

[0005] In order to solve the above technical problems, the present invention provides a large model block transmission method based on semantic segmentation, comprising:

[0006] Analyze the overall structure of the large model and identify key functional areas and data flow paths of the large model;

[0007] Performing semantic segmentation on the large model based on the key functional areas to form a plurality of semantic blocks;

[0008] Defining dependencies of the semantic blocks using the data flow path to form a plurality of model blocks;

[0009] Evaluating the transmission priority of each of the model blocks; wherein the transmission priority is obtained by dynamically weighing the importance of each of the model blocks and the real-time network conditions;

[0010] Based on the transmission priority of each model block, each model block is transmitted through the network.

[0011] Furthermore, the dependency of each semantic block is defined by using the data flow path to form a plurality of model blocks, specifically:

[0012] Building a dependency graph of the large model based on the data flow path;

[0013] Based on the dependency graph of the large model, performing function chain position analysis on each of the semantic blocks to evaluate the dependency depth of each of the semantic blocks;

[0014] Performing data dependency tests on each of the semantic blocks respectively, and evaluating the dependency strength of each of the semantic blocks according to the test results;

[0015] Based on the dependency depth and dependency strength of each semantic block, dependency definitions are performed on each semantic block to form a plurality of model blocks.

[0016] Furthermore, the dependency graph of the large model is constructed based on the data flow path, specifically:

[0017] Analyze the data flow path to obtain the data flow and processing logic of the large model;

[0018] Based on the data flow and the processing logic, identifying data dependencies of the large model;

[0019] A dependency graph is constructed according to the data dependency relationship of the large model and the overall structure of the large model.

[0020] Furthermore, the dependency graph based on the large model performs function chain position analysis on each semantic block to evaluate the dependency depth of each semantic block, specifically:

[0021] Analyzing the path length corresponding to each of the semantic blocks in the dependency graph;

[0022] Analyze the function chain position of each semantic block according to the path length corresponding to each semantic block;

[0023] The dependency depth of each semantic block is evaluated according to the function chain position of each semantic block.

[0024] Furthermore, the data dependency test is performed on each of the semantic blocks respectively, and the dependency strength of each of the semantic blocks is evaluated according to the test result, specifically:

[0025] Determine a set of preceding blocks corresponding to each of the semantic blocks based on the dependency graph;

[0026] Performing data dependency tests on each of the semantic blocks respectively;

[0027] quantifying the degree of dependence of each semantic block on the corresponding set of preceding blocks according to the data dependency test results of each semantic block;

[0028] The dependency strength of each semantic block is evaluated according to the dependency degree of each semantic block on the corresponding set of preceding blocks.

[0029] Furthermore, the evaluation of the transmission priority of each model block is specifically as follows:

[0030] Analyze the functional contribution of each model block to the large model to obtain the influence of each model block;

[0031] Calculating the number of parameters of each model block to obtain the number of parameters of each model block;

[0032] Evaluate the importance of each model block according to the influence and parameter amount of each model block;

[0033] Using a real-time network monitoring tool to obtain current network status data, and calculating the network factor of each of the model blocks based on the current network status data;

[0034] The importance and network factors of each model block are dynamically weighed to evaluate the transmission priority of each model block.

[0035] Furthermore, the functional contribution of each model block to the large model is analyzed to obtain the influence of each model block, specifically:

[0036] Performing a disabling test on each of the model blocks in turn, and evaluating the influence of each of the model blocks by comparing the output results of the large model before and after the disabling of each of the model blocks;

[0037] Alternatively, a model block adjustment test is performed on each of the model blocks in turn, and the influence of each model block is evaluated by comparing the output results of the large model before and after the adjustment of each model block; wherein the model block adjustment test includes a model block parameter adjustment test and a model block output adjustment test.

[0038] Furthermore, the importance of each model block is evaluated according to the influence and parameter amount of each model block, specifically:

[0039] C i =a×A i +β×B i

[0040] In the formula, A i is the influence of the i-th model block; B i is the parameter of the i-th model block; C i is the importance of the i-th model block; a and β are the preset weight coefficients.

[0041] Furthermore, the importance of each model block and the network factor are dynamically weighed to evaluate the transmission priority of each model block, specifically:

[0042] E i =γ×(ρ×D i +(1-ρ)×C i )

[0043] In the formula, Ci is the importance of the i-th model block; D i is the network factor of the ith model block; E i is the transmission priority of the i-th model block; ρ is the preset adjustable trade-off factor; γ is the preset adjustment coefficient.

[0044] The present invention provides a large model block transmission method based on semantic segmentation, which analyzes the overall structure of the large model, identifies the key functional areas and data flow paths of the large model; semantically segments the large model based on the key functional areas to form a number of semantic blocks; defines the dependencies of each semantic block using the data flow path to form a number of model blocks; evaluates the transmission priority of each model block by dynamically weighing the importance of each model block and the real-time network conditions; and transmits each model block through the network based on the transmission priority of each model block. The present invention improves the accuracy and efficiency of model segmentation by subdividing a large machine learning model into multiple independent functional blocks, and dynamically adjusts the transmission priority of each block according to the importance of each block and the network conditions, thereby reducing the network load, optimizing the transmission efficiency and reliability of the large model in the network, ensuring the integrity and functionality of the model during transmission, and greatly improving the operational stability and efficiency of the model.

[0045] Accordingly, the present invention provides a large model block transmission device based on semantic segmentation, comprising: a recognition module, a semantic segmentation module, a dependency definition module, a priority evaluation module and a transmission module;

[0046] The identification module is used to analyze the overall structure of the large model and identify the key functional areas and data flow paths of the large model;

[0047] The semantic segmentation module is used to perform semantic segmentation on the large model based on the key functional areas to form a plurality of semantic blocks;

[0048] The dependency definition module is used to define the dependencies of each semantic block using the data flow path to form a plurality of model blocks;

[0049] The priority evaluation module is used to evaluate the transmission priority of each model block; wherein the transmission priority is obtained by dynamically weighing the importance of each model block and the real-time network conditions;

[0050] The transmission module is used to transmit each of the model blocks through a network based on the transmission priority of each of the model blocks.

[0051] The present invention provides a large model block transmission device based on semantic segmentation. Based on the organic combination of modules, the overall structure of the large model is analyzed to identify the key functional areas and data flow paths of the large model; the large model is semantically segmented based on the key functional areas to form a number of semantic blocks; the dependency of each semantic block is defined using the data flow path to form a number of model blocks; the transmission priority of each model block is evaluated by dynamically weighing the importance of each model block and the real-time network conditions; and each model block is transmitted through the network based on the transmission priority of each model block. The present invention improves the accuracy and efficiency of model segmentation by subdividing a large machine learning model into multiple independent functional blocks, and dynamically adjusts the transmission priority of each block according to the importance of each block and the network conditions, thereby reducing the network load, optimizing the transmission efficiency and reliability of the large model in the network, ensuring the integrity and functionality of the model during transmission, and greatly improving the operational stability and efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of an embodiment of a large model block transmission method based on semantic segmentation provided by the present invention;

[0053] Figure 2 A flowchart of another embodiment of the large model block transmission method based on semantic segmentation provided by the present invention;

[0054] Figure 3 A flowchart of another embodiment of the large model block transmission method based on semantic segmentation provided by the present invention;

[0055] Figure 4 A schematic structural diagram of an embodiment of a large model block transmission device based on semantic segmentation provided by the present invention. DETAILED DESCRIPTION

[0056] 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.

[0057] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0058] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0059] Example 1

[0060] See also Figure 1 , is a flow chart of an embodiment of a large model block transmission method based on semantic segmentation provided by the present invention, the method includes steps 101 to 105, each step is specifically as follows:

[0061] Step 101: Analyze the overall structure of the large model and identify the key functional areas and data flow paths of the large model.

[0062] In the first embodiment of the present invention, a large-scale machine learning model is semantically segmented and can be segmented into multiple independent functional blocks. Specifically, the overall structure of the large model is first analyzed, and the organization of each level and logical unit of the large model is analyzed, so as to identify the key functional areas and data flow paths in the large model, so as to segment the large model into multiple independent parts according to the identified key functional areas and data flow paths of the large model.

[0063] Step 102: semantically segment the large model based on the key functional areas to form a plurality of semantic blocks.

[0064] In the first embodiment of the present invention, the key functional areas obtained by identifying the large model are used for model segmentation, and the large model is divided into multiple independent semantic blocks based on the key functional areas. Each divided semantic block includes a set of functionally interdependent layers or units, so that each semantic block maintains complete functionality and dependency, ensuring that no functionality is lost during block transmission and subsequent operations.

[0065] Step 103: Dependencies are defined for each of the semantic blocks using the data flow path to form a number of model blocks.

[0066] Furthermore, in the first embodiment of the present invention, the data flow path is used to define the dependencies of the semantic blocks to form a plurality of model blocks, specifically:

[0067] Building a dependency graph of the large model based on the data flow path;

[0068] Based on the dependency graph of the large model, performing function chain position analysis on each of the semantic blocks to evaluate the dependency depth of each of the semantic blocks;

[0069] Performing data dependency tests on each of the semantic blocks respectively, and evaluating the dependency strength of each of the semantic blocks according to the test results;

[0070] Based on the dependency depth and dependency strength of each semantic block, dependency definitions are performed on each semantic block to form a plurality of model blocks.

[0071] In the first embodiment of the present invention, the interface and dependency relationship for defining each semantic block are obtained by identifying the large model, thereby ensuring that the data and control flow between blocks can be correctly reorganized and synchronized, so that the function and structure of the entire model can be seamlessly rebuilt at the receiving end later. Among them, the dependency of each block is defined as the dependency relationship of a semantic block on other blocks, which is specifically manifested in that without receiving the preceding semantic block, the block cannot independently complete the predetermined computing task or demonstrate its complete function.

[0072] In the first embodiment of the present invention, the evaluation process of the dependency of each semantic block mainly includes identifying the predecessor block, quantifying the dependency depth and evaluating the dependency strength. Among them, the predecessor block refers to the block that must be processed and transmitted before the current block can perform its function. Identifying the set of predecessor blocks that each semantic block depends on can be achieved by analyzing the data flow and processing logic of the large model. Quantifying the dependency depth is achieved through the path length of the semantic block on the dependency graph. The longer the path, the greater the dependency depth, and the more the corresponding semantic block depends on the predecessor block in terms of functional implementation. Evaluating the dependency strength is achieved through data dependency testing in simulation or actual operation. If the test is greatly affected by the predecessor semantic block, it means that the dependency strength of the semantic block is high.

[0073] Furthermore, in the first embodiment of the present invention, the dependency graph of the large model is constructed based on the data flow path, specifically:

[0074] Analyze the data flow path to obtain the data flow and processing logic of the large model;

[0075] Based on the data flow and the processing logic, identifying data dependencies of the large model;

[0076] A dependency graph is constructed according to the data dependency relationship of the large model and the overall structure of the large model.

[0077] In the first embodiment of the present invention, the dependency evaluation of the semantic block can be achieved by analyzing the dependency graph, so before evaluating the dependency of the semantic block, a dependency graph is first constructed based on the data flow path. Specifically, the data flow and processing logic can be obtained by analyzing the data flow path, and the set of pre-blocks of each semantic block can be identified by analyzing the data flow and processing logic of the large model. The data dependency relationship can be automatically identified using corresponding tools and algorithms. Therefore, the dependency graph of the large model can be constructed by combining the pre-block set of each semantic block and the identified data dependency relationship.

[0078] Furthermore, in the first embodiment of the present invention, based on the dependency graph of the large model, a function chain position analysis is performed on each of the semantic blocks to evaluate the dependency depth of each of the semantic blocks, specifically:

[0079] Analyzing the path length corresponding to each of the semantic blocks in the dependency graph;

[0080] Analyze the function chain position of each semantic block according to the path length corresponding to each semantic block;

[0081] The dependency depth of each semantic block is evaluated according to the function chain position of each semantic block.

[0082] In the first embodiment of the present invention, after the constructed dependency graph is obtained, the position of each semantic block in the function chain of the entire large model is analyzed in the dependency graph, and the path length of each semantic block can be obtained. Since the semantic block with a higher dependency depth is located later in the function chain, the later the semantic block is in the execution path, the more its function realization depends on the processing and data preparation of the previous blocks, so by analyzing the path length of each semantic block, the dependency depth of each semantic block can be quantified.

[0083] Furthermore, in the first embodiment of the present invention, data dependency tests are performed on each of the semantic blocks respectively, and the dependency strength of each of the semantic blocks is evaluated according to the test results, specifically:

[0084] Determine a set of preceding blocks corresponding to each of the semantic blocks based on the dependency graph;

[0085] Performing data dependency tests on each of the semantic blocks respectively;

[0086] quantifying the degree of dependence of each semantic block on the corresponding set of preceding blocks according to the data dependency test results of each semantic block;

[0087] The dependency strength of each semantic block is evaluated according to the dependency degree of each semantic block on the corresponding set of preceding blocks.

[0088] In the first embodiment of the present invention, if a semantic block with high dependency strength does not receive correctly processed predecessor block data, its functional performance or calculation results will be seriously affected. Therefore, by testing the data dependency test, the influence of the predecessor item set corresponding to each semantic block on it can be tested, and the degree of dependency of each semantic block on its predecessor block can be further evaluated. For example, some predecessor blocks of the test semantic block are disabled, and the output results of the test semantic block before and after the predecessor blocks are disabled are analyzed, and the dependency strength of the test semantic block is evaluated through the test results.

[0089] In the first embodiment of the present invention, through detailed large model analysis and precise semantic block definition, the functional integrity and operational dependency of each semantic block after semantic segmentation can be ensured, so that the function and structure of the model can be seamlessly reconstructed at the receiving end, greatly improving the operational stability and efficiency of the model. Moreover, the use of automated tools and algorithms to annotate and segment the model can reduce human errors and improve the accuracy and efficiency of segmentation.

[0090] Step 104: Evaluate the transmission priority of each of the model blocks; wherein the transmission priority is obtained by dynamically weighing the importance of each of the model blocks and the real-time network conditions.

[0091] In the first embodiment of the present invention, the transmission priority of each segmented model block is evaluated, and the key parts are transmitted first during transmission, which can maximize the transmission efficiency and model performance, thereby improving the transmission efficiency and deployment speed of large machine learning models and ensuring the stability and performance of the model when network conditions change. Specifically, the transmission priority of the model block is evaluated based on the importance of the model block and the current network conditions.

[0092] Further, in the first embodiment of the present invention, the transmission priority of each model block is evaluated, specifically:

[0093] Analyze the functional contribution of each model block to the large model to obtain the influence of each model block;

[0094] Calculating the number of parameters of each model block to obtain the number of parameters of each model block;

[0095] Evaluate the importance of each model block according to the influence and parameter amount of each model block;

[0096] Using a real-time network monitoring tool to obtain current network status data, and calculating the network factor of each of the model blocks based on the current network status data;

[0097] The importance and network factors of each model block are dynamically weighed to evaluate the transmission priority of each model block.

[0098] In the first embodiment of the present invention, the transmission priority of the model block is determined by the importance of the model block and the current network conditions, and the importance of the model block is determined by the influence of the model block and the number of parameters of the model block, so that the part that has the greatest impact on the overall model performance can be given priority during the transmission process, thereby improving the model operation efficiency and accuracy.

[0099] Furthermore, in the first embodiment of the present invention, the functional contribution of each model block to the large model is analyzed to obtain the influence of each model block, specifically:

[0100] Performing a disabling test on each of the model blocks in turn, and evaluating the influence of each of the model blocks by comparing the output results of the large model before and after the disabling of each of the model blocks;

[0101] Alternatively, a model block adjustment test is performed on each of the model blocks in turn, and the influence of each model block is evaluated by comparing the output results of the large model before and after the adjustment of each model block; wherein the model block adjustment test includes a model block parameter adjustment test and a model block output adjustment test.

[0102] In the first embodiment of the present invention, the influence of the model block refers to the contribution of the model block to the overall performance of the model, which can be quantitatively determined by analyzing the impact of the model block on the final output accuracy or computational efficiency. Specifically, the influence of the model block can be evaluated based on several aspects such as output accuracy, computational efficiency, number of parameters and functional contribution. Among them, the output accuracy can be analyzed block by block by block to analyze the impact of each model block on the final output accuracy of the model, thereby obtaining the contribution of each model block to the overall prediction accuracy. The computational efficiency can be obtained by analyzing the contribution of each model block in terms of computing resources (such as CPU / GPU time) and memory usage to obtain the impact of each model block on the model efficiency during the calculation process. Since the model block with a large number of parameters contains more learning information, the impact of each model block on the performance and final result of the large model can be obtained by calculating the total number of parameters inside each model block. Functional contribution is to evaluate the role of each model block in the model decision process. For example, if a model block is responsible for the extraction of specific features or the prediction of a specific type, the functional impact of the corresponding model block on the large model can be obtained based on the role.

[0103] In the first embodiment of the present invention, the influence of the model block can be calculated by using a disabling test or a sensitivity analysis. Among them, the disabling test is to disable the model blocks in sequence, run the large model, analyze the output results of the large model before and after disabling the model block, and obtain the influence of the model block. If the performance of the large model drops significantly after disabling the model block, the contribution of the model block is relatively large and the contribution is relatively high. Sensitivity analysis is to change the output or parameters of a certain model block, analyze the output results of the large model before and after changing the model block, and obtain the influence of the model block. If the change in the final output of the model after changing the model block is greater, the contribution of the model block to the large model is greater and the contribution is relatively high.

[0104] Furthermore, in the first embodiment of the present invention, the importance of each model block is evaluated according to the influence and parameter amount of each model block, specifically:

[0105] C i =a×A i +β×B i

[0106] In the formula, Ai is the influence of the i-th model block; B i is the parameter of the i-th model block; C i is the importance of the i-th model block; a and β are the preset weight coefficients.

[0107] In the first embodiment of the present invention, by measuring the number of parameters contained in each model block, the parameter quantity of each model block can be obtained. The influence and parameter quantity of the model block are the determining factors of the importance of the model block. The importance of the model block can be calculated according to the above formula, where a and β are preset weight coefficients, which provide a mechanism to adjust the proportion of performance contribution and number of parameters in the total importance evaluation, so that the model administrator can optimize these weights according to the specific application scenario and the characteristics of the model.

[0108] In the first embodiment of the present invention, by defining a specific scoring mechanism, the influence and the number of parameters are combined to calculate the importance of each model block. By accurately calculating the importance of each model block, limited network resources can be allocated more reasonably to ensure that those blocks that have the greatest impact on model performance can be transmitted first. In the case of limited network conditions or the need for rapid deployment, the priority transmission of model blocks with high importance can quickly restore the core functions of the model and accelerate the deployment and application of the model. By adjusting the value of the weight coefficient, different application requirements and network environments can be flexibly responded to, making the model transmission more adaptable to the current specific conditions.

[0109] Furthermore, in the first embodiment of the present invention, the importance of each model block and the network factor are dynamically weighed to evaluate the transmission priority of each model block, specifically:

[0110] E i =γ×(ρ×D i +(1-ρ)×C i )

[0111] In the formula, C i is the importance of the i-th model block; D i is the network factor of the ith model block; E i is the transmission priority of the i-th model block; ρ is the preset adjustable trade-off factor; γ is the preset adjustment coefficient.

[0112] In the first embodiment of the present invention, the transmission priority of the model block is determined by the importance of the model block and the current network status. The importance of the model block can be calculated based on the influence of the above model block and the parameter amount of the model block, and the current network status can be obtained using real-time network monitoring tools. These tools can obtain parameters related to the current network conditions, such as bandwidth utilization, network delay or packet loss rate, etc. These factors will directly affect the transmission efficiency of the block. ρ in the above transmission priority calculation formula is a trade-off factor between 0 and 1, which is used to adjust the relative importance of network conditions and block importance in the transmission priority; γ is a preset adjustment coefficient, which is used to calibrate the absolute value of the transmission priority to ensure that it is suitable for the priority processing system in practical applications.

[0113] In the first embodiment of the present invention, the parameters ρ and γ can be adjusted according to actual operational requirements to adapt to different network environments and service priorities, ensuring that the most critical model blocks can be transmitted preferentially even when the network status is poor.

[0114] In the first embodiment of the present invention, the transmission priority is evaluated taking into account the functional importance of the model block and the network status (such as bandwidth and latency), which can more effectively utilize network resources and avoid large amounts of data transmission when the network condition is poor, thereby reducing congestion and delays, ensuring that important model blocks can be quickly transmitted and deployed when needed, especially in applications with high real-time requirements, which can significantly improve the system's response speed and user experience.

[0115] Step 105: Based on the transmission priority of each model block, each model block is transmitted through the network.

[0116] In the first embodiment of the present invention, after the large model is semantically segmented to form multiple model blocks and the transmission priority of each model block is evaluated, the model blocks are transmitted one by one or concurrently through the network according to the evaluated transmission priority. Figure 2 , is a flow chart of another embodiment of the large model block transmission method based on semantic segmentation provided by the present invention. Before transmitting the model block, the data packets of each model block are compressed to reduce the required transmission bandwidth and time. After the data packets of each model block are compressed, a suitable protocol is selected to transmit the compressed packet to the target receiving end. After the receiving end receives the compressed packet, it decompresses it to obtain each model block, and then reassembles and verifies each model block, processes the erroneous model block, and obtains the correct reassembled model block.

[0117] In the first embodiment of the present invention, compression of the model block can be achieved through various data compression algorithms, such as ZIP, GZIP or other compression technologies that are more specialized for specific data types, thereby significantly reducing the size of data transmission, thereby reducing network load and related costs, and speeding up data transmission, especially in the case of limited bandwidth, so that the model block can be transmitted in a shorter time.

[0118] In the first embodiment of the present invention, when the model blocks are transmitted to the receiving end, the receiving end verifies and reassembles them to ensure data integrity and model functionality. Figure 3 , is a flow chart of another embodiment of the large model block transmission method based on semantic segmentation provided by the present invention. When the receiving end receives the model block, the received data is verified by digital signature or other integrity verification method to ensure that the data is not damaged or tampered with during the transmission process, and all model blocks can be correctly spliced. Decompress the received data, reorganize the model blocks, and restore the complete structure and function of the model. Perform functional testing and system integration testing on the reorganized model blocks to implement security checks and integrity confirmation, which can avoid functional errors caused by data error propagation, enhance the reliability and stability of the system, and ensure that the model is correctly executed according to the design requirements.

[0119] In the first embodiment of the present invention, after each model block is transmitted, the transmission record is archived, and these records can be used for subsequent performance analysis and network optimization. The archived data may include transmission time, success or failure status, bandwidth used, etc. These data are crucial for analyzing transmission efficiency and identifying possible bottlenecks.

[0120] In the first embodiment of the present invention, the present invention is applicable to online updating, remote deployment or real-time adjustment of machine learning models in edge computing environments. The present invention can support rapid deployment and update operations, adapt to changes in network conditions and limitations of computing resources, and especially in edge computing environments, can achieve rapid response and processing near the user end, making model updates and deployments more flexible and rapid, and is particularly suitable for application scenarios that require rapid iteration and instant response, such as IoT and mobile devices, etc., which enhances the adaptability and real-time performance of the system.

[0121] In summary, the first embodiment of the present invention provides a large model block transmission method based on semantic segmentation, which analyzes the overall structure of the large model, identifies the key functional areas and data flow paths of the large model; semantically segments the large model based on the key functional areas to form a number of semantic blocks; defines the dependencies of each semantic block using the data flow path to form a number of model blocks; evaluates the transmission priority of each model block by dynamically weighing the importance of each model block and the real-time network conditions; and transmits each model block through the network based on the transmission priority of each model block. The present invention improves the accuracy and efficiency of model segmentation by subdividing a large machine learning model into multiple independent functional blocks, and dynamically adjusts the transmission priority of each block according to the importance of each block and the network conditions, thereby reducing the network load, optimizing the transmission efficiency and reliability of the large model in the network, ensuring the integrity and functionality of the model during transmission, and greatly improving the operational stability and efficiency of the model.

[0122] Example 2

[0123] See also Figure 4 , is a structural diagram of an embodiment of a large model block transmission device based on semantic segmentation provided by the present invention, the device includes a recognition module 201, a semantic segmentation module 202, a dependency definition module 203, a priority evaluation module 204 and a transmission module 205;

[0124] The identification module 201 is used to analyze the overall structure of the large model and identify the key functional areas and data flow paths of the large model;

[0125] The semantic segmentation module 202 is used to perform semantic segmentation on the large model based on the key functional areas to form a plurality of semantic blocks;

[0126] The dependency definition module 203 is used to define the dependencies of each semantic block using the data flow path to form a plurality of model blocks;

[0127] The priority evaluation module 204 is used to evaluate the transmission priority of each of the model blocks; wherein the transmission priority is obtained by dynamically weighing the importance of each of the model blocks and the real-time network conditions;

[0128] The transmission module 205 is used to transmit each of the model blocks through a network based on the transmission priority of each of the model blocks.

[0129] Furthermore, in the second embodiment of the present invention, the data flow path is used to define the dependencies of the semantic blocks to form a plurality of model blocks, specifically:

[0130] Building a dependency graph of the large model based on the data flow path;

[0131] Based on the dependency graph of the large model, performing function chain position analysis on each of the semantic blocks to evaluate the dependency depth of each of the semantic blocks;

[0132] Performing data dependency tests on each of the semantic blocks respectively, and evaluating the dependency strength of each of the semantic blocks according to the test results;

[0133] Based on the dependency depth and dependency strength of each semantic block, dependency definitions are performed on each semantic block to form a plurality of model blocks.

[0134] Furthermore, in the second embodiment of the present invention, a dependency graph of the large model is constructed based on the data flow path, specifically:

[0135] Analyze the data flow path to obtain the data flow and processing logic of the large model;

[0136] Based on the data flow and the processing logic, identifying data dependencies of the large model;

[0137] A dependency graph is constructed according to the data dependency relationship of the large model and the overall structure of the large model.

[0138] Furthermore, in the second embodiment of the present invention, based on the dependency graph of the large model, a function chain position analysis is performed on each of the semantic blocks to evaluate the dependency depth of each of the semantic blocks, specifically:

[0139] Analyzing the path length corresponding to each of the semantic blocks in the dependency graph;

[0140] Analyze the function chain position of each semantic block according to the path length corresponding to each semantic block;

[0141] The dependency depth of each semantic block is evaluated according to the function chain position of each semantic block.

[0142] Furthermore, in the second embodiment of the present invention, data dependency tests are performed on each of the semantic blocks respectively, and the dependency strength of each of the semantic blocks is evaluated according to the test results, specifically:

[0143] Determine a set of preceding blocks corresponding to each of the semantic blocks based on the dependency graph;

[0144] Performing data dependency tests on each of the semantic blocks respectively;

[0145] quantifying the degree of dependence of each semantic block on the corresponding set of preceding blocks according to the data dependency test results of each semantic block;

[0146] The dependency strength of each semantic block is evaluated according to the dependency degree of each semantic block on the corresponding set of preceding blocks.

[0147] Furthermore, in the second embodiment of the present invention, the transmission priority of each model block is evaluated, specifically:

[0148] Analyze the functional contribution of each model block to the large model to obtain the influence of each model block;

[0149] Calculating the number of parameters of each model block to obtain the number of parameters of each model block;

[0150] Evaluate the importance of each model block according to the influence and parameter amount of each model block;

[0151] Using a real-time network monitoring tool to obtain current network status data, and calculating the network factor of each of the model blocks based on the current network status data;

[0152] The importance and network factors of each model block are dynamically weighed to evaluate the transmission priority of each model block.

[0153] Furthermore, in the second embodiment of the present invention, the functional contribution of each model block to the large model is analyzed to obtain the influence of each model block, specifically:

[0154] Performing a disabling test on each of the model blocks in turn, and evaluating the influence of each of the model blocks by comparing the output results of the large model before and after the disabling of each of the model blocks;

[0155] Alternatively, a model block adjustment test is performed on each of the model blocks in turn, and the influence of each model block is evaluated by comparing the output results of the large model before and after the adjustment of each model block; wherein the model block adjustment test includes a model block parameter adjustment test and a model block output adjustment test.

[0156] Furthermore, in the second embodiment of the present invention, the importance of each model block is evaluated according to the influence and parameter amount of each model block, specifically:

[0157] C i =a×A i +β×B i

[0158] In the formula, A i is the influence of the i-th model block; B i is the parameter of the i-th model block; C i is the importance of the i-th model block; a and β are the preset weight coefficients.

[0159] Furthermore, in the second embodiment of the present invention, the importance of each model block and the network factor are dynamically weighed to evaluate the transmission priority of each model block, specifically:

[0160] E i =γ×(ρ×D i +(1-ρ)×C i )

[0161] In the formula, C i is the importance of the i-th model block; D i is the network factor of the ith model block; E i is the transmission priority of the i-th model block; ρ is the preset adjustable trade-off factor; γ is the preset adjustment coefficient.

[0162] In summary, the second embodiment of the present invention provides a large model block transmission device based on semantic segmentation, which analyzes the overall structure of the large model based on the organic combination of modules, identifies the key functional areas and data flow paths of the large model; semantically segments the large model based on the key functional areas to form a number of semantic blocks; defines the dependencies of each semantic block using the data flow path to form a number of model blocks; evaluates the transmission priority of each model block by dynamically weighing the importance of each model block and the real-time network conditions; and transmits each model block through the network based on the transmission priority of each model block. The present invention improves the accuracy and efficiency of model segmentation by subdividing a large machine learning model into multiple independent functional blocks, and dynamically adjusts the transmission priority of each block according to the importance of each block and the network conditions, thereby reducing the network load, optimizing the transmission efficiency and reliability of the large model in the network, ensuring the integrity and functionality of the model during transmission, and greatly improving the operational stability and efficiency of the model.

[0163] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A large model block transmission method based on semantic segmentation, characterized in that: include: Analyze the overall structure of the large model and identify key functional areas and data flow paths of the large model; Performing semantic segmentation on the large model based on the key functional areas to form a plurality of semantic blocks; Defining dependencies of the semantic blocks using the data flow path to form a plurality of model blocks; Evaluating the transmission priority of each of the model blocks; wherein the transmission priority is obtained by dynamically weighing the importance of each of the model blocks and the real-time network conditions; Based on the transmission priority of each model block, each model block is transmitted through the network.

2. The large model block transmission method based on semantic segmentation according to claim 1 is characterized in that: The dependency definition of each semantic block is performed by using the data flow path to form a plurality of model blocks, specifically: Building a dependency graph of the large model based on the data flow path; Based on the dependency graph of the large model, performing function chain position analysis on each of the semantic blocks to evaluate the dependency depth of each of the semantic blocks; Performing data dependency tests on each of the semantic blocks respectively, and evaluating the dependency strength of each of the semantic blocks according to the test results; Based on the dependency depth and dependency strength of each semantic block, dependency definitions are performed on each semantic block to form a plurality of model blocks.

3. The large model block transmission method based on semantic segmentation according to claim 2 is characterized in that: The dependency graph of the large model is constructed based on the data flow path, specifically: Analyze the data flow path to obtain the data flow and processing logic of the large model; Based on the data flow and the processing logic, identifying data dependencies of the large model; A dependency graph is constructed according to the data dependency relationship of the large model and the overall structure of the large model.

4. The large model block transmission method based on semantic segmentation according to claim 2 is characterized in that: The dependency graph based on the large model performs function chain position analysis on each semantic block and evaluates the dependency depth of each semantic block, specifically: Analyzing the path length corresponding to each of the semantic blocks in the dependency graph; Analyze the function chain position of each semantic block according to the path length corresponding to each semantic block; The dependency depth of each semantic block is evaluated according to the function chain position of each semantic block.

5. The large model block transmission method based on semantic segmentation according to claim 2 is characterized in that: The data dependency test is performed on each of the semantic blocks respectively, and the dependency strength of each of the semantic blocks is evaluated according to the test results, specifically: Determine a set of preceding blocks corresponding to each of the semantic blocks based on the dependency graph; Performing data dependency tests on each of the semantic blocks respectively; quantifying the degree of dependence of each semantic block on the corresponding set of preceding blocks according to the data dependency test results of each semantic block; The dependency strength of each semantic block is evaluated according to the dependency degree of each semantic block on the corresponding set of preceding blocks.

6. The large model block transmission method based on semantic segmentation according to claim 1 is characterized in that: The evaluation of the transmission priority of each model block is specifically as follows: Analyze the functional contribution of each model block to the large model to obtain the influence of each model block; Calculating the number of parameters of each model block to obtain the number of parameters of each model block; Evaluate the importance of each model block according to the influence and parameter amount of each model block; Using a real-time network monitoring tool to obtain current network status data, and calculating the network factor of each of the model blocks based on the current network status data; The importance and network factors of each model block are dynamically weighed to evaluate the transmission priority of each model block.

7. The large model block transmission method based on semantic segmentation according to claim 6 is characterized in that: The analysis of the functional contribution of each model block to the large model to obtain the influence of each model block is specifically as follows: Performing a disabling test on each of the model blocks in turn, and evaluating the influence of each of the model blocks by comparing the output results of the large model before and after the disabling of each of the model blocks; Alternatively, a model block adjustment test is performed on each of the model blocks in turn, and the influence of each model block is evaluated by comparing the output results of the large model before and after the adjustment of each model block; wherein the model block adjustment test includes a model block parameter adjustment test and a model block output adjustment test.

8. The large model block transmission method based on semantic segmentation according to claim 6 is characterized in that: According to the influence and parameter amount of each model block, the importance of each model block is evaluated, specifically: C i =a×A i +β×B i In the formula, A i is the influence of the i-th model block; B i is the parameter of the i-th model block; C i is the importance of the i-th model block; a and β are the preset weight coefficients.

9. The large model block transmission method based on semantic segmentation according to claim 6 is characterized in that: The importance of each model block and the network factor are dynamically weighed to evaluate the transmission priority of each model block, specifically: E i =γ×(ρ×D i +(1-ρ)×C i ) In the formula, C i is the importance of the i-th model block; D i is the network factor of the ith model block; E i is the transmission priority of the i-th model block; ρ is the preset adjustable trade-off factor; γ is the preset adjustment coefficient.

10. A large model block transmission device based on semantic segmentation, characterized in that: include: Recognition module, semantic segmentation module, dependency definition module, priority evaluation module and transmission module; The identification module is used to analyze the overall structure of the large model and identify the key functional areas and data flow paths of the large model; The semantic segmentation module is used to perform semantic segmentation on the large model based on the key functional areas to form a plurality of semantic blocks; The dependency definition module is used to define the dependencies of each semantic block using the data flow path to form a plurality of model blocks; The priority evaluation module is used to evaluate the transmission priority of each model block; wherein the transmission priority is obtained by dynamically weighing the importance of each model block and the real-time network conditions; The transmission module is used to transmit each of the model blocks through a network based on the transmission priority of each of the model blocks.

Images