Edge AI chip performance evaluation method, device and equipment and storage medium

By using quantile gradient descent self-organized mapping network in chip performance evaluation, handling outliers and noises, and adaptively adjusting parameters, the problems of inaccurate evaluation and insufficient flexibility in the prior art are solved, and AI chips are realized efficient and accurate performance evaluation.

CN118940085APending Publication Date: 2024-11-12XINGHE ZHILIAN AUTOMOBILE TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410947911.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-12

Smart Images

  • Figure CN118940085A_ABST
    Figure CN118940085A_ABST
Patent Text Reader

Abstract

The invention discloses an edge AI chip performance evaluation method, device and equipment and a storage medium, and the method comprises the steps: obtaining performance data from an actually running edge AI device, marking the performance data to obtain sample data, and enabling the performance data to comprise performance indexes of the AI chip when the AI chip processes various calculation tasks; the sample data is input into a self-organizing mapping network based on quantile gradient descent for training, a trained classifier is obtained, and the training process comprises model initialization, input data mapping, node competitive learning, neighborhood node updating, model parameter updating and model iteration training; and obtaining to-be-tested data and inputting the to-be-tested data into the classifier for classification to obtain a performance category classification result of the AI chip. By the adoption of the AI chip evaluation method and device, abnormal values and noise in data can be effectively processed, model parameters can be adjusted in a self-adaptive mode, the robustness and flexibility of the model are improved, and efficient and accurate evaluation of the AI chip is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of chip testing technology, and in particular to an edge AI chip performance evaluation method, device, equipment and storage medium. Background Art

[0002] Existing chip performance evaluation technologies are often not robust enough when dealing with outliers and noise in data, and are easily affected by them, resulting in inaccurate evaluation results. In addition, model parameter adjustments in existing technologies are usually relatively fixed and lack flexibility, which limits the learning efficiency of the model at different stages. Summary of the invention

[0003] The present invention provides an edge AI chip performance evaluation method, device, equipment and storage medium to solve the problems existing in the above-mentioned existing chip performance evaluation technology and improve the accuracy and efficiency of AI chip evaluation.

[0004] To achieve the above objectives, an embodiment of the present invention provides an edge AI chip performance evaluation method, comprising:

[0005] Obtain performance data from the actual running edge AI device and annotate it to obtain sample data, wherein the performance data includes performance indicators of the AI ​​chip when processing various computing tasks;

[0006] Inputting the sample data into a self-organizing map network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training;

[0007] The data to be tested is obtained and input into the classifier for classification to obtain the performance category classification result of the AI ​​chip.

[0008] As an improvement of the above solution, the self-organizing map network is composed of a plurality of nodes, and the process of initializing the model includes:

[0009] Initializing the weights and parameters of the self-organizing map network, wherein the parameters include the location of the nodes, the initial learning rate, and the neighborhood size;

[0010] Based on the random allocation method, the weight of each node is updated by a first formula, wherein the first formula is specifically:

[0011]

[0012] Among them, w i (0) is the initial weight vector of each node i; random_init() is the random allocation function: is the weight vector for each node i, where d is the dimension of the feature input to the classifier; p n is the total number of nodes in the self-organizing map network.

[0013] As an improvement of the above solution, the process of input data mapping includes:

[0014] Inputting the sample data into the self-organizing map network that completes the model initialization;

[0015] Each data point of the sample data will be mapped to a node of the self-organizing map network;

[0016] In each iteration, a sample data is randomly selected from the sample data set after dimensionality reduction and input into the self-organizing map network.

[0017] As an improvement of the above solution, the node competition learning process includes:

[0018] Calculate the similarity between each sample data and each node through a similarity algorithm, and select the node with the greatest similarity as the winner node of the sample data;

[0019] The similarity algorithm includes the Euclidean distance method, and the winner node of the sample data is calculated by the second formula, wherein the second formula is specifically:

[0020]

[0021] Where b is the winning node that is most similar to the sample data x(t); || || is the Euclidean distance; x(t) is the number of nodes in the tth iteration from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node.

[0022] As an improvement of the above solution, the process of updating the neighborhood nodes includes:

[0023] Calculate the distance between each of the nodes and the winner node in the self-organizing map grid by a third formula;

[0024] According to the distance and the perceptual annealing strategy, the weights of the neighboring nodes are calculated and updated by the fourth formula;

[0025] Wherein, the third formula is specifically:

[0026]

[0027] Among them, the neighborhood function h ib(t) is the distance function between node i and the winner node b; r i is the position of node i in the self-organizing map network (x i ,y i ), r b is the position of the winner node b in the self-organizing map network (x b ,y b ); σ(t) is the neighborhood width before updating, which decreases with the number of iterations; t is the index of the number of iterations; Fo() is the distance calculation function, is the distance between node i and the winner node b in the self-organizing map grid;

[0028] The fourth formula is specifically:

[0029] w i (t+1)=w i (t)+α(t)·h ib (t)·(x(t)-w i (t))

[0030] Among them, α(t) is the learning rate before updating, which decreases with the number of iterations; the neighborhood function h ib (t) is the distance function between node i and the winner node b; x(t) is the number of samples from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node; w i (t+1) is the updated weight vector of the i-th node.

[0031] As an improvement of the above solution, the process of updating the model parameters includes:

[0032] Based on the quantile gradient descent and the perceptual annealing strategy, the updated learning rate and neighborhood width are calculated by the fifth formula and dynamically adjusted to optimize the learning process of the self-organizing map network;

[0033] Wherein, the fifth formula is specifically:

[0034]

[0035] Among them, λ α is the time constant for controlling the learning rate annealing rate; σ is the time constant for controlling the annealing rate of the neighborhood width; α(t) is the learning rate before the update; σ(t) is the neighborhood width before the update; α(t+1) is the learning rate after the update; σ(t+1) is the neighborhood width after the update; t is the index of the number of iterations;

[0036] Based on the exponential decay method, the time constant λ is calculated by the sixth formula α And make adjustments, wherein the fifth formula is specifically:

[0037]

[0038] Among them, λ α is the time constant for controlling the learning rate annealing rate; α min is the minimum value of the learning rate, ensuring that the learning rate does not decay to 0; α(0) is the initial learning rate; T is the expected number of training cycles;

[0039] Based on the exponential decay method, the time constant λ is calculated by the seventh formula σ And make adjustments, wherein the seventh formula is specifically:

[0040]

[0041] Among them, λ σ is the time constant for controlling the annealing rate of the neighborhood width; σ min is the minimum value of the neighborhood width, ensuring that the neighborhood function does not shrink to only include the winner node itself; σ(0) is the initial neighborhood width; T is the expected number of training cycles.

[0042] As an improvement of the above solution, the process of iterative training of the model includes:

[0043] The process of node competition learning, neighborhood node updating and model parameter updating is repeated until the self-organizing map network converges or reaches a preset number of iterations.

[0044] To achieve the above objectives, an embodiment of the present invention further provides an edge AI chip performance evaluation device, comprising:

[0045] A data acquisition module is used to obtain performance data from the actual running edge AI device and annotate it to obtain sample data, wherein the performance data includes the performance indicators of the AI ​​chip when processing various computing tasks;

[0046] The model training module is used to input the sample data into the self-organizing map network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training:

[0047] The performance evaluation module is used to obtain the data to be tested and input it into the classifier for classification to obtain the performance category classification result of the AI ​​chip.

[0048] To achieve the above-mentioned purpose, an embodiment of the present invention also provides an edge AI chip performance evaluation device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the above-mentioned edge AI chip performance evaluation method when executing the computer program.

[0049] To achieve the above-mentioned purpose, an embodiment of the present invention also provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the above-mentioned edge AI chip performance evaluation method.

[0050] To achieve the above-mentioned purpose, an embodiment of the present invention also provides a computer program product, which includes a computer program; wherein, when the computer program is running, it controls the device where the program product is located to execute the above-mentioned edge AI chip performance evaluation method.

[0051] Implementing the embodiments of the present invention has the following beneficial effects:

[0052] The edge AI chip performance evaluation method, device, equipment and storage medium provided by the embodiments of the present invention obtain sample data by acquiring performance data from the actually running edge AI device and annotating it, wherein the performance data includes the performance indicators of the AI ​​chip when processing various computing tasks; the sample data is input into a self-organizing mapping network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training; the data to be tested is obtained and input into the classifier for classification to obtain the performance category classification result of the AI ​​chip, which can effectively process outliers and noise in the data and adaptively adjust model parameters, improve the robustness and flexibility of the model, and achieve efficient and accurate evaluation of AI chips. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the edge AI chip performance evaluation method provided by an embodiment of the present invention;

[0054] Figure 2 It is a structural block diagram of an edge AI chip performance evaluation device provided by an embodiment of the present invention;

[0055] Figure 3 It is a structural block diagram of the edge AI chip performance evaluation device provided by an embodiment of 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] In the description of the present invention, it should be noted that the step numbers in the text are only for the convenience of explanation of specific embodiments and do not serve to limit the order in which the steps are executed. The terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0058] See also Figure 1 , Figure 1 is a flow chart of an edge AI chip performance evaluation method provided by an embodiment of the present invention, wherein the edge AI chip performance evaluation method comprises steps S1 to S3, which are specifically as follows:

[0059] S1. Obtain performance data from the actually running edge AI device and annotate it to obtain sample data, wherein the performance data includes performance indicators of the AI ​​chip when processing various computing tasks;

[0060] S2, inputting the sample data into a self-organizing map network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training;

[0061] S3. Obtain the data to be tested and input it into the classifier for classification to obtain the performance category classification result of the AI ​​chip.

[0062] It is understandable that the attributes of the performance data include processing speed (indicating the time it takes for the chip to process the corresponding computing task), energy consumption (indicating the electrical energy consumed to complete the task), temperature (indicating the maximum temperature of the chip during operation), error rate (indicating the frequency of errors during the operation of the chip), throughput (indicating the amount of data that the chip can process per unit time), reliability (evaluating the operating time before the chip fails), compatibility (measuring the adaptability of the chip to cooperate with other devices), size (indicating the physical size of the chip), cost (indicating the production cost of the chip) and maintainability (indicating the difficulty of chip maintenance). It should be emphasized that the above content only illustrates one data format and type of the present invention. In actual applications, the attributes of the data are usually more than these 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0063] The labeling method includes manual labeling, and the labeled performance categories include "high efficiency", "low consumption" and "stable". By classifying the data into "high efficiency", "low consumption" and "stable", an intuitive and effective method is provided to evaluate the multi-dimensional performance indicators of AI chips, making the performance evaluation more comprehensive and accurate.

[0064] In an embodiment of the present invention, performance data is obtained from an actually running edge AI device and annotated to obtain sample data, wherein the performance data includes performance indicators of the AI ​​chip when processing various computing tasks; the sample data is input into a self-organizing mapping network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update, and model iterative training; the data to be tested is obtained and input into the classifier for classification to obtain a performance category classification result of the AI ​​chip, which can effectively process outliers and noise in the data and adaptively adjust model parameters, thereby improving the robustness and flexibility of the model and achieving efficient and accurate evaluation of the AI ​​chip.

[0065] In an optional embodiment, the self-organizing map network is composed of a plurality of nodes, and the process of initializing the model includes:

[0066] Initializing the weights and parameters of the self-organizing map network, wherein the parameters include the location of the nodes, the initial learning rate, and the neighborhood size;

[0067] Based on the random allocation method, the weight of each node is updated by a first formula, wherein the first formula is specifically:

[0068]

[0069] Among them, w i(0) is the initial weight vector of each node i; random_init() is a random allocation function; is the weight vector for each node i, where d is the dimension of the feature input to the classifier; p n is the total number of nodes in the self-organizing map network.

[0070] It can be understood that the size of the self-organizing map network is set to a 10x10 grid with a total of 100 nodes, and the initial weight vector of each node is initialized by uniformly distributing it in the feature space.

[0071] In an optional embodiment, the process of input data mapping includes:

[0072] Inputting the sample data into the self-organizing map network that completes the model initialization;

[0073] Each data point of the sample data will be mapped to a node of the self-organizing map network;

[0074] In each iteration, a sample data is randomly selected from the sample data set after dimensionality reduction and input into the self-organizing map network.

[0075] In an optional embodiment, the node competition learning process includes:

[0076] Calculate the similarity between each sample data and each node through a similarity algorithm, and select the node with the greatest similarity as the winner node of the sample data;

[0077] The similarity algorithm includes the Euclidean distance method, and the winner node of the sample data is calculated by the second formula, wherein the second formula is specifically:

[0078]

[0079] Where b is the winning node that is most similar to the sample data x(t); || || is the Euclidean distance; x(t) is the number of nodes in the tth iteration from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node.

[0080] In an embodiment of the present invention, for each input sample data, the most similar self-organizing map node is determined as the winner node in the self-organizing map network by calculating the Euclidean distance. This process simulates the competition mechanism between neurons and can effectively map the input sample to the appropriate node position to achieve effective classification and accurate mapping of data.

[0081] In an optional embodiment, the process of updating the neighboring nodes includes:

[0082] Calculate the distance between each of the nodes and the winner node in the self-organizing map grid by a third formula;

[0083] According to the distance and the perceptual annealing strategy, the weights of the neighboring nodes are calculated and updated by the fourth formula;

[0084] Wherein, the third formula is specifically:

[0085]

[0086] Among them, the neighborhood function h ib (t) is the distance function between node i and the winner node b; r i is the position of node i in the self-organizing map network (x i ,y i ), r b is the position of the winner node b in the self-organizing map network (x b ,y b ); σ(t) is the neighborhood width before updating, which decreases with the number of iterations; t is the index of the number of iterations; Fo() is the distance calculation function, is the distance between node i and the winner node b in the self-organizing map grid;

[0087] The fourth formula is specifically:

[0088] w i (t+1)=w i (t)+α(t)·h ib (t)·(x(t)-w i (t))

[0089] Among them, α(t) is the learning rate before updating, which decreases with the number of iterations; the neighborhood function h ib (t) is the distance function between node i and the winner node b; x(t) is the number of samples from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node; w i (t+1) is the updated weight vector of the i-th node.

[0090] In an embodiment of the present invention, the neighborhood node update process includes calculating the distance between the node and the winner node in the self-organizing map grid and adjusting the weight of the neighborhood node using the perceptual annealing strategy. Specifically, the third formula calculates the distance between the node and the winner node, uses the neighborhood function to adjust the neighborhood width, and the fourth formula updates the weight vector of the neighborhood node according to the learning rate and the number of neighbors, ensuring that the self-organizing map network adapts to changes in data distribution during training and improves the accuracy of classification and mapping.

[0091] In an optional embodiment, the process of updating the model parameters includes:

[0092] Based on the quantile gradient descent and the perceptual annealing strategy, the updated learning rate and neighborhood width are calculated by the fifth formula and dynamically adjusted to optimize the learning process of the self-organizing map network;

[0093] Wherein, the fifth formula is specifically:

[0094]

[0095] Among them, λ α is the time constant for controlling the learning rate annealing rate; σ is the time constant for controlling the annealing rate of the neighborhood width; α(t) is the learning rate before the update; σ(t) is the neighborhood width before the update; α(t+1) is the learning rate after the update; σ(t+1) is the neighborhood width after the update; t is the index of the number of iterations;

[0096] Based on the exponential decay method, the time constant λ is calculated by the sixth formula α And make adjustments, wherein the fifth formula is specifically:

[0097]

[0098] Among them, λ α is the time constant for controlling the learning rate annealing rate; α min is the minimum value of the learning rate, ensuring that the learning rate does not decay to 0; α(0) is the initial learning rate; T is the expected number of training cycles;

[0099] Based on the exponential decay method, the time constant λ is calculated by the seventh formula σ And make adjustments, wherein the seventh formula is specifically:

[0100]

[0101] Among them, λ σ is the time constant for controlling the annealing rate of the neighborhood width; σ minis the minimum value of the neighborhood width, ensuring that the neighborhood function does not shrink to only include the winner node itself; σ(0) is the initial neighborhood width; T is the expected number of training cycles.

[0102] In an embodiment of the present invention, the fifth formula dynamically calculates the updated learning rate and neighborhood width, controls the annealing rate through the time constant, and ensures that the network can adjust the learning parameters in time during training to improve the convergence efficiency; the sixth and seventh formulas adjust the time constants through exponential decay according to the expected number of training cycles and the initial value to ensure that the learning rate and the neighborhood width decrease steadily during the training process, thereby effectively optimizing the learning and classification capabilities of the model. Therefore, in the process of updating the model parameters, the outliers and noise are processed by quantile gradient descent, which can significantly improve the model's processing ability for unstable data. At the same time, the learning rate and the neighborhood width are dynamically adjusted through the adaptive adjustment mechanism, which can better balance exploration and utilization during the training process, accelerate the convergence speed of the model, and improve learning efficiency.

[0103] In an optional embodiment, the process of iterative training of the model includes:

[0104] The process of node competition learning, neighborhood node updating and model parameter updating is repeated until the self-organizing map network converges or reaches a preset number of iterations.

[0105] It can be understood that after the iterative training of the model is completed, a trained classifier is obtained; the classifier is used to process new samples to be tested, and the performance category results of the predicted classification of each sample to be tested are obtained to complete the AI ​​chip performance evaluation task.

[0106] In an embodiment of the present invention, through continuous competitive learning optimization and neighborhood node weight update, combined with dynamically adjusted model parameters, it is ensured that the self-organizing map network can effectively adapt to data characteristics and improve classification accuracy, thereby achieving the goal of comprehensively evaluating the performance of edge AI chips.

[0107] The edge AI chip performance evaluation method provided by an embodiment of the present invention obtains sample data by acquiring performance data from an actually running edge AI device and annotating the performance data, wherein the performance data includes performance indicators of the AI ​​chip when processing various computing tasks; the sample data is input into a self-organizing mapping network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training; the data to be tested is acquired and input into the classifier for classification to obtain a performance category classification result of the AI ​​chip, which can effectively process outliers and noise in the data and adaptively adjust model parameters, thereby improving the robustness and flexibility of the model and achieving efficient and accurate evaluation of the AI ​​chip.

[0108] See also Figure 2 , Figure 2 : is a structural block diagram of an edge AI chip performance evaluation device 10 provided in an embodiment of the present invention, wherein the edge AI chip performance evaluation device 10 comprises:

[0109] The data acquisition module 11 is used to obtain performance data from the edge AI device in actual operation and annotate it to obtain sample data, wherein the performance data includes performance indicators of the AI ​​chip when processing various computing tasks;

[0110] A model training module 12 is used to input the sample data into a self-organizing map network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training;

[0111] The performance evaluation module 13 is used to obtain the data to be tested and input it into the classifier for classification to obtain the performance category classification result of the AI ​​chip.

[0112] Optionally, the model training module 12 is further used for:

[0113] Initializing the weights and parameters of the self-organizing map network, wherein the parameters include the location of the nodes, the initial learning rate, and the neighborhood size;

[0114] Based on the random allocation method, the weight of each node is updated by a first formula, wherein the first formula is specifically:

[0115]

[0116] Among them, w i (0) is the initial weight vector of each node i; random_init() is a random allocation function; is the weight vector for each node i, where d is the dimension of the feature input to the classifier; p n is the total number of nodes in the self-organizing map network.

[0117] Optionally, the model training module 12 is further used for:

[0118] Inputting the sample data into the self-organizing map network that completes the model initialization;

[0119] Each data point of the sample data will be mapped to a node of the self-organizing map network;

[0120] In each iteration, a sample data is randomly selected from the sample data set after dimensionality reduction and input into the self-organizing map network.

[0121] Optionally, the model training module 12 is further used for:

[0122] Calculate the similarity between each sample data and each node through a similarity algorithm, and select the node with the greatest similarity as the winner node of the sample data;

[0123] The similarity algorithm includes the Euclidean distance method, and the winner node of the sample data is calculated by the second formula, wherein the second formula is specifically:

[0124]

[0125] Where b is the winning node that is most similar to the sample data x(t); || || is the Euclidean distance; x(t) is the number of nodes in the tth iteration from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node.

[0126] Optionally, the model training module 12 is further used for:

[0127] Calculate the distance between each of the nodes and the winner node in the self-organizing map grid by a third formula;

[0128] According to the distance and the perceptual annealing strategy, the weights of the neighboring nodes are calculated and updated by the fourth formula;

[0129] Wherein, the third formula is specifically:

[0130]

[0131] Among them, the neighborhood function h ib(t) is the distance function between node i and the winner node b; r i is the position of node i in the self-organizing map network (x i ,y i ), r b is the position of the winner node b in the self-organizing map network (x b ,y b ); σ(t) is the neighborhood width before updating, which decreases with the number of iterations; t is the index of the number of iterations; Fo() is the distance calculation function, is the distance between node i and the winner node b in the self-organizing map grid;

[0132] The fourth formula is specifically:

[0133] w i (t+1)=w i (t)+α(t)·h ib (t)·(x(t)-w i (t))

[0134] Among them, α(t) is the learning rate before updating, which decreases with the number of iterations; the neighborhood function h ib (t) is the distance function between node i and the winner node b; x(t) is the number of samples from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node; w i (t+1) is the updated weight vector of the i-th node.

[0135] Optionally, the model training module 12 is further used for:

[0136] Based on the quantile gradient descent and the perceptual annealing strategy, the updated learning rate and neighborhood width are calculated by the fifth formula and dynamically adjusted to optimize the learning process of the self-organizing map network;

[0137] Wherein, the fifth formula is specifically:

[0138]

[0139] Among them, λ α is the time constant for controlling the learning rate annealing rate; σ is the time constant for controlling the annealing rate of the neighborhood width; α(t) is the learning rate before the update; σ(t) is the neighborhood width before the update; α(t+1) is the learning rate after the update; σ(t+1) is the neighborhood width after the update; t is the index of the number of iterations;

[0140] Based on the exponential decay method, the time constant λ is calculated by the sixth formula α And make adjustments, wherein the fifth formula is specifically:

[0141]

[0142] Among them, λ α is the time constant for controlling the learning rate annealing rate; α min is the minimum value of the learning rate, ensuring that the learning rate does not decay to 0; α(0) is the initial learning rate; T is the expected number of training cycles;

[0143] Based on the exponential decay method, the time constant λ is calculated by the seventh formula σ And make adjustments, wherein the seventh formula is specifically:

[0144]

[0145] Among them, λ σ is the time constant for controlling the annealing rate of the neighborhood width; σ min is the minimum value of the neighborhood width, ensuring that the neighborhood function does not shrink to only include the winner node itself; σ(0) is the initial neighborhood width; T is the expected number of training cycles.

[0146] Optionally, the model training module 12 is further used for:

[0147] The process of node competition learning, neighborhood node updating and model parameter updating is repeated until the self-organizing map network converges or reaches a preset number of iterations.

[0148] It is worth noting that the working process of each module in the edge AI chip performance evaluation device 10 described in the embodiment of the present invention can refer to the working process of the edge AI chip performance evaluation method described in the above embodiment, and will not be repeated here.

[0149] The edge AI chip performance evaluation device provided by an embodiment of the present invention obtains sample data by acquiring performance data from an actually running edge AI device and annotating the performance data, wherein the performance data includes performance indicators of the AI ​​chip when processing various computing tasks; the sample data is input into a self-organizing mapping network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training; the data to be tested is acquired and input into the classifier for classification to obtain a performance category classification result of the AI ​​chip, which can effectively process outliers and noise in the data and adaptively adjust model parameters, thereby improving the robustness and flexibility of the model and achieving efficient and accurate evaluation of the AI ​​chip.

[0150] In addition, an embodiment of the present invention also provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the above-mentioned edge AI chip performance evaluation method.

[0151] An embodiment of the present invention also provides a computer program product, which includes a computer program; wherein, when the computer program is running, the device where the computer program product is located controls to execute the above-mentioned edge AI chip performance evaluation method.

[0152] See also Figure 3 , Figure 3 : is a structural block diagram of an edge AI chip performance evaluation device 20 provided in an embodiment of the present invention, and the edge AI chip performance evaluation device 20 includes: a processor 21, a memory 22, and a computer program stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program, the steps in the above-mentioned edge AI chip performance evaluation method embodiment are implemented. Alternatively, when the processor 21 executes the computer program, the functions of each module / unit in the above-mentioned device embodiments are implemented.

[0153] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the edge AI chip performance evaluation device 20.

[0154] The edge AI chip performance evaluation device 20 may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will appreciate that the schematic diagram is merely an example of the edge AI chip performance evaluation device 20 and does not constitute a limitation on the edge AI chip performance evaluation device 20. The edge AI chip performance evaluation device 20 may include more or fewer components than shown in the diagram, or may combine certain components, or different components. For example, the edge AI chip performance evaluation device 20 may also include input and output devices, network access devices, buses, and the like.

[0155] The processor 21 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor 21 is the control center of the edge AI chip performance evaluation device 20, and various interfaces and lines are used to connect various parts of the entire edge AI chip performance evaluation device 20.

[0156] The memory 22 can be used to store the computer program and / or module, and the processor 21 realizes various functions of the edge AI chip performance evaluation device 20 by running or executing the computer program and / or module stored in the memory 22, and calling the data stored in the memory 22. The memory 22 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory 22 may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0157] Among them, if the module / unit integrated in the edge AI chip performance evaluation device 20 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor 21, the steps of the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0158] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for evaluating the performance of an edge AI chip, characterized in that: include: Obtain performance data from the actual running edge AI device and annotate it to obtain sample data, wherein the performance data includes performance indicators of the AI ​​chip when processing various computing tasks; Inputting the sample data into a self-organizing map network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training; The data to be tested is obtained and input into the classifier for classification to obtain the performance category classification result of the AI ​​chip.

2. The edge AI chip performance evaluation method according to claim 1, characterized in that: The self-organizing map network is composed of a plurality of nodes, and the process of initializing the model includes: Initializing the weights and parameters of the self-organizing map network, wherein the parameters include the location of the nodes, the initial learning rate, and the neighborhood size; Based on the random allocation method, the weight of each node is updated by a first formula, wherein the first formula is specifically: Among them, w i (0) is the initial weight vector of each node i; random_init() is a random allocation function; is the weight vector for each node i, where d is the dimension of the feature input to the classifier; p n is the total number of nodes in the self-organizing map network.

3. The edge AI chip performance evaluation method according to claim 2, characterized in that: The process of input data mapping includes: Inputting the sample data into the self-organizing map network that completes the model initialization; Each data point of the sample data will be mapped to a node of the self-organizing map network; In each iteration, a sample data is randomly selected from the sample data set after dimensionality reduction and input into the self-organizing map network.

4. The edge AI chip performance evaluation method according to claim 3, characterized in that: The node competition learning process includes: Calculate the similarity between each sample data and each node through a similarity algorithm, and select the node with the greatest similarity as the winner node of the sample data; The similarity algorithm includes the Euclidean distance method, and the winner node of the sample data is calculated by the second formula, wherein the second formula is specifically: Where b is the winning node that is most similar to the sample data x(t); ∥∥ is the calculation of the Euclidean distance; x(t) is the number of nodes in the tth iteration from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node.

5. The edge AI chip performance evaluation method according to claim 4, characterized in that: The process of updating the neighboring nodes includes: Calculate the distance between each of the nodes and the winner node in the self-organizing map grid by a third formula; According to the distance and the perceptual annealing strategy, the weights of the neighboring nodes are calculated and updated by the fourth formula; Wherein, the third formula is specifically: Among them, the neighborhood function h ib (t) is the distance function between node i and the winner node b; r i is the position of node i in the self-organizing map network (x i ,y i ), r b is the position of the winner node b in the self-organizing map network (x b ,y b ); σ(t) is the neighborhood width before updating, which decreases with the number of iterations; t is the index of the number of iterations; Fo() is the distance calculation function, is the distance between node i and the winner node b in the self-organizing map grid; The fourth formula is specifically: w i (t+1)=w i (t)+α(t)·h ib (t)·(x(t)-w i (t)) Among them, α(t) is the learning rate before updating, which decreases with the number of iterations; the neighborhood function h ib (t) is the distance function between node i and the winner node b; x(t) is the number of samples from the reduced-dimensional sample data set X = {x1, x2, ..., x m A sample data randomly selected from}; w i (t) is the weight vector of the i-th node; w i (t+1) is the updated weight vector of the i-th node.

6. The edge AI chip performance evaluation method according to claim 5, characterized in that: The process of updating the model parameters includes: Based on the quantile gradient descent and the perceptual annealing strategy, the updated learning rate and neighborhood width are calculated by the fifth formula and dynamically adjusted to optimize the learning process of the self-organizing map network; Wherein, the fifth formula is specifically: Among them, λ α is the time constant for controlling the learning rate annealing rate; σ is the time constant for controlling the annealing rate of the neighborhood width; α(t) is the learning rate before the update; σ(t) is the neighborhood width before the update; α(t+1) is the learning rate after the update; σ(t+1) is the neighborhood width after the update; t is the index of the number of iterations; Based on the exponential decay method, the time constant λ is calculated by the sixth formula α And make adjustments, wherein the fifth formula is specifically: Among them, λ α is the time constant for controlling the learning rate annealing rate; α min is the minimum value of the learning rate, ensuring that the learning rate does not decay to 0; α(0) is the initial learning rate; T is the expected number of training cycles; Based on the exponential decay method, the time constant λ is calculated by the seventh formula σ And make adjustments, wherein the seventh formula is specifically: Among them, λ σ is the time constant for controlling the annealing rate of the neighborhood width; σ min is the minimum value of the neighborhood width, ensuring that the neighborhood function does not shrink to only include the winner node itself; σ(0) is the initial neighborhood width; T is the expected number of training cycles.

7. The edge AI chip performance evaluation method according to claim 6, characterized in that: The process of iterative training of the model includes: The process of node competition learning, neighborhood node updating and model parameter updating is repeated until the self-organizing map network converges or reaches a preset number of iterations.

8. An edge AI chip performance evaluation device, characterized in that: include: A data acquisition module is used to obtain performance data from the actual running edge AI device and annotate it to obtain sample data, wherein the performance data includes the performance indicators of the AI ​​chip when processing various computing tasks; A model training module, used for inputting the sample data into a self-organizing map network based on quantile gradient descent for training to obtain a trained classifier, wherein the training process includes model initialization, input data mapping, node competition learning, neighborhood node update, model parameter update and model iterative training; The performance evaluation module is used to obtain the data to be tested and input it into the classifier for classification to obtain the performance category classification result of the AI ​​chip.

9. An edge AI chip performance evaluation device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the edge AI chip performance evaluation method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the edge AI chip performance evaluation method as described in any one of claims 1 to 7.