WO2023071766A1 - Model compression method, model compression system, server, and storage medium - Google Patents

Abstract

The present application relates to the field of neural network architecture search, in particular, to a model compression method, a model compression system, a server, and a storage medium. The model compression method in the present application comprises: receiving a candidate model; performing neural network architecture search on the candidate model, and obtaining a plurality of transformed models; performing computing power compression on the candidate model to obtain a compressed model; and using the transformed models and the compressed model as candidate models to perform neural network architecture search and computing power compression again.

Description

Model compression method, model compression system, server and storage medium

priority information

This application claims priority to a Chinese patent application with application number 202111266014.4 filed on October 28, 2021, the entire contents of which are incorporated herein by reference.

technical field

The present application relates to the field of neural network architecture search, in particular to a model compression method, model compression system, server and storage medium.

Background technique

Neural networks have achieved outstanding results in the field of machine learning, such as object classification, target detection, etc. Intelligent algorithms based on neural networks have changed human production and lifestyle. However, due to the high computational complexity of neural networks and the large number of model parameters, model computing power compression technology has become a research hotspot in the academic and industrial circles in recent years.

At present, common methods for model compression include: parameter quantization, network pruning, and artificial design of efficient neural network structures; however, the two methods of parameter quantization and network pruning have great limitations in accuracy; The design of the neural network structure is inseparable from a large amount of human investment and research, and requires a lot of time for manual experiments and parameter adjustments; it can be seen that the current model compression methods cannot balance accuracy and save labor costs.

Contents of the invention

The main purpose of the embodiment of the present application is to provide a model compression method, a model compression system, a server and a storage medium, taking into account the accuracy of obtaining the compressed model and reducing labor costs.

In order to achieve the above purpose, an embodiment of the present application provides a model compression method, including: receiving a candidate model; performing neural network architecture search on the candidate model to obtain multiple transformation models; performing computing power compression on the candidate model to obtain Compression model: using the transformation model and the compression model as the candidate model, re-performing the neural network architecture search and the computing power compression.

The embodiment of the present application also provides a model compression system, including: a receiving module, a model automatic search module, and a computing power compression module; the receiving module is used to receive candidate models; the model automatic search module is used to search for the candidate The model performs a neural network architecture search to obtain a plurality of transformation models and input them into the receiving module as the candidate model; the computing power compression module is used to compress the computing power of the candidate model to obtain a compressed model, and the compressed model is used as The candidate models are input into the receiving module.

The embodiment of the present application also provides a server, including: at least one processor; and a memory connected in communication with the at least one processor; wherein, the memory stores instructions that can be executed by the at least one processor , the instructions are executed by the at least one processor, so that the at least one processor can execute the above-mentioned model compression method.

The embodiment of the present application also provides a computer-readable storage medium storing a computer program, and implementing the above-mentioned model compression method when the computer program is executed by a processor.

The model compression method of this application can obtain as many neural network models as possible through the neural network architecture search, thereby improving the probability of obtaining a better model, and this application can compress multiple neural network models obtained in the previous poll through computing power compression. Transform the model for compression; it can be seen that this application combines model compression with neural network architecture search, and uses the transformed model obtained by neural network architecture search as a candidate model to compress the computing power in the next poll, and again as a candidate Model, the compression model obtained by computing power compression can also be searched for the neural network architecture in the next polling to obtain the transformation model; through continuous polling, after polling reaches a certain level, a high-precision compression model can be obtained, thereby Taking into account the accuracy of the compressed model, and the entire model compression process reduces manual participation and reduces labor costs. In addition, compared with related technologies, the present application can automatically obtain the compressed model, and also speeds up the speed of model compression.

Description of drawings

One or more embodiments are exemplified by the pictures in the corresponding drawings, and these exemplifications do not constitute a limitation to the embodiments. Elements with the same reference numerals in the drawings represent similar elements. Unless otherwise stated, the drawings in the drawings are not limited to scale.

Fig. 1 is a schematic flow chart of a model compression method according to an embodiment of the present application;

FIG. 2 is a schematic flowchart of the sub-steps of step 103 of the model compression method according to an embodiment of the present application;

3 is a schematic flow diagram of a model compression method according to an embodiment of the present application;

Fig. 4 is a schematic flow chart of a model compression method according to an embodiment of the present application;

5 is a schematic structural diagram of a model compression system according to an embodiment of the present application;

6 is a schematic structural diagram of a model compression system according to an embodiment of the present application;

Fig. 7 is a schematic structural diagram of a server according to an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that in each embodiment of the application, many technical details are provided for readers to better understand the application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can also be realized. The division of the following embodiments is for the convenience of description, and should not constitute any limitation to the specific implementation of the present application, and the embodiments can be combined and referred to each other on the premise of no contradiction.

An embodiment of the present application relates to a model compression method, as shown in Figure 1, which is a schematic flow chart of the model compression method of this embodiment, specifically including the following steps:

Step 101, receiving candidate models.

Specifically, the execution subject of this embodiment is the model compression system, and the model compression system needs to receive candidate models first during the process of model compression.

In step 102, a neural network architecture search is performed on the candidate models to obtain multiple transformed models.

Specifically, Neural Architecture Search (NAS, Neural Architecture Search) can effectively liberate human input, and NAS can explore the network structure in a larger space, and has the opportunity to generate high-performance structures beyond human design. The network field has been widely used. Therefore, while the model compression method of this embodiment combines NAS to solve the problem of manpower input, it can also obtain a compressed model with higher accuracy. In one embodiment, performing neural network architecture search on candidate models to obtain multiple transformation models includes: performing neural network architecture search based on a preset search space, performing random transformation on candidate models to obtain multiple transformation models.

Specifically, the process of NAS search is to traverse the preset search space based on the preset search strategy, and randomly transform the candidate models in each NAS search to obtain multiple transformed models. This randomness ensures the diversity of generated model structures, which is conducive to searching for high-performance models. It should be noted that the search space in this embodiment can be set by the user or defaulted by the system, and the search strategy can also be set by the user or defaulted by the system; wherein, the NAS search strategy includes reinforcement learning-based, evolutionary algorithm-based and based on gradient descent etc.

It should be noted that the search strategy involved in this embodiment can be any search strategy, which has strong robustness, and can be set as a corresponding search strategy according to the requirements of the user or the system, which is not specifically limited in this embodiment.

In step 103, computing power is compressed on the candidate model to obtain a compressed model.

The order of the above steps 102 and 103 is not specifically limited in this embodiment. Step 102 may be before step 103, step 102 may be after step 103, and steps 102 and 103 may be performed simultaneously.

In one embodiment, performing computing power compression on the candidate model to obtain the compressed model, that is, step 104 includes the following sub-steps, and the specific flowchart is shown in FIG. 2 , including:

Step 1031, calculate the computing power of each layer in the candidate model.

Specifically, computing power compression is responsible for directional compression of the computing power of candidate models. For example, it will detect model computing power-intensive areas and perform directional compression processing on the layer that consumes the most computing power.

Specifically, the computing power consumption of each layer in the candidate model can be calculated according to the input of the candidate model and the parameters of each layer. The computing power consumption of each layer is mainly determined by the size of the feature map input to the layer and the size of the weight parameters that need to be trained for the operation of the layer, which can be approximately expressed by the following formula:

FLOPs(l)=H×W×C×Params(l);

Among them, l represents an operation layer in the model (such as a convolutional layer), H and W represent the size of the feature map input to this layer, C represents the number of channels of the input feature map, and Params represents the parameter amount of the current layer.

Step 1032, comparing the computing power of each layer to obtain the layer with the largest computing power.

Specifically, by comparing the computing power of each layer obtained in step 1041, it is possible to locate the layer with the largest computing power in the model, that is, the "computing power-intensive area". The computing power compression step mainly performs computing power compression on the "computing power intensive area" of the model.

In step 1033, computing power is compressed on the layer with the largest computing power to obtain a compressed model.

In one embodiment, performing computing power compression on the layer with the largest computing power includes: reducing the parameter amount of the layer with the largest computing power, or removing the layer with the largest computing power, or reducing the size of the input feature map of the layer with the largest computing power .

Specifically, from the above calculation power consumption formula, it can be known that the factors that affect the calculation power mainly include the input feature map of the layer and the parameter quantity of the current layer; therefore, the compression of the calculation power of a certain layer in the model has the following three available directions:

(1) Reduce the parameter amount of the current layer. For example, for a convolution operation, its parameter quantity is determined by the size and number of its convolution kernel, so the size of the convolution kernel can be compressed or the number of filters proposed after starting the search task can be reduced. Reducing the parameter amount of the current layer is equivalent to changing the hyperparameters of the current layer, but the direction of change is along the direction of decreasing the parameter amount.

(2) Remove the current layer, which is the limit case of the first case, that is, reduce the parameter amount of this layer to 0.

(3) Reduce the size of the feature map input into this layer. Reducing the size of the input feature map means that it is necessary to change the operation layer in front of the "computing intensive area", because only by changing the operation layer in front can there be a chance to change the input of the current layer. The specific operations include: ① Add a pooling layer in front of the "computing power-intensive area" to compress the features; The number of channels of the output feature or increase the step size to play the role of downsampling feature map.

In step 104, the transformation model and the compression model are used as candidate models for training.

Specifically, the model generated by neural network architecture search and computing power compression is used as a candidate model, and will return to step 101 to re-perform the neural network architecture search and computing power compression; that is, the above steps 101 to Step 104 is a polling process. After step 104 is completed, step 101 is executed again, thereby forming a closed-loop operation of the system.

Specifically, during the first polling process, the candidate model received by the model compression system is the reference model input by the user.

It should be noted that the candidate models in this embodiment will all be set with an ID (ID, Identity document), the user will input the benchmark model into the system, and the model generated through neural network architecture search and computing power compression will be assigned an ID. This ID is unique, which can ensure that the trained model will not have the problem of repeated training.

The model compression method of this embodiment can obtain as many neural network models as possible through the neural network architecture search, thereby improving the probability of obtaining a better model, and this application can use the computational power compression to obtain as many neural network models as possible in the previous polling. It can be seen that this application combines model compression with neural network architecture search, and uses the transformation model obtained by neural network architecture search as the compression model obtained by computing power compression in the next polling of the candidate model, and again as the compression model The candidate model, the compression model obtained by computing power compression can also be searched for the neural network architecture in the next polling to obtain the transformation model; through continuous polling, after polling reaches a certain level, a high-precision compression model can be obtained. In this way, the accuracy of the compressed model is taken into account, and the entire model compression process reduces manual participation and reduces labor costs. In addition, compared with related technologies, the present application can automatically obtain a compressed model, and also accelerates the speed of model compression.

An embodiment of the present application relates to a model compression method. This embodiment is roughly the same as the previous embodiment. The main difference is that after receiving the candidate model, it also includes: training the candidate model to obtain the trained model; obtaining the trained model Accuracy and computing power: Based on the accuracy and computing power, the evaluation parameters of the candidate models corresponding to the trained model are obtained. For ease of description, the parts of this embodiment that are the same as or corresponding to the previous embodiment will not be described again.

A schematic flow chart of the model compression method of this embodiment is shown in Figure 3, specifically including the following steps:

Step 201, receiving candidate models.

Step 202, train the candidate model to obtain the trained model.

Step 203, acquiring the accuracy and computing power of the trained model.

In step 204, the evaluation parameters of the candidate models corresponding to the trained models are obtained based on the accuracy and computing power.

Step 205, perform neural network architecture search on the candidate models to obtain multiple transformed models.

In step 206, computing power is compressed on the candidate model to obtain a compressed model.

In step 207, the transformation model and the compression model are used as candidate models.

The above step 201, step 205 to step 207 are the same as step 101 to step 104 in the previous embodiment, and will not be repeated here.

It should be noted that the above step 201 to step 207 is a polling process, and the model compression system will enter step 201 again after completing step 207, thus forming a closed-loop operation of the system.

Specifically, after receiving the candidate model, the candidate model is trained to obtain the trained model, and the performance of the trained model is evaluated, so that the candidate model required by the user can be screened out according to the user's needs. Specifically, in the first polling process, the benchmark model input by the user is used as a candidate model, and only the benchmark model input by the user is evaluated for performance. In the subsequent polling process, the number of candidate models is large, and model training needs to be performed on each candidate model, and there are also many trained models. In the process of performance evaluation, it is necessary to traverse all the trained models. And evaluate the performance of each trained model.

Specifically, the performance evaluation of this embodiment requires the use of two parameters, the accuracy of the candidate model and the computing power. Therefore, after the candidate model is obtained in this embodiment, the accuracy and computing power of the candidate model are obtained. Among them, the accuracy It refers to the accuracy rate of the candidate model during use, and computing power refers to the intensity of the calculation amount of the candidate model during use.

Specifically, since accuracy and computing power are different parameters, when evaluating, it is necessary to transform the values of each parameter into the same dimension, that is, to obtain the candidate model corresponding to the trained model based on accuracy and computing power Finally, the comprehensive performance score of the candidate model corresponding to the trained model is given, that is, the evaluation parameter. The evaluation parameter can represent the performance of the candidate model after weighing multiple indicators.

In one embodiment, the evaluation parameters include evaluation scores; the evaluation parameters of the candidate models corresponding to the trained models are obtained based on the accuracy and computing power, including: calculating the accuracy and computing power according to the preset weight ratio to obtain the candidate models evaluation score. Specifically, in this embodiment, different weights are set for accuracy and computing power, so as to calculate the evaluation scores of candidate models, and different weights can be set according to user needs. Set a higher weight for accuracy; or users have higher requirements for computing power, so you can set a higher weight for computing power. Specifically, for the evaluation score, it can be set that the higher the evaluation score, the better the performance of the obtained candidate model.

In one embodiment, the evaluation parameter includes an evaluation score; after obtaining the evaluation parameters of the candidate model corresponding to the trained model based on accuracy and computing power, it also includes: retaining the candidate model whose evaluation score is greater than or equal to the preset score, or retaining Evaluate the N candidate models with the largest scores. Specifically, in addition to the first polling process, there will be multiple candidate models in the subsequent polling process. If each candidate model enters the subsequent neural network architecture search and computing power compression, the system will have a large amount of computation. And it will also cause a waste of resources. Therefore, this embodiment will set a preset rule, that is, retain candidate models whose evaluation scores are greater than or equal to the preset score, or reserve N candidate models with the largest evaluation scores, so as to retain Candidate models that do not meet the preset rules are deleted.

Specifically, in the first polling process, only the benchmark model input by the user is used for training. In the subsequent polling process, the number of candidate models increases and needs to be screened. Each candidate model can be evaluated according to the evaluation score. Sorting, setting a preset score in advance, retaining the candidate models whose evaluation score is greater than or equal to the preset score, or setting a value N in advance, retaining the N candidate models with the largest evaluation score, so as to enter the subsequent neural network architecture search and In the step of computing power compression, candidate models that do not meet the preset rules are deleted, so that models with higher performance can be fully utilized, thereby reducing the amount of computing of the system and avoiding resource waste.

It should be noted that if the user has no special requirements, the default rule is to set a value N, and keep the N candidate models with the largest evaluation scores.

Specifically, in the subsequent polling process, other methods can also be used to retain some candidate models for subsequent neural network architecture search and computing power compression, such as roulette selection method, tournament selection method, etc.

An embodiment of the present application relates to a model compression method. This embodiment is roughly the same as the previous embodiment. The main difference is that in this embodiment, after the evaluation score of the candidate model corresponding to the trained model is obtained based on accuracy and computing power, It also includes: judging whether the maximum evaluation score obtained by this polling is greater than the maximum evaluation score obtained by the previous polling; for the convenience of description, the same or corresponding parts of this embodiment and the previous embodiment will not be described again.

A schematic flow chart of the model compression method of this embodiment is shown in Figure 4, specifically including the following steps:

Step 301, receiving candidate models.

Step 302, train the candidate model to obtain the trained model.

Step 303, acquiring the accuracy and computing power of the trained model.

Step 304: Obtain the evaluation parameters of the candidate model corresponding to the trained model based on the accuracy and computing power.

Step 305, judging whether the maximum evaluation score obtained in this poll is greater than the maximum evaluation score obtained in the previous poll. If yes, go to step 307 and step 308; if not, go to step 306.

Step 306, adding 1 to the counter value, and judging whether the counter value is smaller than a preset threshold. If yes, go to step 307 and step 308; if not, polling ends.

Step 307, perform neural network architecture search on the candidate models to obtain multiple transformed models.

In step 308, computing power is compressed on the candidate model to obtain a compressed model.

In step 309, the transformation model and the compression model are used as candidate models.

The foregoing steps 301 to 304, and steps 307 to 309 are the same as steps 201 to 207 in the previous embodiment, and will not be repeated here.

Specifically, after evaluating the performance of the candidate models in each poll, the evaluation score of each candidate model will be calculated, and the model with the highest evaluation score and the best performance will appear; after that, the maximum evaluation score of this poll will be compared score and the maximum evaluation score obtained in the last poll. If the maximum evaluation score obtained in this poll is greater than the maximum evaluation score obtained in the previous poll, it means that the candidate model is being continuously optimized, and polling again may result in better performance. A good model, at this time, continue to step 307 and step 308; if the maximum evaluation score of this poll is less than or equal to the maximum evaluation score of the previous poll, it means that the model with the best performance has probably been obtained so far , enter step 306 at this time, add 1 to the counter value, and judge whether the counter value is less than the preset threshold value. If the model is high, it is necessary to enter step 307 and step 308 again; if the counter value is greater than or equal to the preset threshold, the number of polling times has reached the set standard, and the polling ends.

Specifically, the model compression system is equipped with a counter. If the maximum evaluation score of this poll is less than or equal to the maximum evaluation score obtained in the previous poll, the counter will be incremented by 1, and the initial value of the counter value is 0.

Specifically, the user can also actively issue a stop command to instruct the system to end polling. When the system has acquired a model with higher performance in steps 303 and 304, the user thinks that there is no need to poll again, and the user can actively end the entire polling process.

In one embodiment, when the polling ends, the target candidate model is obtained from multiple candidate models; the accuracy of the target candidate model is greater than the preset accuracy and the computing power of the target candidate model is smaller than the preset computing power. Specifically, after the polling ends and the search is stopped, the system will organize the log information recorded during the search process, including the structural information, accuracy rate, computing power consumption and comprehensive performance information of the candidate models in each round of iterations. This embodiment will satisfy the Candidate models whose accuracy rate is greater than the preset accuracy and whose computing power is less than the preset computing power are used as target candidate models, and the target candidate models are organized into files and output to the user. After the system completes the output of the results, it releases the computing resources occupied by the search task and ends the entire process. Among them, the system is preset with a benchmark model, the preset accuracy is the accuracy of the benchmark model, and the preset computing power is the computing power of the benchmark model.

Specifically, when the polling ends, the user can also use other methods to obtain target candidate models, for example, after calculating the evaluation score of each candidate model, select the candidate model with the highest evaluation score as the target candidate model, and set The target candidate models are organized into files and output to the user to ensure that the user uses the optimal candidate model.

The step division of the above various methods is only for the sake of clarity of description. During implementation, it can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent. ; Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but not changing the core design of the algorithm and process are all within the scope of protection of this patent.

An embodiment of the present application relates to a model compression system. The specific structural diagram is shown in FIG. 5 , including: a receiving module 401 , an automatic model search module 402 , and a computing power compression module 403 .

Specifically, the receiving module 401 is used to receive candidate models; the automatic model search module 402 is used to search the neural network architecture for the candidate models to obtain multiple transformation models and input them into the receiving module 401 as candidate models; the computing power compression module 403 is used to Computational compression is performed on the candidate model to obtain a compressed model, and the compressed model is input into the receiving module 401 as a candidate model.

In one embodiment, the model compression system further includes: a model training module and a performance evaluation module.

As shown in FIG. 6 , it is a schematic structural diagram of this embodiment, including: a receiving module 401 , an automatic model search module 402 , a computing power compression module 403 , a model training module 404 , and a performance evaluation module 405 .

Specifically, the receiving module 401 is connected to the model training module 404, and the model training module 404 is connected to the performance evaluation module 405. The module 403 is connected to the receiving module 401; the model training module 404 is used to receive the candidate model and train the candidate model to obtain the trained model; the performance evaluation module 405 is used to evaluate the performance of the trained model.

Specifically, the model training module 404 mainly has two functions: ① train candidate models to obtain the trained model; ② verify each candidate model to obtain the accuracy rate and computing power of each candidate model. The model training module will input the obtained two model indicators into the performance evaluation module. In order to efficiently utilize system resources, the model training module adopts a distributed parallel training method.

Specifically, the performance evaluation module 405 will receive the accuracy and computing power of each candidate model, evaluate the comprehensive performance of each candidate model, obtain the evaluation score of each candidate model, and retain the candidate models that meet the preset rules , to delete candidate models that do not satisfy the preset rules.

Specifically, the model automatic search module 402 receives candidate models, and performs neural network architecture search based on the candidate models. Search the network architecture according to the search strategy for the search space set by the user or the default full search space.

Specifically, the computing power compression module 403 receives candidate models. This module is the core module of the system and is responsible for directional compression of the computing power of the candidate models. The layers are subjected to directional compression processing.

It is not difficult to find that this embodiment is a system embodiment corresponding to the previous embodiment, and this embodiment can be implemented in cooperation with the previous embodiment. The relevant technical details mentioned in the previous embodiment are still valid in this embodiment, and will not be repeated here in order to reduce repetition. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied in the previous embodiment.

It is worth mentioning that all the modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, or a part of a physical unit, or multiple physical units. Combination of units. In addition, in order to highlight the innovative part of the present application, units that are not closely related to solving the technical problem proposed in the present application are not introduced in this embodiment, but this does not mean that there are no other units in this embodiment.

An embodiment of the present application relates to a server, as shown in FIG. 7 , including at least one processor 501; and a memory 502 communicatively connected to at least one processor 501; Instructions to be executed, the instructions are executed by at least one processor 501, so that at least one processor 501 can execute the communication control method as described above.

Wherein, the memory 502 and the processor 501 are connected by a bus, and the bus may include any number of interconnected buses and bridges, and the bus connects one or more processors 501 and various circuits of the memory 502 together. The bus may also connect together various other circuits such as peripherals, voltage regulators, and power management circuits, all of which are well known in the art and therefore will not be further described herein. The bus interface provides an interface between the bus and the transceivers. A transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing means for communicating with various other devices over a transmission medium. The data processed by the processor 501 is transmitted on the wireless medium through the antenna, further, the antenna also receives the data and transmits the data to the processor 501 .

Processor 501 is responsible for managing the bus and general processing, and may also provide various functions including timing, peripheral interface, voltage regulation, power management and other control functions. And the memory 502 may be used to store data used by the processor 501 when performing operations.

An embodiment of the present application relates to a computer-readable storage medium storing a computer program. The above method embodiments are implemented when the computer program is executed by the processor.

That is, those skilled in the art can understand that all or part of the steps in the method of the above-mentioned embodiments can be completed by instructing related hardware through a program, the program is stored in a storage medium, and includes several instructions to make a device ( It may be a single-chip microcomputer, a chip, etc.) or a processor (processor) to execute all or part of the steps of the methods in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. .

Those of ordinary skill in the art can understand that the above-mentioned embodiments are specific embodiments for realizing the present application, and in practical applications, various changes can be made to it in form and details without departing from the spirit and spirit of the present application. scope.

Claims (13)

A method for model compression, comprising: receive candidate models; performing a neural network architecture search on the candidate models to obtain a plurality of transformation models; performing computing power compression on the candidate model to obtain a compressed model; Using the transformation model and the compression model as the candidate model, re-perform the neural network architecture search and the computing power compression. The model compression method according to claim 1, wherein said compression of said candidate model to obtain a compressed model comprises: Calculate the computing power of each layer in the candidate model; Comparing the computing power of each of the layers mentioned above, the layer with the largest computing power is obtained; Perform computing power compression on the layer with the largest computing power to obtain the compressed model. The model compression method according to claim 2, wherein said performing computing power compression on the layer with the largest computing power includes: Reduce the parameter amount of the layer with the largest computing power, or remove the layer with the largest computing power, or reduce the size of the input feature map of the layer with the largest computing power. The model compression method according to claim 1, wherein said performing neural network architecture search on said candidate model to obtain a plurality of transformed models, comprising: The neural network architecture search is performed based on a preset search space, and the candidate models are randomly transformed to obtain multiple transformed models. The model compression method according to claim 1, wherein, after receiving the candidate model, further comprising: Train the candidate model to obtain a trained model; Obtain the accuracy and computing power of the trained model; The evaluation parameters of the candidate models corresponding to the trained models are obtained based on the accuracy and the computing power. The model compression method according to claim 5, wherein the evaluation parameters include an evaluation score; after obtaining the evaluation parameters of the candidate model corresponding to the trained model based on the accuracy and the computing power, Also includes: Retaining the candidate models whose evaluation scores are greater than or equal to a preset score, or retaining the N candidate models with the largest evaluation scores. The model compression method according to claim 5, wherein the step of receiving a candidate model to the step of using the transformed model and the compressed model as the candidate model is one polling; the evaluation parameters include Evaluation score; after obtaining the evaluation parameters of the candidate model corresponding to the trained model based on the accuracy and the computing power, it also includes: Determine whether the maximum evaluation score obtained in this poll is greater than the maximum evaluation score obtained in the previous poll; If the maximum evaluation score obtained by this poll is greater than the maximum evaluation score obtained by the previous poll, enter the steps of neural network architecture search and computing power compression; If the maximum evaluation score obtained by this polling is not greater than the maximum evaluation score obtained by the previous polling, then add 1 to the counter value to determine whether the counter value is less than a preset threshold; the initial value of the counter value is 0; If the counter value is less than the preset threshold, enter the steps of neural network architecture search and computing power compression; if the counter value is greater than or equal to the preset threshold, the polling ends. The model compression method according to claim 7, wherein, when polling ends, the target candidate model is obtained from a plurality of candidate models; the accuracy of the target candidate model is greater than a preset accuracy and The computing power of the target candidate model is less than a preset computing power. The model compression method according to claim 1, wherein the candidate model received in the first round of polling is a reference model input by a user. A model compression system, including: a receiving module, an automatic model search module, a computing power compression module, and a selection module; The receiving module is used to receive candidate models; The model automatic search module is used to search the neural network architecture for the candidate model, obtain a plurality of transformed models and input them into the receiving module as the candidate model; The computing power compression module is used to perform computing power compression on the candidate model to obtain a compressed model, and the compressed model is input into the receiving module as the candidate model. The model compression system according to claim 10, further comprising: a model training module and a performance evaluation module; The model training module is connected to the receiving module, the model training module is also connected to the performance evaluation module, and the performance evaluation module is respectively connected to the model automatic search module and the computing power compression module, and the model automatic search The module and the computing power compression module are respectively connected to the receiving module; The model training module is used to train the candidate model to obtain a trained model; The performance evaluation module is used to obtain the accuracy and computing power of the trained model, and obtain the evaluation parameters of the candidate models corresponding to the trained model based on the accuracy and the computing power. A server, including: at least one processor; and, a memory communicatively coupled to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, the instructions are executed by the at least one processor, so that the at least one processor can perform the operation described in any one of claims 1 to 9 The model compression method described above. A computer-readable storage medium storing a computer program, wherein the computer program implements the model compression method according to any one of claims 1 to 9 when executed by a processor.

Images