CN118245811B

CN118245811B - Model parameter management method and device, storage medium and electronic equipment

Info

Publication number: CN118245811B
Application number: CN202410679713.9A
Authority: CN
Inventors: 郑玉会
Original assignee: Suzhou Metabrain Intelligent Technology Co Ltd
Current assignee: Suzhou Metabrain Intelligent Technology Co Ltd
Priority date: 2024-05-29
Filing date: 2024-05-29
Publication date: 2024-09-24
Anticipated expiration: 2044-05-29
Also published as: CN118245811A

Abstract

The embodiment of the application provides a method and a device for managing model parameters, a storage medium and electronic equipment, wherein the method comprises the following steps: detecting a target model corresponding to the first model parameter under the condition that the first model parameter in the first storage space is detected to be subjected to target data adjustment operation; reading second model parameters from the first storage space, and reading third model parameters corresponding to the target model from the second storage space; detecting change description information corresponding to the third model parameter according to the second model parameter and the third model parameter; the object model and the change description information having the correspondence relationship are stored in the second storage space. The application solves the problem of lower management efficiency of the model parameters, thereby achieving the effect of improving the management efficiency of the model parameters.

Description

Model parameter management method and device, storage medium and electronic equipment

Technical Field

The embodiment of the application relates to the field of models, in particular to a method and a device for managing model parameters, a storage medium and electronic equipment.

Background

Corresponding model parameters need to be used during model training, which may be frequently adjusted during model training, and often are large, e.g., on the order of billions of model parameters for large models.

In the related art, only model parameters used for training a model at present are stored in the model training platform, for example, model parameters used for training a model at present are stored in a specified directory in the model training platform. In this way, it is difficult for a user to distinguish between data stored in the model training platform, for example, it is difficult for the user to distinguish which model parameter the parameters stored in the catalog are in particular, and after the training parameters of the model are adjusted, the user may train the model using the model parameters after a long time interval, and the management of the model parameters in the related art may result in the user not being able to train the model using the correct model parameters.

Disclosure of Invention

The embodiment of the application provides a method and a device for managing model parameters, a storage medium and electronic equipment, which are used for at least solving the problem of low management efficiency of the model parameters in the related technology.

According to an embodiment of the present application, there is provided a method for managing model parameters, a model training system having a first storage space for storing current model parameters of a model and a model parameter management service having a second storage space for storing initial model parameters of the model, the method being applied to the model parameter management service, the method comprising: detecting a target model corresponding to a first model parameter in the first storage space under the condition that the first model parameter in the first storage space is detected to be subjected to target data adjustment operation, wherein the target data adjustment operation is a data adjustment operation for updating the parameter in the first storage space; reading a second model parameter from the first storage space, and reading a third model parameter corresponding to the target model from the second storage space, wherein the second model parameter is a model parameter after the target data adjustment operation is executed on the first model parameter; detecting change description information corresponding to the third model parameter according to the second model parameter and the third model parameter, wherein the change description information is used for indicating target change of the third model parameter; and storing the target model and the change description information with the corresponding relation in the second storage space.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the detecting the target model corresponding to the first model parameter includes: acquiring a first detection request, wherein the first detection request is used for requesting to detect the target model corresponding to the first model parameter; and responding to the first detection request, and controlling the storage monitoring service to detect the target model corresponding to the first model parameter.

In an exemplary embodiment, a detector is disposed in the storage monitoring service, and the controlling the storage monitoring service to detect the target model corresponding to the first model parameter includes: acquiring a second detection request, wherein the second detection request is used for requesting to control the storage monitoring service to detect the target model corresponding to the first model parameter; and responding to the second detection request, and controlling the detector to detect the target model corresponding to the first model parameter from model parameters and models with corresponding relations.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the reading the second model parameter from the first storage space includes: acquiring a first reading request, wherein the first reading request is used for requesting to read the second model parameters corresponding to the target model from the first storage space; and responding to the first reading request, and controlling the storage monitoring service to read the second model parameters corresponding to the target model from the first storage space.

In an exemplary embodiment, the controlling the storage monitoring service to read the second model parameter corresponding to the target model from the first storage space includes: controlling the storage monitoring service to detect a first storage address of the second model parameter in the first storage space; and controlling the storage monitoring service to read the data stored in the first storage address to obtain the second model parameters.

In an exemplary embodiment, a model management service is deployed in the model parameter management service, and the reading, from the second storage space, a third model parameter corresponding to the target model includes: acquiring a second read request, wherein the second read request is used for requesting to read the third model parameter corresponding to the target model from the second storage space; and responding to the second reading request, and controlling the model management service to read the third model parameters corresponding to the target model from the second storage space.

In an exemplary embodiment, the controlling the model management service to read the third model parameter corresponding to the target model from the second storage space includes: controlling the model management service to detect a second storage address of the third model parameter in the second storage space; and controlling the model management service to read the data stored in the second storage address to obtain the third model parameters.

In an exemplary embodiment, the detecting, according to the second model parameter and the third model parameter, change description information corresponding to the third model parameter includes: detecting first description information of the second model parameters and second description information of the third model parameters, wherein the first description information is used for describing first update time of the second model parameters and first quantity of the second model parameters, and the second description information is used for describing second update time of the third model parameters and second quantity of the third model parameters; and detecting change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the detecting the first description information of the second model parameter includes: acquiring a third detection request, wherein the third detection request is used for requesting to detect the first description information of the second model parameters; and in response to the third detection request, controlling the storage monitoring service to detect the first update time of the second model parameter and controlling the storage monitoring service to count the first number of the second model parameters.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, the storage monitoring service is deployed with a full-scale statistics table, and the detecting the first description information of the second model parameter includes: and controlling the storage monitoring service to read the first updating time corresponding to the second model parameter from the full-quantity statistical table, and controlling the storage monitoring service to count the first quantity of parameters included in the second model parameter, wherein the full-quantity statistical table is used for storing one or more models, model parameters and updating time with corresponding relations, and the one or more models, model parameters and updating time with corresponding relations comprise the target model, the second model parameter and the first updating time with corresponding relations.

In an exemplary embodiment, the model parameter management service is deployed with a model management service, and the detecting the second description information of the third model parameter includes: acquiring a fourth detection request, wherein the fourth detection request is used for requesting to detect the second description information of the third model parameter; and in response to the fourth detection request, controlling the model management service to detect the second update time of the third model parameter and controlling the model management service to count the second number of the third model parameter.

In an exemplary embodiment, the model parameter management service is configured to deploy a storage monitoring service and a model management service, and the detecting, according to the first update time and the first number indicated by the first description information, the second update time and the second number indicated by the second description information, change description information corresponding to the third model parameter includes: controlling the storage monitoring service to extract the second description information detected by the model management service; and controlling the storage monitoring service to detect the change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

In an exemplary embodiment, the controlling the storage monitoring service to detect the change description information corresponding to the third model parameter according to the first update time and the first number indicated by the first description information and the second update time and the second number indicated by the second description information includes: controlling the storage monitoring service to compare the first updating time with the second updating time to obtain a first comparison result, and controlling the storage monitoring service to compare the first quantity with the second quantity to obtain a second comparison result; and controlling the storage monitoring service to detect the change description information according to the first comparison result and the second comparison result.

In an exemplary embodiment, the controlling the storage monitoring service to detect the change description information according to the first comparison result and the second comparison result includes: controlling the storage monitoring service to detect a fourth model parameter except the third model parameter in the second model parameter and controlling the storage monitoring service to detect that the third model parameter has changed, wherein the target change comprises the new change and the fourth model parameter, when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is greater than the second number; or controlling the storage monitoring service to detect a fifth model parameter of the third model parameters except the second model parameter when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is smaller than the second number, and controlling the storage monitoring service to detect that the third model parameter has reduced variation, wherein the target variation comprises the reduced variation and the fifth model parameter; or when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is equal to the second number, controlling the storage monitoring service to detect a sixth model parameter which is different from the third model parameter in the second model parameter, and controlling the storage monitoring service to detect that the third model parameter is subjected to adjustment change, wherein the target change comprises the adjustment change and the sixth model parameter.

In an exemplary embodiment, the model parameter management service is deployed with a model management service and a storage monitoring service, the model management service and the storage monitoring service are connected, and the storing the target model and the change description information with the correspondence in the second storage space includes: controlling the model management service to extract the change description information detected by the storage monitoring service; and controlling the model management service to store the target model and the change description information with the corresponding relationship in the second storage space.

In an exemplary embodiment, the controlling the model management service to store the target model and the change description information having a correspondence relationship in the second storage space includes: controlling the model management service to store the target model, the new variation and the fourth model parameter having a correspondence in the second storage space, in case the variation description information is used to indicate that the target variation occurring in the third model parameter includes the new variation and the fourth model parameter; or in the case that the change description information is used to indicate that the target change occurring in the third model parameter includes a decrease change and a fifth model parameter, controlling the model management service to store the target model, the decrease change, and the fifth model parameter having a correspondence in the second storage space; or in the case that the change description information is used for indicating that the target change occurring in the third model parameter includes an adjustment change and a sixth model parameter, controlling the model management service to store the target model, the adjustment change and the sixth model parameter having a correspondence in the second storage space.

In an exemplary embodiment, after the storing the target model and the change description information having the correspondence relationship in the second storage space, the method further includes: detecting a target model state of the target model under the condition that a target deletion request is acquired, wherein the target deletion request is used for requesting to delete the second model parameters stored in the first storage space; in the case where the object model state is used to indicate that the object model has been deployed, suspending responding to the object deletion request.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, the storage monitoring service is deployed with a detector, and before the detecting the target model corresponding to the first model parameter, the method further includes: acquiring a fifth detection request, wherein the fifth detection request is used for requesting to detect whether the current model parameters of the model stored in the first storage space are subjected to the target data adjustment operation; and controlling the detector to detect whether the target data adjustment operation is performed on the current model parameters of the model stored in the first storage space in response to the fifth detection request.

According to another embodiment of the present application, there is provided a management apparatus for model parameters, a model training system having a first storage space for storing current model parameters of a model and a model parameter management service having a second storage space for storing initial model parameters of the model, the apparatus being applied to the model parameter management service, the apparatus comprising:

The first detection module is used for detecting a target model corresponding to a first model parameter under the condition that the first model parameter in the first storage space is detected to be subjected to target data adjustment operation, wherein the target data adjustment operation is a data adjustment operation for updating the parameter in the first storage space;

The reading module is used for reading second model parameters from the first storage space and reading third model parameters corresponding to the target model from the second storage space, wherein the second model parameters are model parameters after the target data adjustment operation is performed on the first model parameters;

The second detection module is used for detecting change description information corresponding to the third model parameter according to the second model parameter and the third model parameter, wherein the change description information is used for indicating target change of the third model parameter;

and the storage module is used for storing the target model with the corresponding relation and the change description information in the second storage space.

According to a further embodiment of the application, there is also provided a computer readable storage medium having stored therein a computer program, wherein the computer program is arranged to perform the steps of any of the method embodiments described above when run.

According to a further embodiment of the application there is also provided an electronic device comprising a memory having stored therein a computer program and a processor arranged to run the computer program to perform the steps of any of the method embodiments described above.

According to a further embodiment of the application, there is also provided a computer program product comprising a computer program which, when executed by a processor, implements the steps of any of the method embodiments described above.

According to the application, the model parameter management service is deployed on the model training system, the model parameter used for training the model is stored in the model training system, the model parameter management service stores the initial model parameter of the model, and it can be understood that the model parameter used for training the model and the initial model parameter of the model can be different, and the model parameter management service can store the updated model parameter compared with the change of the initial model parameter under the condition that the model parameter stored in the model training system is updated.

Drawings

FIG. 1 is a block diagram of the hardware architecture of a server device of a method for managing model parameters according to an embodiment of the present application;

FIG. 2 is a flow chart of a method of managing model parameters according to an embodiment of the application;

FIG. 3 is a framework diagram of the management of an alternative model parameter according to an embodiment of the present application;

FIG. 4 is a schematic diagram of an alternative method of managing model parameters according to an embodiment of the application;

FIG. 5 is a schematic diagram of an alternative data protection according to an embodiment of the application;

FIG. 6 is a schematic diagram of the management of an alternative model state implemented in accordance with the present application;

Fig. 7 is a block diagram of a management apparatus of model parameters according to an embodiment of the present application.

Detailed Description

Embodiments of the present application will be described in detail below with reference to the accompanying drawings in conjunction with the embodiments.

It should be noted that the terms "first," "second," and the like in the description and the claims of the present application and the above figures are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order.

The method embodiments provided in the embodiments of the present application may be executed in a server apparatus or similar computing device. Taking the example of running on a server device, fig. 1 is a block diagram of a hardware structure of a server device of a method for managing model parameters according to an embodiment of the present application. As shown in fig. 1, the server device may include one or more (only one is shown in fig. 1) processors 102 (the processor 102 may include, but is not limited to, a microprocessor MCU, a programmable logic device FPGA, or the like processing means) and a memory 104 for storing data, wherein the server device may further include a transmission device 106 for communication functions and an input-output device 108. It will be appreciated by those of ordinary skill in the art that the architecture shown in fig. 1 is merely illustrative and is not intended to limit the architecture of the server apparatus described above. For example, the server device may also include more or fewer components than shown in FIG. 1, or have a different configuration than shown in FIG. 1.

The memory 104 may be used to store a computer program, for example, a software program of application software and a module, such as a computer program corresponding to a method for managing model parameters in an embodiment of the present application, and the processor 102 executes the computer program stored in the memory 104 to perform various functional applications and data processing, that is, implement the above-mentioned method. Memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include memory remotely located with respect to the processor 102, which may be connected to the server device via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The transmission device 106 is used to receive or transmit data via a network. Specific examples of the network described above may include a wireless network provided by a communication provider of a server device. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, simply referred to as a NIC) that can connect to other network devices through a base station to communicate with the internet. In one example, the transmission device 106 may be a Radio Frequency (RF) module, which is configured to communicate with the internet wirelessly.

First, the features involved in the embodiments of the present application are explained as follows:

AI: ARTIFICIAL INTELLIGENCE, artificial intelligence.

LLM: large Language Model, large language models, a model based on machine learning and natural language processing.

The Web: the World Wide Web is a network service established on the Internet, and provides a graphical and easily-accessible visual interface for a browser to search and browse information on the Internet.

Rest API: the Representational STATE TRANSFER Application Programming Interface presents a layer transition application programming interface, a REST-style based Web API, commonly used in Web services.

Pod: the smallest resource management component in the container kubernetes, pod, is the resource object that minimizes running the containerized application. One Pod represents one process running in the cluster.

In this embodiment, a method for managing model parameters is provided, a first storage space and a model parameter management service are deployed on a model training system, the first storage space is used for storing current model parameters of a model, the current model parameters are used for training the model, a second storage space is deployed in the model parameter management service, the second storage space is used for storing initial model parameters of the model, the method is applied to the model parameter management service, fig. 2 is a flowchart of a method for managing model parameters according to an embodiment of the present application, and as shown in fig. 2, the flowchart includes the following steps:

Step S202, detecting a target model corresponding to a first model parameter in the first storage space when detecting that the first model parameter is executed with a target data adjustment operation, wherein the target data adjustment operation is a data adjustment operation for updating parameters in the first storage space;

step S204, reading a second model parameter from the first storage space, and reading a third model parameter corresponding to the target model from the second storage space, wherein the second model parameter is a model parameter after the target data adjustment operation is performed on the first model parameter;

Step S206, detecting change description information corresponding to the third model parameter according to the second model parameter and the third model parameter, wherein the change description information is used for indicating target change of the third model parameter;

Step S208, storing the target model and the change description information having the correspondence relationship in the second storage space.

Through the steps, the model parameter management service is deployed on the model training system, the model parameter used for training the model is stored in the model training system, the initial model parameter of the model is stored in the model parameter management service, and it can be understood that the model parameter used for training the model and the initial model parameter of the model can be different, and in the case that the model parameter stored in the model training system is updated, the model parameter management service can store the updated model parameter which is changed compared with the initial model parameter, in this way, the adjustment of the model parameter can be obtained at any time under the condition that the model parameter is updated, so that the problem that the management efficiency of the model parameter is lower can be solved, and the effect of improving the management efficiency of the model parameter is achieved.

In the solution provided in step S202, the current model parameters may be, but are not limited to, used for training a model, for example, the current model parameters may include, but are not limited to, a currently trained model, parameters required for the current model training, and the like, and the current model parameters may be current model parameters obtained after the initial model parameters are adjusted, for example, values of at least part of model parameters in the initial model parameters are adjusted to obtain the current model parameters, or at least part of model parameters in the initial model parameters are deleted to obtain the current model parameters, and the like.

Alternatively, in this embodiment, in the case where the first model parameter is subjected to the target data adjustment operation, the first model parameter may be, but not limited to, updated such as a new model parameter, a deleted model parameter, or a change in the value of the model parameter.

Alternatively, in the present embodiment, the model training system may be used for training a model, for example, but not limited to, an AI training and reasoning platform, and the like.

In one exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and may, but is not limited to, detect the target model corresponding to the first model parameter by: acquiring a first detection request, wherein the first detection request is used for requesting to detect the target model corresponding to the first model parameter; and responding to the first detection request, and controlling the storage monitoring service to detect the target model corresponding to the first model parameter.

Alternatively, in this embodiment, the target model corresponding to the first model parameter may be detected by, but not limited to, the storage monitoring service, and the first model parameter may be used to train the target model.

In an exemplary embodiment, a detector is disposed in the storage monitoring service, and the storage monitoring service may be controlled to detect the target model corresponding to the first model parameter by, but not limited to: acquiring a second detection request, wherein the second detection request is used for requesting to control the storage monitoring service to detect the target model corresponding to the first model parameter; and responding to the second detection request, and controlling the detector to detect the target model corresponding to the first model parameter from model parameters and models with corresponding relations.

Alternatively, in this embodiment, the control detector may, but is not limited to, detect the target model corresponding to the first model parameter from the model parameters and models having the correspondence relationship, each of the current model parameters of the models stored in the first storage space may, but is not limited to, have a corresponding model, and each model parameter corresponding to the model may, but is not limited to, be the same or different, e.g., model parameter 1 and model parameter 2 may each correspond to model 1.

In an exemplary embodiment, a storage monitoring service is deployed in the model parameter management service, a detector is deployed in the storage monitoring service, and before the target model corresponding to the first model parameter is detected, the method further includes: acquiring a fifth detection request, wherein the fifth detection request is used for requesting to detect whether the current model parameters of the model stored in the first storage space are subjected to the target data adjustment operation; and controlling the detector to detect whether the target data adjustment operation is performed on the current model parameters of the model stored in the first storage space in response to the fifth detection request.

Alternatively, in this embodiment, the storage monitoring service may, but is not limited to, detect, at a target period timing, whether the current model parameters of the model stored in the first storage space are subjected to the target data adjustment operation, and as an alternative example, the target period may, but is not limited to, 5 minutes or 2 minutes or 30 seconds, or the like.

In the solution provided in step S204, the second model parameter may be, but is not limited to, read from the first storage space, and it is understood that the second model parameter is different from the first model parameter, for example, the number of parameters included in the second model parameter is different from the first model parameter, or the value of the parameter included in the second model parameter is different from the value of the corresponding parameter included in the first model.

Alternatively, in the present embodiment, the third model parameter may include, but is not limited to, an initial model parameter of the target model, for example, the third model parameter may include, but is not limited to, an initial model parameter of the target model, which may include, but is not limited to, a model parameter used when the target model is first trained, for example, an initial model parameter of the target model may include, but is not limited to, a model parameter required for training the target model and the model.

In one exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the second model parameter may be read from the first storage space by, but not limited to: acquiring a first reading request, wherein the first reading request is used for requesting to read the second model parameters corresponding to the target model from the first storage space; and responding to the first reading request, and controlling the storage monitoring service to read the second model parameters corresponding to the target model from the first storage space.

Optionally, in this embodiment, a storage monitoring service is deployed in the model parameter management service, which may, but is not limited to, control the storage monitoring service to read, from the first storage space, a second model parameter corresponding to the target model.

In one exemplary embodiment, the storage monitoring service may be controlled to read the second model parameters corresponding to the target model from the first storage space by, but not limited to: controlling the storage monitoring service to detect a first storage address of the second model parameter in the first storage space; and controlling the storage monitoring service to read the data stored in the first storage address to obtain the second model parameters.

Optionally, in this embodiment, the first storage address may, but is not limited to, include a first starting offset address of the second model parameter in the first storage space and a first parameter size, where the first parameter size is used to indicate a data amount of the second model parameter, and may, but is not limited to, control the storage monitoring service to read the data of the first parameter size from the first starting offset address, to obtain the second model parameter.

In an exemplary embodiment, the model parameter management service is deployed with a model management service, and may, but is not limited to, read, from the second storage space, a third model parameter corresponding to the target model by: acquiring a second read request, wherein the second read request is used for requesting to read the third model parameter corresponding to the target model from the second storage space; and responding to the second reading request, and controlling the model management service to read the third model parameters corresponding to the target model from the second storage space.

Optionally, in this embodiment, a model management service is deployed in the model parameter management service, which may, but is not limited to, control the model management service to read, from the second storage space, a third model parameter corresponding to the target model.

In one exemplary embodiment, the model management service may be controlled to read the third model parameters corresponding to the target model from the second storage space by, but not limited to: controlling the model management service to detect a second storage address of the third model parameter in the second storage space; and controlling the model management service to read the data stored in the second storage address to obtain the third model parameters.

Optionally, in this embodiment, the second storage address may, but is not limited to, include a second starting offset address of the third model parameter in the second storage space and a second parameter size, where the second parameter size is used to indicate a data amount of the third model parameter, and may, but is not limited to, control the model management service to read, from the second starting offset address, data of the second parameter size, to obtain the third model parameter.

In the solution provided in step S206, the target change may be, but is not limited to, a change of the third model parameter compared to the second model parameter, and as an alternative example, the target change may be, but is not limited to, an addition change, a deletion change, an adjustment change, etc., for example, the third model parameter is added, at least part of the third model parameters are deleted, or at least part of the third model parameters are adjusted.

In one exemplary embodiment, the change description information corresponding to the third model parameter may be detected according to the second model parameter and the third model parameter by, but not limited to, the following ways: detecting first description information of the second model parameters and second description information of the third model parameters, wherein the first description information is used for describing first update time of the second model parameters and first quantity of the second model parameters, and the second description information is used for describing second update time of the third model parameters and second quantity of the third model parameters; and detecting change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

Alternatively, in the present embodiment, the first number may be, but not limited to, a number of parameters included for indicating the second model parameter, the second number may be, but not limited to, a number of parameters included for indicating the third model parameter, the first update time may be, but not limited to, a time for indicating that the first model parameter is subjected to the target data adjustment operation, and the second update time may be, but not limited to, a time for indicating that the third model parameter is stored in the model parameter management service.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and may, but is not limited to, detect the first description information of the second model parameter by: acquiring a third detection request, wherein the third detection request is used for requesting to detect the first description information of the second model parameters; and in response to the third detection request, controlling the storage monitoring service to detect the first update time of the second model parameter and controlling the storage monitoring service to count the first number of the second model parameters.

Alternatively, in this embodiment, the first update time of the second model parameter read by the storage monitoring service detection may be controlled, and the first number of the second model parameters counted by the storage monitoring service may be controlled, for example, but not limited to, the number of parameters included in the second model parameter counted by the storage monitoring service may be controlled, for example, the second model parameter includes N parameters, and the update time of each parameter in the N parameters counted by the storage monitoring service is controlled to obtain N update times, where N is a positive integer, and the first update time includes N update times.

In one exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the storage monitoring service is deployed with a full-scale statistics table, and the first description information of the second model parameter may be, but is not limited to, detected by: and controlling the storage monitoring service to read the first updating time corresponding to the second model parameter from the full-quantity statistical table, and controlling the storage monitoring service to count the first quantity of parameters included in the second model parameter, wherein the full-quantity statistical table is used for storing one or more models, model parameters and updating time with corresponding relations, and the one or more models, model parameters and updating time with corresponding relations comprise the target model, the second model parameter and the first updating time with corresponding relations.

Optionally, in this embodiment, the update time of the model parameter of each model may be recorded in the full statistics table, in this case, the update time may also be, but not limited to, controlling the storage monitoring service to search the second model parameter corresponding to the target model from the full statistics table, and in the case controlling the storage monitoring service to search the second model parameter corresponding to the target model from the full statistics table, reading the first update time corresponding to the second model parameter from the full statistics table.

In one exemplary embodiment, the model management service is deployed with a model management service, and may, but is not limited to, detect the second description information of the third model parameter by: acquiring a fourth detection request, wherein the fourth detection request is used for requesting to detect the second description information of the third model parameter; and in response to the fourth detection request, controlling the model management service to detect the second update time of the third model parameter and controlling the model management service to count the second number of the third model parameter.

Alternatively, in this embodiment, the control model management service may, but is not limited to, detect the second update time of the read third model parameter, and control the model management service to count the second number of the third model parameter, for example, may, but is not limited to, control the model management service to count the number of parameters included in the third model parameter, for example, the third model parameter includes M parameters, and control the model management service to count the update time of each parameter in the M parameters, to obtain M update times, where M is a positive integer, and the second update time includes M update times.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service and a model management service, and may, but is not limited to, detect the change description information corresponding to the third model parameter according to the first update time and the first number indicated by the first description information, and the second update time and the second number indicated by the second description information by: controlling the storage monitoring service to extract the second description information detected by the model management service; and controlling the storage monitoring service to detect the change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

Optionally, in this embodiment, the change description information corresponding to the third model parameter may be detected according to the first update time and the first number indicated by the first description information, the second update time and the second number indicated by the second description information, and the following manners, but not limited to: controlling the model management service to extract the first description information detected by the storage monitoring service; and controlling the model management service to detect the change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

According to the embodiment of the application, the processing resources of the coordination model management service and the storage monitoring service are realized through the model management service or the storage monitoring service to detect the change description information of the model parameters, for example, under the condition that the storage monitoring service is busy, the change description information of the model parameters can be detected through the model management service, and the flexibility of the change description information of the detection model parameters is improved.

In one exemplary embodiment, the storage monitoring service may be controlled to detect the change description information corresponding to the third model parameter according to the first update time and the first number indicated by the first description information, and the second update time and the second number indicated by the second description information, by, but not limited to: controlling the storage monitoring service to compare the first updating time with the second updating time to obtain a first comparison result, and controlling the storage monitoring service to compare the first quantity with the second quantity to obtain a second comparison result; and controlling the storage monitoring service to detect the change description information according to the first comparison result and the second comparison result.

Alternatively, in this embodiment, the first number may be greater than, less than, or equal to the second number, the second update time may be earlier than, or later than, the first update time, or the second update time may be, but is not limited to, equal to the first update time.

In one exemplary embodiment, the storage monitoring service may be controlled to detect the change description information according to the first comparison result and the second comparison result by, but not limited to, one of the following:

in a first mode, when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is greater than the second number, the storage monitoring service is controlled to detect a fourth model parameter except the third model parameter in the second model parameters, the storage monitoring service is controlled to detect that the third model parameter has changed, and the target change includes the new change and the fourth model parameter.

Optionally, in this embodiment, when the first comparison result is used to indicate that the second update time is earlier than the first update time and the second comparison result is used to indicate that the first number is greater than the second number, it may indicate that the second model parameter is newly added to the third model parameter, and it may be understood that the fourth model parameter is a model parameter that is not included in the third model parameter and is included in the second model parameter, and in this case, the access path of the fourth model parameter in the first storage space may be recorded in the change description information, but not limited to this.

In a second mode, when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is smaller than the second number, the storage monitoring service is controlled to detect a fifth model parameter except the second model parameter in the third model parameters, the storage monitoring service is controlled to detect that the third model parameters have reduced changes, and the target changes include the reduced changes and the fifth model parameters.

Optionally, in this embodiment, in a case where the first comparison result is used to indicate that the second update time is earlier than the first update time and the second comparison result is used to indicate that the first number is smaller than the second number, it may indicate that at least a part of the model parameters in the third model parameters are deleted, and the fifth model parameters are model parameters included in the third model parameters and not included in the second model parameters, where in this case, it may also be but not limited to recording the access path of the fifth model parameters in the second storage space in the change description information.

In a third mode, when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is equal to the second number, the storage monitoring service is controlled to detect a sixth model parameter different from the third model parameter in the second model parameters, the storage monitoring service is controlled to detect that an adjustment change occurs in the third model parameters, and the target change includes the adjustment change and the sixth model parameter.

Optionally, in this embodiment, when the first comparison result is used to indicate that the second update time is earlier than the first update time, and the second comparison result is used to indicate that the first number is equal to the second number, it may be that the value of at least part of the model parameters in the second model parameters is changed compared with the value of the corresponding model parameters in the third model parameters, and it may also be, but not limited to, recording the access address of the sixth model parameters in the first storage space in the change description information.

In the technical scheme provided in the step S208, the target model and the change description information with the corresponding relation can be stored in the second storage space, and in this way, the change of the model parameter after each adjustment compared with the initial model parameter is accurately recorded, the accuracy of adjusting the model parameter in the model training process of the debugger is improved, and the training efficiency of the model is improved.

In one exemplary embodiment, the model parameter management service is deployed with a model management service and a storage monitoring service, and the model management service and the storage monitoring service are connected, and the target model and the change description information with a corresponding relationship may be stored in the second storage space by, but not limited to: controlling the model management service to extract the change description information detected by the storage monitoring service; and controlling the model management service to store the target model and the change description information with the corresponding relationship in the second storage space.

Alternatively, in the present embodiment, the model management service may, but is not limited to, extract and store the change description information detected by the monitoring service through the application program interface, for example, the model management service may, but is not limited to, extract and store the change description information detected by the monitoring service through the rest API.

In one exemplary embodiment, controlling the model management service to store the target model and the change description information having a correspondence relationship in the second storage space may include, but is not limited to, one of the following cases:

In case one, in a case where the change description information is used to indicate that the target change occurring in the third model parameter includes a new change and a fourth model parameter, the model management service is controlled to store the target model, the new change, and the fourth model parameter having a correspondence relationship in the second storage space.

Alternatively, in the present embodiment, in the case where the change description information is used to indicate that the target change occurring in the third model parameter includes the new change and the fourth model parameter, it is also possible, but not limited to, controlling the model management service to store the target model, the new change, the fourth model parameter, and the access path of the fourth model parameter in the first storage space, which have a correspondence relationship, in the second storage space.

In a second case, in a case where the change description information is used to indicate that the target change occurring in the third model parameter includes a reduced change and a fifth model parameter, the model management service is controlled to store the target model, the reduced change, and the fifth model parameter having a correspondence relationship in the second storage space.

Optionally, in this embodiment, in a case where the change description information is used to indicate that the target change occurring in the third model parameter includes a decrease change and a fifth model parameter, the control model management service may also, but is not limited to, store the target model, the decrease change, the fifth model parameter, and an access path of the fifth model parameter in the second storage space, which have a correspondence relationship, in the second storage space.

In a third aspect, in a case where the change description information is used to indicate that the target change occurring in the third model parameter includes an adjustment change and a sixth model parameter, the model management service is controlled to store the target model, the adjustment change, and the sixth model parameter having a correspondence relationship in the second storage space.

Optionally, in this embodiment, in a case where the change description information is used to indicate that the target change occurring in the third model parameter includes the adjustment change and the sixth model parameter, it is also possible, but not limited to, controlling the model management service to store the target model, the adjustment change, the sixth model parameter, and the access path of the sixth model parameter in the first storage space, which have a correspondence relationship, in the second storage space.

In order to better understand the management method of the model parameters in the embodiment of the present application, the following explanation and description of the management method of the model parameters in the embodiment of the present application are provided in connection with an alternative embodiment, which may be, but not limited to, applicable to the embodiment of the present application.

FIG. 3 is a block diagram of an alternative management of model parameters according to an embodiment of the present application, and as shown in FIG. 3, the management method of model parameters in the embodiment of the present application may include, but is not limited to, parameter management, state management, version management, and data protection. A storage monitoring service and a model management service, as well as a model development service, are deployed on a model training model (e.g., AI platform), with model fine-tuning and service deployment deployed in the model development.

1. Parameter management is performed on model data (corresponding to model parameters) downloaded to a user's home directory (e.g., file management) from a remote end through a data download model.

If model parameters of a model are numerous (for example, model parameters of a large model are numerous), it is difficult for a user to clearly memorize the parameters of the model after downloading the model data only in file management, and as time goes long, when the downloaded model is deployed for one or more weeks, the parameters of the model cannot be completely memorized, so that the model data needs to be managed by a nano tube to the model management and the parameter management is needed.

Parameter management for large model management may include, but is not limited to, maintenance of large model meta-information, curing and persisting large model parameters into a database as large models are imported and edited. The data structure of meta-information of large model parameters is as follows:

{

modelFormat: "PyTorch",// model format, string type, support configuration

"ModelSizeInBillions":20,// model parameters in units of billion, numerical type

ModelName ss, name of model, type of character string, global unique under same user

ModelLang [ "Chinese", "English" ],// model language, array type, support configuration

"ContextLength":11,// maximum token length allowed, value type

ModelFamily aichuan-2-chat, model class, string type, support configuration

ModelAbility "chat",// model capabilities, string types, support configuration

"ModelUri"/mnt/xxx ",// model path, string type

"Version":1,// model version, value type

"Quantizations" [ "4-bit", "8-bit", "q4-0" ]// quantization method, array type, support configuration

}

The model format, the model language, the model category, the model capability and the quantization method support configuration. The specific configuration method comprises the following steps: the parameter values supported by the current version are written into the configuration center file, and the model management micro service (equivalent to the model management service) is obtained by reading the configuration center file when being started, so that the selection items of the Web page can be provided in a rest API mode, and the user can flexibly and conveniently obtain the Web page. When the parameters need to be expanded, only the configuration center file needs to be modified, the model management micro service is restarted, and after restarting, the rest API can read new configuration parameters.

The parameters of the data structure are persisted into the database for management.

2. Version management is performed for the same type of model.

The continuous high-speed development of the large model, functions, parameters and the like are iterated rapidly, after a model with a fixed model name is imported into the platform, version management can be performed on the basis of the model, and the model parameters of the model with the same type and rapid iteration are specially used for nanotubes.

Version management can help users to induce the same type of model, filling of redundant parameters of the same type of model is reduced when the model is imported, and version numbers are managed by the platform. After the user can download the model data of the same type to the user home directory, the model is managed to the model management by the function of the newly added version.

The version management of the model brings convenience to the user, including:

1) The models of the same type are classified, so that the user can find the models conveniently.

2) And the filling of redundant parameters of the same type of model is reduced.

3) The version number is automatically managed by the platform without user-defined input.

4) The latest version and the number of versions of the model are displayed in the model list at a glance, so that the user has a clear view of the model.

3. And storing and monitoring the model data managed by the nano tube model, reporting the condition of the model file, and updating the model state.

And reporting the model state through the monitoring result, and carrying out model state management on the model management module. The method reduces shared storage resources and network overhead occupied by large quantity of model data handling, and can accurately provide the user model data through management of model states.

After the monitoring model data (corresponding to the model parameters) is stored and reported, the model management module (corresponding to the model management service) performs state update, and the update of the model state comprises: has failed and has changed. Wherein, failure indicates that the storage monitoring module (equivalent to storage monitoring service) monitors that the model data does not exist, and the model file cannot be found under the model path. The change indicates that the storage monitoring module monitors the change (corresponding to update) of the model data.

FIG. 4 is a schematic diagram of an alternative method for managing model parameters according to an embodiment of the present application, where as shown in FIG. 4, the storage monitoring micro-service and the model management micro-service have been decoupled, and may, but not limited to, separate the storage monitoring module from the AI platform service module into a micro-service (e.g., the storage monitoring micro-service), and reasonably allocate the storage monitoring module resources according to a priori conditions, and decouple from the AI platform service module. The method is characterized in that the storage monitoring is used for monitoring shared storage resources of the whole platform, when the shared storage occupation amount is high, the storage monitoring inevitably occupies resources of a CPU (Central Processing Unit ) and a memory, and after the micro-service is decoupled, the problem that core service micro-service is abnormal due to the fact that the CPU and the memory of the storage monitoring function are too high in occupation ratio is avoided, and the problem that the AI platform cannot be normally used due to the fact that the storage monitoring occupies the service module resources is prevented.

The storage monitoring module (equivalent to the storage monitoring micro-service) stores the file service conditions of all users in the platform in an incremental statistics mode, and the incremental statistics is performed when the catalogue changes. Model data file event listeners (equivalent to detectors) are added to the storage monitoring microservice, and the listening time length is performed once according to the target period (for example, 5 min) of the storage statistics. The method is concretely realized as follows: and monitoring the changes of all the user home directories by using a file monitor, and reporting an event-triggered model state update when the file changes are monitored.

The method for managing model parameters in the embodiment of the application can be implemented by the following steps:

Step 1: the storage monitoring module stores the file use condition of all users in the platform in an incremental statistical mode, when a model is imported, according to the model catalogue selected from the user home catalogue, the update time (corresponding to the first update time) of the model catalogue and the subdirectories thereof can be acquired in a full-volume storage table (corresponding to the full-volume statistical table) through RestAPI call, and can be written into a MariaDB database in model management. And each model parameter is also written into a MariaDB database of model management for lasting management.

Step 2: in the storage monitoring micro-service, a model file event monitor (equivalent to a detector) is defined, interface communication is carried out with the model management micro-service by means of a rest API, and a model catalog of all non-failure models is obtained from the model management micro-service.

Step 3: the method includes the steps that a monitoring interval is defined to be 5min (or 3 min or 30s, and the like, the application is not limited to the above), when a specified 5min monitoring time arrives, the model catalogue in the step 2 is obtained, the model catalogue obtained in the step 2 is searched in a total storage statistical table, and if the model catalogue does not exist, the step 3 is skipped; if a model directory exists, go to step 4.

Step 3: the model catalog does not exist, and the reporting model management micro-service model catalog does not exist an event.

Step 4: acquiring update time (corresponding to second update time) of the model catalogue and all the subdirectories in the step 2, defining the update time as LIST LIST1 for convenience of description, and searching update time (corresponding to first update time) of the model catalogue and all the subdirectories in a full-scale storage statistical table (corresponding to the full-scale statistical table) under the current monitoring event, defining the update time as LIST LIST4 for convenience of description; if the number of LIST4 (corresponding to the first number) is greater than the number of LIST1 (corresponding to the second number), and LIST4 contains LIST1, then step 5 is skipped; if the number of LIST4 is less than the number of LIST1 and LIST4 is included in LIST1, then step 6 is skipped; if the number of LIST1 is equal to the number of LIST4 and the directories are equal, comparing the update time of the corresponding directories, if the update time is unchanged, no change exists in the model directory, reporting of events is not needed, and the step 9 is skipped to finish; if the update time is changed, the process jumps to step 7.

Step 5: the model is added with a model subdirectory under the model directory, and the model management micro-service added directory event and specific directory names (corresponding to the storage path of the added model parameters) are reported.

Step 6: the model subdirectory is deleted under the model directory, and the report model management micro-service delete directory event and specific directory name (corresponding to the storage path of the deleted model parameters).

Step 7: the model directory has file change, and the report model manages the micro-service update directory event and the specific directory name.

Step 8: model management micro-service makes model state change: updating the model state to be invalid when no event exists in the model catalog; a new catalog event, updating the model state to changed, and prompting specific change types and specific catalog 'new catalog' in a state column: the catalog A ", it can be understood that the new change occurs, and the new catalog is catalog A; a catalog deleting event updates the model state to changed, and prompts specific change types and specific catalog deleting catalogs in a state column: catalog B ", it is understood that the deleted catalog is catalog B; updating the catalog event, updating the model state to changed, and prompting specific change types and specific catalog update catalog in a state column: catalog C ", it will be appreciated that updated changes may occur, and that at least some of the parameters stored in catalog C may change in value.

Step 9: model state management ends.

According to the embodiment of the application, the model state is reported by the model file event monitor in the storage monitoring model, and the model state management is carried out in the model management module. The method reduces the shared storage resources and network overhead occupied by the transportation of a large amount of model data, and can accurately provide the situation of the model data for the user through the management of the model state, so that the user can clearly know the situation of the model data which is imported into the model management.

In an exemplary embodiment, after the storing the object model and the change description information having the correspondence relationship in the second storage space, the method further includes: detecting a target model state of the target model under the condition that a target deletion request is acquired, wherein the target deletion request is used for requesting to delete the second model parameters stored in the first storage space; in the case where the object model state is used to indicate that the object model has been deployed, suspending responding to the object deletion request.

Optionally, in this embodiment, in the case that the state of the target model is used to indicate that the target model is deployed, it may indicate that the target model has completed training and has been deployed to the application platform, and in this case, response to the target deletion request may be, but is not limited to, suspended, so that safety of model parameters of the target model is improved, and other personnel are prevented from deleting the model parameters of the model by mistake.

FIG. 5 is a schematic diagram of an alternative data protection according to an embodiment of the present application, and as shown in FIG. 5, the management method of model parameters in the embodiment of the present application may include, but is not limited to, parameter management, state management, version management and data protection.

Management of four model states including unpublished, published, imported as applications, deployed may be included, but is not limited to. Through interface communication with the model deployment micro-service, the model state is managed, and a data protection function is provided for the model being deployed and used, so that other users are prevented from deleting model data, and task failure is caused.

After model data (corresponding to model parameters) is imported into the model management, the model data is in an unpublished state by default, a user can perform tasks such as model test, and then the state of the model is updated to be published through a publishing function. The published model may be imported into an application store of the AI platform, at which point the model state is updated to be imported as an application. Model parameters and paths for large models may be stored in a json file format in a model first layer directory when published, but are not limited to. Since model management is a core business module of the AI platform, after the model is deployed as a service, a Pod (equivalent to a micro service) needs to be newly created. In order to prevent malicious network attack, the AI platform performs security reinforcement, and a network isolation policy is set, so that the port of the core service module cannot be accessed, and when the security is released, model parameters and paths of a large model are stored in a first layer directory of the model in a json file format, so that the security is obtained when the service is deployed.

After the model is deployed as a service, the model state is updated to deployed. For the deployed state model, if the model directory is deleted in file management, data protection measures need to be set to intercept so as to prevent errors of running services. When attempting to delete a file, the file management micro-service and the model deployment micro-service communicate through a Rest API to see if the model file is deployed, and if so, intercept the delete operation to protect the model data.

According to the embodiment of the application, the model state is managed, and the data protection function is provided for the model which is being deployed and used, so that the task failure caused by deleting the model data by other users is prevented.

FIG. 6 is a schematic diagram of the management of an alternative model state implemented in accordance with the application, as shown in FIG. 6, which may include, but is not limited to, unpublished, published, imported as an application, deployed, failed, and altered. Through interface communication with the model deployment micro-service, the model state is managed, and a data protection function is provided for the model being deployed and used, so that other users are prevented from deleting model data, and task failure is caused.

The embodiment of the application provides management and maintenance of a one-stop model state, which is convenient for guiding a user to perform the next operation, and the model catalog downloaded at a far end is selected, the model catalog is in an unpublished state after being imported into the model management, the model catalog is in a published state after being published, the model catalog is imported into an application store after the application is generated, the model catalog is changed into the application catalog, and the model catalog is changed into the deployed model catalog after service deployment. And prompting the disabled and changed states by monitoring the model data.

From the description of the above embodiments, it will be clear to a person skilled in the art that the method according to the above embodiments may be implemented by means of software plus the necessary general hardware platform, but of course also by means of hardware, but in many cases the former is a preferred embodiment. Based on such understanding, the technical solution of the present application may be embodied essentially or in a part contributing to the prior art in the form of a software product stored in a storage medium (e.g. ROM/RAM, magnetic disk, optical disk) comprising instructions for causing a terminal device (which may be a mobile phone, a computer, a server, or a network device, etc.) to perform the method according to the embodiments of the present application.

The embodiment also provides a device for managing model parameters, which is used for implementing the foregoing embodiments and preferred embodiments, and is not described in detail. As used below, the term "module" may be a combination of software and/or hardware that implements a predetermined function. While the means described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and contemplated.

Fig. 7 is a block diagram of a model parameter management apparatus according to an embodiment of the present application, as shown in fig. 7, on a model training system, a first storage space is deployed for storing current model parameters of a model and the current model parameters are used for training the model, and a second storage space is deployed in the model parameter management service and is used for storing initial model parameters of the model, and the apparatus is applied to the model parameter management service, and includes:

A first detection module 702, configured to detect, when it is detected that a first model parameter in the first storage space is subjected to a target data adjustment operation, a target model corresponding to the first model parameter, where the target data adjustment operation is a data adjustment operation that generates an update for a parameter in the first storage space;

A reading module 704, configured to read a second model parameter from the first storage space, and read a third model parameter corresponding to the target model from the second storage space, where the second model parameter is a model parameter after the target data adjustment operation is performed on the first model parameter;

A second detecting module 706, configured to detect, according to the second model parameter and the third model parameter, change description information corresponding to the third model parameter, where the change description information is used to indicate a target change that occurs in the third model parameter;

a storage module 708, configured to store the target model and the change description information that have a correspondence relationship in the second storage space.

By means of the device, the model parameter management service is deployed on the model training system, the model parameter used for training the model is stored in the model training system, the initial model parameter of the model is stored in the model parameter management service, it can be understood that the model parameter used for training the model and the initial model parameter of the model can be different, the model parameter management service can store the updated model parameter to be compared with the change of the initial model parameter under the condition that the model parameter stored in the model training system is updated, and in this way, the user can know the adjustment of the model parameter at any time under the condition that the model parameter is updated, so that the problem that the management efficiency of the model parameter is low can be solved, and the effect of improving the management efficiency of the model parameter is achieved.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the first detection module includes:

the first acquisition unit is used for acquiring a first detection request, wherein the first detection request is used for requesting to detect the target model corresponding to the first model parameter;

and the first control unit is used for responding to the first detection request and controlling the storage monitoring service to detect the target model corresponding to the first model parameter.

In an exemplary embodiment, the storage monitoring service has a detector disposed therein, and the first control unit is configured to:

acquiring a second detection request, wherein the second detection request is used for requesting to control the storage monitoring service to detect the target model corresponding to the first model parameter;

and responding to the second detection request, and controlling the detector to detect the target model corresponding to the first model parameter from model parameters and models with corresponding relations.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the reading module includes:

the second acquisition unit is used for acquiring a first reading request, wherein the first reading request is used for requesting to read the second model parameters corresponding to the target model from the first storage space;

And the second control unit is used for responding to the first reading request and controlling the storage monitoring service to read the second model parameters corresponding to the target model from the first storage space.

In an exemplary embodiment, the second control unit is configured to:

controlling the storage monitoring service to detect a first storage address of the second model parameter in the first storage space;

And controlling the storage monitoring service to read the data stored in the first storage address to obtain the second model parameters.

In one exemplary embodiment, the model parameter management service is deployed with a model management service, and the reading module includes:

A third obtaining unit, configured to obtain a second read request, where the second read request is used to request reading, from the second storage space, the third model parameter corresponding to the target model;

And the third control unit is used for responding to the second reading request and controlling the model management service to read the third model parameters corresponding to the target model from the second storage space.

In an exemplary embodiment, the third control unit is configured to:

Controlling the model management service to detect a second storage address of the third model parameter in the second storage space;

And controlling the model management service to read the data stored in the second storage address to obtain the third model parameters.

In one exemplary embodiment, the second detection module includes:

A first detection unit configured to detect first description information of the second model parameter and detect second description information of the third model parameter, where the first description information is configured to describe a first update time of the second model parameter and a first number of the second model parameter, and the second description information is configured to describe a second update time of the third model parameter and a second number of the third model parameter;

And the second detection unit is used for detecting the change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service, and the first detection unit is configured to:

acquiring a third detection request, wherein the third detection request is used for requesting to detect the first description information of the second model parameters;

And in response to the third detection request, controlling the storage monitoring service to detect the first update time of the second model parameter and controlling the storage monitoring service to count the first number of the second model parameters.

In an exemplary embodiment, a storage monitoring service is deployed in the model parameter management service, a total statistics table is deployed in the storage monitoring service, and the first detection unit is configured to:

And controlling the storage monitoring service to read the first updating time corresponding to the second model parameter from the full-quantity statistical table, and controlling the storage monitoring service to count the first quantity of parameters included in the second model parameter, wherein the full-quantity statistical table is used for storing one or more models, model parameters and updating time with corresponding relations, and the one or more models, model parameters and updating time with corresponding relations comprise the target model, the second model parameter and the first updating time with corresponding relations.

In an exemplary embodiment, the model parameter management service is deployed with a model management service, and the first detection unit is configured to:

acquiring a fourth detection request, wherein the fourth detection request is used for requesting to detect the second description information of the third model parameter;

And in response to the fourth detection request, controlling the model management service to detect the second update time of the third model parameter and controlling the model management service to count the second number of the third model parameter.

In an exemplary embodiment, the model parameter management service is deployed with a storage monitoring service and a model management service, and the second detecting unit is configured to:

Controlling the storage monitoring service to extract the second description information detected by the model management service;

And controlling the storage monitoring service to detect the change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

In an exemplary embodiment, the second detection unit is configured to:

controlling the storage monitoring service to compare the first updating time with the second updating time to obtain a first comparison result, and controlling the storage monitoring service to compare the first quantity with the second quantity to obtain a second comparison result;

and controlling the storage monitoring service to detect the change description information according to the first comparison result and the second comparison result.

In an exemplary embodiment, the second detection unit is configured to:

Controlling the storage monitoring service to detect a fourth model parameter except the third model parameter in the second model parameter and controlling the storage monitoring service to detect that the third model parameter has changed, wherein the target change comprises the new change and the fourth model parameter, when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is greater than the second number; or alternatively

Controlling the storage monitoring service to detect a fifth model parameter of the third model parameters except the second model parameter when the first comparison result is used for indicating that the second update time is earlier than the first update time and the second comparison result is used for indicating that the first number is smaller than the second number, and controlling the storage monitoring service to detect that the third model parameter has reduced variation, wherein the target variation comprises the reduced variation and the fifth model parameter; or alternatively

And under the condition that the first comparison result is used for indicating that the second updating time is earlier than the first updating time and the second comparison result is used for indicating that the first quantity is equal to the second quantity, controlling the storage monitoring service to detect a sixth model parameter which is different from the third model parameter in the second model parameter, controlling the storage monitoring service to detect that the third model parameter is subjected to adjustment change, wherein the target change comprises the adjustment change and the sixth model parameter.

In one exemplary embodiment, the model parameter management service is deployed with a model management service and a storage monitoring service, the model management service and the storage monitoring service are connected, and the storage module includes:

A fourth control unit for controlling the model management service to extract the change description information detected by the storage monitoring service;

and a fifth control unit configured to control the model management service to store the target model and the change description information having a correspondence relationship in the second storage space.

In an exemplary embodiment, the fifth control unit is configured to:

Controlling the model management service to store the target model, the new variation and the fourth model parameter having a correspondence in the second storage space, in case the variation description information is used to indicate that the target variation occurring in the third model parameter includes the new variation and the fourth model parameter; or alternatively

Controlling the model management service to store the target model, the reduced variation, and the fifth model parameter having a correspondence in the second storage space, in a case where the variation description information is used to indicate that the target variation occurring in the third model parameter includes the reduced variation and the fifth model parameter; or alternatively

And controlling the model management service to store the target model, the adjustment change and the sixth model parameter having a correspondence relationship in the second storage space, in the case where the change description information is used to indicate that the target change occurring in the third model parameter includes the adjustment change and the sixth model parameter.

In one exemplary embodiment, the apparatus further comprises:

A third detection module, configured to detect, after the storing the target model and the change description information having the correspondence in the second storage space, a target model state of the target model in a case where a target deletion request is acquired, where the target deletion request is used to request deletion of the second model parameter stored in the first storage space;

And the pause response module is used for pausing responding to the target deletion request under the condition that the target model state is used for representing that the target model is deployed.

In one exemplary embodiment, the apparatus further comprises:

The module is used for deploying a storage monitoring service in the model parameter management service, the storage monitoring service is deployed with a detector, and a fifth detection request is acquired before the target model corresponding to the first model parameter is detected, wherein the fifth detection request is used for requesting to detect whether the current model parameter of the model stored in the first storage space is subjected to the target data adjustment operation;

And a control module for controlling the detector to detect whether the current model parameters of the model stored in the first storage space are subjected to the target data adjustment operation in response to the fifth detection request.

It should be noted that each of the above modules may be implemented by software or hardware, and for the latter, it may be implemented by, but not limited to: the modules are all located in the same processor; or the above modules may be located in different processors in any combination.

Embodiments of the present application also provide a computer readable storage medium having a computer program stored therein, wherein the computer program is arranged to perform the steps of any of the method embodiments described above when run.

In one exemplary embodiment, the computer readable storage medium may include, but is not limited to: a usb disk, a Read-Only Memory (ROM), a random access Memory (Random Access Memory RAM), a removable hard disk, a magnetic disk, or an optical disk, or other various media capable of storing a computer program.

An embodiment of the application also provides an electronic device comprising a memory having stored therein a computer program and a processor arranged to run the computer program to perform the steps of any of the method embodiments described above.

In an exemplary embodiment, the electronic device may further include a transmission device connected to the processor, and an input/output device connected to the processor.

Embodiments of the application also provide a computer program product comprising a computer program which, when executed by a processor, implements the steps of any of the method embodiments described above.

Embodiments of the present application also provide another computer program product comprising a non-volatile computer readable storage medium storing a computer program which, when executed by a processor, implements the steps of any of the method embodiments described above.

Embodiments of the present application also provide a computer program comprising computer instructions stored in a computer-readable storage medium; the processor of the computer device reads the computer instructions from the computer readable storage medium and executes the computer instructions to cause the computer device to perform the steps of any of the method embodiments described above.

Specific examples in this embodiment may refer to the examples described in the foregoing embodiments and the exemplary implementation, and this embodiment is not described herein.

It will be appreciated by those skilled in the art that the modules or steps of the application described above may be implemented in a general purpose computing device, they may be concentrated on a single computing device, or distributed across a network of computing devices, they may be implemented in program code executable by computing devices, so that they may be stored in a storage device for execution by computing devices, and in some cases, the steps shown or described may be performed in a different order than that shown or described herein, or they may be separately fabricated into individual integrated circuit modules, or multiple modules or steps of them may be fabricated into a single integrated circuit module. Thus, the present application is not limited to any specific combination of hardware and software.

The above description is only of the preferred embodiments of the present application and is not intended to limit the present application, but various modifications and variations can be made to the present application by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for managing model parameters is characterized in that,

A first storage space and a model parameter management service are deployed on a model training system, the first storage space is used for storing current model parameters of a model, the current model parameters are used for training the model, a second storage space is deployed in the model parameter management service, the second storage space is used for storing initial model parameters of the model, and the method is applied to the model parameter management service and comprises the following steps:

Detecting a target model corresponding to a first model parameter in the first storage space under the condition that the first model parameter in the first storage space is detected to be subjected to target data adjustment operation, wherein the target data adjustment operation is a data adjustment operation for updating the parameter in the first storage space;

Reading a second model parameter from the first storage space, and reading a third model parameter corresponding to the target model from the second storage space, wherein the second model parameter is a model parameter after the target data adjustment operation is executed on the first model parameter;

detecting change description information corresponding to the third model parameter according to the second model parameter and the third model parameter, wherein the change description information is used for indicating target change of the third model parameter;

and storing the target model and the change description information with the corresponding relation in the second storage space.

2. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The model parameter management service is deployed with a storage monitoring service, and the detecting the target model corresponding to the first model parameter includes:

Acquiring a first detection request, wherein the first detection request is used for requesting to detect the target model corresponding to the first model parameter;

and responding to the first detection request, and controlling the storage monitoring service to detect the target model corresponding to the first model parameter.

3. The method of claim 2, wherein the step of determining the position of the substrate comprises,

The storage monitoring service is provided with a detector, and the controlling the storage monitoring service to detect the target model corresponding to the first model parameter includes:

4. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The model parameter management service is deployed with a storage monitoring service, and the reading of the second model parameter from the first storage space includes:

acquiring a first reading request, wherein the first reading request is used for requesting to read the second model parameters corresponding to the target model from the first storage space;

And responding to the first reading request, and controlling the storage monitoring service to read the second model parameters corresponding to the target model from the first storage space.

5. The method of claim 4, wherein the step of determining the position of the first electrode is performed,

The controlling the storage monitoring service to read the second model parameters corresponding to the target model from the first storage space includes:

6. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The model parameter management service is deployed with a model management service, and the reading of the third model parameter corresponding to the target model from the second storage space includes:

Acquiring a second read request, wherein the second read request is used for requesting to read the third model parameter corresponding to the target model from the second storage space;

And responding to the second reading request, and controlling the model management service to read the third model parameters corresponding to the target model from the second storage space.

7. The method of claim 6, wherein the step of providing the first layer comprises,

The controlling the model management service to read the third model parameter corresponding to the target model from the second storage space includes:

8. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The detecting, according to the second model parameter and the third model parameter, the change description information corresponding to the third model parameter includes:

Detecting first description information of the second model parameters and second description information of the third model parameters, wherein the first description information is used for describing first update time of the second model parameters and first quantity of the second model parameters, and the second description information is used for describing second update time of the third model parameters and second quantity of the third model parameters;

And detecting change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information.

9. The method of claim 8, wherein the step of determining the position of the first electrode is performed,

The model parameter management service is deployed with a storage monitoring service, and the detecting the first description information of the second model parameter includes:

10. The method of claim 8, wherein the step of determining the position of the first electrode is performed,

The model parameter management service is provided with a storage monitoring service, the storage monitoring service is provided with a total statistics table, and the detecting of the first description information of the second model parameter comprises the following steps:

11. The method of claim 8, wherein the step of determining the position of the first electrode is performed,

The model parameter management service is deployed with a model management service, and the detecting the second description information of the third model parameter includes:

12. The method of claim 8, wherein the step of determining the position of the first electrode is performed,

The model parameter management service is configured with a storage monitoring service and a model management service, and the detecting the change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information, the second update time and the second quantity indicated by the second description information includes:

13. The method of claim 12, wherein the step of determining the position of the probe is performed,

The controlling the storage monitoring service to detect the change description information corresponding to the third model parameter according to the first update time and the first quantity indicated by the first description information and the second update time and the second quantity indicated by the second description information includes:

14. The method of claim 13, wherein the step of determining the position of the probe is performed,

The controlling the storage monitoring service to detect the change description information according to the first comparison result and the second comparison result includes:

15. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The model parameter management service is provided with a model management service and a storage monitoring service, the model management service is connected with the storage monitoring service, and the storing of the target model with the corresponding relation and the change description information in the second storage space comprises the following steps:

Controlling the model management service to extract the change description information detected by the storage monitoring service;

And controlling the model management service to store the target model and the change description information with the corresponding relationship in the second storage space.

16. The method of claim 15, wherein the step of determining the position of the probe is performed,

The controlling the model management service to store the target model and the change description information having a correspondence relationship in the second storage space includes:

17. The method of claim 1, wherein the step of determining the position of the substrate comprises,

After the storing of the target model and the change description information having the correspondence relationship in the second storage space, the method further includes:

detecting a target model state of the target model under the condition that a target deletion request is acquired, wherein the target deletion request is used for requesting to delete the second model parameters stored in the first storage space;

in the case where the object model state is used to indicate that the object model has been deployed, suspending responding to the object deletion request.

18. The method of claim 1, wherein the step of determining the position of the substrate comprises,

The model parameter management service is provided with a storage monitoring service, the storage monitoring service is provided with a detector, and before the target model corresponding to the first model parameter is detected, the method further comprises:

acquiring a fifth detection request, wherein the fifth detection request is used for requesting to detect whether the current model parameters of the model stored in the first storage space are subjected to the target data adjustment operation;

And controlling the detector to detect whether the target data adjustment operation is performed on the current model parameters of the model stored in the first storage space in response to the fifth detection request.

19. A management device for model parameters is characterized in that,

A model training system is deployed with a first storage space and a model parameter management service, the first storage space is used for storing current model parameters of a model, the current model parameters are used for training the model, the model parameter management service is deployed with a second storage space, the second storage space is used for storing initial model parameters of the model, and the device is applied to the model parameter management service and comprises:

20. A computer-readable storage medium comprising,

The computer readable storage medium has stored therein a computer program, wherein the computer program when executed by a processor implements the steps of the method of any of claims 1 to 18.

21. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that,

The processor, when executing the computer program, implements the steps of the method as claimed in any one of claims 1 to 18.

22. A computer program product comprising a computer program, characterized in that,

Which computer program, when being executed by a processor, carries out the steps of the method as claimed in any one of claims 1 to 18.