Artificial intelligence model training system and training method

Abstract

The invention provides an artificial intelligence model training system and method, and the system employs a server room parameter, a model computing power parameter, a training large model type, and a corresponding environment parameter regulation and control method to construct an artificial intelligence model training environment control model. The constructed artificial intelligence model training environment control model generates a real-time environment parameter regulation and control method according to real-time server room parameters, real-time model computing power parameters and real-time training large model types, real-time environment parameter regulation and control are carried out on the server room according to the real-time environment parameter regulation and control method, the large model training speed is increased, and the real-time environment parameter regulation and control method is applied to the server room. The consumption generated by model training is reduced; according to the method, the environment parameters are regulated and controlled in real time by utilizing the characteristics, such as the server room parameters, the computing power of the to-be-trained model and the type of the large model, which influence the training efficiency of the large model when the large model is trained, and the constructed training model environment control model realizes real-time control when the large model is trained, so that the model training is helped to save the cost and improve the efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence training and control technology, and in particular relates to an artificial intelligence model training system and training method. Background Art

[0002] With the rapid development of the economy, a large number of large model training server centers have emerged rapidly. The electronic equipment in the large model training server center generates a lot of heat during operation, and corresponding environmental control equipment needs to be configured to cool it down. The environment in the large model training server center directly affects the reliability of the operation of the electronic equipment in the cabinet and the efficiency of large model training. During the operation of the electronic equipment of the large model training server, there will be situations where temperature control needs to be changed at any time. If the control is not good, it is very likely to cause the electronic components to burn out. The model training conditions of each server group in the large model training server center are different, and the power consumed by the calculation is inconsistent, which leads to extremely poor setting of constant environmental parameters, which in turn increases the burden on the electronic equipment in the computer room, and will cause energy waste and increase the operating cost of the computer room.

[0003] Most of the existing systems conduct electronic monitoring of the environment in the computer room, but do not monitor the server operation from the root, that is, to perform real-time regulation and processing at the level of artificial intelligence model training, which can easily lead to environmental out-of-control on the server, thereby affecting the normal use of electronic components.

[0004] Therefore, this field needs an artificial intelligence model training environment control model that can fully consider the relevant source data such as the large model training parameters in the model pre-training process. How does this artificial intelligence model training environment control model generate a real-time environmental parameter control method, so that the server itself can perform real-time environmental parameter control on the server room according to the real-time environmental parameter control method, thereby increasing the speed of training large models and reducing the consumption generated by model training. This is a problem that this field needs to face. Therefore, there is a need for a system and method that can realize real-time control of environmental parameters during large model training, help model training save costs, and improve efficiency.

[0005] Based on the above problems, the present invention designs an artificial intelligence model training system and training method, uses server room parameters, model computing power parameters, training large model types and corresponding environmental parameter control methods to construct an artificial intelligence model training environment control model, and generates a real-time environmental parameter control method according to the real-time server room parameters, real-time model computing power parameters, and real-time training large model types for the constructed artificial intelligence model training environment control model. The server room is controlled by the real-time environmental parameter control method to improve the speed of training large models and reduce the consumption generated by model training. The present invention uses the characteristics that affect the efficiency of large model training when training large models, such as server room parameters, model computing power to be trained, and large model types, to control environmental parameters in real time. In addition, the training data dimension and the model category adopted by the server training large model are used as input vectors, and the improved simple machine learning model is used to construct the training environment control model, which realizes the optimization of real-time control of the environment during large model training, and helps model training save costs and improve efficiency. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes an artificial intelligence model training system and training method.

[0007] A first aspect of the present invention provides an artificial intelligence model training method, characterized in that the method comprises:

[0008] Obtain the parameters of the first computer room where the server is located, the first model computing power parameters used for large model training, the first model type used for training the large model, and the corresponding first environmental parameter control method at this time;

[0009] According to the first computer room parameter, the first model computing power parameter, and the first model type feature, a first environmental control feature is obtained;

[0010] Using the first environmental control feature and the first environmental parameter control method corresponding to the first environmental parameter, an artificial intelligence model is constructed to train an environmental control model;

[0011] Receive the second computer room parameters, second model computing power parameters and second model type of the server room, perform feature processing according to the second computer room parameters, the second model computing power parameters and the second model type to obtain a second environmental control feature, use the artificial intelligence model to train the environmental control model to process the second environmental control feature to generate a second environmental parameter control method, and according to the second environmental parameter control method, the server control device intelligently controls the environment.

[0012] Furthermore, the first computer room parameter or the second computer room parameter includes a server operating temperature parameter and a server operating humidity parameter.

[0013] Furthermore, the first model type or the second model type is divided into three types: unsupervised machine learning method, supervised machine learning method and semi-supervised machine learning method. The feature vector value of the unsupervised machine learning method corresponding to the first model type or the second model type is 0, the feature vector value of the semi-supervised machine learning method corresponding to the first model type or the second model type is 0.5, and the feature vector value of the supervised machine learning method corresponding to the first model type or the second model type is 1.

[0014] Furthermore, according to the first computer room parameter, the first model computing power parameter, and the first model type characteristic processing, a first environmental control characteristic is obtained; or according to the second computer room parameter, the second model computing power parameter, and the second model type characteristic processing, a second environmental control characteristic is obtained.

[0015] Furthermore, the artificial intelligence model training environmental control model is constructed by a machine learning model or a deep learning model.

[0016] Furthermore, the artificial intelligence model trains the environmental control model using a classifier based on the Fischer criterion.

[0017] An artificial intelligence model training system is also provided, the system comprising a computer room data acquisition module, a model computing power parameter acquisition module, a model type determination module, a server control data storage module, an artificial intelligence model training environment control model construction module, and an environmental parameter control module, characterized in that:

[0018] The computer room data acquisition module is used to acquire the parameters of the first computer room or the second computer room where the server is located;

[0019] The model computing power parameter acquisition module is used to acquire the first model computing power parameter or the second model computing power parameter used by the server for large model training;

[0020] The model type determination module is used to obtain the first model type or the second model type used for training the large model. The first model type or the second model type is divided into three types: unsupervised machine learning method, supervised machine learning method and semi-supervised machine learning method. The unsupervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 0, the semi-supervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 0.5, and the supervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 1;

[0021] The server control data storage module is used to receive the first computer room parameter, the first model computing power parameter, and the first model type feature processing to obtain the first environmental control feature, and is also used to receive the second computer room parameter, the second model computing power parameter, and the second model type, perform feature processing according to the second computer room parameter, the second model computing power parameter, and the second model type to obtain the second environmental control feature, and obtain the first environmental parameter control method;

[0022] The artificial intelligence model training environment control model construction module is used to obtain the first environment control feature and the corresponding first environment parameter control method in the server control data storage module, and use the first environment control feature and the corresponding first environment parameter control method to construct an artificial intelligence model training environment control model;

[0023] The environmental parameter control module is connected to the artificial intelligence model training environmental control model construction module, receives the second environmental control feature in the server control data storage module, and calls the artificial intelligence model training environmental control model construction module to obtain the second environmental parameter control method. According to the second environmental parameter control method, the server control room equipment intelligently controls the environment.

[0024] Furthermore, the artificial intelligence model training environment control model is constructed by a machine learning model or a deep learning model;

[0025] The artificial intelligence model training environmental control model adopts a classifier based on the Fischer criterion.

[0026] The present invention uses server room parameters, model computing power parameters, training large model types and corresponding environmental parameter control methods to construct an artificial intelligence model training environment control model. The constructed artificial intelligence model training environment control model generates a real-time environment parameter control method according to the real-time server room parameters, real-time model computing power parameters, and real-time training large model types. The server room is controlled by the real-time environment parameter control method to improve the speed of training large models and reduce the consumption generated by model training. The present invention uses the characteristics that affect the training efficiency of large models when training large models, such as server room parameters, model computing power to be trained, and large model types to control environmental parameters in real time. In addition, the training data dimension and the model category adopted by the server training large model are used as input vectors, and an improved simple machine learning model is used to construct a training environment control model, which realizes the optimization of real-time control of the environment during large model training, helps model training save costs and improve efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of an artificial intelligence model training method of the present invention;

[0028] Figure 2 It is a structural schematic diagram of an artificial intelligence model training system of the present invention;

[0029] Figure 3 It is a classification principle diagram of the classifier based on Fisher's criterion in the present invention;

[0030] Figure 4 This is a schematic diagram of technical parameters of the server used for large model training in an embodiment of the present invention;

[0031] Figure 5 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention provides an artificial intelligence model training system and training method. The invention is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0033] The first embodiment of the present invention:

[0034] A first aspect of the present invention provides an artificial intelligence model training method, characterized in that the method comprises:

[0035] Obtain the parameters of the first computer room where the server is located, the first model computing power parameters used for large model training, the first model type used for training the large model, and the corresponding first environmental parameter control method at this time;

[0036] According to the first computer room parameter, the first model computing power parameter, and the first model type feature, a first environmental control feature is obtained;

[0037] Using the first environmental control feature and the first environmental parameter control method corresponding to the first environmental parameter, an artificial intelligence model is constructed to train an environmental control model;

[0038] A second computer room parameter, a second model computing power parameter and a second model type of the server computer room are received, and feature processing is performed according to the second computer room parameter, the second model computing power parameter and the second model type to obtain a second environmental control feature. The artificial intelligence model training environmental control model processes the second environmental control feature to generate a second environmental parameter control method. According to the second environmental parameter control method, the server controls the computer room equipment to intelligently control the environment.

[0039] Furthermore, the first computer room parameter or the second computer room parameter includes a server operating temperature parameter and a server operating humidity parameter.

[0040] In this embodiment, reference Figure 4As shown in the figure, the technical parameters of a general server will give parameter values ​​such as operating temperature and relative humidity, but this is only a range value. If you want the model training to be more energy-efficient, you need to dynamically adjust environmental parameters such as temperature and humidity in real time.

[0041] Therefore, in this embodiment, the first computer room parameter or the second computer room parameter may be a related vector value of temperature and humidity.

[0042] Furthermore, the first model computing power parameter or the second model computing power parameter is obtained according to the following formula:

[0043]

[0044] In the formula, P represents the computing power parameter of the first model or the computing power parameter of the second model, D represents the dimension of the training data, |T| represents the scientific notation of the amount of data used for the training model rounded up, and k represents the parameter adjustment coefficient.

[0045] In this embodiment, if the amount of training data is 0.2×10 3 The amount of training set data, then |T| takes the value of 10 3 , this is rounded up.

[0046] In this embodiment, when a large model is used for training, a large amount of data is generally retrieved, and these data have corresponding dimensions and quantities to train a large model with relatively accurate prediction. The parameter adjustment coefficient in this embodiment generally takes a value of 0.15 to 0.55, which is the optimal model computing power parameter correction coefficient obtained by experiments in the present invention. The corrected model computing power parameter is more suitable for the input vector of the model training in this scenario.

[0047] Furthermore, the first model type or the second model type is divided into three types: unsupervised machine learning method, supervised machine learning method and semi-supervised machine learning method. The feature vector value of the unsupervised machine learning method corresponding to the first model type or the second model type is 0, the feature vector value of the semi-supervised machine learning method corresponding to the first model type or the second model type is 0.5, and the feature vector value of the supervised machine learning method corresponding to the first model type or the second model type is 1.

[0048] In this embodiment, unsupervised and semi-supervised model classification is adopted. This is because different models consume different time for training and have different model generalization capabilities. According to the theoretical basis of pattern recognition, the present invention is only divided into the above three types.

[0049] Further, according to the first computer room parameter, the first model computing power parameter, and the first model type feature processing, a first environmental control feature is obtained; or according to the second computer room parameter, the second model computing power parameter, and the second model type feature processing, a second environmental control feature is obtained, and the following formula is used for feature processing to obtain the environmental control feature G:

[0050]

[0051] Wherein, N1, N2, N3 are weight coefficients corresponding to the computer room parameter, model computing power parameter and model type respectively, N1+N2+N3=1, |M| is the first computer room parameter or the second computer room parameter, P is the first model computing power parameter or the second model computing power parameter, and T is the first model type or the second model type.

[0052] In this embodiment, the vector eigenvalue of |M| is 25*20%=5, where 25 represents a temperature of 25° C. 20% represents a relative humidity of 20% in the computer room.

[0053] Furthermore, the artificial intelligence model training environmental control model is constructed by a machine learning model or a deep learning model.

[0054] Furthermore, the artificial intelligence model training environment control model adopts a classifier based on the Fisher criterion, and the calculation formula is as follows:

[0055]

[0056] G is the first environmental control feature or the second environmental control feature, F(G) is the output first environmental parameter control method or the second environmental parameter control method, W T is the normal vector perpendicular to the hyperplane, which is obtained by training the first environmental control feature and the first environmental parameter control method, D is the dimension of the training data, N1, N2, and N3 are the weight coefficients corresponding to the computer room parameters, model computing power parameters, and model types, respectively.

[0057] An artificial intelligence model training system is also provided, the system comprising a computer room data acquisition module, a model computing power parameter acquisition module, a model type determination module, a server control data storage module, an artificial intelligence model training environment control model construction module, and an environmental parameter control module, characterized in that:

[0058] The computer room data acquisition module is used to acquire the parameters of the first computer room or the second computer room where the server is located;

[0059] The model computing power parameter acquisition module is used to acquire the first model computing power parameter or the second model computing power parameter used by the server for large model training;

[0060] The model type determination module is used to obtain the first model type or the second model type used for training the large model. The first model type or the second model type is divided into three types: unsupervised machine learning method, supervised machine learning method and semi-supervised machine learning method. The unsupervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 0, the semi-supervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 0.5, and the supervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 1;

[0061] The server control data storage module is used to receive the first computer room parameter, the first model computing power parameter, and the first model type feature processing to obtain the first environmental control feature, and is also used to receive the second computer room parameter, the second model computing power parameter, and the second model type, perform feature processing according to the second computer room parameter, the second model computing power parameter, and the second model type to obtain the second environmental control feature, and obtain the first environmental parameter control method;

[0062] The artificial intelligence model training environment control model construction module is used to obtain the first environment control feature and the corresponding first environment parameter control method in the server control data storage module, and use the first environment control feature and the corresponding first environment parameter control method to construct an artificial intelligence model training environment control model;

[0063] The environmental parameter control module is connected to the artificial intelligence model training environmental control model construction module, receives the second environmental control feature in the server control data storage module, and calls the artificial intelligence model training environmental control model construction module to obtain the second environmental parameter control method. According to the second environmental parameter control method, the server controls the computer room equipment to intelligently control the environment.

[0064] Furthermore, the first model computing power parameter or the second model computing power parameter is obtained according to the following formula:

[0065]

[0066] In the formula, P represents the computing power parameter of the first model or the computing power parameter of the second model, D represents the dimension of the training data, |T| represents the scientific notation of the amount of data used for the training model rounded up, and k represents the parameter adjustment coefficient.

[0067] Furthermore, the artificial intelligence model training environment control model is constructed by a machine learning model or a deep learning model;

[0068] The artificial intelligence model training environment control model adopts a classifier based on the Fischer criterion, and the calculation formula is as follows:

[0069]

[0070] G is the first environmental control feature or the second environmental control feature, F(G) is the output first environmental parameter control method or the second environmental parameter control method, W T is the normal vector perpendicular to the hyperplane, which is obtained by training the first environmental control feature and the first environmental parameter control method, D is the dimension of the training data, N1, N2, and N3 are the weight coefficients corresponding to the computer room parameters, model computing power parameters, and model types, respectively.

[0071] In this embodiment, the dimension of the training data is generally 1-3 dimensions, and multi-dimensional data training is not excluded. At this time, N1, N2, and N3 are 0.6, 0.3, and 0.1 respectively. Generally, large model training will adopt a semi-supervised training method or a supervised training method. In this embodiment, a semi-supervised training method is adopted.

[0072] The environmental control features input in this embodiment are:

[0073]

[0074] The P value varies with the amount and dimension of the training data. In the present invention, the accuracy of subsequent regulation model training can be improved based on this value.

[0075] In this embodiment, F(G) obtained is the output of the first environmental parameter control method or the second environmental parameter control method. The control method is determined by the value of F(G). When the value of F(G) is greater than 1, the low-level environmental parameter control method is adopted. When the value of F(G) is less than 1, the high-level environmental parameter control method is adopted. When the value of F(G) is equal to 1, the intermediate environmental parameter control method is adopted. The environmental parameter control methods at each level respectively control the parameter adjustment of the computer room control equipment, including the parameters of the computer room control equipment such as the air conditioning temperature and the humidifier humidity.

[0076] The present invention uses server room parameters, model computing power parameters, training large model types and corresponding environmental parameter control methods to construct an artificial intelligence model training environment control model. The constructed artificial intelligence model training environment control model generates a real-time environment parameter control method according to the real-time server room parameters, real-time model computing power parameters, and real-time training large model types. The server room is controlled by the real-time environment parameter control method to improve the speed of training large models and reduce the consumption generated by model training. The present invention uses the characteristics that affect the training efficiency of large models when training large models, such as server room parameters, model computing power to be trained, and large model types to control environmental parameters in real time. In addition, the training data dimension and the model category adopted by the server training large model are used as input vectors, and an improved simple machine learning model is used to construct a training environment control model, which realizes the optimization of real-time control of the environment during large model training, helps model training save costs and improve efficiency.

[0077] The present invention does not particularly specify the structure of some modules, which shall be subject to the contents recorded in the prior art. The prior art mentioned in the above background technology section and the specific embodiment section of the present invention can be used as part of the present invention to understand the meaning of some technical features or parameters. The scope of protection of the present invention shall be subject to the contents actually recorded in the claims.

Claims

1. An artificial intelligence model training method, characterized in that: The method comprises: Obtain the parameters of the first computer room where the server is located, the first model computing power parameters used for large model training, the first model type used for training the large model, and the corresponding first environmental parameter control method at this time; According to the first computer room parameter, the first model computing power parameter, and the first model type feature, a first environmental control feature is obtained; Using the first environmental control feature and the first environmental parameter control method corresponding to the first environmental parameter, an artificial intelligence model is constructed to train an environmental control model; A second computer room parameter, a second model computing power parameter and a second model type of the server computer room are received, and feature processing is performed according to the second computer room parameter, the second model computing power parameter and the second model type to obtain a second environmental control feature. The artificial intelligence model training environmental control model processes the second environmental control feature to generate a second environmental parameter control method. According to the second environmental parameter control method, the server controls the computer room equipment to intelligently control the environment.

2. An artificial intelligence model training method as claimed in claim 1, characterized in that: The first computer room parameter or the second computer room parameter includes a server operating temperature parameter and a server operating humidity parameter.

3. An artificial intelligence model training method as claimed in claim 2, characterized in that: The first model type or the second model type is divided into three types: unsupervised machine learning method, supervised machine learning method and semi-supervised machine learning method. The feature vector value of the unsupervised machine learning method corresponding to the first model type or the second model type is 0.

4. An artificial intelligence model training method as claimed in claim 3, characterized in that: The feature vector value of the first model type or the second model type corresponding to the semi-supervised machine learning method is 0.5, and the feature vector value of the first model type or the second model type corresponding to the supervised machine learning method is 1.

5. The artificial intelligence model training method according to claim 1, characterized in that: The first environment control feature is obtained by processing the first computer room parameter, the first model computing power parameter, and the first model type feature, or the second environment control feature is obtained by processing the second computer room parameter, the second model computing power parameter, and the second model type feature.

6. An artificial intelligence model training method as claimed in claim 5, characterized in that: The artificial intelligence model training environmental control model is constructed by a machine learning model or a deep learning model.

7. An artificial intelligence model training method as claimed in claim 6, characterized in that: The artificial intelligence model training environment control model adopts a classifier based on the Fisher criterion.

8. An artificial intelligence model training system, the system includes a computer room data acquisition module, a model computing power parameter acquisition module, a model type determination module, a server control data storage module, an artificial intelligence model training environment control model construction module, and an environmental parameter control module, characterized in that: The computer room data acquisition module is used to acquire the parameters of the first computer room or the second computer room where the server is located; The model computing power parameter acquisition module is used to acquire the first model computing power parameter or the second model computing power parameter used by the server for large model training; The model type determination module is used to obtain the first model type or the second model type used for training the large model. The first model type or the second model type is divided into three types: unsupervised machine learning method, supervised machine learning method and semi-supervised machine learning method. The unsupervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 0, the semi-supervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 0.5, and the supervised machine learning method corresponds to the first model type or the second model type with a feature vector value of 1; The server control data storage module is used to receive the first computer room parameter, the first model computing power parameter, and the first model type feature processing to obtain the first environmental control feature, and is also used to receive the second computer room parameter, the second model computing power parameter, and the second model type, perform feature processing according to the second computer room parameter, the second model computing power parameter, and the second model type to obtain the second environmental control feature, and obtain the first environmental parameter control method; The artificial intelligence model training environment control model construction module is used to obtain the first environment control feature and the corresponding first environment parameter control method in the server control data storage module, and use the first environment control feature and the corresponding first environment parameter control method to construct an artificial intelligence model training environment control model; The environmental parameter control module is connected to the artificial intelligence model training environmental control model construction module, receives the second environmental control feature in the server control data storage module, and calls the artificial intelligence model training environmental control model construction module to obtain the second environmental parameter control method. According to the second environmental parameter control method, the server controls the computer room equipment to intelligently control the environment.

9. An artificial intelligence model training system as claimed in claim 8, characterized in that: The artificial intelligence model training environmental control model is constructed by a machine learning model or a deep learning model.

Images