WO2024239927A1 - Model training method and related device - Google Patents

Abstract

The present application discloses a model training method and a related device, capable of reducing the total time costs required for a model training process. The method of the present application comprises: in a t-th iteration of a model to be trained, once a gradient of an N-th layer of said model in the t-th iteration is obtained, the gradient of the N-th layer in the t-th iteration can be immediately normalized on the basis of a moving average of global gradient norm (MGGN) in a (t-1)-th iteration, so that a normalized gradient of the N-th layer in the t-th iteration is obtained; and by analogy, once a gradient of a first layer in the t-th iteration is obtained, the gradient of the first layer in the t-th iteration can be immediately normalized on the basis of the MGGN in the (t-1)-th iteration, so that a normalized gradient of the first layer in the t-th iteration is obtained. Thus, parameters of the first layer to the N-th layer can be updated on the basis of the normalized gradients of the first layer to the N-th layer in the t-th iteration, so that the t-th iteration of said model is completed.

Description

A model training method and related equipment

This application claims the priority of the Chinese patent application filed with the State Intellectual Property Office on May 4, 2023, with application number 202310492136.8 and invention name “A model training method and related equipment”, the entire contents of which are incorporated by reference in this application.

Technical Field

The embodiments of the present application relate to the field of artificial intelligence (AI) technology, and in particular to a model training method and related equipment.

Background Art

The bidirectional encoder representations from transformers (BERT) model has excellent performance in machine reading comprehension and other tasks due to its powerful language representation ability. However, during the training process of the BERT model, the convergence speed of the loss is often slow, resulting in a high training cost for the BERT model.

In order to speed up the convergence of loss during model training, related technologies propose a method based on global gradient normalization to complete the training of the BERT model. Specifically, after obtaining the loss based on the training data input to the model to be trained, the back propagation for the model to be trained can be completed based on the loss, thereby obtaining the gradients of each layer in the model to be trained. Next, the global gradient norm can be calculated based on the gradients of each layer, and then the gradients of each layer can be normalized using the global gradient norm. Then, the normalized gradients of each layer are used to update the parameters of each layer. Continuously repeating the above process can accelerate the convergence of the loss and complete the entire training process of the model to be trained, thereby obtaining the BERT model.

However, the calculation and use of the global gradient norm can only be performed after the backpropagation of the model is completed, which undoubtedly increases the tail time of the model training process and causes the time cost of the model training process to be too high.

Summary of the invention

The embodiments of the present application provide a model training method and related equipment, which can effectively shorten the tail time in each iteration of the model, thereby reducing the total time cost required for the entire training process of the model.

A first aspect of an embodiment of the present application provides a model training method, the method comprising:

Assume that the entire training process of the model to be trained includes K iterations, K ≥ 2. After completing the first iteration of the model to be trained, enter the second iteration of the model to be trained to the Kth iteration of the model to be trained.

In the second iteration to the Kth iteration of the model to be trained, since the process of each iteration is similar, any one of the iterations is schematically introduced below, and the iteration is called the tth iteration, K≥t≥2. In the t-th iteration of the model to be trained, based on the training data input to the model to be trained, after obtaining the gradient of the N-th layer of the model to be trained in the t-th iteration, the gradient of the N-th layer in the t-th iteration can be immediately normalized based on the moving average of the global gradient norm (MGGN) in the t-1-th iteration, so as to obtain the normalized gradient of the N-th layer in the t-th iteration, and then, after obtaining the gradient of the N-1-th layer of the model to be trained in the t-th iteration, the gradient of the N-1-th layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-1-th layer in the t-th iteration, ..., after obtaining the gradient of the 1st layer of the model to be trained in the t-th iteration, the gradient of the 1st layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the 1st layer in the t-th iteration. Then, the parameters of the 1st to Nth layers can be updated based on the normalized gradients of the 1st to Nth layers in the tth iteration. At this point, the tth iteration of the model to be trained is completed, so the t+1th iteration of the model to be trained can be entered.

By repeating the above process continuously, the second iteration to the Kth iteration of the model to be trained can be gradually completed, and the target model can be obtained.

From the above method, it can be seen that: since the MGGN in the t-1th iteration can be obtained by the gradient of the 1st layer to the Nth layer in the t-1th iteration, that is, the MGGN in the t-1th iteration has been obtained in advance in the t-1th iteration of the model to be trained, after entering the tth iteration of the model to be trained, once the gradient of any layer in the model to be trained in the tth iteration is obtained, the MGGN in the t-1th iteration can be immediately used to normalize the gradient of the layer in the tth iteration, thereby obtaining the normalized value of the layer in the t-1th iteration. In this way, the back propagation and gradient normalization in the t-th iteration of the model to be trained can be performed simultaneously, which can effectively shorten the tail time of the t-th iteration of the model to be trained, thereby reducing the total time cost required for the entire training process of the model to be trained.

In a possible implementation, the method further includes: obtaining the MGGN in the t-th iteration based on the gradient of the N layer in the t-th iteration. In the above implementation, since the gradients of the 1st layer to the N-th layer in the t-th iteration have been obtained, the gradients of the 1st layer to the N-th layer in the t-th iteration can be used to obtain the MGGN in the t-th iteration.

In a possible implementation, based on the gradient of the N layer in the t-th iteration, obtaining the MGGN in the t-th iteration includes: performing a first calculation on the gradient of the N layer in the t-th iteration to obtain the global gradient norm (global gradient norm, GGN) in the t-th iteration; performing a second calculation on the MGGN in the t-1-th iteration and the GGN in the t-th iteration to obtain the MGGN in the t-th iteration. In the above implementation, since the gradients of the 1st layer to the Nth layer in the t-th iteration have been obtained, the gradients of the 1st layer to the Nth layer in the t-th iteration can be calculated to obtain the GGN in the t-th iteration, and the MGGN in the t-1-th iteration and the GGN in the t-th iteration can be calculated to obtain the MGGN in the t-th iteration.

In a possible implementation, the first calculation includes at least one of the following: exponentiation, summation, and square root calculation. In the above implementation, the square of the gradient of the first layer in the t-th iteration, ..., and the square of the gradient of the N-th layer in the t-th iteration can be obtained first, and then the square of the gradient of the first layer in the t-th iteration, ..., and the square of the gradient of the N-th layer in the t-th iteration are added, and finally the square root of the sum of these squares is taken to obtain the GGN in the t-th iteration.

In a possible implementation, the second calculation is a weighted sum calculation. In the above implementation, the GGN in the t-th iteration and the MGGN in the t-1-th iteration may be weighted summed using preset weights to obtain the MGGN in the t-th iteration.

In one possible implementation, based on the training data input to the model to be trained, obtaining the gradient of the i-th layer of the model to be trained in the t-th iteration includes: based on the M training data input to the model to be trained, obtaining M to-be-processed gradients of the i-th layer of the model to be trained in the t-th iteration, M≥2; averaging the M to-be-processed gradients to obtain the gradient of the i-th layer in the t-th iteration. In the aforementioned implementation, in the t-th iteration of the model to be trained, there are M training data input to the model to be trained. Then, based on the M data, M candidate gradients of the N-th layer in the t-th iteration can be obtained accordingly, and the M candidate gradients of the N-th layer in the t-th iteration can be directly averaged to obtain the gradient of the N-th layer in the t-th iteration, and the gradient of the N-th layer in the t-th iteration is immediately normalized based on the MGGN in the t-1-th iteration to obtain the normalized gradient of the N-th layer in the t-th iteration. Then, based on the M data, the M candidate gradients of the N-1th layer in the t iteration can be obtained accordingly, and the M candidate gradients of the N-1th layer in the t iteration are directly averaged to obtain the gradient of the N-1th layer in the t iteration, and the gradient of the N-1th layer in the t iteration is immediately normalized based on the MGGN in the t-1 iteration to obtain the normalized gradient of the N-1th layer in the t iteration, ..., finally, based on the M data, the M candidate gradients of the 1st layer in the t iteration can be obtained accordingly, and the M candidate gradients of the 1st layer in the t iteration are directly averaged to obtain the gradient of the 1st layer in the t iteration, and the gradient of the 1st layer in the t iteration is immediately normalized based on the MGGN in the t-1 iteration to obtain the normalized gradient of the 1st layer in the t iteration.

In one possible implementation, the method further includes: in the first iteration of the model to be trained, based on the training data input to the model to be trained, obtaining the gradient of the N layer in the first iteration; based on the gradient of the N layer in the first iteration, obtaining the GGN in the first iteration, and using the GGN in the first iteration as the MGGN in the first iteration; based on the GGN in the first iteration, normalizing the gradient of the N layer in the first iteration to obtain the normalized gradient of the N layer in the first iteration; based on the normalized gradient of the N layer in the first iteration, updating the parameters of the N layer to enter the second iteration of the model to be trained. In the aforementioned implementation, in the first iteration of the model to be trained, the training data can be input to the model to be trained, then, based on the training data, the gradients of the first layer to the N layer of the model to be trained in the first iteration can be gradually obtained. Then, since the gradients of the first layer to the N layer in the first iteration have been obtained, the gradients of the first layer to the N layer in the first iteration can be calculated to obtain the GGN in the first iteration, and the GGN in the first iteration can be directly used as the MGGN in the first iteration. Subsequently, the gradients of the first layer to the N layer in the first iteration can be normalized based on the GGN in the first iteration to obtain the normalized gradients of the first layer to the N layer in the first iteration. Finally, the parameters of the first layer to the N layer can be updated based on the normalized gradients of the first layer to the N layer in the first iteration. At this point, the first iteration of the model to be trained is completed, so the second iteration of the model to be trained can be entered.

A second aspect of an embodiment of the present application provides a model training device, which includes: a first acquisition module, used to obtain, in the t-th iteration of the model to be trained, the gradient of the i-th layer of the model to be trained in the t-th iteration based on the training data input to the model to be trained, t≥2, i＝1,...,N, N≥1; a normalization module, used to normalize the gradient of the i-th layer in the t-th iteration based on the MGGN in the t-1-th iteration The gradient of the i-th layer is normalized to obtain the normalized gradient of the i-th layer in the t-th iteration, and the MGGN in the t-1-th iteration is obtained based on the gradient of the N-th layer of the model to be trained in the t-1-th iteration; an updating module is used to update the parameters of the i-th layer based on the normalized gradient of the i-th layer in the t-th iteration to enter the t+1-th iteration of the model to be trained.

It can be seen from the above device that: since the MGGN in the t-1th iteration can be obtained by the gradient of the 1st layer to the Nth layer in the t-1th iteration, that is, the MGGN in the t-1th iteration has been obtained in advance in the t-1th iteration of the model to be trained, after entering the tth iteration of the model to be trained, once the gradient of any layer in the model to be trained in the tth iteration is obtained, the MGGN in the t-1th iteration can be immediately used to normalize the gradient of the layer in the tth iteration, so as to obtain the normalized gradient of the layer in the t-1th iteration, so that the back propagation and gradient normalization in the tth iteration of the model to be trained can be carried out simultaneously, which can effectively shorten the tail time of the tth iteration of the model to be trained, thereby reducing the total time cost required for the entire training process of the model to be trained.

In a possible implementation, the device further includes: a second acquisition module, configured to acquire the MGGN in the t-th iteration based on the gradient of the N layer in the t-th iteration.

In one possible implementation, the second acquisition module is used to: perform a first calculation on the gradient of the N layer in the t-th iteration to obtain the global gradient norm GGN in the t-th iteration; perform a second calculation on the MGGN in the t-1-th iteration and the GGN in the t-th iteration to obtain the MGGN in the t-th iteration.

In a possible implementation manner, the first calculation includes at least one of the following: a power calculation, a sum calculation, and a square root calculation.

In a possible implementation manner, the second calculation is a weighted sum calculation.

In one possible implementation, the first acquisition module is used to: obtain M unprocessed gradients of the i-th layer of the model to be trained in the t-th iteration based on M training data input into the model to be trained, where M≥2; and average the M unprocessed gradients to obtain the gradient of the i-th layer in the t-th iteration.

In one possible implementation, the first acquisition module is also used to obtain the gradient of the N layer in the first iteration based on the training data input into the model to be trained in the first iteration of the model to be trained; the second acquisition module is used to obtain the GGN in the first iteration based on the gradient of the N layer in the first iteration, and the GGN in the first iteration is used as the MGGN in the first iteration; the normalization module is also used to normalize the gradient of the N layer in the first iteration based on the GGN in the first iteration to obtain the normalized gradient of the N layer in the first iteration; the update module is also used to update the parameters of the N layer based on the normalized gradient of the N layer in the first iteration to enter the second iteration of the model to be trained.

A third aspect of an embodiment of the present application provides a model training device, which includes a memory and a processor; the memory stores code, and the processor is configured to execute the code. When the code is executed, the model training device executes the method described in the first aspect or any possible implementation method of the first aspect.

A fourth aspect of an embodiment of the present application provides a circuit system, which includes a processing circuit, and the processing circuit is configured to execute the method described in the first aspect or any possible implementation manner of the first aspect.

A fifth aspect of an embodiment of the present application provides a chip system, which includes a processor for calling a computer program or computer instructions stored in a memory so that the processor executes the method described in the first aspect or any possible implementation method of the first aspect.

In a possible implementation manner, the processor is coupled to the memory through an interface.

In a possible implementation, the chip system also includes a memory, in which a computer program or computer instructions are stored.

A sixth aspect of an embodiment of the present application provides a computer storage medium, which stores a computer program. When the program is executed by a computer, the computer implements the method described in the first aspect or any possible implementation method of the first aspect.

A seventh aspect of the embodiments of the present application provides a computer program product, which stores instructions. When the instructions are executed by a computer, the computer implements the method described in the first aspect or any possible implementation method of the first aspect.

In an embodiment of the present application, in the t-th iteration of the model to be trained, after obtaining the gradient of the N-th layer of the model to be trained in the t-th iteration, the gradient of the N-th layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-th layer in the t-th iteration. Similarly, after obtaining the gradient of the N-1-th layer in the t-th iteration, the gradient of the N-1-th layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-1-th layer in the t-th iteration, ..., after obtaining the gradient of the 1st layer in the t-th iteration, the gradient of the 1st layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the 1st layer in the t-th iteration. In this way, the parameters of the 1st to Nth layers can be updated based on the normalized gradients of the 1st to Nth layers in the t-th iteration. At this point, the t-th iteration of the model to be trained is completed, so the t+1-th iteration of the model to be trained can be entered. In the above process, the MGGN in the t-1-th iteration can be obtained from the gradient of the 1st layer to the Nth layer in the t-1-th iteration, that is, the MGGN in the t-1-th iteration has been obtained in advance in the t-1-th iteration of the model to be trained. Therefore, after entering the t-th iteration of the model to be trained, once the gradient of any layer in the model to be trained in the t-th iteration is obtained, the MGGN in the t-1-th iteration can be immediately used to normalize the gradient of the layer in the t-th iteration, so as to obtain the normalized gradient of the layer in the t-1-th iteration. In this way, the back propagation and gradient normalization in the t-th iteration of the model to be trained can be carried out simultaneously, which can effectively shorten the tail time of the t-th iteration of the model to be trained, thereby reducing the total time cost required for the entire training process of the model to be trained.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a structure of an artificial intelligence main framework;

FIG2a is a schematic diagram of a structure of an information processing system provided in an embodiment of the present application;

FIG2b is another schematic diagram of the structure of the information processing system provided in an embodiment of the present application;

FIG2c is a schematic diagram of a device related to information processing provided in an embodiment of the present application;

FIG3 is a schematic diagram of the architecture of the system 100 provided in an embodiment of the present application;

FIG4 is a flow chart of a model training method provided in an embodiment of the present application;

FIG5 is a schematic diagram of a comparison result provided in an embodiment of the present application;

FIG6 is another schematic diagram of the comparison results provided in the embodiment of the present application;

FIG7 is a schematic diagram of an application example of the model training method provided in an embodiment of the present application;

FIG8 is another schematic diagram of a comparative structure provided in an embodiment of the present application;

FIG9 is a flow chart of an information processing method provided in an embodiment of the present application;

FIG10 is a schematic diagram of a structure of a model training device provided in an embodiment of the present application;

FIG11 is a schematic diagram of a structure of an information processing device provided in an embodiment of the present application;

FIG12 is a schematic diagram of a structure of an execution device provided in an embodiment of the present application;

FIG13 is a schematic diagram of a structure of a training device provided in an embodiment of the present application;

FIG. 14 is a schematic diagram of the structure of a chip provided in an embodiment of the present application.

DETAILED DESCRIPTION

The embodiments of the present application provide a model training method and related equipment, which can effectively shorten the tail time in each iteration of the model, thereby reducing the total time cost required for the entire training process of the model.

The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, which is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

The BERT model has excellent performance in machine reading comprehension and other tasks due to its powerful language representation capabilities. However, during the training process of the BERT model, the convergence speed of the loss is often slow, resulting in a high training cost for the BERT model.

In order to speed up the convergence speed of loss during model training, the relevant technology proposes a method based on global gradient normalization to complete the training of the BERT model. Specifically, the training process of the BERT model includes multiple iterations. In a certain iteration of the model to be trained, after the loss is obtained based on the training data input to the model to be trained, the back propagation for the model to be trained can be completed based on the loss, thereby obtaining the gradients of each layer in the model to be trained. Then, the global gradient norm can be calculated based on the gradients of each layer, and the gradients of each layer can be normalized using the global gradient norm. Then, the normalized gradients of each layer are used to update the parameters of each layer. At this point, the iteration of the model to be trained is completed, and the next iteration of the model to be trained can be entered. In the process of multiple iterations, the convergence speed of the loss can be accelerated until the loss converges, and the entire training process of the model to be trained can be completed, thereby obtaining the BERT model.

However, the calculation of the global gradient norm and gradient normalization can only be performed after the back propagation of the model is completed. That is to say, in each iteration of the model to be trained, the global gradient norm can only be calculated after the gradients of all layers in the model to be trained are obtained. This norm is used to normalize the gradients of all layers, which undoubtedly increases the tail time of each iteration during the model training process, resulting in an excessively high total time cost for the model training process.

In order to solve the above problems, the embodiment of the present application provides a model training method, which can be implemented in combination with artificial intelligence (AI) technology. AI technology is a technical discipline that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence. AI technology obtains the best results by sensing the environment, acquiring knowledge and using knowledge. In other words, artificial intelligence technology is a branch of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Using artificial intelligence for data processing is a common application of artificial intelligence.

First, the overall workflow of the artificial intelligence system is described. Please refer to Figure 1. Figure 1 is a structural diagram of the main framework of artificial intelligence. The following is an explanation of the above artificial intelligence theme framework from the two dimensions of "intelligent information chain" (horizontal axis) and "IT value chain" (vertical axis). Among them, the "intelligent information chain" reflects a series of processes from data acquisition to processing. For example, it can be a general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, intelligent execution and output. In this process, the data has undergone a condensation process of "data-information-knowledge-wisdom". The "IT value chain" reflects the value that artificial intelligence brings to the information technology industry from the underlying infrastructure of human intelligence, information (providing and processing technology implementation) to the industrial ecology process of the system.

(1) Infrastructure

The infrastructure provides computing power support for the AI system, enables communication with the outside world, and supports it through the basic platform. It communicates with the outside world through sensors; computing power is provided by smart chips (CPU, NPU, GPU, ASIC, FPGA and other hardware acceleration chips); the basic platform includes distributed computing frameworks and networks and other related platform guarantees and support, which can include cloud storage and computing, interconnected networks, etc. For example, sensors communicate with the outside world to obtain data, and these data are provided to the smart chips in the distributed computing system provided by the basic platform for calculation.

(2) Data

The data on the upper layer of the infrastructure is used to represent the data sources in the field of artificial intelligence. The data involves graphics, images, voice, text, and IoT data of traditional devices, including business data of existing systems and perception data such as force, displacement, liquid level, temperature, and humidity.

(3) Data processing

Data processing usually includes data training, machine learning, deep learning, search, reasoning, decision-making and other methods.

Among them, machine learning and deep learning can symbolize and formalize data for intelligent information modeling, extraction, preprocessing, and training.

Reasoning refers to the process of simulating human intelligent reasoning in computers or intelligent systems, using formalized information to perform machine thinking and solve problems based on reasoning control strategies. Typical functions are search and matching.

Decision-making refers to the process of making decisions after intelligent information is reasoned, usually providing functions such as classification, sorting, and prediction.

(4) General capabilities

After the data has undergone the data processing mentioned above, some general capabilities can be further formed based on the results of the data processing, such as an algorithm or a general system, for example, translation, text analysis, computer vision processing, speech recognition, image recognition, etc.

(5) Smart products and industry applications

Smart products and industry applications refer to the products and applications of artificial intelligence systems in various fields. They are the encapsulation of the overall artificial intelligence solution, which productizes intelligent information decision-making and realizes practical applications. Its application areas mainly include: smart terminals, smart transportation, smart medical care, autonomous driving, smart cities, etc.

Next, several application scenarios of this application are introduced.

FIG2a is a schematic diagram of a structure of an information processing system provided in an embodiment of the present application, wherein the information processing system includes a user device and a data processing device. The user device includes an intelligent terminal such as a mobile phone, a personal computer or an information processing center. The user device is the initiator of information processing, and as the initiator of the information processing request, the request is usually initiated by the user through the user device.

The above-mentioned data processing device can be a device or server with data processing function such as a cloud server, a network server, an application server and a management server. The data processing device receives information processing requests from the intelligent terminal through an interactive interface, and then performs information processing in the form of machine learning, deep learning, search, reasoning, decision-making, etc. through the memory storing the data and the processor link of the data processing. The memory in the data processing device can be a general term, including local storage and database storing historical data. The database can be in the data processing It can be located on the management device or on other network servers.

In the information processing system shown in FIG2a, the user device can receive the user's instruction, for example, the user device can obtain an information input/selected by the user, and then initiate a request to the data processing device, so that the data processing device performs information processing on the information obtained by the user device, thereby obtaining a processing result for the information. Exemplarily, the user device can obtain the information input by the user (for example, a question to be answered, a prompt to be subjected to sentiment analysis, an incomplete text to be supplemented, and other types of language information to be processed), and then initiate an information processing request to the data processing device, so that the data processing device performs a series of processing on the information based on the information processing request, thereby obtaining a processing result of the information (for example, an answer to a question, a sentiment belonging to a prompt, a supplementary part of an incomplete text, etc.).

In FIG. 2 a , the data processing device may execute the information processing method of the embodiment of the present application.

Figure 2b is another structural diagram of the information processing system provided in an embodiment of the present application. In Figure 2b, the user device directly serves as a data processing device. The user device can directly obtain input from the user and directly process it by the hardware of the user device itself. The specific process is similar to that of Figure 2a. Please refer to the above description and will not be repeated here.

In the information processing system shown in Figure 2b, the user device can receive instructions from the user. For example, the user device can obtain information input by the user (for example, a question to be answered, a prompt to be subjected to sentiment analysis, an incomplete text to be supplemented, and other types of language information to be processed), and then perform a series of processing on the information to obtain the processing result of the information (for example, an answer to a question, the sentiment belonging to a prompt, a supplementary part of an incomplete text, etc.).

In FIG. 2b , the user equipment itself can execute the information processing method of the embodiment of the present application.

FIG. 2c is a schematic diagram of an information processing related device provided in an embodiment of the present application.

The user device in the above Figures 2a and 2b can specifically be the local device 301 or the local device 302 in Figure 2c, and the data processing device in Figure 2a can specifically be the execution device 210 in Figure 2c, wherein the data storage system 250 can store the data to be processed of the execution device 210, and the data storage system 250 can be integrated on the execution device 210, and can also be set on the cloud or other network servers.

The processors in Figures 2a and 2b can perform data training/machine learning/deep learning through a neural network model or other models (for example, a model based on a support vector machine), and use the model finally trained or learned from the data to execute information processing applications on the image, thereby obtaining corresponding processing results.

Figure 3 is a schematic diagram of the system 100 architecture provided in an embodiment of the present application. In Figure 3, the execution device 110 is configured with an input/output (I/O) interface 112 for data interaction with an external device. The user can input data to the I/O interface 112 through the client device 140. The input data may include: various tasks to be scheduled, callable resources and other parameters in the embodiment of the present application.

When the execution device 110 preprocesses the input data, or when the computing module 111 of the execution device 110 performs calculation and other related processing (such as implementing the function of the neural network in the present application), the execution device 110 can call the data, code, etc. in the data storage system 150 for the corresponding processing, and can also store the data, instructions, etc. obtained by the corresponding processing in the data storage system 150.

Finally, the I/O interface 112 returns the processing result to the client device 140 so as to provide it to the user.

It is worth noting that the training device 120 can generate corresponding target models/rules based on different training data for different goals or tasks, and the corresponding target models/rules can be used to achieve the above goals or complete the above tasks, thereby providing the user with the desired results. The training data can be stored in the database 130 and come from the training samples collected by the data collection device 160.

In the case shown in FIG. 3 , the user can manually give input data, and the manual giving can be operated through the interface provided by the I/O interface 112. In another case, the client device 140 can automatically send input data to the I/O interface 112. If the client device 140 is required to automatically send input data and needs to obtain the user's authorization, the user can set the corresponding authority in the client device 140. The user can view the results output by the execution device 110 on the client device 140, and the specific presentation form can be a specific method such as display, sound, action, etc. The client device 140 can also be used as a data acquisition terminal to collect the input data of the input I/O interface 112 and the output results of the output I/O interface 112 as shown in the figure as new sample data, and store them in the database 130. Of course, it is also possible not to collect through the client device 140, but the I/O interface 112 directly stores the input data of the input I/O interface 112 and the output results of the output I/O interface 112 as new sample data in the database 130.

It is worth noting that FIG. 3 is only a schematic diagram of a system architecture provided by an embodiment of the present application, and the positional relationship between the devices, components, modules, etc. shown in the figure does not constitute any limitation. For example, in FIG. 3, the data storage system 150 is an external memory relative to the execution device 110. In other cases, the data storage system 150 can also be placed in the execution device 110. As shown in FIG. 3, the training device 110 can be used to perform the training of the training device 110. 120 training to obtain the neural network.

The embodiment of the present application also provides a chip, which includes a neural network processor NPU. The chip can be set in the execution device 110 as shown in Figure 3 to complete the calculation work of the calculation module 111. The chip can also be set in the training device 120 as shown in Figure 3 to complete the training work of the training device 120 and output the target model/rule.

Neural network processor NPU, NPU is mounted on the main central processing unit (CPU) (host CPU) as a coprocessor, and the main CPU assigns tasks. The core part of NPU is the operation circuit, and the controller controls the operation circuit to extract data from the memory (weight memory or input memory) and perform operations.

In some implementations, the arithmetic circuit includes multiple processing units (process engines, PEs) internally. In some implementations, the arithmetic circuit is a two-dimensional systolic array. The arithmetic circuit can also be a one-dimensional systolic array or other electronic circuits capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit is a general-purpose matrix processor.

For example, suppose there is an input matrix A, a weight matrix B, and an output matrix C. The operation circuit takes the corresponding data of matrix B from the weight memory and caches it on each PE in the operation circuit. The operation circuit takes the matrix A data from the input memory and performs matrix operations with matrix B. The partial results or final results of the matrix are stored in the accumulator.

The vector calculation unit can further process the output of the operation circuit, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc. For example, the vector calculation unit can be used for network calculations of non-convolutional/non-FC layers in neural networks, such as pooling, batch normalization, local response normalization, etc.

In some implementations, the vector computation unit can store the processed output vector to a unified buffer. For example, the vector computation unit can apply a nonlinear function to the output of the computation circuit, such as a vector of accumulated values, to generate an activation value. In some implementations, the vector computation unit generates a normalized value, a merged value, or both. In some implementations, the processed output vector can be used as an activation input to the computation circuit, such as for use in a subsequent layer in a neural network.

The unified memory is used to store input data and output data.

The weight data is directly transferred from the external memory to the input memory and/or the unified memory through the direct memory access controller (DMAC), the weight data in the external memory is stored in the weight memory, and the data in the unified memory is stored in the external memory.

The bus interface unit (BIU) is used to enable interaction between the main CPU, DMAC and instruction fetch memory through the bus.

An instruction fetch buffer connected to the controller, used to store instructions used by the controller;

The controller is used to call the instructions cached in the memory to control the working process of the computing accelerator.

Generally, the unified memory, input memory, weight memory and instruction fetch memory are all on-chip memories, and the external memory is a memory outside the NPU, which can be a double data rate synchronous dynamic random access memory (DDR SDRAM), a high bandwidth memory (HBM) or other readable and writable memory.

Since the embodiments of the present application involve the application of a large number of neural networks, in order to facilitate understanding, the relevant terms and related concepts such as neural networks involved in the embodiments of the present application are first introduced below.

(1) Neural Network

A neural network may be composed of neural units, and a neural unit may refer to an operation unit with xs and intercept 1 as input, and the output of the operation unit may be:

Where s=1, 2, ...n, n is a natural number greater than 1, Ws is the weight of xs, and b is the bias of the neural unit. f is the activation function of the neural unit, which is used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into the output signal. The output signal of the activation function can be used as the input of the next convolutional layer. The activation function can be a sigmoid function. A neural network is a network formed by connecting many of the above-mentioned single neural units together, that is, the output of one neural unit can be the input of another neural unit. The input of each neural unit can be connected to the local receptive field of the previous layer to extract the characteristics of the local receptive field. The local receptive field can be an area composed of several neural units.

The work of each layer in the neural network can be described by the mathematical expression y=a(Wx+b): From a physical level, the work of each layer in the neural network can be understood as completing the transformation from the input space to the output space (i.e., the row space to the column space of the matrix) through five operations on the input space (the set of input vectors). These five operations include: 1. Dimension increase/reduction; 2. Zoom in/out; 3. Rotation; 4. Translation; 5. "Bending". Among them, operations 1, 2, and 3 are completed by Wx, operation 4 is completed by +b, and operation 5 is implemented by a(). The word "space" is used here because the classified object is not a single thing, but a class of things, and space refers to the collection of all individuals of this class of things. Among them, W is a weight vector, and each value in the vector represents the weight value of a neuron in the neural network of this layer. The vector W determines the spatial transformation from the input space to the output space described above, that is, the weight W of each layer controls how to transform the space. The purpose of training a neural network is to finally obtain the weight matrix of all layers of the trained neural network (the weight matrix formed by many layers of vectors W). Therefore, the training process of a neural network is essentially about learning how to control spatial transformations, or more specifically, learning the weight matrix.

Because we want the output of the neural network to be as close as possible to the value we really want to predict, we can compare the current network's predicted value with the target value we really want, and then update the weight vector of each layer of the neural network based on the difference between the two (of course, there is usually an initialization process before the first update, that is, pre-configuring parameters for each layer in the neural network). For example, if the network's predicted value is high, adjust the weight vector to make it predict a lower value, and keep adjusting until the neural network can predict the target value we really want. Therefore, it is necessary to predefine "how to compare the difference between the predicted value and the target value", which is the loss function or objective function, which are important equations used to measure the difference between the predicted value and the target value. Among them, taking the loss function as an example, the higher the output value (loss) of the loss function, the greater the difference, so the training of the neural network becomes a process of minimizing this loss as much as possible.

(2) Back propagation algorithm

Neural networks can use the error back propagation (BP) algorithm to correct the size of the parameters in the initial neural network model during the training process, so that the reconstruction error loss of the neural network model becomes smaller and smaller. Specifically, the forward transmission of the input signal to the output will generate error loss, and the error loss information is back-propagated to update the parameters in the initial neural network model, so that the error loss converges. The back propagation algorithm is a back propagation movement dominated by error loss, which aims to obtain the optimal parameters of the neural network model, such as the weight matrix.

(3) Global gradient normalization

In the training process of the neural network model, after completing the back propagation of the neural network model, the gradients of each layer in the neural network model can be obtained. In order to prevent the gradient explosion during the model training process and speed up the convergence of the loss in the process, the gradients of each layer can be normalized. Specifically, after obtaining the gradients of each layer, a global gradient norm can be calculated using the gradients of each layer, and then the gradients of each layer can be divided by the global gradient norm to obtain the normalized gradients of each layer, so as to update the parameters of each layer in the neural network model, and then complete the entire training process of the model.

The method provided in the present application is described below from the training side of the neural network and the application side of the neural network.

The model training method provided in the embodiment of the present application involves the processing of data sequences, and can be specifically applied to data training, machine learning, deep learning and other methods to perform symbolic and formalized intelligent information modeling, extraction, preprocessing, training, etc. on the training data (for example, the training data in the model training method provided in the embodiment of the present application), and finally obtain a trained neural network (for example, the target model obtained after training the training model based on the model training method provided in the embodiment of the present application); and, the information processing method provided in the embodiment of the present application can use the above-mentioned trained neural network to input the input data (for example, the information in the information processing method provided in the embodiment of the present application) into the trained neural network to obtain output data (for example, the processing result of the information in the information processing method provided in the embodiment of the present application). It should be noted that the model training method and information processing method provided in the embodiment of the present application are inventions based on the same concept, and can also be understood as two parts in a system, or two stages of an overall process: such as the model training stage and Model application phase.

FIG4 is a flow chart of a model training method provided in an embodiment of the present application. As shown in FIG4 , the method includes:

401. After completing the first iteration of the model to be trained, enter the second iteration of the model to be trained to the Kth iteration of the model to be trained, where K≥2;

In this embodiment, it is assumed that the training process of the model to be trained (for example, the BERT model to be trained) includes K iterations (K is a positive integer greater than or equal to 2), and the K iterations of the model to be trained can be divided into two parts, one part is the first iteration of the model to be trained, and the other part is the second iteration to the Kth iteration of the model to be trained. It should be noted that since the training process of the model to be trained includes K iterations, before training the model to be trained, a data set can be obtained first, and the data set includes K batches of training data (for example, these training data can be language information in various tasks, such as questions in question-answering tasks, prompts in sentiment analysis tasks, incomplete texts in text supplementation tasks, etc.), and among these K batches of training data, a batch of training data can be used to complete one iteration of the model to be trained.

Specifically, the first iteration of the model to be trained can be completed in the following ways:

(1) In the first iteration of the model to be trained, the first batch of training data can be input into the model to be trained so that the first batch of training data can be processed by the model to be trained to obtain the processing results of the first batch of training data, and the loss in the first iteration can be obtained based on the processing results of the first batch of training data. Then, based on the loss in the first iteration, the gradients of the first layer to the Nth layer of the model to be trained in the first iteration can be gradually obtained (N is a positive integer greater than or equal to 1). Then, since the gradients of the first layer to the Nth layer in the first iteration have been obtained, the gradients of the first layer to the Nth layer in the first iteration can be calculated to obtain the GGN in the first iteration, and the GGN in the first iteration can be directly used as the MGGN in the first iteration. Subsequently, based on the GGN in the first iteration, the gradients of the first layer to the Nth layer in the first iteration can be normalized to obtain the normalized gradients of the first layer to the Nth layer in the first iteration. Finally, the parameters of the 1st to Nth layers can be updated based on the normalized gradients of the 1st to Nth layers in the 1st iteration. At this point, the 1st iteration of the model to be trained is completed, so the 2nd iteration of the model to be trained can be entered.

More specifically, the gradients of the first layer to the Nth layer in the first iteration can be calculated by a preset formula (for example, the formula can implement at least one of the following calculations: exponentiation calculation, sum calculation, and square root calculation), so as to obtain the GGN in the first iteration and the MGGN in the first iteration. For example, the formula can be presented as:

In the above formula, GGN ₁ is the GGN in the first iteration, grad _1,1 is the gradient of the first layer in the first iteration, and grad _1,N is the gradient of the Nth layer in the first iteration.

(2) In the first iteration of the model to be trained, the first batch of training data can be input into the model to be trained so that the first batch of training data can be processed by the model to be trained, thereby obtaining the processing result of the first batch of training data, and based on the processing result of the first batch of training data, the loss in the first iteration is obtained. Next, based on the loss in the first iteration, the gradient of the Nth layer of the model to be trained in the first iteration is obtained, and the gradient of the Nth layer in the first iteration is immediately normalized based on the preset MGGN, so as to obtain the normalized gradient of the Nth layer in the first iteration. Next, based on the loss in the first iteration, the gradient of the N-1th layer of the model to be trained in the first iteration can be obtained, and the gradient of the N-1th layer in the first iteration is immediately normalized based on the preset MGGN, so as to obtain the normalized gradient of the N-1th layer in the first iteration, ..., finally, based on the loss in the first iteration, the gradient of the first layer of the model to be trained in the first iteration can be obtained, and the gradient of the first layer in the first iteration is immediately normalized based on the preset MGGN, so as to obtain the normalized gradient of the first layer in the first iteration. Then, since the gradients of the 1st layer to the Nth layer in the 1st iteration have been obtained, the gradients of the 1st layer to the Nth layer in the 1st iteration can be calculated to obtain the GGN in the 1st iteration, and the preset MGGN and the GGN in the 1st iteration are calculated to obtain the MGGN in the 1st iteration. Finally, the parameters of the 1st layer to the Nth layer can be updated based on the normalized gradients of the 1st layer to the Nth layer in the 1st iteration. At this point, the 1st iteration of the model to be trained is completed, so the 2nd iteration of the model to be trained can be entered.

More specifically, a preset formula (for example, the formula can implement at least one of the following calculations: exponentiation, summation, opening, The gradient of the first layer to the Nth layer in the first iteration is calculated by square root calculation and weighted sum calculation, so as to obtain the GGN in the first iteration and the MGGN in the first iteration. For example, the formula can be presented as:

In the above formula, MGGN _p is the preset MGGN, and β is the preset weight.

Furthermore, the model to be trained can be deployed on M graphics processing units (GPUs). Then, in the first iteration of the model to be trained, the first batch of training data input to the model to be trained may include M training data (also referred to as M training data, M is a positive integer greater than or equal to 2), and these M training data can be input into the model to be trained on the M GPUs accordingly. After the models to be trained on the M GPUs process the M training data respectively, the processing results of the M training data can be obtained accordingly. Then, based on the processing results of the M training data, the M losses in the first iteration can be obtained accordingly. At this time, there are multiple situations:

Based on the above situation (1), we can first obtain M candidate gradients of the Nth layer in the first iteration based on the M losses, and directly average the M candidate gradients of the Nth layer in the first iteration to obtain the gradient of the Nth layer in the first iteration. Then, we can obtain M candidate gradients of the N-1th layer in the first iteration based on the M losses, and directly average the M candidate gradients of the N-1th layer in the first iteration to obtain the gradient of the N-1th layer in the first iteration, ..., and finally, we can obtain M candidate gradients of the 1st layer in the first iteration based on the M losses, and directly average the M candidate gradients of the 1st layer in the first iteration to obtain the gradient of the 1st layer in the first iteration. Since the gradients of the 1st to Nth layers in the first iteration have been obtained, the gradients of the 1st to Nth layers in the first iteration can be calculated to obtain the GGN in the first iteration. Subsequently, the gradients from the 1st layer to the Nth layer in the 1st iteration may be normalized based on the GGN in the 1st iteration, thereby obtaining the normalized gradients from the 1st layer to the Nth layer in the 1st iteration.

Based on the above situation (2), M candidate gradients of the Nth layer in the first iteration can be obtained based on the M losses, and the M candidate gradients of the Nth layer in the first iteration are directly averaged to obtain the gradient of the Nth layer in the first iteration, and the gradient of the Nth layer in the first iteration is immediately normalized based on the preset MGGN to obtain the normalized gradient of the Nth layer in the first iteration. Then, based on the M losses, M candidate gradients of the N-1th layer in the first iteration can be obtained accordingly, and the M candidate gradients of the N-1th layer in the first iteration are directly averaged to obtain the gradient of the N-1th layer in the first iteration, and the gradient of the N-1th layer in the first iteration is immediately normalized based on the preset MGGN to obtain the normalized gradient of the N-1th layer in the first iteration, ..., finally, based on the M losses, M candidate gradients of the 1st layer in the first iteration can be obtained accordingly, and the M candidate gradients of the 1st layer in the first iteration are directly averaged to obtain the gradient of the 1st layer in the first iteration, and the gradient of the 1st layer in the first iteration is immediately normalized based on the preset MGGN to obtain the normalized gradient of the 1st layer in the first iteration.

402. In the tth iteration of the model to be trained, based on the training data input to the model to be trained, obtain the gradient of the i-th layer of the model to be trained in the tth iteration, K≥t≥2, i＝1,...,N, N≥1.

403. Based on the MGGN in the t-1th iteration, the gradient of the i-th layer in the t-th iteration is normalized to obtain the normalized gradient of the i-th layer in the t-1th iteration, and the MGGN in the t-1th iteration is obtained based on the gradient of the N layer of the model to be trained in the t-1th iteration.

404. Based on the normalized gradient of the i-th layer in the t-th iteration, update the parameters of the i-th layer to enter the t+1-th iteration of the model to be trained.

After completing the first iteration of the model to be trained, the second iteration of the model to be trained can be entered, and after completing the second iteration of the model to be trained, the third iteration of the model to be trained can be entered, and so on, until the Kth iteration of the model to be trained is completed. At this time, the model training conditions are met (assuming that the loss reaches convergence in the Kth iteration), and the target model (for example, the trained BERT model) can be obtained.

Specifically, the second iteration to the Kth iteration of the model to be trained can be completed in the following manner:

In the second iteration to the Kth iteration of the model to be trained, since the process of each iteration is similar, any one of the iterations is schematically introduced below, and the iteration is called the tth iteration (K≥t≥2).

In the tth iteration of the model to be trained, the tth batch of training data can be input into the model to be trained, so that the tth batch of training data can be processed by the model to be trained, thereby obtaining the processing result of the tth batch of training data, and based on the processing result of the tth batch of training data, the loss in the tth iteration is obtained. Next, based on the loss in the t-th iteration, the gradient of the N-th layer of the model to be trained in the t-th iteration can be obtained, and the gradient of the N-th layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-th layer in the t-th iteration. Next, based on the loss in the t-th iteration, the gradient of the N-1-th layer of the model to be trained in the t-th iteration can be obtained, and the gradient of the N-1-th layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration (the MGGN in the t-1-th iteration can be obtained from the gradients from the 1st layer to the N-th layer in the t-1 iteration), so as to obtain the normalized gradient of the N-1-th layer in the t-th iteration, ..., finally, based on the loss in the t-th iteration, the gradient of the 1st layer of the model to be trained in the t-th iteration can be obtained, and the gradient of the 1st layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the 1st layer in the t-th iteration. Then, since the gradients of the 1st to Nth layers in the tth iteration have been obtained, the gradients of the 1st to Nth layers in the tth iteration can be calculated to obtain the GGN in the tth iteration, and the MGGN in the t-1th iteration and the GGN in the tth iteration can be calculated to obtain the MGGN in the tth iteration. Finally, the parameters of the 1st to Nth layers can be updated based on the normalized gradients of the 1st to Nth layers in the tth iteration. At this point, the tth iteration of the model to be trained is completed, so the t+1th iteration of the model to be trained can be entered.

It should be noted that when t=K, the loss in the Kth iteration has converged, so there is no need to calculate the GGN in the Kth iteration and the MGGN in the Kth iteration. After updating the parameters of the 1st to Nth layers of the model to be trained, there is no need to enter the K+1th iteration, and the final target model can be obtained directly.

More specifically, the gradients of the first layer to the Nth layer in the tth iteration can be calculated by a preset formula (for example, the formula can implement at least one of the following calculations: exponentiation calculation, sum calculation, square root calculation, and weighted sum calculation), so as to obtain the GGN in the tth iteration and the MGGN in the tth iteration. For example, the formula can be presented as:

In the above formula, GGN _t is the GGN in the t-th iteration, MGGN _t-1 is the MGGN in the t-1-th iteration, MGGN _t is the MGGN in the t-th iteration, and β is a preset weight.

Furthermore, the model to be trained may be deployed on M GPUs. Then, in the t-th iteration of the model to be trained, the t-th batch of training data input to the model to be trained may include M training data, and the M training data may be input into the model to be trained on the M GPUs accordingly. After the models to be trained on the M GPUs process the M training data respectively, the processing results of the M training data may be obtained accordingly. Then, based on the processing results of the M training data, the M losses in the t-th iteration may be obtained accordingly.

Then, based on the M losses, M candidate gradients of the N-th layer in the t-th iteration can be obtained accordingly, and the M candidate gradients of the N-th layer in the t-th iteration are directly averaged to obtain the gradient of the N-th layer in the t-th iteration, and the gradient of the N-th layer in the t-th iteration is immediately normalized based on the MGGN in the t-1-th iteration to obtain the normalized gradient of the N-th layer in the t-th iteration. Next, based on the M losses, the M candidate gradients of the N-1th layer in the t iteration can be obtained accordingly, and the M candidate gradients of the N-1th layer in the t iteration are directly averaged to obtain the gradient of the N-1th layer in the t iteration, and the gradient of the N-1th layer in the t iteration is immediately normalized based on the MGGN in the t-1 iteration to obtain the normalized gradient of the N-1th layer in the t iteration, ..., finally, based on the M losses, the M candidate gradients of the 1st layer in the t iteration can be obtained accordingly, and the M candidate gradients of the 1st layer in the t iteration are directly averaged to obtain the gradient of the 1st layer in the t iteration, and the gradient of the 1st layer in the t iteration is immediately normalized based on the MGGN in the t-1 iteration to obtain the normalized gradient of the 1st layer in the t iteration.

In order to better understand the model training method provided by the embodiment of the present application, the method provided by the embodiment of the present application is compared with the method provided by the related art one below, and the comparison results are shown in Figures 5 and 6 (Figure 5 is a schematic diagram of the comparison results provided by the embodiment of the present application, and Figure 6 is another schematic diagram of the comparison results provided by the embodiment of the present application). In the embodiment of the present application, the MGGN in the t-1th iteration is used to complete the gradient normalization in the tth iteration, and the related art one uses the GGN in the tth iteration to complete the gradient normalization in the tth iteration. Based on Figures 5 and 6, it can be seen that the degree of fit between the MGGN in the t-1th iteration and the GGN in the tth iteration is very high, and the loss curves obtained by the two methods are also almost overlapping. It can be seen that the degree of equivalent matching of the two methods is quite high.

In the method provided by the embodiment of the present application, gradient normalization and back propagation are performed simultaneously, as shown in FIG7 (FIG. 7 is a schematic diagram of an application example of the model training method provided by the embodiment of the present application). Assume that in a certain iteration of the model, when the back propagation reaches a certain layer in the model, GPU0 can obtain the gradient G0 of the layer, GPU1 can obtain the gradient G1 of the layer, GPU2 can obtain the gradient G2 of the layer, and GPU3 can obtain the gradient G3 of the layer. After the four GPUs obtain the four gradients G0, G1, G2, and G3 of the layer through mutual communication, they can all average G0, G1, G2, and G3, and divide them by the MGGN in the previous iteration to obtain the normalized gradient of the layer. Then, the four GPUs can perform similar operations on the previous layer of the layer until the entire back propagation stage and the gradient normalization stage are completed, and then the MGGN in the iteration can be updated, and the normalized gradients of each layer can be used to complete the parameter update of each layer. At this point, the iteration is completed and the next iteration can be entered.

In addition, the method provided in the embodiment of the present application can also be compared with the method provided in the related technology 2. The comparison result is shown in Figure 8 (Figure 8 is another schematic diagram of the comparison structure provided in the embodiment of the present application). Both of them have the purpose of training the BERT model. Based on Figure 8, compared with the loss curve of the method provided in the related technology 2, the loss curve of the method provided in the embodiment of the present application can significantly accelerate the loss convergence, and when achieving the same accuracy, the method provided in the embodiment of the present application can save about 25% of the number of iterations.

In an embodiment of the present application, in the t-th iteration of the model to be trained, after obtaining the gradient of the N-th layer of the model to be trained in the t-th iteration, the gradient of the N-th layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-th layer in the t-th iteration; similarly, after obtaining the gradient of the N-1-th layer in the t-th iteration, the gradient of the N-1-th layer in the t-th iteration is immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-1-th layer in the t-th iteration, ..., after obtaining the gradient of the 1st layer in the t-th iteration, the gradient of the 1st layer in the t-th iteration is immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the 1st layer in the t-th iteration. In this way, the parameters of the 1st to Nth layers can be updated based on the normalized gradients of the 1st to Nth layers in the tth iteration. At this point, the tth iteration of the model to be trained is completed, so the t+1th iteration of the model to be trained can be entered. In the above process, the MGGN in the t-1th iteration can be obtained by the gradients of the 1st to Nth layers in the t-1th iteration. That is to say, the MGGN in the t-1th iteration has been obtained in advance in the t-1th iteration of the model to be trained. Therefore, after entering the tth iteration of the model to be trained, once the gradient of any layer in the model to be trained in the tth iteration is obtained, the MGGN in the t-1th iteration can be immediately used to normalize the gradient of the layer in the tth iteration, so as to obtain the normalized gradient of the layer in the t-1th iteration. In this way, the back propagation and gradient normalization in the tth iteration of the model to be trained can be performed simultaneously, which can effectively shorten the tail time of the tth iteration of the model to be trained, thereby reducing the total time cost required for the entire training process of the model to be trained.

The above is a detailed description of the model training method provided in the embodiment of the present application. The following is a brief introduction to the information processing method provided in the embodiment of the present application. As shown in Figure 9 (Figure 9 is a flow chart of the information processing method provided in the embodiment of the present application), as shown in Figure 9, the method is implemented by the target model obtained by training in the embodiment shown in Figure 4, and the method includes:

901. Obtain information to be processed.

902. Input the information to be processed into the target model so as to process the information through the target model, thereby obtaining a processing result of the information.

For example, suppose the target model is a trained BERT model. After a question is input into the BERT model, the BERT model can output the answer to the question.

For another example, suppose the target model is a trained BERT model. After a prompt is input into the BERT model, the BERT model can output the emotion to which the prompt belongs.

For example, assuming that the target model is a trained BERT model, after inputting an incomplete text into the BERT model, the BERT model can output the supplementary part of the incomplete text, and so on.

In addition, the target model trained by the model training method provided in the embodiment of the present application and the model trained by the method of the related technology 2 can be fine-tuned and tested on the four downstream tasks of SQUAD, MRPC, SST-2, and MNLI, and the test results can be compared. As shown in Table 1:

Table 1

Based on Table 1, it can be seen that the model trained in the embodiment of the present application has better accuracy, that is, it has better performance in various downstream tasks.

The above is a detailed description of the model training method and information processing method provided in the embodiment of the present application. The following is an introduction to the model training device and information processing device provided in the embodiment of the present application. FIG10 is a structural diagram of the model training device provided in the embodiment of the present application. As shown in FIG10 , the device includes:

A first acquisition module 1001 is used to acquire, in the tth iteration of the model to be trained, the gradient of the i-th layer of the model to be trained in the tth iteration based on the training data input to the model to be trained, t≥2, i＝1,...,N, N≥1;

A normalization module 1002 is used to normalize the gradient of the i-th layer in the t-th iteration based on the MGGN in the t-1-th iteration to obtain the normalized gradient of the i-th layer in the t-th iteration, and the MGGN in the t-1-th iteration is obtained based on the gradient of the N layer of the model to be trained in the t-1-th iteration;

The updating module 1003 is used to update the parameters of the i-th layer based on the normalized gradient of the i-th layer in the t-th iteration to enter the t+1-th iteration of the model to be trained.

In an embodiment of the present application, in the t-th iteration of the model to be trained, after obtaining the gradient of the N-th layer of the model to be trained in the t-th iteration, the gradient of the N-th layer in the t-th iteration can be immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-th layer in the t-th iteration; similarly, after obtaining the gradient of the N-1-th layer in the t-th iteration, the gradient of the N-1-th layer in the t-th iteration is immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the N-1-th layer in the t-th iteration, ..., after obtaining the gradient of the 1st layer in the t-th iteration, the gradient of the 1st layer in the t-th iteration is immediately normalized based on the MGGN in the t-1-th iteration, so as to obtain the normalized gradient of the 1st layer in the t-th iteration. In this way, the parameters of the first layer to the Nth layer can be updated based on the normalized gradient of the first layer to the Nth layer in the tth iteration. At this point, the tth iteration of the model to be trained is completed, so the t+1th iteration of the model to be trained can be entered. In the above process, the MGGN in the t-1th iteration can be obtained by the gradient of the first layer to the Nth layer in the t-1th iteration. That is to say, the MGGN in the t-1th iteration has been obtained in advance in the t-1th iteration of the model to be trained. Therefore, after entering the tth iteration of the model to be trained, once the gradient of a certain layer in the model to be trained in the tth iteration is obtained, the MGGN in the t-1th iteration can be immediately used to normalize the gradient of the layer in the tth iteration, so as to obtain the normalized gradient of the layer in the t-1th iteration. In this way, the back propagation and gradient normalization in the tth iteration of the model to be trained can be performed simultaneously, which can effectively shorten the tail time of the tth iteration of the model to be trained, thereby reducing the total time cost required for the entire training process of the model to be trained.

In a possible implementation manner, the device further includes: a second acquisition module, configured to acquire the MGGN in the t-th iteration based on the gradient of the N layer in the t-th iteration.

In one possible implementation, the second acquisition module is used to: perform a first calculation on the gradient of the N layer in the t-th iteration to obtain the global gradient norm GGN in the t-th iteration; perform a second calculation on the MGGN in the t-1-th iteration and the GGN in the t-th iteration to obtain the MGGN in the t-th iteration.

In a possible implementation manner, the first calculation includes at least one of the following: a power calculation, a sum calculation, and a square root calculation.

In a possible implementation manner, the second calculation is a weighted sum calculation.

In one possible implementation, the first acquisition module 1001 is used to: obtain M unprocessed gradients of the i-th layer of the model to be trained in the t-th iteration based on M training data input into the model to be trained, where M≥2; and average the M unprocessed gradients to obtain the gradient of the i-th layer in the t-th iteration.

In a possible implementation, the first acquisition module 1001 is further used to obtain the gradient of the N layer in the first iteration based on the training data input to the model to be trained in the first iteration of the model to be trained; the second acquisition module is used to obtain the GGN in the first iteration based on the gradient of the N layer in the first iteration, and the GGN in the first iteration is used as the MGGN in the first iteration; the normalization module 1002 is further used to normalize the gradient of the N layer in the first iteration based on the GGN in the first iteration, and obtain the GGN of the N layer in the first iteration. The normalized gradient of the update module 1003 is also used to update the parameters of the N layer based on the normalized gradient of the N layer in the first iteration to enter the second iteration of the model to be trained.

FIG. 11 is a schematic diagram of a structure of an information processing device provided in an embodiment of the present application. As shown in FIG. 11 , the device includes a target model obtained by training in the embodiment shown in FIG. 10 . The device includes:

The acquisition module 1101 is used to acquire information to be processed.

The processing module 1102 is used to input the information to be processed into the target model so as to process the information through the target model, thereby obtaining the processing result of the information.

It should be noted that the information interaction, execution process, etc. between the modules/units of the above-mentioned device are based on the same concept as the method embodiment of the present application, and the technical effects they bring are the same as those of the method embodiment of the present application. For specific contents, please refer to the description in the method embodiment shown above in the embodiment of the present application, and will not be repeated here.

The embodiment of the present application also relates to an execution device, and FIG. 12 is a structural schematic diagram of the execution device provided by the embodiment of the present application. As shown in FIG. 12, the execution device 1200 can be specifically manifested as a mobile phone, a tablet, a laptop computer, an intelligent wearable device, a server, etc., which is not limited here. Among them, the information processing device described in the corresponding embodiment of FIG. 11 can be deployed on the execution device 1200 to implement the function of information processing in the corresponding embodiment of FIG. 9. Specifically, the execution device 1200 includes: a receiver 1201, a transmitter 1202, a processor 1203 and a memory 1204 (wherein the number of processors 1203 in the execution device 1200 can be one or more, and FIG. 12 takes one processor as an example), wherein the processor 1203 may include an application processor 12031 and a communication processor 12032. In some embodiments of the present application, the receiver 1201, the transmitter 1202, the processor 1203 and the memory 1204 may be connected via a bus or other means.

The memory 1204 may include a read-only memory and a random access memory, and provides instructions and data to the processor 1203. A portion of the memory 1204 may also include a non-volatile random access memory (NVRAM). The memory 1204 stores processor and operation instructions, executable modules or data structures, or subsets thereof, or extended sets thereof, wherein the operation instructions may include various operation instructions for implementing various operations.

The processor 1203 controls the operation of the execution device. In a specific application, the various components of the execution device are coupled together through a bus system, wherein the bus system includes not only a data bus but also a power bus, a control bus, and a status signal bus, etc. However, for the sake of clarity, various buses are referred to as bus systems in the figure.

The method disclosed in the above embodiment of the present application can be applied to the processor 1203, or implemented by the processor 1203. The processor 1203 can be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit in the processor 1203 or the instruction in the form of software. The above processor 1203 can be a general processor, a digital signal processor (digital signal processing, DSP), a microprocessor or a microcontroller, and can further include an application specific integrated circuit (application specific integrated circuit, ASIC), a field programmable gate array (field-programmable gate array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The processor 1203 can implement or execute the various methods, steps and logic block diagrams disclosed in the embodiment of the present application. The general processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to be executed, or the hardware and software modules in the decoding processor can be combined and executed. The software module may be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 1204, and the processor 1203 reads the information in the memory 1204 and completes the steps of the above method in combination with its hardware.

The receiver 1201 can be used to receive input digital or character information, and generate signal input related to the relevant settings and function control of the execution device. The transmitter 1202 can be used to output digital or character information through the first interface; the transmitter 1202 can also be used to send instructions to the disk group through the first interface to modify the data in the disk group; the transmitter 1202 can also include a display device such as a display screen.

In an embodiment of the present application, in one case, the processor 1203 is used to obtain the processing result of the information through the target model in the embodiment corresponding to Figure 9.

The embodiment of the present application also relates to a training device. FIG13 is a schematic diagram of the structure of the training device provided by the embodiment of the present application. As shown in FIG13, the training device 1300 is implemented by one or more servers. The training device 1300 may have relatively large differences due to different configurations or performances. It may include one or more central processing units (CPU) 1313 (for example, one or more processors) and memory 1332, and one or more storage media 1330 (for example, one or more mass storage devices) storing application programs 1342 or data 1344. Among them, the memory 1332 and the storage medium 1330 may be short-term storage or persistent storage. The program stored in the storage medium 1330 may include one or more modules (not shown), each of which may include a series of instruction operations in the training device. Furthermore, the central processor 1313 may be configured to communicate with the storage medium 1330 and execute a series of instruction operations in the storage medium 1330 on the training device 1300.

The training device 1300 may also include one or more power supplies 1326, one or more wired or wireless network interfaces 1350, one or more input and output interfaces 1358; or, one or more operating systems 1341, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.

Specifically, the training device can execute the model training method in the embodiment corresponding to Figure 4 to obtain the target model.

An embodiment of the present application also relates to a computer storage medium, in which a program for signal processing is stored. When the program is run on a computer, the computer executes the steps executed by the aforementioned execution device, or the computer executes the steps executed by the aforementioned training device.

An embodiment of the present application also relates to a computer program product, which stores instructions, which, when executed by a computer, enable the computer to execute the steps executed by the aforementioned execution device, or enable the computer to execute the steps executed by the aforementioned training device.

The execution device, training device or terminal device provided in the embodiments of the present application may specifically be a chip, and the chip includes: a processing unit and a communication unit, wherein the processing unit may be, for example, a processor, and the communication unit may be, for example, an input/output interface, a pin or a circuit, etc. The processing unit may execute the computer execution instructions stored in the storage unit so that the chip in the execution device executes the data processing method described in the above embodiment, or so that the chip in the training device executes the data processing method described in the above embodiment. Optionally, the storage unit is a storage unit in the chip, such as a register, a cache, etc. The storage unit may also be a storage unit located outside the chip in the wireless access device end, such as a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM), etc.

Specifically, please refer to FIG. 14 , which is a schematic diagram of the structure of a chip provided in an embodiment of the present application. The chip can be expressed as a neural network processor NPU 1400. NPU 1400 is mounted on the host CPU (Host CPU) as a coprocessor, and tasks are assigned by the Host CPU. The core part of the NPU is the operation circuit 1403, which is controlled by the controller 1404 to extract matrix data from the memory and perform multiplication operations.

In some implementations, the operation circuit 1403 includes multiple processing units (Process Engine, PE) inside. In some implementations, the operation circuit 1403 is a two-dimensional systolic array. The operation circuit 1403 can also be a one-dimensional systolic array or other electronic circuits capable of performing mathematical operations such as multiplication and addition. In some implementations, the operation circuit 1403 is a general-purpose matrix processor.

For example, assume there is an input matrix A, a weight matrix B, and an output matrix C. The operation circuit takes the corresponding data of matrix B from the weight memory 1402 and caches it on each PE in the operation circuit. The operation circuit takes the matrix A data from the input memory 1401 and performs matrix operation with matrix B, and the partial result or final result of the matrix is stored in the accumulator 1408.

The unified memory 1406 is used to store input data and output data. The weight data is directly transferred to the weight memory 1402 through the direct memory access controller (DMAC) 1405. The input data is also transferred to the unified memory 1406 through the DMAC.

BIU stands for Bus Interface Unit, that is, the bus interface unit 1413, which is used for the interaction between AXI bus and DMAC and instruction fetch buffer (IFB) 1409.

The bus interface unit 1413 (Bus Interface Unit, BIU for short) is used for the instruction fetch memory 1409 to obtain instructions from the external memory, and is also used for the storage unit access controller 1405 to obtain the original data of the input matrix A or the weight matrix B from the external memory.

DMAC is mainly used to transfer input data in the external memory DDR to the unified memory 1406 or to transfer weight data to the weight memory 1402 or to transfer input data to the input memory 1401.

The vector calculation unit 1407 includes multiple operation processing units, and further processes the output of the operation circuit 1403 when necessary, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc. It is mainly used for non-convolutional/fully connected layer network calculations in neural networks, such as Batch Normalization, pixel-level summation, upsampling of predicted label planes, etc.

In some implementations, the vector calculation unit 1407 can store the processed output vector to the unified memory 1406. For example, the vector calculation unit 1407 can apply a linear function; or a nonlinear function to the output of the operation circuit 1403, such as linear interpolation of the predicted label plane extracted by the convolution layer, and then, for example, a vector of accumulated values to generate an activation value. In some implementations, the vector calculation unit 1407 generates a normalized value, a pixel-level summed value, or both. In some implementations, the processed output vector can be used as an activation input to the operation circuit 1403, for example, for use in a subsequent layer in a neural network.

An instruction fetch buffer 1409 connected to the controller 1404 is used to store instructions used by the controller 1404;

Unified memory 1406, input memory 1401, weight memory 1402 and instruction fetch memory 1409 are all on-chip memories. External memories are private to the NPU hardware architecture.

The processor mentioned in any of the above places may be a general-purpose central processing unit, a microprocessor, an ASIC, or one or more integrated circuits for controlling the execution of the above program.

It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

Through the description of the above implementation mode, the technicians in the field can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present application, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.

In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

Claims (17)

A model training method, characterized in that the method comprises: In the t-th iteration of the model to be trained, based on the training data input into the model to be trained, the gradient of the i-th layer of the model to be trained in the t-th iteration is obtained, t≥2, i＝1,...,N, N≥1; Based on the moving average MGGN of the global gradient norm in the t-1th iteration, the gradient of the i-th layer in the t-th iteration is normalized to obtain the normalized gradient of the i-th layer in the t-th iteration, and the MGGN in the t-1th iteration is obtained based on the gradient of the N layer of the model to be trained in the t-1th iteration; Based on the normalized gradient of the i-th layer in the t-th iteration, the parameters of the i-th layer are updated to enter the t+1-th iteration of the model to be trained. The method according to claim 1, characterized in that the method further comprises: Based on the gradient of the N layers in the t-th iteration, an MGGN in the t-th iteration is obtained. The method according to claim 2, characterized in that the obtaining of the MGGN in the t-th iteration based on the gradient of the N layer in the t-th iteration comprises: Performing a first calculation on the gradient of the N layer in the t-th iteration to obtain a global gradient norm GGN in the t-th iteration; A second calculation is performed on the MGGN in the t-1th iteration and the GGN in the tth iteration to obtain the MGGN in the tth iteration. The method according to claim 3 is characterized in that the first calculation includes at least one of the following: exponentiation calculation, summation calculation and square root calculation. The method according to claim 3 or 4 is characterized in that the second calculation is a weighted sum calculation. The method according to any one of claims 1 to 5, characterized in that the step of obtaining the gradient of the i-th layer of the model to be trained in the t-th iteration based on the training data input into the model to be trained comprises: Based on the M training data input to the model to be trained, obtaining M to-be-processed gradients of the i-th layer of the model to be trained in the t-th iteration, where M≥2; The M to-be-processed gradients are averaged to obtain the gradient of the i-th layer in the t-th iteration. The method according to any one of claims 1 to 6, characterized in that the method further comprises: In a first iteration of the model to be trained, based on the training data input into the model to be trained, obtaining the gradient of the N layers in the first iteration; Based on the gradient of the N layer in the first iteration, obtaining the GGN in the first iteration, and using the GGN in the first iteration as the MGGN in the first iteration; Based on the GGN in the first iteration, normalizing the gradients of the N layers in the first iteration to obtain the normalized gradients of the N layers in the first iteration; Based on the normalized gradient of the N layer in the first iteration, the parameters of the N layer are updated to enter the second iteration of the model to be trained. A model training device, characterized in that the device comprises: A first acquisition module is used to acquire, in the t-th iteration of the model to be trained, the gradient of the i-th layer of the model to be trained in the t-th iteration based on the training data input into the model to be trained, t≥2, i＝1,...,N, N≥1; A normalization module, configured to normalize the gradient of the i-th layer in the t-th iteration based on the MGGN in the t-1-th iteration to obtain the normalized gradient of the i-th layer in the t-th iteration, wherein the MGGN in the t-1-th iteration is obtained based on the gradient of the N layer of the model to be trained in the t-1-th iteration; An updating module is used to update the parameters of the i-th layer based on the normalized gradient of the i-th layer in the t-th iteration to enter the t+1-th iteration of the model to be trained. The device according to claim 8, characterized in that the device further comprises: The second acquisition module is used to acquire the MGGN in the t-th iteration based on the gradient of the N layer in the t-th iteration. The device according to claim 9, characterized in that the second acquisition module is used to: Performing a first calculation on the gradient of the N layer in the t-th iteration to obtain a global gradient norm GGN in the t-th iteration; A second calculation is performed on the MGGN in the t-1th iteration and the GGN in the tth iteration to obtain the MGGN in the tth iteration. The device according to claim 10 is characterized in that the first calculation includes at least one of the following: exponentiation calculation, summation calculation and square root calculation. The device according to claim 10 or 11 is characterized in that the second calculation is a weighted sum calculation. The device according to any one of claims 8 to 12, characterized in that the first acquisition module is used to: Based on the M training data input to the model to be trained, obtaining M to-be-processed gradients of the i-th layer of the model to be trained in the t-th iteration, where M≥2; The M to-be-processed gradients are averaged to obtain the gradient of the i-th layer in the t-th iteration. The device according to any one of claims 8 to 13, characterized in that the first acquisition module is further used to obtain the gradient of the N layers in the first iteration of the model to be trained based on the training data input to the model to be trained; A second acquisition module is configured to acquire a GGN in the first iteration based on the gradient of the N layer in the first iteration, and the GGN in the first iteration is used as the MGGN in the first iteration; The normalization module is further used to normalize the gradient of the N layer in the first iteration based on the GGN in the first iteration to obtain the normalized gradient of the N layer in the first iteration; The updating module is also used to update the parameters of the N layers based on the normalized gradients of the N layers in the first iteration to enter the second iteration of the model to be trained. A model training device, characterized in that the device includes a memory and a processor; the memory stores code, and the processor is configured to execute the code. When the code is executed, the model training device executes the method as described in any one of claims 1 to 7. A computer storage medium, characterized in that the computer storage medium stores one or more instructions, and when the instructions are executed by one or more computers, the one or more computers implement the method described in any one of claims 1 to 7. A computer program product, characterized in that the computer program product stores instructions, and when the instructions are executed by a computer, the computer implements the method according to any one of claims 1 to 7.