WO2022109779A1

WO2022109779A1 - Data processing method and related device

Info

Publication number: WO2022109779A1
Application number: PCT/CN2020/131120
Authority: WO
Inventors: 王宇; 顾宝成; 苗方正; 马兵; 范照云
Original assignee: 华为技术有限公司
Priority date: 2020-11-24
Filing date: 2020-11-24
Publication date: 2022-06-02
Also published as: CN116670688A

Abstract

A data processing method and a related device, which relate to the field of artificial intelligence, and are used for filling in the gap of it being impossible to determine a target arrangement feature of target data when a rank is unknown or the rank is known but a dimension value of at least one dimension is unknown. The method comprises: acquiring at least one piece of input information, wherein each piece of input information is used for representing an arrangement feature of target data, the arrangement feature comprises a first feature, the value of the first feature is a first value or the value of the first feature is a second value, the first value is used for representing that the rank of the target data is unknown, and the second value is used for representing that the rank of the target data is known and a dimension value corresponding to at least one dimension is unknown; acquiring a constraint condition corresponding to a target operator; and processing the at least one piece of input information on the basis of the constraint condition, so as to obtain at least one piece of output information, wherein each piece of output information is used for representing a target arrangement feature of the target data, and the value of the target arrangement feature comprises a target value of the first feature.

Description

A data processing method and related equipment

technical field

The present application relates to the field of communication technologies, and in particular, to a data processing method and related equipment.

Background technique

Artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a branch of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that responds in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making. Research in the field of artificial intelligence includes robotics, natural language processing, computer vision, decision-making and reasoning, human-computer interaction, recommendation and search, and basic AI theory. With the rapid development of AI and the further expansion of deep learning applications, it is not only necessary to implement some customized operators, but also a set of derivation methods for realizing the shape of the output data of the operators, so as to compile or run images on images. The derivation method can be called during the process.

In the related art, an unknown dimension value is only represented by -1, and the range of the dimension value is represented by the maximum value and the minimum value, so as to complete the derivation of the shape of the output data. However, in a neural network (NN) model, the dimension value and rank value of the input data are often unknown. Therefore, a new method is proposed for how to better derive the shape of the output data. challenge.

SUMMARY OF THE INVENTION

The embodiments of the present application provide a data processing method and related equipment, which aim to fill in the blank that the target arrangement feature of target data cannot be determined when the rank is unknown or the rank is known but the dimension value of at least one dimension is unknown. .

A first aspect of the embodiments of the present application provides a data processing method, the method may include: acquiring at least one piece of input information, wherein each input information is used to represent an arrangement feature of target data, and the arrangement feature includes a first feature, the value of the first feature is the first value or the value of the first feature is the second value, the first value is used to represent the rank of the unknown target data, and the second value is used to represent the rank of the known target data and unknown The dimension value corresponding to at least one dimension; obtain the constraints corresponding to the target operator, and the constraints are used to represent the execution logic of the target operator; process at least one input information based on the constraints to obtain at least one output information, each output information Both are used to characterize the target arrangement feature of the target data, and the value of the target arrangement feature includes the target value of the first feature. In the above manner, in the case where the value of the first feature in the arrangement features reflects the rank of the unknown target data or the value of the known target data but the dimension value corresponding to at least one dimension is unknown, the constraint conditions of the target operator are used to determine At least one input information is processed, so that at least one output information obtained can reflect the target arrangement characteristics of the target data, filling the situation where the rank is unknown or the rank is known but the dimension value of at least one dimension is unknown. The blank of the target layout feature of the target data.

In some embodiments, the arrangement feature further includes a second feature and a third feature, the second feature is a third value, the third value is used to represent a rank range, and the third feature is a fourth value Or the value of the third feature is the fifth value, the fourth value is used to represent the total dimension range of the target data when the value of the first feature is the first value, and the fifth value is used to represent the value of the first feature. When the second value is the second value, the dimension range corresponding to each dimension; obtaining at least one output information for at least one input information based on the constraints, including: processing the value of the second feature in the at least one input information based on the constraints to obtain the first a target value, the first target value is used to reflect the target range of the rank in the target data; the value of the first feature and the first target value in at least one input information are processed based on the constraint condition to obtain the second target value, the first target value The second target value is used to reflect the target value of the rank; the value of the third feature and the second target value in at least one input information are processed based on the constraint condition to obtain the third target value, and the third target value is used to reflect the target data At least one output information is obtained based on the first target value, the second target value and the third target value.

In some embodiments, after the at least one input information is processed based on the constraint condition to obtain at least one output information, the method may further include: determining a pre-allocated memory range of the corresponding output information based on the first target value and the third target value . In the above manner, the required memory range can be pre-allocated in advance based on the first target value and the third target value during image compilation. Better optimize memory allocation before running, so as to allocate memory reasonably and reduce unnecessary memory waste.

In some embodiments, the method may further include: selecting a target template based on the first target value, where the target template is used to optimize the target operator. Through the above method, the target template matching the first target value can be selected from the database, and the unmatched template can be eliminated, so that the target operator can be optimized and compiled based on the target template, so as to reduce the processing of the target operator branch, so as to select an appropriate execution algorithm to improve the compilation and execution performance of the target operator.

In other embodiments, the method may further include: obtaining a target arrangement feature of the target data based on any one of the at least one output information, wherein the target arrangement feature includes a first feature, a second feature and a third feature feature, the value of the first feature in the target arrangement feature is the second target value, the value of the second feature in the target arrangement feature is the first target value, and the value of the third feature in the target arrangement feature is the third value target value.

In other embodiments, the first value is a sign or a numerical value for indicating that the rank is unknown.

A second aspect of the embodiments of the present application provides a data processing device, and the data processing device may include: a programming interface module for acquiring at least one input information, wherein each input information is used to characterize the arrangement feature of target data , the arrangement feature includes a first feature, the value of the first feature is a first value or the value of the first feature is a second value, the first value is used to indicate the rank of the unknown target data, and the second value is used to indicate that the The rank of the target data is known and the dimension value corresponding to at least one dimension is unknown; the programming interface module is used to obtain the constraints corresponding to the target operator, and the constraints are used to represent the execution logic of the target operator; the processing module is used based on the constraints At least one input information is processed to obtain at least one output information, each output information is used to represent a target arrangement feature of the target data, and the value of the target arrangement feature includes the target value of the first feature.

In some embodiments, the arrangement feature further includes a second feature and a third feature, the second feature is a third value, the third value is used to represent a rank range, and the third feature is a fourth value Or the value of the third feature is the fifth value, the fourth value is used to represent the total dimension range of the target data when the value of the first feature is the first value, and the fifth value is used to represent the value of the first feature. When the second value is the second value, the dimension range corresponding to each dimension; the processing module is specifically configured to: process the value of the second feature in at least one input information according to the constraint condition to obtain the first target value, and the first target value is used for Reflect the target range of the rank in the target data; process the value of the first feature and the first target value in at least one input information according to the constraint condition to obtain a second target value, and the second target value is used to reflect the target value of the rank ; Process the value of the third feature in the at least one input information and the second target value according to the constraint condition to obtain the third target value, and the third target value is used for the target dimension range of the range target data; According to the first target value , the second target value and the third target value to obtain at least one output information.

In other embodiments, the processing module is further specifically configured to, after processing at least one piece of information based on a constraint condition to obtain at least one piece of output information, determine a pre-allocated memory corresponding to the output information based on the first target value and the third target value scope.

In other embodiments, the processing module is specifically configured to select a target template according to the first target value, and the target template is used to optimize the target operator.

In other embodiments, the processing module is further specifically configured to obtain a target arrangement feature of the target data based on any one of the at least one output information, wherein the target arrangement feature includes a first feature, a second feature, and a first feature. Three features, the value of the first feature in the target arrangement feature is the second target value, the value of the second feature in the target arrangement feature is the first target value, and the value of the third feature in the target arrangement feature is the first target value Three target values.

A third aspect of the present application provides a data processing device, which may include: a memory for storing computer-readable instructions. It may further comprise a processor coupled to the memory for executing computer readable instructions in the memory to perform the method as described in the first aspect or any of the possible embodiments of the first aspect.

A fourth aspect of the present application provides a computer-readable storage medium, when instructions are executed on a computer device, the computer device causes the computer device to perform the method described in the first aspect or any possible implementation manner of the first aspect.

A fifth aspect of the present application provides a computer program product, which, when run on a computer, enables the computer to execute the method described in the first aspect or any possible implementation manner of the first aspect.

A sixth aspect of the present application provides a chip system, where the chip system may include a processor for supporting a terminal device or a server to implement the methods involved in the first aspect or any possible implementation manner of the first aspect. function.

Optionally, in combination with the above sixth aspect, in a first possible implementation manner, the chip system may further include a memory, which is used to store necessary program instructions and data of the terminal device. The chip system may be composed of chips, or may include chips and other discrete devices. Wherein, the chip system may include an application specific integrated circuit (ASIC), an off-the-shelf programmable gate array (field programmable gate array, FPGA), or other programmable logic devices. Further, the chip system may also include interface circuits and the like.

It should be noted that, the beneficial effects brought by the implementations of the second aspect to the sixth aspect of the present application can be understood with reference to the implementation of the first aspect, which will not be repeated here.

In the technical solutions provided by the embodiments of the present application, since the value of the first feature in the arrangement features can reflect the rank of the unknown target data or the situation that the value of the target data is known but the dimension value corresponding to at least one dimension is unknown, the value of the dimension corresponding to at least one dimension is unknown. The constraints of the target operator process at least one input information, so that at least one output information obtained can reflect the target arrangement characteristics of the target data, and fill in the rank unknown or the rank is known but the dimension value of at least one dimension is unknown. In the case of , the blank of the target arrangement feature of the target data cannot be determined.

Description of drawings

1 is a flowchart of a data processing provided in an embodiment of the present application;

2 is a schematic diagram of a data processing method provided by an embodiment of the present application;

3 is a schematic diagram of a memory allocation scenario applied to an image provided in an embodiment of the present application;

4 is a schematic diagram of a compilation scene applied to an image provided in an embodiment of the present application;

5 is a schematic diagram of a hardware structure of a data processing device provided by an embodiment of the present application;

FIG. 6 is a schematic structural diagram of an execution device provided by an embodiment of the present application;

FIG. 7 is a schematic structural diagram of another data processing device according to an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings in the embodiments of the present application. It is to be understood that the terms "comprising" and "comprising" used in the description and claims of this application indicate the presence of the described features, integers, steps, operations, elements and/or components, but do not exclude one or more other The presence or addition of features, integers, steps, operations, elements, components and/or sets thereof. It should also be understood that the terms used in the specification of the present application are only for the purpose of describing particular embodiments, and are not intended to limit the present application.

In order to better understand the technical solutions described in this application, the key technical terms involved in the embodiments of this application are explained below:

Since the embodiments of the present application involve the application of neural networks, for ease of understanding, related 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 can be composed of neural units, and a neural unit can refer to an operation unit that takes x _s and an intercept 1 as input, and the output of the operation unit can be shown in the following formula:

Among them, 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 an output signal. The output signal of the activation function can be used as the input of the next convolutional layer, and the activation function can be a sigmoid function. A neural network is a network formed by connecting a plurality of the above 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 with the local receptive field of the previous layer to extract the features of the local receptive field, and the local receptive field can be an area composed of several neural units.

There are many types of neural networks, for example, deep neural network (DNN), also known as multi-layer neural network, that is, a neural network with multiple layers of hidden layers; another example, convolutional neural network (convolutional neuron network) , CNN), is a deep neural network with a convolutional structure. This application does not limit the specific types of neural networks involved.

(2) Operator

An operator is a function that implements a specific function. For example, take the reshape operator, which is used to reinterpret the shape of tensor data. For another example, take the transpose operator as an example, the operator is used to adjust the dimension order of tensor data. In this application, commonly used functions for building deep learning model algorithms are collectively referred to as operators, and any operation performed on any function can be considered as an operator. For example, convolution is a mathematical method of integral transformation. For example, if the function f3 is generated by two functions f1 and f2, then f1, f2 and the convolution result f3 can be regarded as an operator respectively.

(3) Arrangement features

Arrangement features can be understood as being used to characterize the arrangement of target data, for example, the target data may be composed of several-dimensional pixel points, and the specific value or specific value range of the corresponding dimension value of each dimension, and so on. The arrangement features described above may include, but are not limited to, features such as the rank, dimension range, rank range, and/or shape of the target data, etc. of the target data.

(4) Rank

Rank can be understood as the number of dimensions of the target data. For example, if the target data consists of two-dimensional pixel points, the rank of the target data can be 2.

(5) Artificial intelligence processor

The artificial intelligence processor is also referred to as a dedicated processor. In this embodiment of the present application, the artificial intelligence processor refers to a processor for a specific application or field. For example: graphics processor (graphics processing unit, GPU), also known as display core, visual processor, display chip, is a kind of special equipment used in personal computers, workstations, game consoles and some mobile devices (such as tablet computers, smart phones, etc.) A dedicated processor for image computing work. Another example: the neural network processor (neural processing unit, NPU) is a special processor for matrix multiplication in the field of artificial intelligence. It adopts the architecture of "data-driven parallel computing" and is especially good at processing video and images. class of massive multimedia data.

(6) Deep Learning Framework

In order to meet the growing demand for neural networks, deep learning frameworks came into being. Through deep learning frameworks, researchers only need to pay attention to the network structure of deep learning algorithms, and write simple python (a cross-platform computer programming language) scripting The network structure can complete a complex deep learning network task, thereby realizing model inference and training on hardware. In other words, the deep learning framework is used to lower the development threshold in the field of deep learning, and provide the basic computing framework of deep learning for rapidly building deep learning applications. The current mainstream deep learning frameworks in the industry mainly include Tensorflow, Torch, Mxnet, Thenao, Caffe, etc. Taking the convolutional neural network framework Caffe as an example, in practical applications, Caffe supports various types of deep learning architectures, image classification and image segmentation, and can also support convolutional neural networks (CNN), used for target Detection of convolutional neural network (region-CNN, RCNN), long short-term memory neural network (long short-term memory, LSTM) and fully connected neural network design. The deep learning framework can support various types of basic operators. Specifically, the various types of basic operators involved here can include: common neural network operators. For example, common neural network operators are: convolution/deconvolution operator, pooling operator, activation operator, classifier (softmax) operator, and fully connected operator. The activation operator may include, but is not limited to, ReLU, Sigmoid, Tanh, and other operators that can be implemented by interpolation.

In the related art, the unknown dimension value is only represented by -1, and the maximum value and the minimum value are used to represent the range of the dimension value, so as to complete the shape derivation of the output data. However, in a neural network (NN) model, the dimension value and rank value of the data are often unknown. In this case, the target arrangement characteristics of the output data that can be derived are extremely high. It is probably not accurate enough, resulting in poor optimization results in scenarios such as subsequent optimization operator compilation.

In order to solve the above problems, the embodiments of the present application provide a situation where the rank of the target data is in an unknown state, or in a situation where the rank is known but the dimension value corresponding to at least one dimension is also in an unknown state, a more accurate deduction can be made. The target arrangement feature of the target data is obtained, so that the data processing device can solve the problem that the compilation time may be uncontrollable due to the long traversal time in the process of image optimization in the subsequent image data compilation and other scenarios based on the target arrangement feature. And improve the optimization effect. FIG. 1 shows a flowchart of a data processing provided in an embodiment of the present application. As can be seen from Figure 1, it includes n input information, n is an integer greater than or equal to 1, and each input information includes a shape that can represent the input information, dimensions range, rank (range), and rank range. After performing rank range derivation, range derivation, shape derivation and dimensions range derivation on the n input information through the constraints of the operator, m pieces of input information are obtained, where m is an integer greater than or equal to 1, and each output information also contains All include the derived shape, dimensions range, range, and rank range and other information.

It should be noted that the data processing described above can be applied to application scenarios such as image, speech, or natural language processing. In practical applications, it can also be applied to scenarios such as videos, and no limited description is made here.

In order to facilitate a better understanding of the present application, please refer to FIG. 2 , which is a schematic diagram of a data processing method provided in an embodiment of the present application, as follows:

201. Acquire at least one input information, wherein each input information is used to represent an arrangement feature of the target data, the arrangement feature includes a first feature, and the value of the first feature is a first value or a value of the first feature is the second value, the first value is used to represent the rank of the unknown target data, the second value is used to represent the rank of the known target data and the dimension value corresponding to at least one dimension is unknown.

In the embodiment, the described target data may include, but not limited to, data such as images and voices.

Since each input information can represent different arrangement features of the same target data, and different arrangement features reflect different shapes of the target data, for example, only the rank of the target data is given in some input information, And the value of the rank is 2. At this time, after receiving the input information, the data processing device can know that the target data is 2-dimensional data; for another example, the value of the rank of the target data is given in some information. 4, then the data processing device should know that the target data is 4-dimensional data.

It should be noted that the arrangement feature may include a first feature, and the first feature may be understood as the rank feature of the target data described above. When the value of the first feature in a certain input information is the first value, it means that the specific value of the rank of the target data cannot be directly obtained from the input information at this time, that is, the data processing device cannot know the specific value of the rank of the target data. Exactly how many dimensions of the target data are. It should be understood that the first value described above is used to represent the rank of the unknown target data, and the first value may take a value of -1, or a range, such as [1,100]. In practical applications, the value of the first value can also be -2, -3, -4, etc. It cannot be determined that the target data is a numerical value of several-dimensional data, or it can be a symbol such as #, %, or &, or Take other ranges, such as [2,80], etc., which will not be specifically limited here.

Or, when the value of the first feature in a certain input information is a second value, and the second value can be used to represent the rank of the known target data but the dimension value corresponding to at least one dimension is unknown, it also means that the data The processing device can know from the input information that the target data is data of several dimensions, but the target operator cannot know the dimension value corresponding to each dimension.

For example, one of the shapes of the target data is given in some input information, such as [-1,80]. At this time, the data processing device can also know the shape of the target data after receiving the input information. The rank is 2, that is, the target data is 2-dimensional data, and the dimension value corresponding to the second dimension is 80, while the dimension value corresponding to the other first dimension is in an unknown state. For another example, if the value of the first feature is [-1,-1], the data processing device can know that the rank of the target data is 2, but the dimension values corresponding to the first dimension and the second dimension are unknown. state.

Based on this, the data processing device can obtain at least one input information from the operator at the upper level, for example: obtain input information 1 from operator 1, obtain input information 2 from operator 2, ... obtain input information from operator n Input information n (n≥1, and an integer) and so on. In this way, after obtaining the at least one input information, the data processing device can obtain the constraint condition corresponding to the target operator according to the user's requirement, so as to process the at least one input information based on the constraint condition of the target operator.

202. Acquire a constraint condition corresponding to the target operator, where the constraint condition is used to characterize the execution logic of the target operator.

That is to say, it is understood that, after obtaining at least one input information for characterizing the arrangement feature of the target data, the data processing device can obtain the constraint condition corresponding to the target operator based on the user's requirement.

For example, if the user wants to obtain a target arrangement feature that can better indicate the target data after taking the intersection from the at least one input information described above, then the data processing device can determine that the target operator is an addition (add). ) operator, and obtain the constraints corresponding to the addition operator, that is, the data processing device needs to know the execution logic of the target operator, such as broadcasting, shrinking and other operations. It is worth noting that, in addition to the add operator described above, the target operator in the embodiment of the present application may also be other operators in practical applications, such as a subtraction operator (sub) and a multiplication operator (mul). , the exponential operator (exp), and so on.

203. Process the value of the second feature in at least one input information based on the constraint condition to obtain a first target value, where the first target value is used to reflect the target range of the rank in the target data, wherein the arrangement feature further includes a second target value. feature and third feature, the value of the second feature is the third value, the third value is used to indicate the range of ranks, the value of the third feature is the fourth value or the value of the third feature is the fifth value, the third value is The four values are used to represent the total dimension range of the target data when the value of the first feature is the first value, and the fifth value is used to represent the dimension range corresponding to each dimension when the value of the first feature is the second value.

In the embodiment, since the arrangement feature may also include a second feature and a third feature, the described second feature can be understood as a rank range feature, and the third feature can be understood as a dimension range feature. . It should be understood that the value of the second feature may be a third value, that is, the range of the rank of the target data is represented by the third value. For example, the third value may be in the range of [2, 7], [3, 8], etc., which is not limited here.

For the third feature, when the values of the first feature are different, the third feature also has correspondingly different values. Specifically, when the value of the first feature in a certain input information is the first value, since the first value can be used to represent the rank of the unknown target data, it is difficult to know each dimension from the corresponding input information. Therefore, the value of the third feature may be a fourth value, and the fourth value represents the total dimension range of the target data. For example, when the first value is -2, the fourth value may be [50, 200], [30, 150], etc., indicating the total dimension range when the rank of the target data is unknown, which is not limited here.

Or, when the value of the first feature in a certain input information is the second value, since the second value can be used to indicate the rank of the known target data but the dimension value corresponding to at least one dimension is unknown, the third feature is The value of can be a fifth value, and the fifth value indicates the corresponding dimension range of each dimension. For example, when the second value is [-1,-1], the fifth value can be [[1,100], [80,100]]; or when the second value is [-1,80], then The fifth value may be [[1,100], [80,80]], etc., which is not limited here.

In this way, the data processing device can process the value of the second feature in the at least one input information based on the constraint condition to obtain the first target value, and the first target value can indicate the target range of the rank.

204. Process the value of the first feature and the first target value in the at least one input information based on the constraint condition to obtain a second target value, where the second target value is used to reflect the target value of the rank.

205. Process the value of the third feature and the second target value in the at least one input information based on the constraint condition to obtain a third target value, where the third target value is used to reflect the target dimension range of the target data.

206. Obtain at least one piece of output information based on the first target value, the second target value, and the third target value.

In an embodiment, after the data processing device obtains the target range of the rank, the data processing device may further process the value of the first feature and the first target value in at least one input information based on the constraint condition to obtain the second target value. , the target value of the rank can be represented by the second target value, and then the data processing device also determines the third target value based on the constraint condition on the value of the third feature in the at least one input information and the above-mentioned second target value, and the third target value is determined by the second target value. The third target value reflects the target dimension range of the target data. In this way, the data processing device can obtain at least one output information based on the first target value, the second target value and the third target value, so that each output information can represent the target arrangement feature of the target data.

For example, suppose that the target data is an image as an example, and the data processing device obtains two input information used to characterize the arrangement features of the image, respectively:

Input information 1: shape_1: [-1,80], rank_1: 2, rank_range_1: [2,2], dimensions range_1: [[1,100],[80,80]];

Input information 2: shape_2: none, rank_2: -2, rank_range_2: [2,7], dimensions range_2: [50,200]. It should be noted that the value of the first feature can be interpreted as the first value of rank_2:-2 in the input information 2, and the second value can be understood as [-1,80] in the input information 1; in addition, the first The value of the second feature as the third value can be understood as rank_range_1: [2,2] in the input information 1, and rank_range_2: [2,7] in the input information 2; the value of the third feature is the fourth value can be It is understood as the dimensions range_2 in the input information 2: [50, 200], and the value of the third feature is the fifth value can be understood as the dimensions range_1 in the input information 1: [[1, 100], [80, 80]].

Therefore, when the target operator is the add operator, the data processing device can take the intersection between rank_range_1:[2,2] and rank_range_2:[2,7] based on the rank range derivation constraint in the add operator, and obtain The first target value is rank range_out: [2,2]. Then, the data processing device processes the first target value [2, 2] in combination with rank_2: -2 and rank_1: 2 based on the rank derivation constraint in the add operator, and obtains the second target value rank_out: 2. At this time, you can It is determined that the target data should be two-dimensional data. Then, the data processing device can also process the dimensions range_1:[[1,100],[80,80]], the dimensions range_2:[50,200] and the second target value based on the dimension range derivation constraints in the add operator. Specifically, The data processing device intersects the dimension range [1,100] of the first dimension in the dimensions range_1:[[1,100],[80,80]] with the dimensions range_2:[50,200] to obtain [50,100], and then the dimensions range_1:[ The dimension range [80,80] of the second dimension in [1,100],[80,80]] and the dimensions range_2:[50,200] take the intersection to get [80,80]. In this way, the final third target value is dimensions range_out: [[50,100], [80,80]].

Based on this, the output information of the image can be: rank range_out: [2, 2], rank_out: 2, and dimensions range_out: [[50, 100], [80, 80]]. Therefore, it can be known from the output information that the image is a two-dimensional image, and the dimension range corresponding to each dimension of the image can be reduced to [[50,100], [80,80]] based on the constraints of the target operator, which is the subsequent image. The optimization scene of the engine provides an optimization basis and reduces the traversal time and compilation time in the process of image optimization.

It is worth noting that the output information may also include shape_out: [-1, 80]. In addition, the above only takes two input information and the add operator as an example for description. Specifically, in practical applications, one or more input information can also be processed based on the constraints of the target operator to obtain at least one output information. The details can be understood with reference to the foregoing example of the add operator, which will not be repeated here.

It is further explained that, for obtaining at least one output information described above, the number of the output information will depend on the constraints or constraints of the target operator. For example, for the relu_grad_v2 operator, if two input information is processed, Then the obtained output information is two, which will not be specifically limited here.

Exemplarily, in some embodiments, after the at least one input information is processed based on the constraint condition to obtain at least one output information, the method further includes:

The pre-allocated memory range of the corresponding output information is determined based on the first target value and the third target value.

It can be understood that, since the first target value can reflect the target range of ranks in the target data, and the third target value can reflect the target dimension range of the target data, the data processing device can also be based on the first target value and the third target value. The target value calculates the memory size required to pre-store the corresponding output information, that is, the pre-allocated memory range. In the above manner, the required memory range can be pre-allocated in advance based on the first target value and the third target value during image compilation. Better optimize memory allocation before running, so as to allocate memory reasonably and reduce unnecessary memory waste. For example, referring to FIG. 3 , which is a schematic diagram of a memory allocation scenario applied to an image provided in an embodiment of the present application. As can be seen from FIG. 3 , after acquiring the input information of the image, the data processing device processes the input information based on the constraints of the target operator described above to obtain output information, and performs memory pre-allocation according to the output information.

Exemplarily, in other embodiments, the aforementioned data processing method may further include: selecting a target template based on the first target value, and the target template is used to optimize the target operator.

In the embodiment, since the first target value can reflect the target range of the rank in the target data, the data processing device can also select a target template that matches the first target value from the database, and eliminate unmatched templates, so as to facilitate The target operator can be optimized and compiled based on the target template to reduce the processing branches of the target operator, so as to select an appropriate execution algorithm and improve the compilation and execution performance of the target operator. For example, refer to FIG. 4 , which is a schematic diagram of a compilation scene applied to an image provided in an embodiment of the present application. It can be seen from Figure 4 that the data processing device obtains the input information, processes the input information based on the constraints of the target operator described above, obtains the output information, and further calls the compilation interface provided by the target operator, so as to obtain the output information based on the constraints of the target operator described above. A target value selects an appropriate template for compilation, and finally outputs the compilation result.

207. Obtain a target arrangement feature of the target data based on any one of the at least one output information.

In the embodiment, after obtaining at least one output information, the target arrangement feature of the target data can be obtained based on any output information, which fills the situation where the rank is unknown or the rank is known but the dimension value of at least one dimension is unknown. The blanks of the target arrangement features of the target data are determined. It should be noted that the target arrangement feature also includes the aforementioned first feature, the second feature and the third feature, and the value of the first feature in the target arrangement feature can be the aforementioned second target value, and the target The value of the second feature in the arrangement features is the aforementioned first target value, and the value of the third feature in the target arrangement feature is the aforementioned third target value.

In the above, a data processing method provided by the embodiments of the present application has been introduced. Through the solutions provided by the embodiments of the present application, at least one output information that can be used to characterize the target arrangement feature of the target data can be obtained, which fills the gap in the rank When the rank is unknown or the rank is known but the dimension value of at least one dimension is unknown, the blank of the target arrangement feature of the target data cannot be determined.

It can be understood that, in order to realize the above-mentioned functions, the above-mentioned data processing device includes corresponding hardware structures and/or software modules for executing each function. Those skilled in the art should easily realize that the present application can be implemented in hardware or in the form of a combination of hardware and computer software. Whether a function is performed by hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of this application.

Describing from the hardware structure, the data processing device in FIG. 1 to FIG. 4 may be implemented by one entity device, may also be implemented jointly by multiple entity devices, or may be a logic function module in one entity device. This embodiment of the present application This is not specifically limited.

For example, FIG. 5 shows a schematic diagram of a hardware structure of a data processing device provided by an embodiment of the present application. It includes: a communication interface 501 and a processor 502 , and may also include a memory 503 .

The communication interface 501 can use any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc. .

The processor 502 includes but is not limited to a central processing unit (CPU), a network processor (NP), an application-specific integrated circuit (ASIC) or a programmable logic device (programmable logic device, PLD) one or more. The above-mentioned PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a general-purpose array logic (generic array logic, GAL) or any combination thereof. Processor 502 is responsible for communication lines 504 and general processing, and may also provide various functions including timing, peripheral interface, voltage regulation, power management, and other control functions. Memory 503 may be used to store data used by processor 502 in performing operations.

Memory 503 may be read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (RAM), or other types of information and instructions It can also be an electrically erasable programmable read-only memory (electrically programmable read-only memory, EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, Optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or capable of carrying or storing desired program code in the form of instructions or data structures and capable of Any other medium that can be accessed by a computer, but is not limited to this. The memory may exist independently and be connected to the processor 502 through the communication line 504 . The memory 503 may also be integrated with the processor 502 . If the memory 503 and the processor 502 are separate devices from each other, the memory 503 and the processor 502 are connected, for example, the memory 503 and the processor 502 may communicate through a communication line 504. The communication interface 501 and the processor 502 can communicate through a communication line 504 , and the communication interface 501 can also be directly connected to the processor 502 .

Communication lines 504 , which may include any number of interconnected buses and bridges, link together various circuits including one or more processors 502 , represented by processor 502 , and memory, represented by memory 503 . Communication lines 504 may also link together various other circuits such as peripherals, voltage regulators, and power management circuits, etc., which are well known in the art and, therefore, will not be described further herein.

In a specific embodiment, the data processing device may include: a memory for storing computer-readable instructions. It may also include a communication interface coupled with the memory for acquiring at least one input information, wherein each input information is used to characterize the arrangement feature of the target data, the arrangement feature includes a first feature, and the value of the first feature is The value of the first value or the first feature is a second value, the first value is used to represent the rank of the unknown target data, the second value is used to represent the rank of the known target data and the dimension value corresponding to at least one dimension is unknown; and Obtain the constraints corresponding to the target operator, and the constraints are used to represent the execution logic of the target operator. Also included: a processor coupled with the communication interface for executing computer-readable instructions in the memory to perform the following operations: process at least one input message based on constraints to obtain at least one output message, each output message for use in The target arrangement feature of the target data is represented, and the value of the target arrangement feature includes the target value of the first feature.

In a specific embodiment, the arrangement feature further includes a second feature and a third feature, the value of the second feature is a third value, the third value is used to represent the range of ranks, and the value of the third feature is the third value The value of the fourth value or the third feature is the fifth value, the fourth value is used to represent the total dimension range of the target data when the value of the first feature is the first value, and the fifth value is used to represent the value of the first feature. When the value is the second value, the dimension range corresponding to each dimension; the processor is specifically used for: processing the value of the second feature in at least one input information based on the constraint condition to obtain the first target value, the first target value It is used to reflect the target range of the rank in the target data; the value of the first feature and the first target value in at least one input information are processed based on the constraints to obtain a second target value, and the second target value is used to reflect the rank. target value; process the value of the third feature and the second target value in at least one input information based on the constraint condition to obtain a third target value, and the third target value is used to reflect the target dimension range of the target data; based on the first At least one output information is obtained from the target value, the second target value and the third target value.

In a specific embodiment, the processor is further configured to: process at least one input information based on a constraint condition, and after obtaining at least one output information, determine a prediction of the corresponding output information based on the first target value and the third target value Allocate a memory range.

In a specific implementation manner, the processor is specifically configured to: select a target template based on the first target value, and the target template is used to optimize the target operator.

In a specific embodiment, the processor is specifically configured to: obtain a target arrangement feature of the target data based on any one of the at least one output information, wherein the target arrangement feature includes a first feature, a second feature and The third feature, the value of the first feature in the target arrangement feature is the second target value, the value of the second feature in the target arrangement feature is the first target value, and the value of the third feature in the target arrangement feature is the value third target value.

In a specific embodiment, the first value is a symbol or a numerical value used to indicate that the rank is unknown.

Referring to FIG. 6 , it is a schematic structural diagram of an execution device provided by an embodiment of the present application. As shown in FIG. 6 , the execution device may include a processor 601 , a memory 602 , a communication bus 603 , and a communication interface 604 , and the artificial intelligence processor 605 is connected to the memory 602 and the communication interface 604 through the communication bus 603 .

The processor 601 may be a central processing unit (CPU), and the processor 601 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs) ), off-the-shelf programmable gate array (field-programmable gate array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 601 may be a microprocessor or the processor 601 may be any conventional processor or the like.

The processor 601 may also be an integrated circuit chip, which has signal processing capability. In the implementation process, each step of the data processing method of the present application may be completed by an integrated logic circuit of hardware in the processor 601 or instructions in the form of software.

The memory 602 may be read-only memory (ROM), random access memory (RAM), or other memory. In this embodiment of the present application, the memory 602 is used for storing data and various software programs.

Optionally, in this embodiment of the present application, the memory 602 may include a physical device for storing information, and usually the information is digitized and then stored in a medium using electrical, magnetic, or optical methods. The memory of this embodiment may further include: devices that use electrical energy to store information, such as RAM, ROM, etc.; devices that use magnetic energy to store information, such as hard disks, floppy disks, magnetic tapes, magnetic core memory, magnetic bubble memory, U disk; use A device that stores information optically, such as a CD or DVD. Of course, there are other ways of memory, such as quantum memory, graphene memory, and so on.

The communication interface 604 uses transceiving means such as, but not limited to, a transceiver to enable communication between the executing device and other devices or a communication network. For example, at least one input message or the like may be received via the communication interface 604 .

Optionally, the execution device may further include at least one artificial intelligence processor 605 .

The artificial intelligence processor 605 can be mounted on the main CPU (host CPU) as a co-processor, and the main CPU assigns tasks to it. In practical applications, the artificial intelligence processor 605 may implement one or more operations. For example, taking a neural network processor (network processing unit, NPU) NPU as an example, the core part of the NPU is an arithmetic circuit, and the controller controls the arithmetic circuit to extract the matrix data in the memory 602 and perform multiplication and addition operations.

Optionally, the artificial intelligence processor 605 may include 8 clusters (clusters), and each cluster includes 4 artificial intelligence processor cores.

Optionally, the artificial intelligence processor 605 may be an artificial intelligence processor with a reconfigurable architecture. Here, the reconfigurable architecture refers to that if an artificial intelligence processor can utilize reusable hardware resources and flexibly change its own architecture according to different application requirements, in order to provide corresponding requirements for each specific application If it matches the architecture, the artificial intelligence processor is called a reconfigurable computing system, and its architecture is called a reconfigurable architecture.

It should be understood that the execution device is only an example provided by the embodiments of the present application, and the execution device may have more or less components than those shown, may combine two or more components, or may have a combination of components Different configurations are implemented.

The above description mainly describes the data processing device provided in the embodiments of the present application from the perspective of entity functions. From the perspective of functional units, the present application may divide the data processing device into functional units according to the above method embodiments. For example, each functional unit may be divided corresponding to each function, or two or more functions may be integrated into one function. in the unit. The above-mentioned integrated functional units may be implemented in the form of hardware, and may also be implemented in the form of software functional units.

For example, in the case of dividing each functional unit in an integrated manner, FIG. 7 shows a schematic structural diagram of a data processing device provided by an embodiment of the present application. As shown in FIG. 7, an embodiment of the data processing device of the present application may include:

The programming interface module 701 is used to obtain at least one input information, wherein each input information is used to represent the arrangement feature of the target data, the arrangement feature includes a first feature, and the value of the first feature is a first value or a first value. The value of a feature is a second value, the first value is used to represent the rank of the unknown target data, the second value is used to represent the rank of the known target data and the dimension value corresponding to at least one dimension is unknown;

The programming interface module 701 is used to obtain the constraint condition corresponding to the target operator, and the constraint condition is used to characterize the execution logic of the target operator;

The processing module 702 is configured to process at least one input information based on the constraint condition to obtain at least one output information, each output information is used to characterize the target arrangement feature of the target data, and the value of the target arrangement feature includes the first feature target value.

In some embodiments, the arrangement feature further includes a second feature and a third feature, the second feature is a third value, the third value is used to represent a rank range, and the third feature is a fourth value Or the value of the third feature is the fifth value, the fourth value is used to represent the total dimension range of the target data when the value of the first feature is the first value, and the fifth value is used to represent the value of the first feature. When the second value is the dimension range corresponding to each dimension; the processing module 702 is specifically used for:

The value of the second feature in at least one input information is processed according to the constraint condition to obtain a first target value, and the first target value is used to reflect the target range of the rank in the target data;

The value of the first feature and the first target value in the at least one input information are processed according to the constraint condition to obtain the second target value, and the second target value is used to reflect the target value of the rank;

The value of the third feature and the second target value in the at least one input information are processed according to the constraint condition to obtain the third target value, and the third target value is used for the target dimension range of the range target data;

At least one output information is obtained according to the first target value, the second target value and the third target value.

In some embodiments, the processing module 702 is further specifically configured to: after the at least one piece of information is processed based on the constraint condition to obtain at least one piece of output information, determine the pre-allocation of the corresponding output information based on the first target value and the third target value memory range.

In some embodiments, the processing module 702 is configured to select a target template according to the first target value, and the target template is used to optimize the target operator.

In other embodiments, the processing module 702 is further configured to obtain a target arrangement feature of the target data based on any one of the at least one output information, wherein the target arrangement feature includes a first feature, a second feature, and a first feature. Three features, the value of the first feature in the target arrangement feature is the second target value, the value of the second feature in the target arrangement feature is the first target value, and the value of the third feature in the target arrangement feature is the first target value Three target values.

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

A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the procedures or functions according to the embodiments of the present application are generated in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable device. Computer instructions may be stored on 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, computer, server, or data center over a wire (e.g. coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.) to another website site, computer, server, or data center. The computer-readable storage medium can be any available medium that can be stored by a computer or a data storage device such as a server, a data center, etc. that includes one or more available media integrated. Useful media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVD), or semiconductor media (eg, Solid State Disk (SSD)), among others.

Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium can include: ROM, RAM, magnetic disk or optical disk, etc.

The data processing method and related equipment provided by the embodiments of the present application have been described in detail above. The principles and implementations of the present application are described with specific examples in this article. The method of the application and its core idea; meanwhile, for those skilled in the art, according to the idea of the application, there will be changes in the specific embodiments and application scope. Application restrictions.

Claims

A method for data processing, comprising:

Acquire at least one input information, wherein each of the input information is used to represent an arrangement feature of the target data, the arrangement feature includes a first feature, and the value of the first feature is a first value or the The value of the first feature is a second value, the first value is used to represent the unknown rank of the target data, the second value is used to represent the known rank of the target data and at least one dimension is unknown the corresponding dimension value;

Obtaining constraints corresponding to the target operator, where the constraints are used to characterize the execution logic of the target operator;

The at least one input information is processed based on the constraint condition to obtain at least one output information, each of the output information is used to represent the target arrangement feature of the target data, and the value of the target arrangement feature A target value for the first feature is included.
The method according to claim 1, wherein the arrangement feature further comprises a second feature and a third feature, the value of the second feature is a third value, and the third value is used to represent the The range of the rank, the value of the third feature is the fourth value or the value of the third feature is the fifth value, and the fourth value is used to indicate that the value of the first feature is the first value value, the total dimension range of the target data, and the fifth value is used to indicate the dimension range corresponding to each dimension when the value of the first feature is the second value;

Obtain at least one output information for the at least one input information based on the constraint condition, including:

processing the value of the second feature in the at least one input information based on the constraint condition to obtain a first target value, where the first target value is used to reflect the target range of ranks in the target data;

The value of the first feature in the at least one input information and the first target value are processed based on the constraint condition to obtain a second target value, where the second target value is used to reflect the target of the rank value;

Based on the constraint condition, the value of the third feature in the at least one input information and the second target value are processed to obtain a third target value, and the third target value is used to reflect the target data. target dimension range;

At least one output information is obtained based on the first target value, the second target value and the third target value.
The method according to claim 2, wherein after processing the at least one input information based on the constraint condition to obtain at least one output information, the method further comprises:

The pre-allocated memory range of the corresponding output information is determined based on the first target value and the third target value.
The processing method according to claim 2 or 3, wherein the processing method further comprises:

A target template is selected based on the first target value, and the target template is used to optimize the target operator.
The processing method according to any one of claims 2-4, wherein the processing method further comprises:

A target arrangement feature of the target data is obtained based on any one of the at least one output information, wherein the target arrangement feature includes the first feature, the second feature, and the third feature, and the target The value of the first feature in the arrangement feature is the second target value, the value of the second feature in the target arrangement feature is the first target value, and the value of the third feature in the target arrangement feature is The value is the third target value.
The method according to any one of claims 1-5, wherein the first value is a symbol or a numerical value used to indicate that the rank is unknown.
A data processing device, comprising:

A programming interface module for acquiring at least one input information, wherein each of the input information is used to represent an arrangement feature of the target data, the arrangement feature includes a first feature, and the value of the first feature is The first value or the value of the first feature is a second value, where the first value is used to represent the rank of the unknown target data, and the second value is used to represent the known rank of the target data. rank and the dimension value corresponding to at least one dimension is unknown;

The programming interface module is used to obtain constraints corresponding to the target operator, where the constraints are used to characterize the execution logic of the target operator;

A processing module, configured to process the at least one input information based on the constraint condition to obtain at least one output information, each of the output information is used to characterize the target arrangement feature of the target data, the target arrangement The value of the cloth feature includes the target value of the first feature.
The data processing device according to claim 7, wherein the arrangement feature further comprises a second feature and a third feature, and the value of the second feature is a third value, and the third value is used for Indicates the range of the rank, the value of the third feature is the fourth value or the value of the third feature is the fifth value, and the fourth value is used to indicate that the value of the first feature is the value When the first value is the total dimension range of the target data, the fifth value is used to indicate the dimension range corresponding to each dimension when the value of the first feature is the second value;

The processing module is used for:

The value of the second feature in the at least one input information is processed according to the constraint condition to obtain a first target value, where the first target value is used to reflect the target range of ranks in the target data;

The value of the first feature in the at least one input information and the first target value are processed according to the constraint condition to obtain a second target value, where the second target value is used to reflect the target of the rank value;

The value of the third feature in the at least one input information and the second target value are processed according to the constraint condition to obtain a third target value, and the third target value is used to range the target data. target dimension range;

At least one output information is obtained according to the first target value, the second target value and the third target value.
The data processing device according to claim 8, wherein the processing module is further configured to process the at least one piece of information based on the constraint condition to obtain at least one output piece of information, based on the first The target value and the third target value determine a pre-allocated memory range corresponding to the output information.
The data processing device according to claim 8 or 9, wherein the processing module is configured to select a target template according to the first target value, and the target template is used to optimize the target operator.
The data processing device according to any one of claims 8-10, wherein the processing module is further configured to obtain a target arrangement of the target data based on any one of the at least one output information. distribution feature, wherein the target distribution feature includes the first feature, the second feature and the third feature, the value of the first feature in the target distribution feature is the second target value, the target The value of the second feature in the arrangement features is the first target value, and the value of the third feature in the target arrangement feature is the third target value.
The data processing device according to any one of claims 8-11, wherein the first value is a symbol or a numerical value used to represent an unknown rank.
A data processing device, comprising:

memory for storing computer-readable instructions;

Also included, a processor coupled to the memory for executing computer readable instructions in the memory to perform the method as described in any one of claims 1-6.
A computer-readable storage medium, wherein the instructions, when executed on a computer device, cause the computer device to perform the method described in any one of claims 1 to 6.
A computer program product which, when run on a computer, enables the computer to perform a method as described in any one of claims 1 to 6.