Super-resolution based summary correction method, device and equipment and storage medium

Abstract

The application relates to a summary error correction method, device and equipment based on super resolution and a storage medium, and the method comprises the following steps: constructing a summary; the summary uses the addressing and insertion of the summary, increases the function of the summary addressing, and outputs the bucket information of the summary and the addressing information of the data stream; a summary error correction model based on super resolution is constructed; the summary error correction model is used to reconstruct a single summary in a compressed representation into an original data stream element based on the super resolution reconstruction idea; the summary error correction model is pre-trained by using an offline data set; the pre-trained summary error correction model is used to reconstruct the data stream sequence of the summary in an online manner through a client-server mode, and the summary error correction model is fine-tuned. The method does not change the actually deployed summary system, the recorded summary information is reconstructed in an online manner through the super resolution model, the pre-training is performed through a real or synthetic data set, the summary error correction is supported in a distributed environment, and the query accuracy of the data stream summary is improved.

Description

Technical Field

[0001] This application relates to the field of network communication technology, and in particular to a super-resolution-based summary error correction method, apparatus, device, and storage medium. Background Technology

[0002] Sketching is a mainstream method in data stream mining and analysis. A sketch consists of an array of m (positive integer system parameters, such as 128 or 1024) buckets contiguously distributed in memory. For elements in a key-value pair data stream, the sketch does not store the keys of the elements; instead, it equips each bucket with a fixed-size counter (such as an algebraic accumulator counter) to aggregate the value ranges of different elements. Sketching has the advantages of fixed space resources and low computational overhead.

[0003] The summary technology is deployed as an independent service or a library in fields such as cloud computing, big data, edge computing network management, data stream mining and analysis, network security, and distributed training of artificial intelligence. It provides simplified computing and storage resource compression for high-dimensional or dynamic data and has the unique advantage of fast real-time response.

[0004] Elements from the data stream are aggregated into the summary using insertion functions provided by the summary. Each element determines one or more buckets using an addressing function and is then inserted into the corresponding bucket's counter. When a query is needed, the same addressing function as the insertion process is used to select the bucket in the array, and the bucket's counter is returned as the predicted value for the corresponding element. When data stream elements are inserted into the summary, the addressing function may cause collisions, i.e., multiple elements are inserted into the same bucket. Common addressing functions use hash functions, which output uniformly random bucket indices; however, collisions are unpredictable and cannot be adjusted later. Query errors caused by addressing collisions reduce the accuracy of the summary technique and affect its application value.

[0005] Existing research mainly avoids inserting elements with large differences into the same bucket by classifying the elements of the data stream: (1) Data stream clustering: clustering the data stream according to the value range and inserting data streams of the same type into the same bucket; (2) Tail separation: removing the singular values ​​at the tail of the data stream and storing them separately in a dictionary; (3) Data stream downsampling: downsampling the data stream to reduce the number of elements inserted into the summary. However, existing methods require the transformation of the data stream analysis workflow and the introduction of an additional data stream classification module to perceive the value distribution of the data stream, which cannot be applied to the already deployed summary system. Summary of the Invention

[0006] Therefore, it is necessary to provide a super-resolution-based summary error correction method, apparatus, device, and storage medium to address the aforementioned technical problems.

[0007] A super-resolution-based summary error correction method, the method comprising:

[0008] Construct a summary; the summary adopts the addressing and insertion of the summary, adds the summary addressing function, and outputs the bucket information of the summary and the addressing information of the data stream.

[0009] A summary error correction model based on super-resolution is constructed; the summary error correction model is used to reconstruct a single summary from a compressed representation into the original data stream element based on the idea of ​​super-resolution reconstruction.

[0010] The summary error correction model was pre-trained using an offline dataset.

[0011] The summary data stream sequence is reconstructed online using a pre-trained summary error correction model via a client-server approach, and the summary error correction model is fine-tuned based on the summary information from the client.

[0012] In one embodiment, the construction summary includes:

[0013] Initialize the summary data structure and address dictionary.

[0014] The summary data structure is updated online based on dynamic data stream elements, and the addressing information of the data stream elements is recorded using an addressing dictionary.

[0015] In one embodiment, the super-resolution-based summary error correction model is constructed as a super-resolution reconstruction model based on the data flow summary of a self-attention encoder; the super-resolution reconstruction model includes an input layer, a hidden layer, and an output layer.

[0016] The input layer is used to construct two features of the same dimension from an array I of individual summary information and an addressing dictionary using fully connected layers, and then concatenates the two features to obtain a concatenated feature.

[0017] The hidden layer is used to obtain encoded features based on a self-attention encoder architecture according to the splicing features.

[0018] The output layer is used to process the encoded features using a fully connected layer to obtain a sequence estimate of the data stream elements.

[0019] In one embodiment, the hidden layer includes: a summary location encoding module, a summary self-attention module, and a summary encoder.

[0020] The summary position encoding module is used to calculate the position encoding representation of the splicing feature using a sinusoidal encoding method, and accumulate it to the splicing feature to obtain the position-encoded splicing representation.

[0021] The summary self-attention module is used to calculate the self-attention based on the concatenated representation after position encoding:

[0022]

[0023] Where f represents self-attention; Z' represents the concatenated feature; d represents the dimension of the feature before concatenation in the input layer; and the superscript T represents matrix transpose.

[0024] The summary encoder is used to process data using a multi-level recursive coding layer based on self-attention, to obtain the coded features as follows:

[0025]

[0026] Where L is the level, l is the result of the current coding layer, and l+1 is the result of the next coding layer.

[0027] In one embodiment, the summary error correction model is pre-trained using an offline dataset, including:

[0028] Obtain the actual data stream sequence or synthesize the data stream sequence using statistical distribution methods, generate summary information based on the data stream sequence, and obtain the summary dataset.

[0029] The parameters of the summary error correction model are randomly initialized.

[0030] The summary error correction model is pre-trained based on the summary dataset.

[0031] Output the pre-trained error correction model to the file system.

[0032] In one embodiment, the summary data stream sequence is reconstructed online using a pre-trained summary correction model via a client-server approach, and the summary correction model is fine-tuned based on the summary information from the client, including:

[0033] The model is deployed using a client-server approach.

[0034] The client records summary and addressing information to the server, outputs a compressed file, and submits the summary information compressed file through a remote network communication protocol.

[0035] The server uses a pre-trained summary error correction model to reconstruct the summary data stream sequence online based on the compressed file containing the summary information, and then feeds back the data stream reconstruction results to the client.

[0036] Based on the summary information from the client, the summary error correction model is fine-tuned.

[0037] In one embodiment, a client-server approach is used for model deployment, including:

[0038] Online reconstruction is provided using a client-server approach. The pre-trained summary error correction model is deployed on the server using the existing model runtime environment, and no additional deployment is required on the client.

[0039] A super-resolution-based summary error correction device, the device comprising:

[0040] The summary construction module is used to construct summaries; the summaries adopt the summary addressing and insertion, add summary addressing function, and output summary bucket information and data stream addressing information.

[0041] The summary error correction model construction module is used to construct a summary error correction model based on super-resolution; the summary error correction model is used to reconstruct a single summary from a compressed representation into the original data stream element based on the super-resolution reconstruction idea.

[0042] The summary error correction model offline training module is used to pre-train the summary error correction model using an offline dataset.

[0043] The summary data stream sequence online reconstruction module is used to reconstruct the summary data stream sequence online using a pre-trained summary error correction model via a client-server approach, and to fine-tune the summary error correction model based on the summary information from the client.

[0044] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0045] Construct a summary; the summary adopts the addressing and insertion of the summary, adds the summary addressing function, and outputs the bucket information of the summary and the addressing information of the data stream.

[0046] A summary error correction model based on super-resolution is constructed; the summary error correction model is used to reconstruct a single summary from a compressed representation into the original data stream element based on the idea of ​​super-resolution reconstruction.

[0047] The summary error correction model was pre-trained using an offline dataset.

[0048] The summary data stream sequence is reconstructed online using a pre-trained summary error correction model via a client-server approach, and the summary error correction model is fine-tuned based on the summary information from the client.

[0049] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0050] Construct a summary; the summary adopts the addressing and insertion of the summary, adds the summary addressing function, and outputs the bucket information of the summary and the addressing information of the data stream.

[0051] A summary error correction model based on super-resolution is constructed; the summary error correction model is used to reconstruct a single summary from a compressed representation into the original data stream element based on the idea of ​​super-resolution reconstruction.

[0052] The summary error correction model was pre-trained using an offline dataset.

[0053] The summary data stream sequence is reconstructed online using a pre-trained summary error correction model via a client-server approach, and the summary error correction model is fine-tuned based on the summary information from the client.

[0054] The aforementioned summary correction method, apparatus, device, and storage medium based on super-resolution include the following steps: constructing a summary; utilizing summary addressing and insertion methods while adding summary addressing functionality, outputting summary bucket information and data stream addressing information; constructing a super-resolution-based summary correction model; using the summary correction model to reconstruct compressed representations of individual summaries into original data stream elements based on the super-resolution reconstruction concept; pre-training the summary correction model using an offline dataset; online reconstructing the summary data stream sequence using the pre-trained summary correction model via a client-server approach, and fine-tuning the summary correction model based on the client's summary information. This method can reconstruct recorded summary information online using a super-resolution model without modifying the actual deployed summary system, pre-training using real or synthetic datasets, supporting summary correction in a distributed environment, and improving the query accuracy of data stream summaries. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a super-resolution-based summary error correction method in one embodiment.

[0056] Figure 2 Here is a summary block diagram of the error correction model structure in another embodiment;

[0057] Figure 3 This is a structural block diagram of a super-resolution-based summary error correction device in one embodiment;

[0058] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] In one embodiment, such as Figure 1As shown, a super-resolution-based summary error correction method is provided, which includes the following steps:

[0061] Step 100: Construct a summary; the summary retains the addressing and insertion of the summary, adds the summary addressing function, and outputs the bucket information of the summary and the addressing information of the data stream.

[0062] Specifically, (1) Initialization: Initialize the summary data structure and addressing dictionary. For example, the most common count-sum summary includes an array of m buckets. Each element selects k buckets through k independent hash functions and appends the element's value to the bucket's counter field (denoted as SUM) in an algebraic summation manner.

[0063] (2) Insertion: Update the summary data structure online for dynamic data stream elements and use the addressing dictionary to record the addressing information of the data stream elements.

[0064] Step 102: Construct a super-resolution-based summary error correction model; the summary error correction model is used to reconstruct a single summary from a compressed representation into the original data stream element based on the super-resolution reconstruction idea.

[0065] Specifically, summary resolution: Super-resolution technology originates from the field of digital imaging, where the resolution of an image is related to the number of pixels in the digital image. For example, a 128*128 image contains 16384 pixels. Similarly, the resolution of a summary is defined as the number of buckets it contains; for example, a summary with 1024 buckets has a resolution of 1024.

[0066] Objective function: Let the original data stream sequence be represented as X (with n elements), and the summary representation be I (with m buckets), where m is usually less than n to achieve compressed representation of the data stream; otherwise, the storage overhead will be higher than that of the original data stream. Define the super-resolution reconstruction function f: I[m], P[m*n]→X[n], n≤m, that is, to reconstruct the original data stream sequence using the summary representation and the addressing dictionary.

[0067] By using bucket information and addressing information as input to the summarizing error correction model, the constructed super-resolution-based summarizing error correction model can be, but is not limited to, a self-attention architecture-based super-resolution model, or it can employ other existing super-resolution reconstruction models. The self-attention architecture-based super-resolution model is a super-resolution reconstruction method based on data flow summarization using a self-attention encoder.

[0068] Step 104: Pre-train the summary error correction model using an offline dataset.

[0069] Specifically, the summary error correction model is trained using real or synthetic datasets, and its parameters are optimized to improve its generalization ability.

[0070] Step 106: Reconstruct the summary data stream sequence online using the pre-trained summary error correction model via a client-server approach, and fine-tune the summary error correction model based on the summary information from the client.

[0071] The aforementioned summary correction method, apparatus, device, and storage medium based on super-resolution include the following steps: constructing a summary; utilizing summary addressing and insertion methods while adding summary addressing functionality, outputting summary bucket information and data stream addressing information; constructing a super-resolution-based summary correction model; using the summary correction model to reconstruct compressed representations of individual summaries into original data stream elements based on the super-resolution reconstruction concept; pre-training the summary correction model using an offline dataset; online reconstructing the summary data stream sequence using the pre-trained summary correction model via a client-server approach, and fine-tuning the summary correction model based on the client's summary information. This method can reconstruct recorded summary information online using a super-resolution model without modifying the actual deployed summary system, pre-training using real or synthetic datasets, supporting summary correction in a distributed environment, and improving the query accuracy of data stream summaries.

[0072] In one embodiment, step 100 includes: initializing a summary data structure and an addressing dictionary; updating the summary data structure online based on dynamic data stream elements; and using the addressing dictionary to record the addressing information of the data stream elements.

[0073] Specifically, the addressing and insertion functions of the summary are reused to update the summary structure; for element pairs of the data stream...<key,value> Bucket Information<i1,…,ik> Add a record to the addressing dictionary <key,<i1,…,ik> >Record. If a key appears multiple times in different locations in the data stream, the new bucket location information is appended to the existing key in the addressing dictionary.

[0074] In one embodiment, step 102 includes: constructing a super-resolution-based summary error correction model as a super-resolution reconstruction model of a data stream summary based on a self-attention encoder; the super-resolution reconstruction model includes an input layer, a hidden layer, and an output layer; the input layer is used to construct two features of the same dimension for an array I of individual summary information and an addressing dictionary using fully connected layers, and concatenates the two features to obtain concatenated features; the hidden layer is used to obtain encoded features based on a self-attention encoder architecture according to the concatenated features; the output layer is used to process the encoded features using fully connected layers to obtain a sequence estimate of data stream elements.

[0075] In one embodiment, the hidden layer includes: a summary position encoding module, a summary self-attention module, and a summary encoder; the summary position encoding module is used to calculate the position encoding representation of the concatenated features using a sinusoidal encoding method, and accumulate it to the concatenated features to obtain the position-encoded concatenated representation; the summary self-attention module is used to calculate the self-attention based on the position-encoded concatenated representation.

[0076]

[0077] Where f is self-attention; Z′ is the concatenated feature; d is the dimension of the feature before concatenation in the input layer; and the superscript T represents matrix transpose.

[0078] The summary encoder, which processes data using multiple levels of recursive coding layers based on self-attention, yields the following encoded features:

[0079]

[0080] Where L is the level, l is the result of the current coding layer, and l+1 is the result of the next coding layer.

[0081] Specifically, a super-resolution reconstruction model based on a self-attention encoder-based data flow summarization is used as the summarization error correction model. This model specifically includes:

[0082] (1) Input layer: The input of the model is a single summary information: containing an array I (m-dimensional) and an addressing dictionary P (m*n-dimensional);

[0083] 1) Array I Embedding: Constructing an m*d dimensional representation Y, i.e., g, using a fully connected layer. I :I→Y I ;

[0084] 2) Addressing dictionary P embedding: Constructing an m*d-dimensional representation Z, i.e., g, using a fully connected layer. P :P→Y P ;

[0085] 3) Embedded splicing: Construct a (m*(2*d)) dimensional splicing Z, i.e., g P，I Y I Y P →Z, Z=[Y I Y P ].

[0086] (2) Output layer: The output of the model is the estimated sequence of reconstructed data stream elements, X'.

[0087] 1) Output layer: Construct an n-dimensional sequence X′ using a fully connected layer, i.e., g P : Y′→X′;

[0088] 2) Data stream element sequence estimation: Output X′ as the sequence estimate of data stream elements.

[0089] (3) Hidden layer: A self-attention encoder-based architecture is adopted, with the input layer represented by a 2*d-dimensional concatenation of [Y]. I Y P ] is the input, and the output is a d-dimensional encoded representation of Y′.

[0090] 1) Summary positional encoding: To improve the distinction between different buckets, a sinusoidal encoding method is used to calculate the positional encoding representation of the concatenated representation Z, which is accumulated into the concatenated representation Z, and the positional encoded concatenated representation Z′ is output.

[0091] 2) Summary of Self-Attention: To obtain the mutual influence between different buckets, self-attention is calculated using the concatenated representation of positional encoding, i.e., calculating the mutual influence between any two buckets: f: (Z′, Z′, Z′) → Z″, that is...

[0092] 3) Summary Encoder: Constructs a multi-level recursive coding layer (the number of levels is denoted as L), and outputs the encoded result: i.e.

[0093] (4) Forward operation of the model: Given the input array I and the addressing dictionary P, the input layer, hidden layer and output layer are called recursively according to the hierarchical relationship of the model, and finally the reconstruction result of the data stream sequence is output.

[0094] In one embodiment, step 104 includes: acquiring a real data stream sequence or synthesizing a data stream sequence using a statistical distribution method; generating summary information based on the data stream sequence to obtain a summary dataset; randomly initializing the parameters of the summary error correction model; pre-training the summary error correction model based on the summary dataset; and outputting the pre-trained error correction model to the file system.

[0095] Specifically, the pre-training of the summary error correction model includes the following steps:

[0096] (1) Summary sample collection: Under the premise of conditions permitting, collect real data stream sequences, or directly synthesize data stream sequences through typical statistical distributions (such as normal distribution, long-tailed distribution), generate summary information, and output summary dataset;

[0097] (2) Pre-training of the model: The super-resolution reconstruction model is trained on the collected summary dataset. The parameters of the model are randomly initialized. Then, the standardized Adam stochastic gradient descent method is used, with the default learning rate set to 0.001 and the number of training epochs set to 100.

[0098] (3) Model export: Output the pre-trained error correction model to the file system.

[0099] In one embodiment, step 106 includes: deploying the model using a client-server approach; the client records summary and addressing information to the server, outputs a compressed file, and submits the summary information compressed file through a remote network communication protocol; the server reconstructs the summary data stream sequence online using a pre-trained summary error correction model based on the summary information compressed file, and feeds back the data stream reconstruction result to the client; and fine-tunes the summary error correction model based on the summary information from the client.

[0100] Specifically, online reconstruction of the data stream and model fine-tuning include the following steps:

[0101] (1) Model deployment: Online reconstruction is provided through a client-server approach. The server uses the existing model runtime environment (such as PyTorch runtime) to deploy the pre-trained error correction model, and the client does not need to deploy it separately.

[0102] (2) Summary collection: The client records summary and addressing information to the server, outputs a compressed file, and submits the summary information compressed file through a remote network communication protocol (the default is the HTTP network protocol);

[0103] (3) Model inference: The server-side model runtime environment receives summary information, outputs the estimated data stream sequence through the forward operation of the model, and feeds back the data stream reconstruction result through remote network communication; where the forward operation of the model refers to: given the input array I and the addressing dictionary P, the input layer, hidden layer and output layer are recursively called in sequence according to the hierarchical relationship of the model, and finally the reconstruction result of the data stream sequence is output.

[0104] (4) Model fine-tuning: Fine-tune the pre-trained model based on the client's summary information.

[0105] In one embodiment, model deployment is performed using a client-server approach, which includes: providing online reconstruction using a client-server approach, deploying a pre-trained summary error correction model on the server using the existing model runtime environment, and eliminating the need for additional deployment on the client.

[0106] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0107] In one embodiment, such as Figure 2 As shown, a super-resolution-based summarization error correction device is provided, comprising: a summarization construction module, a summarization error correction model construction module, a summarization error correction model offline training module, and a summarization data stream sequence online reconstruction module, wherein:

[0108] The summary building module is used to build summaries. The summary retains the summary addressing and insertion, and adds the summary addressing function to output the bucket information and data stream addressing information of the summary.

[0109] The summary error correction model building module is used to build a summary error correction model based on super-resolution; the summary error correction model is used to reconstruct a single summary from a compressed representation into the original data stream element based on the super-resolution reconstruction idea;

[0110] The summary error correction model offline training module is used to pre-train the summary error correction model using an offline dataset;

[0111] The summary data stream sequence online reconstruction module is used to reconstruct the summary data stream sequence online using a pre-trained summary error correction model via a client-server approach, and to fine-tune the summary error correction model based on the summary information from the client.

[0112] In one embodiment, the summary construction module is further configured to initialize the summary data structure and the addressing dictionary; update the summary data structure online according to the dynamic data flow elements; and use the addressing dictionary to record the addressing information of the data flow elements.

[0113] In one embodiment, the summary error correction model construction module is further used to construct a super-resolution reconstruction model of the data stream summary based on a self-attention encoder. The super-resolution reconstruction model includes an input layer, a hidden layer, and an output layer. The input layer is used to construct two features of the same dimension for the array I of individual summary information and the addressing dictionary respectively using a fully connected layer, and then concatenates the two features to obtain concatenated features. The hidden layer is used to obtain encoded features based on the concatenated features using a self-attention encoder-based architecture. The output layer is used to process the encoded features using a fully connected layer to obtain the sequence estimate of the data stream elements.

[0114] In one embodiment, the hidden layer of the summary error correction model construction module includes: a summary position encoding module, a summary self-attention module, and a summary encoder; the summary position encoding module is used to calculate the position encoding representation of the concatenated features using a sinusoidal encoding method, and accumulate it to the concatenated features to obtain the position-encoded concatenated representation; the summary self-attention module is used to calculate the self-attention based on the position-encoded concatenated representation.

[0115]

[0116] Where f is self-attention; Z' is the concatenated feature; d is the dimension of the feature before concatenation in the input layer; and the superscript T represents matrix transpose.

[0117] The summary encoder, which processes data using multiple levels of recursive coding layers based on self-attention, yields the following encoded features:

[0118]

[0119] Where L is the level, l is the result of the current coding layer, and l+1 is the result of the next coding layer.

[0120] In one embodiment, the summary error correction model offline training module is further used to acquire real data stream sequences or synthesize data stream sequences in a manner that combines statistical distribution, generate summary information based on the data stream sequences to obtain a summary dataset; randomly initialize the parameters of the summary error correction model; pre-train the summary error correction model based on the summary dataset; and output the pre-trained error correction model to the file system.

[0121] In one embodiment, the summary data stream sequence online reconstruction module is also used to deploy the model in a client-server manner; the client records summary and addressing information to the server, outputs a compressed file, and submits the summary information compressed file through a remote network communication protocol; the server reconstructs the summary data stream sequence online using a pre-trained summary error correction model based on the summary information compressed file, and feeds back the data stream reconstruction result to the client; the summary error correction model is fine-tuned based on the summary information from the client.

[0122] In one embodiment, the summary data stream sequence online reconstruction module is also used to provide online reconstruction in a client-server manner, deploying a pre-trained summary error correction model on the server using the existing model runtime environment, without the need for additional deployment on the client.

[0123] Specific limitations regarding the super-resolution-based summary error correction device can be found in the limitations of the super-resolution-based summary error correction method described above, and will not be repeated here. Each module in the aforementioned super-resolution-based summary error correction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0124] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a super-resolution-based summary error correction method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0125] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0126] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiment.

[0127] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0128] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0129] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A summary error correction method based on super-resolution, characterized in that, The method includes: Construct a summary; the summary adopts the addressing and insertion logic of the existing summary, adds the summary addressing function, and outputs the bucket information of the summary and the addressing information of the data stream; specifically, it includes: initializing the summary data structure and addressing dictionary; updating the summary data structure online according to the dynamic data stream elements, and using the addressing dictionary to record the addressing information of the data stream elements; A super-resolution-based summary correction model is constructed. This model reconstructs compressed summaries into original data stream elements based on the super-resolution reconstruction concept. Specifically, the super-resolution-based summary correction model is a super-resolution reconstruction model of data stream summaries based on a self-attention encoder. The super-resolution reconstruction model includes an input layer, a hidden layer, and an output layer. The input layer constructs two features of the same dimension from an array I of individual summary information and an addressing dictionary using fully connected layers, and concatenates these two features to obtain concatenated features. The hidden layer uses a self-attention encoder-based architecture to obtain encoded features based on the concatenated features. The output layer processes the encoded features using fully connected layers to obtain sequence estimates of the data stream elements. The summary error correction model was pre-trained using an offline dataset; The summary error correction model is used to reconstruct the summary data stream sequence online using a pre-trained summary error correction model via a client-server approach, and the summary error correction model is fine-tuned based on the summary information from the client. Specifically, this includes: deploying the model using a client-server approach; the client recording summary and addressing information to the server, outputting a compressed file, and submitting the summary information compressed file via a remote network communication protocol; the server reconstructing the summary data stream sequence online using the pre-trained summary error correction model based on the summary information compressed file, and feeding back the data stream reconstruction result to the client; and fine-tuning the summary error correction model based on the summary information from the client.

2. The method according to claim 1, characterized in that, The hidden layer includes: a summary location encoding module, a summary self-attention module, and a summary encoder; The summary position encoding module is used to calculate the position encoding representation of the splicing feature using a sinusoidal encoding method, and accumulate it to the splicing feature to obtain the position-encoded splicing representation; The summary self-attention module is used to calculate the self-attention based on the concatenated representation after position encoding: in, f For self-attention; For splicing features, d The superscript represents the dimension of the features before concatenation in the input layer. T Represents the transpose of a matrix; The summary encoder is used to process data using a multi-level recursive coding layer based on self-attention, to obtain the coded features as follows: in, l For the current coding layer result, l +1 represents the result of the next coding layer.

3. The method according to claim 1, characterized in that, The summary error correction model is pre-trained using an offline dataset, including: Obtain a real data stream sequence or synthesize a data stream sequence through statistical distribution, generate summary information based on the data stream sequence, and obtain a summary dataset; The parameters of the summary error correction model are randomly initialized; The summary error correction model is pre-trained based on the summary dataset; Output the pre-trained error correction model to the file system.

4. The method according to claim 1, characterized in that, The model is deployed using a client-server approach, including: Online reconstruction is provided using a client-server approach. The pre-trained summary error correction model is deployed on the server using the existing model runtime environment, and no additional deployment is required on the client.

5. A summary error correction device based on super-resolution, characterized in that, The device includes: The summary construction module is used to construct the summary; the summary adopts the summary addressing and insertion, adds the summary addressing function, and outputs the bucket information of the summary and the addressing information of the data stream; specifically, it includes: initializing the summary data structure and addressing dictionary; updating the summary data structure online according to the dynamic data stream elements, and using the addressing dictionary to record the addressing information of the data stream elements; A summary error correction model construction module is used to construct a summary error correction model based on super-resolution. This model reconstructs compressed summaries into original data stream elements based on the super-resolution reconstruction concept. Specifically, the super-resolution summary error correction model is a super-resolution reconstruction model of data stream summaries based on a self-attention encoder. The super-resolution reconstruction model includes an input layer, a hidden layer, and an output layer. The input layer constructs two features of the same dimension for each summary information array I and the addressing dictionary using fully connected layers, and concatenates these two features to obtain concatenated features. The hidden layer uses a self-attention encoder-based architecture to obtain encoded features based on the concatenated features. The output layer processes the encoded features using fully connected layers to obtain sequence estimates of the data stream elements. The summary error correction model offline training module is used to pre-train the summary error correction model using an offline dataset; The summary data stream sequence online reconstruction module is used to reconstruct the summary data stream sequence online using a pre-trained summary error correction model via a client-server approach, and to fine-tune the summary error correction model based on the summary information from the client. Specifically, it includes: deploying the model using a client-server approach; the client recording summary and addressing information to the server, outputting a compressed file, and submitting the summary information compressed file via a remote network communication protocol; the server reconstructing the summary data stream sequence online using the pre-trained summary error correction model based on the summary information compressed file, and feeding back the data stream reconstruction result to the client; and fine-tuning the summary error correction model based on the summary information from the client.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.

Images