A data processing method, device, equipment and readable storage medium

Abstract

The application discloses a data processing method, device and equipment and a readable storage medium. The method comprises the following steps: an in-memory computing chip acquires target data to be calculated; a digital signal corresponding to the target data is modulated and converted to obtain a corresponding carrier signal, and a classification label of the carrier signal is determined; a plurality of intelligent agents corresponding to the classification label are created; an on-chip total routing table of the in-memory computing chip is maintained based on value-based reinforcement learning and an adaptive Kalman filter; and the operation of the plurality of intelligent agents in the in-memory computing chip is guided by using the on-chip total routing table, so as to complete the calculation of the target data. In the application, the storage and calculation of data are performed in the in-memory computing chip, thereby eliminating the data moving operation between storage and calculation, and in combination with the label-based processing mode, the data calculation efficiency can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application, in particular to a data processing method and device, equipment and readable storage medium. BACKGROUND

[0002] With the increasing amount of data and the increase of computing, the computer based on traditional von Neumann architecture (including its server cluster): has been unable to meet the demand of computing. Because the traditional computer based on von Neumann architecture and its server cluster, whether based on CPU or based on GPU, its computing unit and storage unit are designed separately, that is, storage separation, data needs to be moved frequently between computing unit and storage unit, and the data moving time may be hundreds of times of the computing time: in this process, about 60%-90% of power consumption may be caused, resulting in the decline of computing performance (efficiency).

[0003] To sum up, how to effectively solve the data computing data handling and other problems is a technical problem that the technical personnel in the field urgently need to solve. SUMMARY

[0004] The purpose of the present application is to provide a data processing method, device, equipment and readable storage medium, which can break the limitation of storage wall to power wall, can save the data moving back and forth between storage and computing in the data computing process, and can greatly improve the computing speed and computing amount of the storage and computing integrated chip.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] A data processing method, comprising:

[0007] The storage and computing integrated chip obtains target data to be calculated;

[0008] The target data corresponding digital signal is modulated and converted to obtain the corresponding carrier signal, and the classification label of the carrier signal is determined;

[0009] The classification label corresponding multi-agent is created;

[0010] Based on the value of reinforcement learning and adaptive Kalman filter, the on-chip total routing table of the storage and computing integrated chip is maintained;

[0011] Using the on-chip total routing table, the operation of the multi-agent in the storage and computing integrated chip is guided to complete the calculation of the target data.

[0012] Preferably, the value-based reinforcement learning and adaptive Kalman filter maintain the on-chip total routing table of the storage and computing integrated chip, comprising:

[0013] initializing parameters of the value-based reinforcement learning, and maintaining the on-chip total routing table;

[0014] creating the adaptive Kalman filter with multiple fading factors, and recording the signal quantity by using the adaptive Kalman filter;

[0015] finding a new path of the multi-agent by using the value-based reinforcement learning;

[0016] judging the training progress of the value-based reinforcement learning by using the signal quantity;

[0017] when the training progress meets the reporting condition, stopping finding the new path of the multi-agent, and recording the current found new path into the on-chip total routing table.

[0018] Preferably, when the training progress meets the reporting condition, stopping finding the new path of the multi-agent includes:

[0019] when the training starts, giving the multi-agent a part of the stage reward;

[0020] when the total reward rate reaches the specified amount of the reporting condition, stopping finding the new path of the multi-agent.

[0021] Preferably, recording the signal quantity by using the adaptive Kalman filter includes:

[0022] recording the signal quantity at each time of the value-based reinforcement learning process by using the adaptive Kalman filter.

[0023] Preferably, judging the training progress of the value-based reinforcement learning by using the signal quantity includes:

[0024] judging whether the parameters will continue to generate signals at each training round number by using the signal quantity;

[0025] obtaining network path parameters based on the judgment result;

[0026] determining the training progress of the value-based reinforcement learning by using the network path parameters.

[0027] Preferably, determining the classification label of the carrier signal includes:

[0028] classifying the carrier signal by using a classification algorithm to obtain the classification label.

[0029] Preferably, determining the classification label of the carrier signal includes:

[0030] Input the carrier signal into the trained label comparison network for recognition to obtain the classification label;

[0031] The training process of the label comparison network comprises:

[0032] Freeze the first 15 layers of the convolutional neural network VGG-16, and set a signal classifier in the 16th layer of the convolutional neural network VGG-16 to obtain an initial network;

[0033] Input the training sample into the initial network for training;

[0034] After training, the label comparison network is obtained.

[0035] A data processing apparatus applied to a memory-computing integrated chip, comprising:

[0036] A data receiving module configured to obtain target data to be calculated;

[0037] A data conversion module configured to modulate and convert a digital signal corresponding to the target data to obtain a corresponding carrier signal and determine a classification label of the carrier signal;

[0038] A multi-agent management module configured to create a multi-agent corresponding to the classification label;

[0039] An on-chip routing determination module configured to maintain an on-chip total routing table of the memory-computing integrated chip based on value-based reinforcement learning and an adaptive Kalman filter;

[0040] A data computing module configured to guide the operation of the multi-agent in the memory-computing integrated chip by using the on-chip total routing table to complete the calculation of the target data.

[0041] An electronic device, comprising:

[0042] A memory configured to store a computer program;

[0043] A processor configured to implement the steps of the above data processing method when executing the computer program.

[0044] A readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the above data processing method.

[0045] The method provided in the embodiment of the application is applied to obtain target data to be calculated by the storage-computing integrated chip; a digital signal corresponding to the target data is modulated and converted to obtain a corresponding carrier signal, and a classification label of the carrier signal is determined; a multi-agent corresponding to the classification label is created; a value-based reinforcement learning and an adaptive Kalman filter are used to maintain an on-chip total routing table of the storage-computing integrated chip; and the on-chip total routing table is used to guide the operation of the multi-agent in the storage-computing integrated chip to complete the calculation of the target data.

[0046] In the application, after the storage-computing integrated chip receives the target data to be calculated, the digital signal corresponding to the target data is first converted into a carrier signal, and then the classification label of the carrier signal is determined. In the subsequent calculation process, the data of the traditional computer system is no longer involved in the transmission and calculation, but is processed based on the label, which can greatly reduce the calculation amount. After obtaining the classification label, the multi-agent corresponding to the classification label is created, the value-based reinforcement learning and the adaptive Kalman filter are used to maintain the on-chip total routing table of the storage-computing integrated chip, and the on-chip total routing table is used to guide the operation of the multi-agent in the storage-computing integrated chip to complete the calculation of the target data. As can be seen, in the application, the storage and calculation of data are performed in the storage-computing integrated chip, thereby eliminating the data migration operation between storage and calculation, and in combination with the label-based processing mode, the data calculation efficiency can be effectively improved.

[0047] Correspondingly, the application also provides a data processing apparatus, device and readable storage medium corresponding to the above-mentioned data processing method, which have the above-mentioned technical effects, and details are not described herein. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the application or the related art, the following will briefly introduce the drawings needed to be used in the embodiments or the related art description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0049] Figure 1 The embodiment of the application is a data processing method implementation flowchart;

[0050] Figure 2 The embodiment of the application is a signal conversion schematic diagram;

[0051] Figure 3 The embodiment of the application is a label determination flowchart;

[0052] Figure 4 The embodiment of the application is a specific implementation schematic diagram of a data processing method;

[0053] Figure 5 FIG. 1 is a structural schematic diagram of a data processing device according to an embodiment of the present application;

[0054] Figure 6 FIG. 1 is a structural schematic diagram of a data processing device according to an embodiment of the present application;

[0055] Figure 7 FIG. 1 is a structural schematic diagram of a data processing device according to an embodiment of the present application; DETAILED DESCRIPTION

[0056] In order to make the personnel in the technical field better understand the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person of ordinary skill in the art without making creative labor, belong to the scope of protection of the present application.

[0057] Please refer to Figure 1 , Figure 1 FIG. 1 is a flowchart of a data processing method according to an embodiment of the present application, which can be applied to a memory-compute integrated chip. The memory-compute integrated chip, i.e., the memory-compute integrated chip on which storage and computation are completed (storage and computation are both completed in the memory), uses the memory end to compute data (no longer separates storage and computation), thereby breaking the limitation of the storage wall and the power wall, and improving the data computing efficiency.

[0058] The method comprises the following steps:

[0059] S101, obtaining target data to be computed.

[0060] In the embodiment, the target data itself is not limited. Specifically, the target data can be obtained by an external loading mode.

[0061] S102, modulating and converting a digital signal corresponding to the target data to obtain a corresponding carrier signal, and determining a classification label of the carrier signal.

[0062] The target data to be computed is loaded from the outside to the storage unit (Memory Unit) with a specified serial number in the memory-compute integrated chip, and then the digital signal loaded into the memory-compute integrated chip is modulated and converted accordingly. Since the carrier transmission has the advantages of fast transmission rate, small transmission bandwidth (large amount of signals passing per unit time), the digital signal can be converted into a carrier signal.

[0063] Specifically, the modulation mode is: modulation into a carrier signal (generally a sinusoidal signal with equal amplitude) for transmission, at which time the corresponding input original data carrier signal waveform diagram can be obtained.

[0064] Please refer to Figure 2 After obtaining the carrier signal, the classification label (hereinafter referred to as label) of the carrier waveform can be determined according to the carrier signal waveform diagram. For example, the carrier signal is classified according to the phase and labeled.

[0065] Specifically, the label of the carrier signal can be determined by classifying it based on the specific parameters or shape of its signal waveform.

[0066] In a specific embodiment of the present application, the classification label of the carrier signal can be obtained by using a classification algorithm to classify the carrier signal. Specifically, the signal class can be labeled, i.e. labeled, using the K-means classification algorithm (of course, other classification algorithms can also be used) according to the carrier signal waveform diagram.

[0067] For example, the source data * (multiplication symbol): 01A (signal class label); source data / (division symbol): 01B.

[0068] In another specific embodiment of the present application, the classification label of the carrier signal can also be determined by means of an automatic classification model. Please refer to Figure 3 The specific implementation includes:

[0069] The carrier signal is input into the trained label matching network for recognition to obtain the classification label;

[0070] The training process of the label matching network includes:

[0071] The first 15 layers of the convolutional neural network VGG-16 are frozen, and a signal classifier is set in the 16th layer of the convolutional neural network VGG-16 to obtain an initial network;

[0072] The training sample is input into the initial network for training;

[0073] After training, the label matching network is obtained.

[0074] For ease of description, the above model recognition process and model training process will be described together.

[0075] Specifically, a feature-based transfer learning (an algorithm in deep learning) algorithm can be embedded, which functions to replace actual real data with labels in subsequent data addressing and data calculation processes, thereby increasing speed.

[0076] After obtaining the label of the carrier signal waveform graph (Label: Carrier Signal), a classic convolutional neural network VGG-16 (a total of 16 layers) can be introduced as an initial network of the label comparison network of the feature-based transfer learning. Then, the network of all layers of the VGG-16 network is retrained, and then the first 15 layers of the network are frozen, and a classifier is defined at the 16th layer of the network. During the definition process, some features of the carrier signal waveform graph, such as wave peak, wave trough, frequency, and the like, can be added. The initial network is retrained to obtain a label comparison network.

[0077] That is, the label (Label: Carrier Signal) of the carrier signal waveform graph can be used as input to identify its category (i.e., which label it belongs to). Test & verification: input any data to be identified (to be calculated later), and obtain its category, i.e., the label.

[0078] After determining the classification label, in subsequent all calculations, the data and the carrier signal are not directly calculated, and only the label is calculated: because the label only occupies a small amount of memory space in the computer, the movement and calculation of the label only consume a small amount of computing resources.

[0079] S103, create a multi-agent corresponding to the classification label.

[0080] Multi-agent, namely Actor. Specifically, the label of the carrier signal waveform graph can be set as the Actor.

[0081] Specifically, how to create a multi-agent can be specifically referred to the related definition of the multi-agent, which will not be described here.

[0082] S104, based on value-based reinforcement learning and adaptive Kalman filter, maintain the on-chip total routing table of the storage and calculation integrated chip.

[0083] Under the traditional computer architecture, in the process of addressing data, many algorithms, such as FIFO algorithm: need to wait for the release of computing resources in the queue before they can get the resources; if the previous storage or computing resources are not released, they will be suspended. In this embodiment, the introduction of value-based reinforcement learning is mainly to change from passive to active, and actively seek idle resources.

[0084] At the same time of the initialization of the RL model (the beginning of the RL activation), the signal quantity of each parameter (variable) of the model passing through the network at each time is recorded. In this embodiment, a set of adaptive Kalman filter schemes with multiple fading factors are created to separate the real network signal (the above-mentioned variable participating in the training) and the network noise. The network noise referred to herein includes: the instantaneous traffic passing through the network instantaneously, network abnormal data and other index parameters.

[0085] In a specific embodiment of the present application, please refer to Figure 4 , the value-based reinforcement learning and the adaptive Kalman filter, the on-chip total routing table of the storage-computing integrated chip, including:

[0086] Step one, initialize the parameters of the value-based reinforcement learning, and maintain the on-chip total routing table.

[0087] Step two, create an adaptive Kalman filter with multiple fading factors, and record the signal quantity by using the adaptive Kalman filter.

[0088] Among them, recording the signal quantity by using the adaptive Kalman filter includes: recording the signal quantity at each time in the value-based reinforcement learning process by using the adaptive Kalman filter. Specifically, the adaptive Kalman filter can be used to filter signals, so that different signals pass through different regions, and different signal quantities (at different times) are recorded in the parameters of the reinforcement learning.

[0089] Step three, find a new path of the multi-agent by using the value-based reinforcement learning.

[0090] Preferably, the multi-agent can be given a part of the phased reward at the beginning of the training; when the total reward rate reaches the specified amount of the reporting condition, the search for the new path of the multi-agent is stopped. In this way, there is no need to wait for a long time to find the optimal path.

[0091] Step four, determine the training progress of the value-based reinforcement learning by using the signal quantity.

[0092] Specifically, the signal quantity can be used to determine whether the parameters will continue to generate signals at each training round; based on the determination result, the network path parameters are obtained; and the training progress of the value-based reinforcement learning is determined by using the network path parameters.

[0093] Step five, stop searching for the new path of the multi-agent when the training progress meets the reporting condition, and record the current found new path condition to the on-chip total routing table.

[0094] For ease of description, the above five steps will be described in combination.

[0095] The value-based reinforcement learning (RL) method is introduced, which changes from passive to active and actively seeks idle resources. RL initialization is set, including the following variables:

[0096] Environment (E: current in-chip memory communication channels, including idle / in-use; block address in-chip memory); signal congestion coefficient; current voltage, temperature, and other factors.

[0097] S (state), T (time), Actor (multi-agent), a (learning rate), discount rate γ, total reward (Rewards), etc. Among them, multi-agent corresponds to the label of the carrier signal waveform diagram.

[0098] With the increase of data movement (or called transportation) and calculation, a dynamic OSPF (Open Shortest Path First, i.e. on-chip total routing table) can be maintained (as time passes or with a certain task end as a time window), and the initial value in this table is given by reinforcement learning iteration.

[0099] Create multiple fading factor adaptive Kalman filters (multiple). Multi-agent has an OSPF-like table (value table, a total table, only one), and the information obtained by the filter is fed back to this table; based on this table, the multi-agent selects the next hop road (next address).

[0100] Specifically, at the same time of RL model initialization (RL activation begins), the amount of signal passing through the network at each time is recorded. At this point, a set of multiple fading factor adaptive Kalman filter schemes can be created to separate real network signals (the above-mentioned variables participating in training) and network noise, which includes: network instantaneous traffic, network abnormal data, and other indicator parameters. The specific implementation process is as follows:

[0101] Step 1, collect the communication signals in the network channel at the time of model initialization (same as setting a time point S): according to the Kalman filter, the state prediction estimate value of each multi-dimensional signal in the channel can be obtained (alternative: at the same time, draw a visualization diagram as an aid), because the traditional Kalman filter has memory, if the early data measurement is inaccurate (here mainly refers to all the data carried by the network communication at a certain time node): refers to the statistical characteristics are not accurate: it may directly lead to inaccurate state estimation of the present state and future state, and in severe cases may cause the filter to diverge.

[0102] Step 2, the filter divergence can be handled by directly modifying the prediction error covariance matrix of the next step of state estimation by choosing a suitable fading factor, thereby playing a role in suppressing filter divergence.

[0103] Step 3, because the transmission speed of the semaphore in network communication in model learning is slower than that of actual data, the semaphore at each time through reinforcement learning can be estimated. Because training reaches each layer: the transmission channel inside the on-chip chip will transmit some training parameters back and forth, and the memory or cache space occupied by these specific parameters is relatively large, the back-and-forth transmission will inevitably cause congestion. Now, by using the filter to determine whether the parameters will continue to generate signals at each training round (some signals will be dropped: disappear), it can be inferred that the model is trapped in a local optimum (referring to training stagnation at a certain position of the RL model).

[0104] At this step, a self-created multi-fading factor adaptive Kalman filter is embedded in the architecture of value-based reinforcement learning to filter different semaphore in network path to determine the training progress of the model. At this point, a parameter for value-based reinforcement learning is generated: network path parameter (which can be understood as the arrival path inside the memory-computing integrated chip), which can be used as an environmental parameter (Environment-Kalman, E_K) in the parameters of reinforcement learning.

[0105] At the beginning of training, a non-full-range reward value can be set, for example, finding an optimal path, giving the agent a partial reward within a certain period. When the total reward reaches a certain amount (for example, the total full-range reward value is 100 points, and if it reaches 80 points, learning can be stopped), the new path is updated to the OSPF routing table: for the data participating in movement and calculation in the future (at this step, the input data has been converted into labels, so here refers to labels).

[0106] After migrating learning + reinforcement learning (Kalman filter) as described above, a total routing table (dynamic update) inside the memory-computing integrated chip, i.e., on-chip total routing table, can be obtained.

[0107] S105, using the on-chip total routing table, guiding the operation of multiple agents in the memory-computing integrated chip to complete the calculation of target data.

[0108] All data transfers, data calculations, and other data operations will be automatically linked to this on-chip total routing table and initialized at initialization to guide the operation through the on-chip chip.

[0109] It can be seen that in the embodiment, the storage and calculation of data are always carried out in the chip, breaking the storage and calculation barrier of the traditional Von Neumann architecture, and effectively improving the data calculation efficiency.

[0110] By applying the method provided in the embodiment of the application, the memory-compute chip obtains target data to be calculated; a digital signal corresponding to the target data is modulated and converted to obtain a corresponding carrier signal, and a classification label of the carrier signal is determined; a multi-agent corresponding to the classification label is created; a value-based reinforcement learning and an adaptive Kalman filter are used to maintain an on-chip total routing table of the memory-compute chip; and the multi-agent is guided to operate in the memory-compute chip by using the on-chip total routing table, so as to complete the calculation of the target data.

[0111] In the application, after the memory-compute chip receives target data to be calculated, a digital signal corresponding to the target data is first converted into a carrier signal, and then a classification label of the carrier signal is determined. In the subsequent calculation process, the data of the traditional computer system is no longer involved in transmission and calculation, but is processed based on the label, which can greatly reduce the calculation amount. After the classification label is obtained, a multi-agent corresponding to the classification label is created, a value-based reinforcement learning and an adaptive Kalman filter are used to maintain an on-chip total routing table of the memory-compute chip, and the multi-agent is guided to operate in the memory-compute chip by using the on-chip total routing table, so as to complete the calculation of the target data. It can be seen that in the application, the storage and calculation of data are carried out in the memory-compute chip, thereby eliminating the data migration operation between storage and calculation, and in combination with the label-based processing mode, the data calculation efficiency can be effectively improved.

[0112] Corresponding to the above method embodiment, the embodiment of the application further provides a data processing apparatus. The data processing apparatus described below can be mutually referred to with the data processing method described above.

[0113] Referring to Figure 5 The apparatus is applied to a memory-compute chip and includes the following modules.

[0114] The data receiving module 101 is configured to obtain target data to be calculated.

[0115] The data conversion module 102 is configured to modulate and convert a digital signal corresponding to the target data to obtain a corresponding carrier signal, and determine a classification label of the carrier signal.

[0116] The multi-agent management module 103 is configured to create a multi-agent corresponding to the classification label.

[0117] The on-chip routing determination module 104 is configured to use a value-based reinforcement learning and an adaptive Kalman filter to maintain an on-chip total routing table of the memory-compute chip.

[0118] The data computing module 105 is configured to guide the operation of the plurality of agents in the in-memory computing chip by using the on-chip global routing table, so as to complete the computation of the target data.

[0119] By using the device provided in the embodiments of the present application, the in-memory computing chip obtains target data to be computed; a digital signal corresponding to the target data is modulated and converted to obtain a corresponding carrier signal, and a classification label of the carrier signal is determined; a plurality of agents corresponding to the classification label are created; a value-based reinforcement learning and an adaptive Kalman filter are used to maintain an on-chip global routing table of the in-memory computing chip; and the operation of the plurality of agents in the in-memory computing chip is guided by using the on-chip global routing table, so as to complete the computation of the target data.

[0120] In the present application, after the in-memory computing chip receives target data to be computed, a digital signal corresponding to the target data is first converted into a carrier signal, and then a classification label of the carrier signal is determined. In the subsequent computation process, the data of the traditional computer system is no longer involved in transmission and computation, but is processed based on the label, which can greatly reduce the computation amount. After obtaining the classification label, a plurality of agents corresponding to the classification label are created, a value-based reinforcement learning and an adaptive Kalman filter are used to maintain an on-chip global routing table of the in-memory computing chip, and the operation of the plurality of agents in the in-memory computing chip is guided by using the on-chip global routing table, so as to complete the computation of the target data. As can be seen, in the present application, the storage and computation of data are performed in the in-memory computing chip, thereby eliminating the data migration operation between storage and computation, and in combination with the label-based processing mode, the data computation efficiency can be effectively improved.

[0121] In one specific embodiment of the present application, the on-chip routing determination module 104 is specifically configured to initialize parameters of the value-based reinforcement learning, and maintain the on-chip global routing table.

[0122] An adaptive Kalman filter with multiple fading factors is created, and the adaptive Kalman filter is used to record the signal amount.

[0123] The value-based reinforcement learning is used to find a new path of the plurality of agents.

[0124] The signal amount is used to determine the training progress of the value-based reinforcement learning.

[0125] When the training progress meets the reporting condition, the finding of the new path of the plurality of agents is stopped, and the currently found new path is recorded in the on-chip global routing table.

[0126] In one specific embodiment of the present application, the on-chip routing determination module 104 is specifically configured to give a part of the stage rewards to the plurality of agents at the beginning of the training.

[0127] When the total return rate reaches the specified quota of the reporting condition, the search for a new path of the multi-agent is stopped.

[0128] In an embodiment of the present application, the on-chip routing determination module 104 is specifically configured to record the signal quantity at each time in the value-based reinforcement learning process by using an adaptive Kalman filter.

[0129] In an embodiment of the present application, the on-chip routing determination module 104 is specifically configured to determine whether the parameter will continue to generate a signal at each training round by using the signal quantity.

[0130] Based on the determination result, the network path parameter is obtained.

[0131] The training progress of the value-based reinforcement learning is determined by using the network path parameter.

[0132] In an embodiment of the present application, the data conversion module 102 is specifically configured to classify the carrier signal by using a classification algorithm to obtain a classification label.

[0133] In an embodiment of the present application, the data conversion module 102 is specifically configured to input the carrier signal into the trained label comparison network for identification to obtain a classification label.

[0134] The training process of the label comparison network includes:

[0135] The first 15 layers of the convolutional neural network VGG-16 are frozen, and a signal classifier is set in the 16th layer of the convolutional neural network VGG-16 to obtain an initial network.

[0136] The training sample is input into the initial network for training.

[0137] After the training is completed, the label comparison network is obtained.

[0138] Corresponding to the above method embodiment, the present embodiment also provides an electronic device. The electronic device described below can be mutually corresponding and referred to with the data processing method described above.

[0139] Referring to Figure 6 As shown in the figure, the electronic device includes:

[0140] The memory 332 is configured to store a computer program.

[0141] The processor 322 is configured to execute the computer program to implement the steps of the data processing method of the above method embodiment.

[0142] Specifically, please refer to Figure 7 , Figure 7A specific structural diagram of an electronic device is provided for the embodiment. The electronic device can have great differences due to different configurations or performances, and can include one or more central processing units (CPUs) 322 (for example, one or more processors) and a memory 332 storing one or more computer applications 342 or data 344. The memory 332 can be temporary storage or persistent storage. The programs stored in the memory 332 can include one or more modules (not shown in the figure), and each module can include a series of instruction operations in the data processing device. Further, the central processing unit 322 can be configured to communicate with the memory 332 and execute the series of instruction operations in the memory 332 on the electronic device 301.

[0143] The electronic device 301 can also include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input / output interfaces 358, and / or one or more operating systems 341.

[0144] The steps in the data processing method described above can be implemented by the structure of the electronic device.

[0145] Corresponding to the above method embodiments, the embodiments of the present application also provide a readable storage medium. The readable storage medium described below can be correspondingly referred to the data processing method described above.

[0146] A readable storage medium, the readable storage medium storing a computer program, the computer program being executed by a processor to implement the steps of the data processing method of the above method embodiments.

[0147] The readable storage medium can be a Flash, a DDRAM, a SDRAM, a random access memory (RAM), or various readable storage media that can store program codes.

[0148] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the device disclosed in the embodiments, the description is relatively simple because it corresponds to the method disclosed in the embodiments. The relevant parts can be referred to the description of the method part.

[0149] Those skilled in the art can further understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in a general manner. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0150] The steps of the method or algorithm described in combination with the embodiments disclosed herein can be directly implemented by hardware, a software module executed by a processor, or a combination of both. The software module can be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0151] Finally, it should be noted that, in this document, terms such as including, including or any other variants are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such a process, method, article or device.

[0152] The principles and implementation manners of the present application are described in this document by applying specific examples, and the above description of the examples is only for the purpose of helping to understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in specific implementation manners and application ranges; in summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A data processing method, characterized by, The method comprises the steps of: an in-memory computing chip obtains target data to be calculated; modulating and converting a digital signal corresponding to the target data to obtain a corresponding carrier signal, and determining a classification label of the carrier signal; creating a plurality of agents corresponding to the classification label; maintaining an on-chip total routing table of the in-memory computing chip based on value-based reinforcement learning and an adaptive Kalman filter; using the on-chip total routing table to guide the operation of the plurality of agents in the in-memory computing chip to complete the calculation of the target data; wherein the maintenance of the on-chip total routing table of the in-memory computing chip based on value-based reinforcement learning and the adaptive Kalman filter comprises: initializing parameters of the value-based reinforcement learning and maintaining the on-chip total routing table; creating a plurality of fading factors of the adaptive Kalman filter and recording signal quantities using the adaptive Kalman filter; finding a new path of the plurality of agents using the value-based reinforcement learning; judging the training progress of the value-based reinforcement learning using the signal quantities; when the training progress meets a reporting condition, stopping finding the new path of the plurality of agents and updating the currently found new path to the on-chip total routing table; wherein recording signal quantities using the adaptive Kalman filter comprises: recording signal quantities at each time during the value-based reinforcement learning process using the adaptive Kalman filter; wherein judging the training progress of the value-based reinforcement learning using the signal quantities comprises: judging whether parameters will continue to generate signals at each training round using the signal quantities; obtaining network path parameters based on the judgment result; determining the training progress of the value-based reinforcement learning using the network path parameters.

2. The data processing method according to claim 1, characterized in that, When the training progress meets the reporting condition, stopping finding the new path of the plurality of agents comprises: providing the plurality of agents with part of the phased reward reward at the beginning of training; stopping finding the new path of the plurality of agents when the total reward rate reaches the specified amount of the reporting condition.

3. The data processing method according to claim 1 or 2, characterized in that, Determining the classification label of the carrier signal comprises: classifying the carrier signal using a classification algorithm to obtain the classification label.

4. The data processing method according to claim 1 or 2, characterized by, Determining the classification label of the carrier signal comprises: inputting the carrier signal into a trained label comparison network for identification to obtain the classification label; wherein the training process of the label comparison network comprises: freezing the first 15 layers of a convolutional neural network VGG-16, and setting a signal classifier in the 16th layer of the convolutional neural network VGG-16 to obtain an initial network; inputting training samples into the initial network for training; after training, obtaining the label comparison network.

5. A data processing apparatus, characterized by, The in-memory computing chip comprises: a data receiving module for obtaining target data to be calculated; a data conversion module for modulating and converting a digital signal corresponding to the target data to obtain a corresponding carrier signal, and determining a classification label of the carrier signal; a plurality of agent management modules for creating a plurality of agents corresponding to the classification label; The on-chip routing determination module is configured to maintain an on-chip total routing table of the memory-compute integrated chip based on a value-based reinforcement learning and an adaptive Kalman filter, including: initializing parameters of the value-based reinforcement learning and maintaining the on-chip total routing table; creating a multi-exponential-factor adaptive Kalman filter and recording a signal amount by using the adaptive Kalman filter; finding a new path of the multi-agent by using the value-based reinforcement learning; judging a training progress of the value-based reinforcement learning by using the signal amount; when the training progress meets a reporting condition, stopping finding the new path of the multi-agent and updating the currently found new path to the on-chip total routing table; wherein recording the signal amount by using the adaptive Kalman filter includes recording the signal amount at each time in the value-based reinforcement learning process by using the adaptive Kalman filter; wherein judging the training progress of the value-based reinforcement learning by using the signal amount includes judging whether the parameters will continue to generate signals at each training round by using the signal amount; obtaining network path parameters based on a judgment result; and determining the training progress of the value-based reinforcement learning by using the network path parameters; The data computation module is configured to guide operations of the multi-agent in the memory-compute integrated chip by using the on-chip total routing table to complete computation on the target data.

6. An electronic device, comprising: The data processing method comprises the following steps: a memory configured to store a computer program; a processor configured to implement the steps of the data processing method according to any one of claims 1 to 4 when executing the computer program.

7. A readable storage medium, characterized by, The computer program is stored on the readable storage medium and is configured to implement the steps of the data processing method according to any one of claims 1 to 4 when executed by the processor.

Images