Training method, resistance state adjusting method, image processing method and device

Abstract

The invention provides a training method, a resistance state adjusting method and an image processing method and device. The training method comprises the following steps: acquiring a simulation data set generated by a micro-magnetic simulation technology and an experiment data set obtained based on an experiment; a mixed data set is constructed according to the simulation data set and the experiment data set, the mixed data set comprises a plurality of sample pairs, and each sample pair comprises at least two mixed training samples; for each sample pair, processing the sample pair by using an initial memory model to obtain a first prediction performance parameter corresponding to each mixed training sample, and calculating a first residual error of at least two first prediction performance parameters; under the condition that the first residual error meets a residual error threshold value, testing the to-be-tested sample pair through an experiment so as to update the initial sample label, and obtaining a new sample label; and under the condition that a training set iteration termination condition is satisfied, training the initial memory model by using the updated mixed data set to obtain a magnetoresistive memory model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of magnetoresistive memory, and more particularly, to a training method, a resistance state adjustment method, an image processing method and device. BACKGROUND

[0002] Magnetic Random Access Memory (MRAM) is a non-volatile memory that stores data using the magnetoresistance effect, and records binary data through the difference in magnetization direction of a magnetic tunnel junction (MTJ).

[0003] In the process of implementing the concept of the present application, it is found that the magnetoresistive memory model constructed by the related technology has room for improvement in solving speed, and there is a large difference between its performance prediction and the performance of the real device. SUMMARY

[0004] Therefore, the present application provides a training method of a magnetoresistive memory model, a resistance state adjustment method, an image processing method, a training device of a magnetoresistive memory model, a resistance state adjustment device, an image processing device, an electronic device, a computer readable storage medium, and a computer program product.

[0005] One aspect of the present application provides a training method of a magnetoresistive memory model, comprising:

[0006] obtaining a simulation data set generated by a micromagnetic simulation technology and an experimental data set obtained based on an experiment, wherein the simulation data set and the experimental data set each include a plurality of initial training samples and an initial sample label corresponding to each initial training sample, the initial training sample represents an input parameter of a magnetic random access memory, and the initial sample label represents an energy storage performance parameter of the magnetic random access memory;

[0007] iteratively performing operations 1-3:

[0008] Operation 1: constructing a mixed data set according to the simulation data set and the experimental data set, wherein the mixed data set includes a plurality of sample pairs, and each sample pair includes at least two mixed training samples;

[0009] Operation 2: for each sample pair, processing the sample pair using an initial memory model to obtain a first predicted performance parameter corresponding to each mixed training sample, and calculating a first residual of at least two first predicted performance parameters;

[0010] Operation 3, in a case where the first residual satisfies a residual threshold, determining the sample pair as a to-be-tested sample pair, and testing the sample pair through an experiment to update an initial sample label of each mixed training sample in the sample pair to obtain a new sample label.

[0011] In a case where a training set iteration termination condition is satisfied, training the initial memory model by using the updated mixed data set to obtain a trained magnetoresistive memory model.

[0012] According to an embodiment of the present application, the training method further comprises:

[0013] In a case of the i th iteration, a second residual of the mixed data set is calculated according to a plurality of first prediction performance parameters of a plurality of sample pairs.

[0014] According to the second residual, a proportioning weight corresponding to the simulation data set and the experimental data set respectively is adjusted.

[0015] In a case of performing the i+1 th iteration, a plurality of target training samples are selected according to the proportioning weight of the simulation data set and the experimental data set respectively to construct a new mixed data set.

[0016] According to an embodiment of the present application, training the initial memory model by using the updated mixed data set to obtain a trained magnetoresistive memory model comprises:

[0017] According to the proportioning weight of the updated simulation data set and the updated experimental data set respectively determined in the last iteration, a plurality of updated training samples are selected to construct the updated mixed data set.

[0018] Training the initial memory model by using the updated mixed data set to obtain a trained magnetoresistive memory model.

[0019] According to an embodiment of the present application, the updated mixed data set comprises a plurality of target training samples and a target sample label corresponding to each target training sample.

[0020] According to an embodiment of the present application, training the initial memory model by using the updated mixed data set to obtain a trained magnetoresistive memory model comprises:

[0021] Operations 4-6 are iteratively executed.

[0022] Operation 4, for each target training sample, inputting the target training sample into an initial neural network to solve a magnetodynamics process of the magnetoresistive random access memory and output a set of magnetodynamics parameters.

[0023] Operation 5, inputting the above magnetic dynamics parameter set into the full connection network, and outputting a second predicted performance parameter;

[0024] Operation 6, adjusting the model parameters of the initial memory model according to the above second predicted performance parameter and the above target sample label, wherein the initial memory model includes the initial neural network and the full connection network;

[0025] In the case of meeting the training iteration termination condition, the initial memory model of the last iteration is determined as the above magnetic resistance memory model.

[0026] According to an embodiment of the present application, adjusting the model parameters of the initial memory model according to the above second predicted performance parameter and the above target sample label comprises:

[0027] Based on the target loss function, a target loss value is calculated according to the above second predicted performance parameter and the above target sample label, wherein the target loss function is constructed according to a cross-entropy loss function and a magnetic body dynamics behavior equation;

[0028] The model parameters of the initial memory model are adjusted based on the above target loss value.

[0029] According to an embodiment of the present application, the input parameters of the magnetic resistance random memory include input current and magnetic field; the energy storage performance parameters include tunneling magnetoresistance, threshold current and coercive field; and the magnetic dynamics parameter set includes magnetic resistance state, damping constant, saturation magnetization and anisotropy constant.

[0030] Another aspect of the present application provides a magnetic resistance random memory resistance state adjustment method, comprising:

[0031] Obtaining a magnetic resistance memory model corresponding to the above magnetic resistance random memory, wherein the magnetic resistance memory model has continuous weights;

[0032] Iteratively performing the following operations:

[0033] According to the above continuous weights and the device resistance state of the above magnetic resistance random memory, difference information is calculated;

[0034] According to the above difference information, the write voltage of the above magnetic resistance random memory is adjusted to obtain a new write voltage;

[0035] The above new write voltage is applied to the above magnetic resistance random memory to adjust the resistance state of the above magnetic resistance random memory.

[0036] Another aspect of the present application provides an image processing method, comprising:

[0037] Obtaining a to-be-recognized image;

[0038] inputting the image to be identified into an input layer of an image recognition model, and outputting image features, the image features including current features and magnetic field features;

[0039] inputting the image features into a magnetoresistive memory model of the image recognition model, and outputting third predicted performance parameters;

[0040] inputting the third predicted performance parameters into a hidden layer of the image recognition model, and outputting pulse features, the pulse features including voltage pulse features and current pulse features;

[0041] inputting the pulse features into an output layer of the image recognition model, and outputting region identification parameters corresponding to different regions in the image to be identified.

[0042] Another aspect of the present application provides a training device of a magnetoresistive memory model, comprising:

[0043] a first obtaining module configured to obtain a simulation data set generated by a micromagnetic simulation technology and an experimental data set based on experimental data obtained by experiments, wherein the simulation data set and the experimental data set each include a plurality of initial training samples and initial sample labels corresponding to each of the initial training samples, the initial training samples represent input parameters of a magnetoresistive random memory, and the initial sample labels represent energy storage performance parameters of the magnetoresistive random memory;

[0044] an iteration module configured to iteratively execute operations of a construction unit, a processing unit and a determination unit:

[0045] the construction unit is configured to construct a mixed data set according to the simulation data set and the experimental data set, wherein the mixed data set includes a plurality of sample pairs, and each sample pair includes at least two mixed training samples;

[0046] the processing unit is configured to process each sample pair by using an initial memory model to obtain first predicted performance parameters corresponding to each of the mixed training samples, and to calculate first residuals of the at least two first predicted performance parameters;

[0047] the determination unit is configured to determine the sample pair as a sample pair to be tested if the first residuals satisfy a residual threshold, and to test the sample pair by experiments to update initial sample labels of each of the mixed training samples in the sample pair to obtain new sample labels;

[0048] a training module configured to train the initial memory model by using the updated mixed data set to obtain a trained magnetoresistive memory model if a training set iteration termination condition is satisfied.

[0049] Another aspect of the present application provides an image processing device, comprising:

[0050] a second obtaining module, configured to obtain an image to be recognized;

[0051] a feature extraction module, configured to input the image to be recognized into an input layer of an image recognition model, and output image features, wherein the image features comprise current features and magnetic field features;

[0052] a prediction module, configured to input the image features into a magnetoresistive random access memory model of the image recognition model, and output a third predicted performance parameter;

[0053] a processing module, configured to input the third predicted performance parameter into a hidden layer of the image recognition model, and output pulse features, wherein the pulse features comprise voltage pulse features and current pulse features;

[0054] an output module, configured to input the pulse features into an output layer of the image recognition model, and output region recognition parameters corresponding to different regions in the image to be recognized.

[0055] Another aspect of the present application provides a resistance state adjustment device of a magnetoresistive random access memory, comprising:

[0056] a third obtaining module, configured to obtain a magnetoresistive random access memory model corresponding to the magnetoresistive random access memory, wherein the magnetoresistive random access memory model has continuous weights;

[0057] a second iteration module, configured to iteratively perform operations of a calculation unit, an adjustment unit and an application unit:

[0058] the calculation unit, configured to calculate difference information according to the continuous weights and a device resistance state of the magnetoresistive random access memory;

[0059] the adjustment unit, configured to adjust a write voltage of the magnetoresistive random access memory according to the difference information, to obtain a new write voltage;

[0060] the application unit, configured to apply the new write voltage to the magnetoresistive random access memory, to adjust the resistance state of the magnetoresistive random access memory.

[0061] Another aspect of the present application provides an electronic device, comprising:

[0062] one or more processors;

[0063] a memory, configured to store one or more programs,

[0064] Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described above.

[0065] Another aspect of the present application provides a computer readable storage medium storing computer executable instructions for implementing the method as described above when executed.

[0066] Another aspect of the present application provides a computer program product comprising computer executable instructions for implementing the method as described above when executed.

[0067] According to the embodiment of the present application, the simulation data set and the experimental data set are mixed to obtain a mixed data set, and the calculation of the first residual of the obtained first prediction performance parameter is performed based on the initial memory model processing each sample pair in the mixed data set, and it is determined whether to perform an experimental test on the sample pair to update the initial sample label based on the first residual, so as to complete the training of the magnetoresistive memory model by using the updated mixed data set. Since the embodiment models the magnetoresistive memory by using the neural network, the solving speed of the magnetoresistive memory is improved, and in the training process of the model, the sample mixing and the experimental updating mode based on the residual are used to improve the number of training samples and further ensure the quality of the training samples, thereby effectively improving the prediction accuracy of the magnetoresistive memory model. BRIEF DESCRIPTION OF DRAWINGS

[0068] The above and other objects, features and advantages of the present application will become more apparent from the following description of the embodiments of the present application taken with reference to the accompanying drawings, in which:

[0069] Figure 1 An exemplary system architecture to which the training method, the resistance state adjustment method, and the image processing method according to embodiments of the present application can be applied is shown;

[0070] Figure 2 A flowchart of a training method of a magnetoresistive memory model according to an embodiment of the present application is shown;

[0071] Figure 3 A flowchart of a training method of a magnetoresistive memory model according to another embodiment of the present application is shown;

[0072] Figure 4 A flowchart of a resistance state adjustment method of a magnetoresistive random access memory according to an embodiment of the present application is shown;

[0073] Figure 5 A flowchart of an image processing method according to an embodiment of the present application is shown;

[0074] Figure 6A specific flowchart of image recognition according to an embodiment of the present application is shown.

[0075] Figure 7 A block diagram of a training apparatus of a magnetoresistive memory model according to an embodiment of the present application is shown.

[0076] Figure 8 A block diagram of an image processing apparatus according to an embodiment of the present application is shown.

[0077] Figure 9 A block diagram of a resistance state adjustment apparatus of a magnetoresistive random access memory according to an embodiment of the present application is shown.

[0078] Figure 10 A block diagram of an electronic device suitable for implementing the above-described method according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0079] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. It is to be understood, however, that the description is merely exemplary of the present application, and is not intended to limit the scope of the present application. In the following detailed description of the embodiments of the present application, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without these specific details. In other instances, well-known structures and functions have not been described in detail in order to avoid obscuring aspects of the present application.

[0080] The terms used herein are merely used to describe specific embodiments, and are not intended to limit the present application. The terms "include" and "have" and the like used herein indicate the presence of the described features, steps, operations, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, or components.

[0081] All terms used herein, including technical and scientific terms, have the same meanings as those generally understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0082] In the case of using expressions similar to "at least one of A, B, and C, etc.", it is generally to be interpreted as including one or more of the same. For example, "a system having at least one of A, B, and C" should be interpreted as including a system having A alone, a system having B alone, a system having C alone, a system having A and B together, a system having A and C together, a system having B and C together, and / or a system having A, B, and C together, etc.

[0083] In the embodiments of the present application, the collection, updating, analysis, processing, use, transmission, provision, disclosure, storage, etc. of the data involved (for example, including but not limited to user personal information) comply with the relevant legal regulations, are used for legal purposes, and do not violate public order and good customs. In particular, necessary measures have been taken for user personal information to prevent illegal access to user personal information data and to maintain user personal information security and network security.

[0084] In the embodiments of the present application, the authorization or consent of the user is obtained before the user's personal information is acquired or collected.

[0085] Embodiments of the present application provide a training method, a resistance state adjustment method, an image processing method and apparatus. The training method includes obtaining a simulation data set generated by a micromagnetic simulation technology and an experimental data set obtained based on an experiment; constructing a mixed data set according to the simulation data set and the experimental data set, the mixed data set including a plurality of sample pairs, each sample pair including at least two mixed training samples; for each sample pair, processing the sample pair using an initial memory model to obtain a first prediction performance parameter corresponding to each mixed training sample, and calculating a first residual of the at least two first prediction performance parameters; in a case where the first residual satisfies a residual threshold, testing the to-be-tested sample pair by an experiment to update an initial sample label to obtain a new sample label; in a case where a training set iteration termination condition is satisfied, training the initial memory model using the updated mixed data set to obtain a magnetic resistance memory model.

[0086] Figure 1 An exemplary system architecture to which the training method, resistance state adjustment method, and image processing method according to embodiments of the present application can be applied is shown. It should be noted that, Figure 1 The system architecture shown is merely an example of a system architecture to which embodiments of the present application can be applied, to help those skilled in the art understand the technical content of the present application, but does not mean that embodiments of the present application cannot be used in other devices, systems, environments or scenarios.

[0087] As Figure 1 shown, the system architecture 100 according to this embodiment can include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104 and a server 105. The network 104 is a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103 and the server 105. The network 104 can include various connection types, such as wired and / or wireless communication links, etc.

[0088] The user can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send messages, etc. The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be installed with various communication client applications, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (only as an example).

[0089] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smartphones, tablet computers, laptop computers, desktop computers, etc.

[0090] The server 105 can be a server providing various services, such as a background management server providing support for websites browsed by the user using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (only as an example). The background management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data, etc. obtained or generated according to user requests) to the terminal device.

[0091] It should be noted that the training method of the magnetoresistive memory model, the resistance state adjustment method, and the image processing method provided in the embodiments of the present application can generally be executed by the server 105. Correspondingly, the training device of the magnetoresistive memory model, the resistance state adjustment device, and the image processing device provided in the embodiments of the present application can generally be arranged in the server 105. The training method of the magnetoresistive memory model, the resistance state adjustment method, and the image processing method provided in the embodiments of the present application can also be executed by a server or a server cluster different from the server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or the server 105. Correspondingly, the training device of the magnetoresistive memory model, the resistance state adjustment device, and the image processing device provided in the embodiments of the present application can also be arranged in a server or a server cluster different from the server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or the server 105. Alternatively, the training method of the magnetoresistive memory model, the resistance state adjustment method, and the image processing method provided in the embodiments of the present application can also be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103, or by other terminal devices different from the first terminal device 101, the second terminal device 102, or the third terminal device 103. Correspondingly, the training device of the magnetoresistive memory model, the resistance state adjustment device, and the image processing device provided in the embodiments of the present application can also be arranged in the first terminal device 101, the second terminal device 102, or the third terminal device 103, or in other terminal devices different from the first terminal device 101, the second terminal device 102, or the third terminal device 103.

[0092] It should be understood that Figure 1 The number of terminal devices, networks, and servers in the system is only illustrative. Any number of terminal devices, networks, and servers can be provided according to implementation needs.

[0093] Figure 2 A flowchart of the training method of the magnetoresistive memory model according to the embodiments of the present application is shown.

[0094] As Figure 2 The training method of the magnetoresistive memory model includes operations S210-S230.

[0095] In operation S210, a simulation data set generated by a micromagnetic simulation technique and an experimental data set based on experimental data obtained by experiments are acquired, where the simulation data set and the experimental data set each include a plurality of initial training samples and an initial sample label corresponding to each initial training sample, the initial training sample representing an input parameter of the magnetoresistive random access memory, and the initial sample label representing an energy storage performance parameter of the magnetoresistive random access memory.

[0096] In operation S220, operations S221-S223 are iteratively performed:

[0097] In operation S221, a hybrid dataset is constructed according to the simulation dataset and the experimental dataset, where the hybrid dataset includes a plurality of sample pairs, and each sample pair includes at least two hybrid training samples.

[0098] In operation S222, for each sample pair, the initial memory model is used to process the sample pair to obtain a first prediction performance parameter corresponding to each hybrid training sample, and a first residual of the at least two first prediction performance parameters is calculated.

[0099] In operation S223, in a case where the first residual satisfies a residual threshold, the sample pair is determined as a to-be-tested sample pair, and the sample pair is tested through an experiment to update the initial sample label of each hybrid training sample in the sample pair to obtain a new sample label.

[0100] In operation S230, in a case where a training set iteration termination condition is satisfied, the initial memory model is trained using the updated hybrid dataset to obtain a trained magnetoresistive memory model.

[0101] According to an embodiment of the present application, micro-magnetic simulation technology is a computational method based on quantum mechanics and classical electromagnetic theory, which is used to study the behavior of magnetic fields at the microscopic scale. This technology establishes a mathematical model of the micro-magnetic system and uses a computer to solve it numerically, thus achieving accurate prediction of the magnetic field distribution and macroscopic magnetic properties in magnetic materials. Micro-magnetic simulation technology usually uses numerical methods such as finite element method (FEM) or molecular dynamics (MD) to solve complex problems in micro-magnetic systems. These methods iteratively calculate the variation of magnetic moments to gradually approach the real physical process. For example, in a two-dimensional micro-magnetic system, numerical integration of equations can be used to simulate the influence of an external magnetic field on the boundary conditions of magnetic domains, while in a three-dimensional case, more complex multi-layer structures can be considered to simulate the stacking of multiple magnetic materials in practical applications.

[0102] According to an embodiment of the present application, the formation of the simulation dataset can not only be generated by micro-magnetic simulation technology, but also by other physical simulation models, such as spin transport models, non-equilibrium Green's function tunneling magnetoresistance models, electromagnetic field models, magnetic dynamics models, and micro-magnetic simulation models.

[0103] According to the embodiments of the present application, because the time and material cost required by the experiment is high, the expansion of the sample can be performed through the micro-magnetic simulation technology, so that the model trained by more samples has higher accuracy. Specifically, the initial training samples and the corresponding initial sample labels about the magnetic resistance memory are obtained through the micro-magnetic simulation technology and the experimental method respectively, so as to form the simulation data set and the experimental data set respectively.

[0104] According to the embodiments of the present application, before the training of the model is performed, the initial training samples in the simulation data set and the experimental data set are first mixed according to a certain proportion to obtain a mixed data set, for example, m initial training samples are selected from the simulation data set, and n initial training samples are selected from the experimental data set, so that the number of initial training samples in the mixed data set is m+n, wherein the corresponding initial sample labels are also selected when the initial training samples are selected. Thereafter, the initial training samples in the mixed data set can be randomly paired to form a plurality of sample pairs, wherein m and n are both positive integers greater than 0.

[0105] According to the embodiments of the present application, for each sample pair, each initial training sample in the sample pair is input into the initial memory model to obtain a first prediction performance parameter corresponding to each initial training sample, and the two first prediction performance parameters are calculated by residual, specifically by calculating the residual through the first prediction performance parameter and the corresponding initial sample label, so as to obtain the first residual of each sample in the sample pair. Wherein, the energy storage performance parameter and the prediction performance parameter can refer to the quality factor (Figure of Merit, FoM) of the tunneling magnetoresistance (Tunnel MagnetoResistance, TMR) of the magnetic resistance memory.

[0106] According to the embodiments of the present application, if the difference between the two first residuals satisfies the residual threshold, at this time the sample pair can be determined as a to-be-tested sample pair, so that each mixed training sample in the to-be-tested sample pair is verified through the experimental method, and the initial sample label of the mixed training sample is updated through the experimental result to obtain a new sample label. Wherein, the specific value of the residual threshold can be set according to actual needs, for example, it can be 2.

[0107] According to embodiments of this application, after iteratively executing operations S221 to S223, the updated hybrid dataset can be used as the final training set after the iteration terminates to train the initial memory model, thereby obtaining the magnetoresistive memory model. The initial memory model can include multiple sub-networks, with different sub-networks used to construct different physical processes or nodes of the magnetoresistive memory, thus ensuring that the constructed model incorporates real-world process variations and defect distributions, making it easier to map the model to real magnetoresistive random access memory. The sub-networks can be gated recurrent networks, convolutional neural networks, graph neural networks, or fully connected networks.

[0108] According to an embodiment of this application, a hybrid dataset is obtained by mixing a simulation dataset and an experimental dataset. A first residual is calculated for each sample pair in the hybrid dataset based on the first prediction performance parameter obtained from the initial memory model. Based on the first residual, it is determined whether to conduct experimental testing on the sample pairs to update the initial sample labels. Thus, the updated hybrid dataset is used to complete the training of the magnetoresistive memory model. Since this embodiment models the magnetoresistive memory using a neural network, the solution speed for the magnetoresistive memory is improved. Furthermore, during model training, the sample mixing and residual-based experimental update method increase the number of training samples while further ensuring the quality of the training samples, thereby effectively improving the prediction accuracy of the magnetoresistive memory model.

[0109] According to an embodiment of this application, the training method further includes: in the i-th iteration, calculating a second residual of the hybrid dataset based on multiple first prediction performance parameters of multiple sample pairs; adjusting the matching weights corresponding to the simulation dataset and the experimental dataset respectively based on the second residual; and in the i+1-th iteration, selecting multiple target training samples according to the matching weights of the simulation dataset and the experimental dataset respectively to construct a new hybrid dataset.

[0110] According to an embodiment of this application, in each iteration of the mixed training set, after all sample pairs in the mixed training set have completed the prediction of the first prediction performance parameter, the second residual of the overall mixed training set can be calculated. For example, the first residual of each mixed training sample in the mixed training set can be summed.

[0111] According to an embodiment of this application, the weighting of the simulation dataset and the experimental dataset is adjusted based on the sum of the residuals. For example, when the sum of the residuals is greater than the upper limit of the residuals, it indicates that the number of initial training samples selected from the simulation dataset is large. In this case, the weighting of the simulation dataset can be appropriately reduced to obtain a new weighting.

[0112] According to an embodiment of this application, during the next iteration of training, multiple target training samples are selected from the simulation dataset and the experimental dataset based on the weight ratio determined in the previous iteration to construct a new hybrid dataset for this training. For example, if the weight ratios of the simulation dataset and the experimental dataset determined in the previous iteration are 0.4 and 0.6 respectively, and the initial number of training samples in the simulation dataset and the experimental dataset are A and B respectively, then the number of samples in the new hybrid dataset is C = 0.4 × A + 0.6 × B.

[0113] According to embodiments of this application, by adjusting the weighting of different datasets based on the second residual, the samples in the mixed dataset used for training are more representative and of higher quality, thereby effectively improving the accuracy of the model.

[0114] According to an embodiment of this application, training an initial memory model using an updated hybrid dataset to obtain a trained magnetoresistive memory model includes: selecting multiple updated training samples according to the weight ratios determined by the updated simulation dataset and the updated experimental dataset in the last iteration to construct an updated hybrid dataset; and training the initial memory model using the updated hybrid dataset to obtain a trained magnetoresistive memory model.

[0115] According to embodiments of this application, when forming the final hybrid dataset for training the model, multiple updated training samples are selected from the updated simulation dataset and the updated experimental dataset based on the weighting of the different datasets determined in the last iteration, to construct the updated hybrid dataset. Using this hybrid dataset to iteratively train the initial memory model yields a magnetoresistive memory model that more closely resembles the actual performance of the magnetoresistive memory.

[0116] It should be noted that during the construction of any mixed dataset, not only are the corresponding training samples selected, but also the sample labels corresponding to those training samples are selected, including the unupdated initial sample labels or the new sample labels obtained after the update.

[0117] Figure 3 A flowchart of a training method for a magnetoresistive memory model according to another embodiment of this application is shown.

[0118] According to embodiments of this application, the updated hybrid dataset includes multiple target training samples and target sample labels corresponding to each target training sample.

[0119] According to an embodiment of this application, an initial memory model is trained using an updated hybrid dataset to obtain a trained magnetoresistive memory model, including iteratively executing operations 301 to 303: In operation 301, for each target training sample, the target training sample is input into an initial neural network to solve the magnetodynamic process of the magnetoresistive random access memory and output a set of magnetodynamic parameters; In operation 301, the set of magnetodynamic parameters is input into a fully connected network and outputs a second prediction performance parameter; In operation 303, the model parameters of the initial memory model are adjusted according to the second prediction performance parameter and the target sample label, wherein the initial memory model includes an initial neural network and a fully connected network; If the training iteration termination condition is met, the initial memory model of the last iteration is determined as the magnetoresistive memory model.

[0120] According to an embodiment of this application, see Figure 3 For each target training sample in the updated hybrid dataset, which may be a training sample obtained through a physical simulation model or a training sample obtained through experimental simulation, the target training sample in the hybrid dataset is input into an initial neural network (such as a gated recurrent network or a convolutional neural network). The initial neural network is used to solve the magnetodynamic process of the magnetoresistive random access memory, thereby obtaining a set of magnetodynamic parameters.

[0121] According to embodiments of this application, a gated recurrent unit (GRU) is preferably used as the initial neural network. The GRU is used to process time-series inputs and capture magnetization dynamics. The GRU addresses the vanishing gradient problem by introducing update and reset gates and efficiently captures long-term dependencies. Furthermore, the GRU structure is relatively simple and training is fast. By selectively retaining or forgetting historical information through the gating mechanism, it is suitable for modeling physical processes with temporal evolution.

[0122] According to an embodiment of this application, see Figure 3 By processing static parameters such as the set of magnetoresistive parameters through a fully connected network, a second prediction performance parameter is obtained. Based on the second prediction performance parameter and the target sample label, the model parameters of the initial memory model are iteratively adjusted. When the training iteration termination condition is met, the initial memory model of the last iteration can be determined as the magnetoresistive memory model.

[0123] According to an embodiment of this application, adjusting the model parameters of the initial memory model based on the second prediction performance parameter and the target sample label includes: calculating a target loss value based on the target loss function, the second prediction performance parameter, and the target sample label, wherein the target loss function is constructed based on the cross-entropy loss function and the magnet dynamics behavior equation; and adjusting the model parameters of the initial memory model based on the target loss value.

[0124] According to embodiments of this application, in the iterative parameter tuning of the model, the target loss function used is obtained by weighted summation of the cross-entropy loss function and the magnetic dynamics equation. The magnetic dynamics equation can refer to the Landau-Lifshitz-Gilbert (LLG) equation, which describes the dynamic behavior of magnetic moments in magnetic materials. The LLG equation depicts the classical motion of magnetic moments, including the effect of an applied magnetic field on the magnetic moments and the interactions between magnetic moments, while also considering the coupling effect between magnetic moments. Therefore, it can describe materials with different lattice structures and different magnetic moment magnitudes. Furthermore, the LLG equation can also describe the domain structure and domain wall motion of magnetic materials, playing a crucial role in understanding the macroscopic structure and dynamic behavior of magnetic materials.

[0125] According to an embodiment of this application, see Figure 3 After each model prediction yields the second prediction performance parameter (i.e., the quality factor FoMs), a loss calculation is performed on the second prediction performance parameter and the corresponding target sample label based on the target loss function to obtain the target loss value. Based on this target loss value, the model parameters of the initial memory model are adjusted, thus determining the initial memory model for the last iteration as a magnetoresistive memory model if the iteration termination condition is met.

[0126] According to embodiments of this application, since the objective loss function considers the cross-entropy loss and the magnetic moment dynamics of the magnetic body in the field of magnetoresistive memory, the resulting magnetoresistive memory model is closer to the magnetoresistive memory in the real environment, thereby improving the accuracy of the model in performance prediction. At the same time, the physical constraints of the LLG equations can improve the solution speed.

[0127] According to embodiments of this application, the input parameters of the magnetoresistive random access memory include the input current I(t) and the magnetic field H(t); the energy storage performance parameters include the quality factor (FoMs) such as tunneling magnetoresistance (TMR), threshold current Ic, and coercive field Hc; and the set of magnetodynamic parameters includes the magnetoresistive state, damping constant α, saturation magnetization Ms, and anisotropy constant Ku.

[0128] According to an embodiment of this application, after the model training is completed, the model can be verified using experimental data (i.e., experimental samples). If the residual between the verified performance parameters obtained by the model and the true performance parameters is less than the verification threshold, it indicates that the model has been sufficiently trained, and the model can then be used as the final magnetoresistive memory model.

[0129] Figure 4 A flowchart of a resistive state adjustment method for a magnetoresistive random access memory according to an embodiment of this application is shown.

[0130] like Figure 4 As shown, the resistive state adjustment method of the magnetoresistive random access memory includes operations S410 to S420:

[0131] In operation S410, a magnetoresistive memory model corresponding to the magnetoresistive random access memory is obtained, wherein the magnetoresistive memory model has continuous weights.

[0132] In operation S420, operations S421 to S423 are executed iteratively:

[0133] In operation S421, difference information is calculated based on continuous weights and the device resistance state of the magnetoresistive random access memory.

[0134] In operation S422, the write voltage of the magnetoresistive random access memory is adjusted according to the difference information to obtain a new write voltage.

[0135] In operation S423, a new write voltage is applied to the magnetoresistive random access memory to adjust the resistive state of the magnetoresistive random access memory.

[0136] According to an embodiment of this application, in the process of adjusting the resistance state of a magnetoresistive random access memory (MRRAM) based on a magnetoresistive memory model, the MRRAM model corresponding to the MRRAM model is first obtained. The difference between the continuous weights in the model and the device resistance state is calculated. Based on this difference, the write voltage is adjusted to obtain a new write voltage. Applying this new write voltage to the MRRAM adjusts its resistance state. Through iterative operations, the resistance state of the MRRAM can be adjusted to a suitable range.

[0137] According to embodiments of this application, a hybrid dataset is obtained by mixing a simulation dataset and an experimental dataset. A first residual is calculated for the first predicted performance parameter obtained from each sample pair in the hybrid dataset based on the initial memory model. Based on the first residual, it is determined whether to perform experimental testing on the sample pair to update the initial sample labels, thereby completing the training of the magnetoresistive memory model using the updated hybrid dataset. By calculating the difference between the continuous weights of the magnetoresistive memory model and the device resistance state of the magnetoresistive random access memory (MRRAM), the resistance state of the MRRAM is adjusted by adjusting the write voltage, thus achieving precise adjustment of the MRRAM's resistance state.

[0138] Figure 5 A flowchart of an image processing method according to an embodiment of this application is shown. Figure 6 A detailed flowchart of image recognition according to an embodiment of this application is shown.

[0139] like Figure 5 As shown, the image processing method includes the following operations S510 to S550.

[0140] The S510 is used to acquire the image to be recognized.

[0141] When operating the S520, the image to be recognized is input into the input layer of the image recognition model, and the image features are output, including current features and magnetic field features.

[0142] In operation S530, image features are input into the magnetoresistive memory model of the image recognition model, and the third prediction performance parameter is output.

[0143] In operation S540, the third prediction performance parameter is input into the hidden layer of the image recognition model, and pulse features are output, including voltage pulse features and current pulse features.

[0144] When operating the S550, the pulse features are input to the output layer of the image recognition model, and the corresponding region recognition parameters for different regions in the image to be recognized are output.

[0145] According to embodiments of this application, the image to be identified can be any type of image, such as a face recognition image, a vehicle recognition image, or an industrial defect detection image.

[0146] According to embodiments of this application, such as Figure 6As shown, for any type of image to be recognized, the image is first converted into image features in vector form by the input layer of the image recognition model. The image features are then input into the trained magnetoresistive memory model to obtain the corresponding third prediction performance parameter. This third prediction performance parameter is processed by the hidden layer to obtain impulse features. After being transformed by the output layer, the region recognition parameters of different regions in the image to be recognized can be obtained. The region recognition parameters can refer to the key features of faces, vehicles, or local recognition features of products.

[0147] In one specific embodiment, taking the production of magnetoresistive random access memory (MRRAM) in a factory as an example, image data of the memory can be captured by a camera or other imaging device and sent to an electronic device. The electronic device then uses an image recognition model to identify the image data. If the identification parameters in the identified area are abnormal, an alarm can be triggered to alert workers, or the production line can be directly stopped to reduce production costs. If the identification parameters in the identified area are normal, the production line can continue operating normally to complete the production of the MRAM.

[0148] According to an embodiment of this application, a hybrid dataset is obtained by mixing a simulation dataset and an experimental dataset. A first residual is calculated for each sample pair in the hybrid dataset based on the first prediction performance parameter obtained from the initial memory model. Based on the first residual, it is determined whether to conduct experimental testing on the sample pairs to update the initial sample labels. Thus, the updated hybrid dataset is used to complete the training of the magnetoresistive memory model. Since this embodiment models the magnetoresistive memory using a neural network, the solution speed for the magnetoresistive memory is improved. Furthermore, during model training, the sample mixing and residual-based experimental update method increase the number of training samples while further ensuring the quality of the training samples, thereby effectively improving the prediction accuracy of the magnetoresistive memory model.

[0149] Figure 7 A block diagram of a training apparatus for a magnetoresistive memory model according to an embodiment of this application is shown.

[0150] like Figure 7 As shown, the training device 700 for the magnetoresistive memory model includes a first acquisition module 710, an iteration module 720, and a training module 730.

[0151] The first acquisition module 710 is used to acquire a simulation dataset generated by micromagnetic simulation technology and an experimental dataset obtained through experiments. Both the simulation dataset and the experimental dataset include multiple initial training samples and initial sample labels corresponding to each initial training sample. The initial training samples represent the input parameters of the magnetoresistive random access memory, and the initial sample labels represent the energy storage performance parameters of the magnetoresistive random access memory.

[0152] Iteration module 720 is used to iteratively perform the operations of building units, processing units, and determining units:

[0153] The building unit is used to construct a hybrid dataset based on the simulation dataset and the experimental dataset, wherein the hybrid dataset includes multiple sample pairs, and each sample pair includes at least two hybrid training samples.

[0154] The processing unit is used to process the sample pair using the initial memory model for each sample pair, obtain the first prediction performance parameter corresponding to each mixed training sample, and calculate the first residual of at least two first prediction performance parameters.

[0155] The determination unit is used to determine the sample pair as the test sample pair when the first residual meets the residual threshold, and to test the sample pair through experiments to update the initial sample label of each mixed training sample in the sample pair to obtain a new sample label.

[0156] The training module 730 is used to train the initial memory model using the updated hybrid dataset, provided that the training set iteration termination condition is met, to obtain the trained magnetoresistive memory model.

[0157] According to an embodiment of this application, a hybrid dataset is obtained by mixing a simulation dataset and an experimental dataset. A first residual is calculated for each sample pair in the hybrid dataset based on the first prediction performance parameter obtained from the initial memory model. Based on the first residual, it is determined whether to conduct experimental testing on the sample pairs to update the initial sample labels. Thus, the updated hybrid dataset is used to complete the training of the magnetoresistive memory model. Since this embodiment models the magnetoresistive memory using a neural network, the solution speed for the magnetoresistive memory is improved. Furthermore, during model training, the sample mixing and residual-based experimental update method increase the number of training samples while further ensuring the quality of the training samples, thereby effectively improving the prediction accuracy of the magnetoresistive memory model.

[0158] Figure 8 A block diagram of an image processing apparatus according to an embodiment of this application is shown.

[0159] like Figure 7 As shown, the image processing device 800 includes a second acquisition module 810, a feature extraction module 820, a prediction module 830, a processing module 840, and an output module 850.

[0160] The second acquisition module 810 is used to acquire the image to be recognized.

[0161] The feature extraction module 820 is used to input the image to be recognized into the input layer of the image recognition model and output image features, including current features and magnetic field features.

[0162] The prediction module 830 is used to input image features into the magnetoresistive memory model of the image recognition model and output a third prediction performance parameter.

[0163] The processing module 840 is used to input the third prediction performance parameter into the hidden layer of the image recognition model and output pulse features, wherein the pulse features include voltage pulse features and current pulse features.

[0164] The output module 850 is used to input pulse features into the output layer of the image recognition model and output the region recognition parameters corresponding to different regions in the image to be recognized.

[0165] According to an embodiment of this application, a hybrid dataset is obtained by mixing a simulation dataset and an experimental dataset. A first residual is calculated for each sample pair in the hybrid dataset based on the first prediction performance parameter obtained from the initial memory model. Based on the first residual, it is determined whether to conduct experimental testing on the sample pairs to update the initial sample labels. Thus, the updated hybrid dataset is used to complete the training of the magnetoresistive memory model. Since this embodiment models the magnetoresistive memory using a neural network, the solution speed for the magnetoresistive memory is improved. Furthermore, during model training, the sample mixing and residual-based experimental update method increase the number of training samples while further ensuring the quality of the training samples, thereby effectively improving the prediction accuracy of the magnetoresistive memory model.

[0166] Figure 9 A block diagram of a resistive state adjustment device for a magnetoresistive random access memory according to an embodiment of this application is shown.

[0167] like Figure 9 As shown, the resistance state adjustment device 900 of the magnetoresistive random access memory includes a third acquisition module 910 and a second iteration module 920.

[0168] The third acquisition module 910 is used to acquire the magnetoresistive memory model corresponding to the magnetoresistive random access memory, wherein the magnetoresistive memory model has continuous weights;

[0169] The second iteration module 920 iteratively executes the operations of the calculation unit, the adjustment unit, and the application unit: the calculation unit is used to calculate the difference information based on the continuous weights and the device resistance state of the magnetoresistive random access memory; the adjustment unit is used to adjust the write voltage of the magnetoresistive random access memory based on the difference information to obtain a new write voltage; the application unit is used to apply the new write voltage to the magnetoresistive random access memory to adjust the resistance state of the magnetoresistive random access memory.

[0170] According to an embodiment of this application, a hybrid dataset is obtained by mixing a simulation dataset and an experimental dataset. A first residual is calculated for each sample pair in the hybrid dataset based on the first prediction performance parameter obtained from the initial memory model. Based on the first residual, it is determined whether to conduct experimental testing on the sample pairs to update the initial sample labels. Thus, the updated hybrid dataset is used to complete the training of the magnetoresistive memory model. Since this embodiment models the magnetoresistive memory using a neural network, the solution speed for the magnetoresistive memory is improved. Furthermore, during model training, the sample mixing and residual-based experimental update method increase the number of training samples while further ensuring the quality of the training samples, thereby effectively improving the prediction accuracy of the magnetoresistive memory model.

[0171] Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be implemented by dividing them into multiple modules. Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or implemented by hardware or firmware in any other reasonable manner by integrating or packaging circuits, or implemented in any one of software, hardware, and firmware, or in a suitable combination of any of these. Alternatively, one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0172] For example, any and multiple of the following modules can be implemented in one module / unit / subunit: the first acquisition module 710, the iteration module 720, the training module 730; or the second acquisition module 810, the feature extraction module 820, the prediction module 830, the processing module 840, the output module 850; or the third acquisition module 910, the second iteration module 920. Alternatively, any one of these modules / units / subunits can be split into multiple modules / units / subunits. Or, at least some of the functionality of one or more of these modules / units / subunits can be combined with at least some of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this application, at least one of the following can be implemented at least partially as hardware circuits, such as a first acquisition module 710, an iteration module 720, a training module 730, a second acquisition module 810, a feature extraction module 820, a prediction module 830, a processing module 840, an output module 850, or a third acquisition module 910, a second iteration module 920: a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuit, or as hardware or firmware, or as any one of software, hardware, and firmware implementations or any appropriate combination thereof. Alternatively, at least one of the following can be implemented, at least partially, as a computer program module: the first acquisition module 710, the iteration module 720, the training module 730; the second acquisition module 810, the feature extraction module 820, the prediction module 830, the processing module 840, the output module 850; or the third acquisition module 910, the second iteration module 920. When the computer program module is run, it can perform the corresponding function.

[0173] It should be noted that the training device, resistance adjustment device, and image processing device of the magnetoresistive memory model in the embodiments of this application correspond to the training method, resistance adjustment method, and image processing method of the magnetoresistive memory model in the embodiments of this application, respectively. The specific descriptions of the training device, resistance adjustment device, and image processing device of the magnetoresistive memory model can be found in the training method, resistance adjustment method, and image processing method of the magnetoresistive memory model, respectively, and will not be repeated here.

[0174] Figure 10 A block diagram of an electronic device suitable for implementing the methods described above, according to an embodiment of this application, is shown. Figure 10 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0175] like Figure 10As shown, an electronic device 1000 according to an embodiment of this application includes a processor 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage portion 1008 into a random access memory (RAM) 1003. The processor 1001 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1001 may also include onboard memory for caching purposes. The processor 1001 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0176] RAM 1003 stores various programs and data required for the operation of electronic device 1000. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Processor 1001 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 1002 and / or RAM 1003. It should be noted that the programs may also be stored in one or more memories other than ROM 1002 and RAM 1003. Processor 1001 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0177] According to embodiments of this application, the electronic device 1000 may further include an input / output (I / O) interface 1005, which is also connected to a bus 1004. The electronic device 1000 may also include one or more of the following components connected to the input / output (I / O) interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the input / output (I / O) interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1010 as needed so that computer programs read from it can be installed into the storage section 1008 as needed.

[0178] According to embodiments of this application, the method flow according to embodiments of this application can be implemented as a computer software program. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by processor 1001, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0179] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0180] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0181] For example, according to embodiments of this application, a computer-readable storage medium may include the ROM 1002 and / or RAM 1003 described above and / or one or more memories other than ROM 1002 and RAM 1003.

[0182] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of this application. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the methods provided in the embodiments of this application.

[0183] When the computer program is executed by the processor 1001, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc. described above can be implemented by computer program modules.

[0184] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1009, and / or installed from a removable medium 1011. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0185] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0186] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments of this application can be combined and / or combined in various ways. All such combinations and / or combinations fall within the scope of this application.

[0187] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. This application does not depart from its scope, and those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A training method for a magnetoresistive memory model, characterized in that, include: A simulation dataset generated by micromagnetic simulation technology and an experimental dataset obtained through experiments are obtained. Both the simulation dataset and the experimental dataset include multiple initial training samples and initial sample labels corresponding to each initial training sample. The initial training samples represent the input parameters of the magnetoresistive random access memory, and the initial sample labels represent the energy storage performance parameters of the magnetoresistive random access memory. Iteratively execute operations 1 through 3: Operation 1: Construct a hybrid dataset based on the simulation dataset and the experimental dataset, wherein the hybrid dataset includes multiple sample pairs, and each sample pair includes at least two hybrid training samples; Operation 2: For each sample pair, process the sample pair using the initial memory model to obtain the first prediction performance parameter corresponding to each of the mixed training samples, and calculate the first residual of at least two of the first prediction performance parameters; Operation 3: If the first residual satisfies the residual threshold, the sample pair is determined as the test sample pair, and the test sample pair is tested through experiments to update the initial sample label of each mixed training sample in the sample pair and obtain a new sample label. If the training set iteration termination condition is met, the initial memory model is trained using the updated hybrid dataset to obtain the trained magnetoresistive memory model.

2. The method according to claim 1, characterized in that, Also includes: In the case of the i-th iteration, the second residual of the mixed dataset is calculated based on multiple first prediction performance parameters of multiple sample pairs; Based on the second residual, adjust the weighting ratios corresponding to the simulation dataset and the experimental dataset, respectively; In the (i+1)th iteration, multiple target training samples are selected according to the respective weight ratios of the simulation dataset and the experimental dataset to construct a new hybrid dataset.

3. The method according to claim 2, characterized in that, The initial memory model is trained using the updated hybrid dataset to obtain a trained magnetoresistive memory model, including: Based on the weighted ratios determined in the last iteration of the updated simulation dataset and the updated experimental dataset, multiple updated training samples are selected to construct the updated hybrid dataset. The initial memory model is trained using the updated hybrid dataset to obtain a trained magnetoresistive memory model.

4. The method according to claim 1 or 3, characterized in that, The updated hybrid dataset includes multiple target training samples and target sample labels corresponding to each target training sample; The initial memory model is trained using the updated hybrid dataset to obtain a trained magnetoresistive memory model, including: Perform operations 4 through 6 iteratively: Operation 4: For each target training sample, input the target training sample into the initial neural network to solve the magnetoresistive random access memory's magnetodynamic process and output a set of magnetodynamic parameters; Operation 5: Input the set of magnetodynamic parameters into a fully connected network and output the second predicted performance parameter; Operation 6: Adjust the model parameters of the initial memory model according to the second prediction performance parameters and the target sample label, wherein the initial memory model includes the initial neural network and the fully connected network; If the training iteration termination condition is met, the initial memory model of the last iteration is determined as the magnetoresistive memory model.

5. The method according to claim 4, characterized in that, Based on the second prediction performance parameter and the target sample label, the model parameters of the initial memory model are adjusted, including: Based on the target loss function, the target loss value is calculated according to the second prediction performance parameter and the target sample label, wherein the target loss function is constructed based on the cross-entropy loss function and the magnet dynamics behavior equation; The model parameters of the initial memory model are adjusted based on the target loss value.

6. The method according to claim 4, characterized in that, The input parameters of the magnetoresistive random access memory include input current and magnetic field; the energy storage performance parameters include tunneling magnetoresistance, threshold current and coercive field; the set of magnetodynamic parameters includes magnetoresistive state, damping constant, saturation magnetization and anisotropy constant.

7. A method for adjusting the resistive state of a magnetoresistive random access memory, characterized in that, include: Obtain a magnetoresistive memory model corresponding to the magnetoresistive random access memory, wherein the magnetoresistive memory model is trained by the method of any one of claims 1 to 6, and the magnetoresistive memory model has continuous weights; Perform the following operations iteratively: The difference information is calculated based on the continuous weights and the device resistance state of the magnetoresistive random access memory; The write voltage of the magnetoresistive random access memory is adjusted based on the difference information to obtain a new write voltage; The new write voltage is applied to the magnetoresistive random access memory to adjust the resistive state of the magnetoresistive random access memory.

8. An image processing method, characterized in that, include: Acquire the image to be recognized; The image to be identified is input into the input layer of the image recognition model, and the image features are output, including current features and magnetic field features. The image features are input into the magnetoresistive memory model of the image recognition model, and a third prediction performance parameter is output, wherein the magnetoresistive memory model is trained by the method of any one of claims 1 to 6; The third prediction performance parameter is input into the hidden layer of the image recognition model, and pulse features are output, wherein the pulse features include voltage pulse features and current pulse features; The pulse features are input to the output layer of the image recognition model, which outputs region recognition parameters corresponding to different regions in the image to be recognized.

9. A training device for a magnetoresistive memory model, characterized in that, include: The first acquisition module is used to acquire a simulation dataset generated by micromagnetic simulation technology and an experimental dataset obtained through experiments. The simulation dataset and the experimental dataset each include multiple initial training samples and initial sample labels corresponding to each initial training sample. The initial training samples represent the input parameters of the magnetoresistive random access memory, and the initial sample labels represent the energy storage performance parameters of the magnetoresistive random access memory. The iteration module is used to iteratively execute the operations of the building units, processing units, and determining units: A construction unit is configured to construct a hybrid dataset based on the simulation dataset and the experimental dataset, wherein the hybrid dataset includes multiple sample pairs, and each sample pair includes at least two hybrid training samples; The processing unit is configured to process each sample pair using an initial memory model to obtain a first prediction performance parameter corresponding to each of the mixed training samples, and to calculate a first residual of at least two of the first prediction performance parameters. The determining unit is configured to determine the sample pair as a test sample pair when the first residual satisfies the residual threshold, and to test the sample pair through experiments in order to update the initial sample label of each mixed training sample in the sample pair and obtain a new sample label. The training module is used to train the initial memory model using the updated hybrid dataset, provided that the training set iteration termination condition is met, to obtain a trained magnetoresistive memory model.

10. An image processing apparatus, characterized in that, include: The second acquisition module is used to acquire the image to be recognized; The feature extraction module is used to input the image to be identified into the input layer of the image recognition model and output image features, including current features and magnetic field features; A prediction module is used to input the image features into the magnetoresistive memory model of the image recognition model and output a third prediction performance parameter, wherein the magnetoresistive memory model is trained by the method of any one of claims 1 to 6; The processing module is used to input the third prediction performance parameter into the hidden layer of the image recognition model and output pulse features, wherein the pulse features include voltage pulse features and current pulse features; The output module is used to input the pulse features into the output layer of the image recognition model and output the region recognition parameters corresponding to different regions in the image to be recognized.

Images