Method and device for neural network model training based on storage and calculation integrated chip

Abstract

The invention discloses a method and device for training a neural network model based on a storage and calculation integrated chip, and the method comprises the steps: initializing a weight parameter of an initial neural network model, obtaining an initial weight value of each network layer, and adjusting a conductivity value of the storage and calculation integrated chip according to the initial weight value; inputting the training set into the initial neural network model, outputting a predicted value of each training sample, and determining a current loss value according to a difference between the predicted value and a true value; calculating an error gradient of the current loss value to the weight parameter, and determining a weight adjustment value according to the error gradient; according to the relationship between the initial weight value and the weight value boundary, quantizing the weight adjustment value to a preset integer range to obtain a target quantized value; and calling a weight updating interface, adjusting the initial weight value along the adjustment direction represented by the target quantized value, adjusting the conductivity value of the storage and calculation integrated chip according to the adjusted weight value of each network layer, obtaining an updated neural network model, and continuing training until a target neural network model is obtained.

Description

Technical Field

[0001] The present application relates to the technical field of storage-computing integrated chips, and more specifically, to a method and device for training a neural network model based on a storage-computing integrated chip. Background Art

[0002] Traditional storage-computing integrated chips are only used as inference devices in neural network training. Model training must first be performed on a traditional AI (Artificial Intelligence) accelerator on the PC (Personal Computer) side, and then the trained model is placed on the storage-computing integrated chip for inference. Training and inference are separate.

[0003] Most of the existing neural network model training methods are designed for traditional computing architectures. Not only are the training methods based on floating-point numbers, which have a large dynamic range and high precision, but the storage-computing integrated chip is an integer parameter with limited precision and a small range. Therefore, the traditional training methods are not suitable for the storage-computing integrated structure. In addition, the current-voltage characteristics of the storage-computing integrated chip are usually nonlinear, which makes it complicated to calculate the gradient and weight update when training the neural network, especially in the back-propagation algorithm, and it is difficult to ensure the accuracy and stability of the training.

[0004] In addition, the noise characteristics of the integrated storage and computing chip itself lead to differences between different chips, which need to be fitted during PC training. In addition, the offset of each chip ADC (Analog to Digital Converter) is different, and the output leakage current of each array is different. The differences lead to slightly different weight storage and calculation values ​​of each chip. Therefore, if an on-chip training method based on the integrated storage and computing chip can be implemented, these characteristic difference noises can be eliminated. At the same time, for the scenario of neural network inference on the terminal device, the integrated storage and computing chip itself can dynamically update the model without the need for retraining in the cloud during deployment.

[0005] Based on the above, how to implement neural network training on a storage and computing integrated chip is a problem that needs to be solved urgently. Summary of the invention

[0006] The present application provides a method and device for training a neural network model based on a storage-computing integrated chip, which can implement neural network training on the storage-computing integrated chip and ensure the accuracy and stability of the neural network model training.

[0007] The specific technical solutions are as follows:

[0008] In a first aspect, an embodiment of the present application provides a method for training a neural network model based on a storage-computing integrated chip. The method is applied to the storage-computing integrated chip. The method includes:

[0009] Initializing the weight parameters of the initial neural network model, obtaining the initial weight value of each network layer of the initial neural network model, and adjusting the conductance value of the storage-computing integrated chip according to the initial weight value of each network layer, wherein the weight value corresponding to the same weight parameter is positively correlated with the corresponding conductance value;

[0010] Inputting a training set into the initial neural network model, outputting a predicted value of each training sample, and determining a current loss value according to a difference between the predicted value and the true value of each training sample, wherein the training set includes a plurality of the training samples and a true value corresponding to each of the training samples;

[0011] Calculating the error gradient of the current loss value to the weight parameter of each network layer, and determining the weight adjustment value of each network layer according to the error gradient of each network layer, wherein the weight adjustment value represents the weight change to be adjusted;

[0012] For each network layer, according to the relationship between the initial weight value and the weight value boundary of the network layer, the weight adjustment value is quantized into a preset integer range to obtain a target quantization value corresponding to the weight adjustment value, wherein different quantization values ​​within the preset integer range represent different adjustment directions respectively;

[0013] For the initial weight value of each network layer, a weight update interface is called to adjust the initial weight value along the adjustment direction represented by the target quantization value corresponding to the initial weight value to obtain the adjusted weight value, and the conductance value of the storage-computing integrated chip is adjusted according to the adjusted weight value of each network layer to obtain an updated neural network model;

[0014] The training set is continuously trained based on the updated neural network model until the current loss value satisfies a preset convergence condition or reaches a preset threshold of iteration times, and the obtained neural network model is used as the target neural network model.

[0015] In a possible implementation, for each network layer, according to the relationship between the initial weight value of the network layer and the weight value boundary, the weight adjustment value is quantized into a preset integer range to obtain a target quantization value corresponding to the weight adjustment value, including:

[0016] When the initial weight value of the current network layer reaches any weight value boundary, the weight adjustment value of the current network layer is quantized into the preset integer range to obtain an initial quantization value, and the initial quantization value is adjusted to a quantization value in the opposite direction to obtain the target quantization value, wherein the current network layer is the network layer of the weight adjustment value to be quantized;

[0017] When the initial weight value of the current network layer does not reach any weight value boundary, the weight adjustment value of the current network layer is quantized to the preset integer range to obtain an initial quantization value, and the initial quantization value is directly used as the target quantization value.

[0018] In a possible implementation, quantizing the weight adjustment value of the current network layer into the preset integer range to obtain an initial quantization value includes:

[0019] When the weight adjustment value of the current network layer is greater than 0, quantizing the weight adjustment value of the current network layer to a first integer value as the initial quantization value, wherein the first integer value is a preset positive integer, and the first integer value indicates that the weight adjustment value is adjusted in a positive increment direction;

[0020] When the weight adjustment value of the current network layer is equal to 0, 0 is used as the initial quantization value, and when the initial quantization value is 0, it indicates that the weight adjustment value is not adjusted;

[0021] When the weight adjustment value of the current network layer is less than 0, the weight adjustment value of the current network layer is quantized to a second integer value as the initial quantization value, wherein the second integer value is a preset negative integer, and the second integer value indicates that the weight adjustment value is adjusted in the reverse incremental direction.

[0022] In a possible implementation, each training sample in the training set includes an input vector;

[0023] The training set is input into the initial neural network model, and the predicted value of each training sample is output, including:

[0024] Inputting the training set into the input layer of the initial neural network model, and performing a matrix multiplication operation on the input vector in the training set and a matrix consisting of a plurality of initial weight values ​​of the input layer to obtain an output result of the input layer;

[0025] Inputting the output result of the input layer into the hidden layer of the initial neural network model, performing nonlinear transformation on the output result of the input layer through a V-type relu activation function, performing matrix multiplication operation on the output result after the nonlinear transformation and a matrix composed of multiple initial weight values ​​of the hidden layer, and obtaining the output result of the hidden layer;

[0026] The output result of the hidden layer is input into the output layer of the initial neural network model, and the output result of the hidden layer is normalized to obtain the probability distribution of each category, and the category with the largest probability is selected as the prediction value of the initial neural network model for the input vector.

[0027] In a second aspect, an embodiment of the present application provides a device for training a neural network model based on a storage-computing integrated chip, the device being applied to the storage-computing integrated chip, the device comprising:

[0028] An initialization unit, used to initialize the weight parameters of the initial neural network model and obtain the initial weight value of each network layer of the initial neural network model;

[0029] A conductance adjustment unit, used to adjust the conductance value of the storage-computing integrated chip according to the initial weight value of each network layer, wherein the weight value corresponding to the same weight parameter is positively correlated with the corresponding conductance value;

[0030] A prediction unit, used to input a training set into the initial neural network model and output a predicted value of each training sample, wherein the training set includes a plurality of the training samples and a true value corresponding to each of the training samples;

[0031] A determination unit, configured to determine a current loss value according to a difference between a predicted value and a true value of each of the training samples;

[0032] A calculation unit, used to calculate the error gradient of the current loss value to the weight parameter of each network layer, and determine the weight adjustment value of each network layer according to the error gradient of each network layer, wherein the weight adjustment value represents the weight change to be adjusted;

[0033] a quantization unit, configured to quantize the weight adjustment value into a preset integer range for each network layer according to the relationship between the initial weight value of the network layer and the weight value boundary, and obtain a target quantization value corresponding to the weight adjustment value, wherein different quantization values ​​within the preset integer range represent different adjustment directions;

[0034] A weight adjustment unit, configured to call a weight update interface for the initial weight value of each network layer, adjust the initial weight value along an adjustment direction represented by the target quantization value corresponding to the initial weight value, and obtain an adjusted weight value;

[0035] The conductance adjustment unit is used to adjust the conductance value of the storage-computation integrated chip according to the adjusted weight value of each network layer to obtain an updated neural network model;

[0036] An iterative training unit is used to continue training the training set based on the updated neural network model until the current loss value meets a preset convergence condition or reaches a preset threshold of iteration times, and the obtained neural network model is used as a target neural network model.

[0037] In a possible implementation, the quantization unit includes:

[0038] A quantization module, configured to quantize the weight adjustment value of the current network layer within the preset integer range to obtain an initial quantization value when the initial weight value of the current network layer reaches any weight value boundary, and quantize the weight adjustment value of the current network layer within the preset integer range to obtain an initial quantization value when the initial weight value of the current network layer does not reach any weight value boundary, wherein the current network layer is the network layer of the weight adjustment value to be quantized;

[0039] A setting module is used to adjust the initial quantization value to a quantization value in the opposite direction to obtain the target quantization value when the initial weight value of the current network layer reaches any weight value boundary; when the initial weight value of the current network layer does not reach any weight value boundary, directly use the initial quantization value as the target quantization value.

[0040] In a possible implementation manner, the quantization module is used to:

[0041] When the weight adjustment value of the current network layer is greater than 0, quantizing the weight adjustment value of the current network layer to a first integer value as the initial quantization value, wherein the first integer value is a preset positive integer, and the first integer value indicates that the weight adjustment value is adjusted in a positive increment direction;

[0042] When the weight adjustment value of the current network layer is equal to 0, 0 is used as the initial quantization value, and when the initial quantization value is 0, it indicates that the weight adjustment value is not adjusted;

[0043] When the weight adjustment value of the current network layer is less than 0, the weight adjustment value of the current network layer is quantized to a second integer value as the initial quantization value, wherein the second integer value is a preset negative integer, and the second integer value indicates that the weight adjustment value is adjusted in the reverse incremental direction.

[0044] In a possible implementation, each training sample in the training set includes an input vector;

[0045] The prediction unit comprises:

[0046] An input layer calculation module, used to input the training set into the input layer of the initial neural network model, and perform matrix multiplication operation on the input vector in the training set and the matrix composed of multiple initial weight values ​​of the input layer to obtain the output result of the input layer;

[0047] A hidden layer calculation module, used for inputting the output result of the input layer into the hidden layer of the initial neural network model, performing nonlinear transformation on the output result of the input layer through a V-type relu activation function, performing matrix multiplication operation on the output result after the nonlinear transformation and a matrix composed of multiple initial weight values ​​of the hidden layer, to obtain the output result of the hidden layer;

[0048] The output layer calculation module is used to input the output result of the hidden layer into the output layer of the initial neural network model, and normalize the output result of the hidden layer to obtain the probability distribution of each category, and select the category with the largest probability as the prediction value of the initial neural network model for the input vector.

[0049] In a third aspect, an embodiment of the present application provides a storage-computing integrated chip, which is at least used to implement the method described in any possible implementation method of the first aspect.

[0050] In a fourth aspect, an embodiment of the present application provides an electronic device, comprising at least one storage and computing integrated chip as described in the third aspect.

[0051] The method and device for training a neural network model based on a storage-computing integrated chip provided in an embodiment of the present application can train a neural network model on a storage-computing integrated chip. Specifically, the storage-computing integrated chip can initialize the weight parameters of the initial neural network model, obtain the initial weight value of each network layer of the initial neural network model, adjust the conductance value of the storage-computing integrated chip according to the initial weight value of each network layer, input the training set into the initial neural network model, output the predicted value of each training sample, and determine the current loss value according to the difference between the predicted value and the true value of each training sample, calculate the error gradient of the current loss value to the weight parameter of each network layer, and determine the error gradient of each network layer according to the error gradient of each network layer. After the weight adjustment value of the layer is obtained, the weight adjustment value can be quantized to a preset integer range for each network layer according to the relationship between the initial weight value of the network layer and the weight value boundary, and the target quantization value corresponding to the weight adjustment value can be obtained. Then the weight update interface is called to adjust the initial weight value along the adjustment direction represented by the target quantization value corresponding to the initial weight value to obtain the adjusted weight value, and the conductance value of the storage and computing integrated chip is adjusted according to the adjusted weight value of each network layer to obtain the updated neural network model. Iterative training is performed through the above method until the current loss value meets the preset convergence condition or reaches the preset iteration number threshold, and the final target neural network model required can be obtained. It can be seen from this that in the model training process, similar to adjusting the weights of the neural network by gradient descent, the conductance value of the storage and computing integrated array will also be adjusted according to the error feedback. By applying an appropriate voltage (such as a pulse), the conductance of the storage and computing integrated chip can be changed toward the desired value, thereby optimizing the performance of the neural network. Moreover, by quantizing the weight adjustment value to a preset integer range that only represents the adjustment direction, it can not only meet the integer characteristics of the storage and computing integrated chip, but also because the conductance value only needs to be updated according to the adjustment direction and does not require an accurate update value, it can significantly reduce the time for updating the weight link during subsequent on-chip training.

[0052] In addition, the technical effects that can be achieved by the embodiments of the present application include:

[0053] 1. When quantizing the weight adjustment value to a preset integer range, the embodiment of the present application does not directly quantize it, but first determines whether the initial weight value of each network layer reaches the weight value boundary. If the initial weight value of a network layer reaches any weight value boundary, the weight adjustment value of the network layer will be quantized to a preset integer range to obtain the initial quantization value, and then the initial quantization value will be adjusted in the opposite direction to obtain the final target quantization value, thereby realizing the weight constraint method of extreme value rebound, avoiding the problem of nonlinear and uncontrollable weight data, and thereby reducing the risk of chip damage.

[0054] 2. When training the weights of the neural network, in the process of forward propagation, when calculating the output part of each layer unit in the hidden layer, by adopting a new neuron activation function, namely the V-type relu function, it can be ensured that the weights will not shift in the negative direction as a whole after multiple updates, making full use of the representation range of the storage and computing integrated chip. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art description are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work.

[0056] Figure 1 A flowchart of a method for training a neural network model based on a storage-computing integrated chip provided in an embodiment of the present application;

[0057] Figure 2 A flowchart of another method for training a neural network model based on a storage-computing integrated chip provided in an embodiment of the present application;

[0058] Figure 3 An example diagram of the internal network layer structure of a neural network model provided in an embodiment of the present application;

[0059] Figure 4 A curve diagram of a V-type relu activation function provided in an embodiment of the present application;

[0060] Figure 5 A block diagram of the composition of a device for training a neural network model based on a storage and computing integrated chip provided in an embodiment of the present application. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0062] It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The terms "including" and "having" in the embodiments of the present application and the accompanying drawings and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices.

[0063] The embodiments of the present application can be applied to scenes of neural network models such as image recognition, text recognition, and speech recognition. For example, a face recognition model is used to recognize faces in images, a text recognition model is used to recognize text content in text, and a speech recognition model is used to recognize content in speech. The neural network model trained in the storage and computing integrated chip of the embodiments of the present application can be a neural network model in any scenario, and the embodiments of the present application do not limit this.

[0064] Combine the following Figure 1 and Figure 2 The method of training neural network models based on storage and computing integrated chips is elaborated in detail:

[0065] S110: Initialize the weight parameters of the initial neural network model, obtain the initial weight value of each network layer of the initial neural network model, and adjust the conductance value of the storage and computing integrated chip according to the initial weight value of each network layer.

[0066] Among them, the weight value corresponding to the same weight parameter is positively correlated with the corresponding conductance value.

[0067] In the early stage of initial neural network model training, the weight parameters can be randomly initialized to obtain the initial weight value of each network layer, and then the conductance value can be adjusted according to the positive correlation between the weight parameters and the conductance value of the storage and computing integrated chip.

[0068] In a neural network, weights usually represent the numerical value of the strength of the connection between neurons. The integrated storage and computing chip uses conductance to store and represent this "connection strength". Specifically, different conductance states of the integrated storage and computing chip (such as high conductance or low conductance) can represent different weight values ​​in the neural network:

[0069] A high conductance represents a stronger connection, similar to a larger weight;

[0070] A low conductance represents a weaker connection, similar to a smaller weight.

[0071] During the training process, the conductance of the storage and computing chip will be adjusted as the input signal changes, thereby updating the weights.

[0072] S120: Input the training set into the initial neural network model, output the predicted value of each training sample, and determine the current loss value based on the difference between the predicted value and the true value of each training sample.

[0073] The training set includes multiple training samples and a true value corresponding to each training sample, and in the embodiment of the present application, each training sample in the training set includes an input vector. The input vector can be a vector generated based on original data such as sample images, sample texts, sample voices, etc.

[0074] like Figure 2 As shown, the specific implementation process of inputting the training set into the initial neural network model and outputting the predicted value of each training sample includes: inputting the training set into the input layer of the initial neural network model, and performing matrix multiplication operation on the input vector in the training set and the matrix composed of multiple initial weight values ​​of the input layer to obtain the output result of the input layer; inputting the output result of the input layer into the hidden layer of the initial neural network model, and performing nonlinear transformation on the output result of the input layer through the V-type relu activation function, and performing matrix multiplication operation on the output result after the nonlinear transformation and the matrix composed of multiple initial weight values ​​of the hidden layer to obtain the output result of the hidden layer; inputting the output result of the hidden layer into the output layer of the initial neural network model, and normalizing the output result of the hidden layer to obtain the probability distribution of each category, and selecting the category with the largest probability as the predicted value of the initial neural network model for the input vector.

[0075] Among them, Figure 3 As shown, the neural network model includes an input layer, a hidden layer and an output layer, and the hidden layer can include at least one layer. In this figure, x represents input data, w represents weight parameters, z represents the result of weight matrix calculation of each layer, f() represents activation function, and y is the final output result.

[0076] like Figure 4 As shown in the figure, when training the weights of the neural network, in the process of forward propagation, a new neuron activation function, namely the V-type relu function, is used in the output part of each layer unit in the hidden layer. The details are as follows:

[0077] The hidden layer neuron activation function uses a variant of relu. The difference from relu is that the negative semi-axis slope is -1, forming a V-shaped function symmetrical on the Y axis, with the formula y=|x|. This activation function can ensure that the weights will not shift in the negative direction as a whole after multiple updates, making full use of the representation range of the storage and computing integrated chip.

[0078] After obtaining the predicted value of each training sample, the difference between the predicted value and the true value of each training sample can be calculated, and the mean square error of all the differences can be used as the current loss value. The cross entropy can also be used as the current loss value, or other algorithms can be used. The embodiment of the present application does not limit the method for calculating the loss value.

[0079] S130: Calculate the error gradient of the current loss value with respect to the weight parameters of each network layer, and determine the weight adjustment value of each network layer according to the error gradient of each network layer.

[0080] The weight adjustment value represents the weight change to be adjusted.

[0081] After obtaining the current loss value, the back propagation method (such as the chain rule) can be used to calculate the error gradient of the current loss value to the weight parameters of each network layer, and the optimization algorithm (such as SGD (Stochastic Gradient Descent) momentum, Adam (Adaptive Moment Estimation), etc.) can be used to determine the weight adjustment value of each network layer according to the error gradient of each network layer.

[0082] S140: For each network layer, according to the relationship between the initial weight value of the network layer and the weight value boundary, the weight adjustment value is quantized into a preset integer range to obtain a target quantization value corresponding to the weight adjustment value.

[0083] Among them, different quantization values ​​within the preset integer range represent different adjustment directions, including a positive increment direction, a reverse increment direction, and a 0 direction (ie, no adjustment, the value remains unchanged).

[0084] In the storage-computing integrated chip, the data is of int type, and the data range is usually [-8, 7]. When adjusting the weight parameters during neural network training, the write parameters may reach the extreme values ​​of -8 and 7. If a large number of extreme values ​​appear, there is a risk of damaging the chip. In addition, values ​​that exceed the allowable range of the storage-computing integrated chip will also cause the weight data to be nonlinear and uncontrollable. A neural network with a large number of extreme values ​​is not conducive to training. This is different from the wide dynamic range of floating-point numbers. For the training of such integer weights, it is necessary to limit the boundary constraint weights.

[0085] This weight constraint method can be called extreme value rebound, that is, when the adjusted weight is already an extreme value when updating the weight, the update forces the extreme value parameter to be adjusted in the opposite direction. Therefore, when the initial weight value of the current network layer reaches any weight value boundary, the weight adjustment value of the current network layer is quantized to a preset integer range to obtain the initial quantization value, and the initial quantization value is adjusted to the quantization value in the opposite direction to obtain the target quantization value, where the current network layer is the network layer of the current weight adjustment value to be quantized. For example, when the weight value of a certain position is already -8, when the weight is updated again, the adjustment direction is set to 1 when calling the weight update interface. When the weight parameter is 7, -1 is passed in.

[0086] When the initial weight value of the current network layer does not reach any weight value boundary, the weight adjustment value of the current network layer is quantized to a preset integer range to obtain an initial quantization value, and the initial quantization value is directly used as the target quantization value.

[0087] The specific implementation of “quantizing the weight adjustment value of the current network layer into a preset integer range to obtain an initial quantization value” may include:

[0088] (1) When the weight adjustment value of the current network layer is greater than 0, the weight adjustment value of the current network layer is quantized to a first integer value as an initial quantization value.

[0089] The first integer value is a preset positive integer, and the first integer value indicates that the weight adjustment value is adjusted along the positive increment direction. The first integer value can be 1, or 2, etc.

[0090] (2) When the weight adjustment value of the current network layer is equal to 0, 0 is used as the initial quantization value, and when the initial quantization value is 0, it means that the weight adjustment value is not adjusted.

[0091] (3) When the weight adjustment value of the current network layer is less than 0, the weight adjustment value of the current network layer is quantized to a second integer value as an initial quantization value.

[0092] The second integer value is a preset negative integer, and the second integer value indicates that the weight adjustment value is adjusted in the direction of the reverse increment. The second integer value can be -1, or -2, etc.

[0093] S150: For the initial weight value of each network layer, call the weight update interface, adjust the initial weight value along the adjustment direction represented by the target quantization value corresponding to the initial weight value, obtain the adjusted weight value, and adjust the conductance value of the storage and computing integrated chip according to the adjusted weight value of each network layer to obtain an updated neural network model.

[0094] The target quantization value only indicates the adjustment direction, and the specific adjustment amount can be random.

[0095] For example, assuming that the 2x2 weight matrix on the memristor array needs to be adjusted, the initial weight values ​​of the memristors are as follows:

[0096] [-8,5]

[0097] [-3,7]

[0098] The weight adjustment value of each network layer is determined according to the error gradient of each network layer:

[0099] [-0.03,0.0]

[0100] [0.01,0.05]

[0101] The initial quantized value obtained after the weight adjustment value is quantized to a preset integer range (including -1, 0, 1) is:

[0102] [-1,0]

[0103] [1,1]

[0104] Since there are weight value boundary values ​​in the initial weight values, the extreme value rebound method is used to process the initial quantization value to obtain the target quantization value, including:

[0105] [1,0]

[0106] [1,-1]

[0107] Pass the target quantized value processed by the extreme rebound constraint to the weight update interface for actual weight adjustment.

[0108] -1 decreases the weight, 1 increases the weight, and 0 remains unchanged. The operation is as follows (+1, -1 indicates the adjustment direction):

[0109] [-8+1,5+0]

[0110] [-3+1,7-1]

[0111] Due to the noise characteristics of the memristor itself, the +1 and -1 update values ​​do not represent a change of 1 unit value, but a random value change, but the direction is consistent, so the actual value obtained may be as follows:

[0112] [-4,5]

[0113] [-1,5].

[0114] The storage and computing chip supports a weight parameter storage precision of int4, that is, each storage unit can have 16 resistance states (or conductance states), which can only represent an integer in the int4 range (-8 to 7), and the two are linearly mapped. The weight of the memristor is updated through block programming, but only the update operation direction (increase or decrease the weight) is given, without giving the specific change value. The conductivity of an area is increased (set) and decreased (reset) at the same time to complete the programming of the storage and computing chip and realize the weight update of the memristor. The operation process of this block programming is implemented through software control. The software increases and decreases the conductivity value of the storage and computing chip in the entire area (set, reset), and only needs to be updated according to the change direction, without requiring accurate update values, which can significantly reduce the time of updating the weight link during subsequent on-chip training.

[0115] During the training process, similar to adjusting the weights of a neural network through gradient descent, the conductance value of the storage-computing chip is also adjusted based on the error feedback. By applying an appropriate voltage (such as a pulse), the conductance of the storage-computing chip can be changed toward the desired value, thereby optimizing the performance of the neural network.

[0116] It should be added that the reasons why the embodiment of the present application only focuses on the weight adjustment direction rather than the specific amount of data to be adjusted include the following:

[0117] 1. The optimization goal is the local minimum:

[0118] The goal of a neural network is to find a point that minimizes the loss function. In multidimensional space, this point is usually not unique, but a local minimum or a global minimum. We only focus on the direction of the gradient, because the gradient points to the direction where the loss function rises fastest, and we need to adjust the weights in the opposite direction of the gradient to reduce the loss.

[0119] 2. Working principle of gradient descent method:

[0120] The update rule of gradient descent is based on the gradient information of the weight to the loss function. Refer to the following formula and explanation:

[0121]

[0122] Among them, ω old is the weight value before adjustment, ω new is the adjusted weight value, η is the learning rate, is the question gradient of the current weight. The direction of the update is determined. The specific value of the weight will be adjusted according to the learning rate, but the most important thing is the direction of the gradient. As long as the direction is correct, the weight will be adjusted in the direction of minimizing the loss.

[0123] 3. Scale insensitivity:

[0124] Minimizing the loss function is not intrinsically dependent on the exact value of the weights, but rather on the direction in which the weights are adjusted. If the learning rate is appropriate, a slight adjustment that is too large or too small will not significantly affect the final result. The important thing is to update in the right direction.

[0125] 4. Dynamics of the training process:

[0126] During the training process, the weight values ​​will continue to change, and the weight update process focuses more on gradual adjustments, making the model more accurate after each update, without having to pay too much attention to the absolute value of each update. Excessive attention to specific values ​​may lead to inefficient model training or overfitting.

[0127] In general, the weight update in neural network training focuses on the adjustment direction, that is, adjusting the model parameters along the direction of the loss function, rather than the specific updated value. The goal of optimization is to gradually approach the minimum value of the loss function, and the update direction is the most important information in this process.

[0128] S160: Continue training the training set based on the updated neural network model until the current loss value meets the preset convergence condition or reaches the preset iteration number threshold, and use the obtained neural network model as the target neural network model.

[0129] The preset convergence condition includes but is not limited to: the current loss value is less than or equal to a preset loss threshold. The preset loss threshold and the preset iteration number threshold can both be empirical values.

[0130] The method for training a neural network model based on a storage-computing integrated chip provided in an embodiment of the present application can train a neural network model on a storage-computing integrated chip. Specifically, the storage-computing integrated chip can initialize the weight parameters of the initial neural network model, obtain the initial weight value of each network layer of the initial neural network model, adjust the conductance value of the storage-computing integrated chip according to the initial weight value of each network layer, and input the training set into the initial neural network model, output the predicted value of each training sample, and determine the current loss value according to the difference between the predicted value and the true value of each training sample, calculate the error gradient of the current loss value to the weight parameters of each network layer, and determine the error gradient of each network layer according to the error gradient of each network layer. After the weight adjustment value is obtained, the weight adjustment value can be quantized to a preset integer range for each network layer according to the relationship between the initial weight value and the weight value boundary of the network layer, and the target quantization value corresponding to the weight adjustment value can be obtained. Then, the weight update interface is called to adjust the initial weight value along the adjustment direction represented by the target quantization value corresponding to the initial weight value to obtain the adjusted weight value, and the conductance value of the storage-computing integrated chip is adjusted according to the adjusted weight value of each network layer to obtain the updated neural network model. Iterative training is performed through the above method until the current loss value meets the preset convergence condition or reaches the preset iteration number threshold, and the final target neural network model required can be obtained. It can be seen from this that in the model training process, similar to adjusting the weights of the neural network by gradient descent, the conductance value of the storage-computing integrated array will also be adjusted according to the error feedback. By applying an appropriate voltage (such as a pulse), the conductance of the storage-computing integrated chip can be changed toward the desired value, thereby optimizing the performance of the neural network. Moreover, by quantizing the weight adjustment value to a preset integer range that only represents the adjustment direction, not only the integer characteristics of the storage and computing integrated chip are met, but also because the conductance value only needs to be updated according to the adjustment direction and does not require an accurate update value, the time for updating the weight during subsequent on-chip training can be significantly reduced.

[0131] In addition, when quantizing the weight adjustment value to a preset integer range, the embodiment of the present application does not directly quantize it, but first determines whether the initial weight value of each network layer reaches the weight value boundary. If the initial weight value of a network layer reaches any weight value boundary, the weight adjustment value of the network layer will be quantized to a preset integer range to obtain the initial quantization value, and then the initial quantization value will be adjusted in the opposite direction to obtain the final required target quantization value, thereby realizing the weight constraint method of extreme value rebound, avoiding the problem of nonlinear uncontrollable weight data, and thus reducing the risk of chip damage. When performing weight training for neural network training, in the process of forward propagation, when calculating the output part of each layer unit in the hidden layer, by adopting a new neuron activation function, namely the V-type relu function, it can be ensured that the weight will not be offset in the negative direction as a whole after multiple updates, and the representation range of the storage and computing integrated chip is fully utilized.

[0132] Based on the above method embodiment, another embodiment of the present application provides a device for training a neural network model based on a storage-computing integrated chip, wherein the device is applied to a storage-computing integrated chip, such as Figure 5 As shown, the device comprises:

[0133] An initialization unit 210 is used to initialize the weight parameters of the initial neural network model and obtain the initial weight value of each network layer of the initial neural network model;

[0134] The conductance adjustment unit 220 is used to adjust the conductance value of the storage-computation integrated chip according to the initial weight value of each network layer, wherein the weight value corresponding to the same weight parameter is positively correlated with the corresponding conductance value;

[0135] A prediction unit 230, configured to input a training set into the initial neural network model and output a predicted value of each training sample, wherein the training set includes a plurality of the training samples and a true value corresponding to each of the training samples;

[0136] A determination unit 240, configured to determine a current loss value according to a difference between a predicted value and a true value of each of the training samples;

[0137] A calculation unit 250, configured to calculate an error gradient of the current loss value with respect to a weight parameter of each network layer, and determine a weight adjustment value of each network layer according to the error gradient of each network layer, wherein the weight adjustment value represents a weight change amount to be adjusted;

[0138] A quantization unit 260 is used to quantize the weight adjustment value into a preset integer range for each network layer according to the relationship between the initial weight value of the network layer and the weight value boundary, and obtain a target quantization value corresponding to the weight adjustment value, wherein different quantization values ​​within the preset integer range represent different adjustment directions;

[0139] The weight adjustment unit 270 is used to call a weight update interface for the initial weight value of each network layer, adjust the initial weight value along the adjustment direction represented by the target quantization value corresponding to the initial weight value, and obtain an adjusted weight value;

[0140] The conductance adjustment unit 220 is used to adjust the conductance value of the storage-computation integrated chip according to the adjusted weight value of each network layer to obtain an updated neural network model;

[0141] The iterative training unit 280 is used to continue training the training set based on the updated neural network model until the current loss value meets the preset convergence condition or reaches the preset iteration number threshold, and the obtained neural network model is used as the target neural network model.

[0142] In a possible implementation manner, the quantization unit 260 includes:

[0143] A quantization module, configured to quantize the weight adjustment value of the current network layer within the preset integer range to obtain an initial quantization value when the initial weight value of the current network layer reaches any weight value boundary, and quantize the weight adjustment value of the current network layer within the preset integer range to obtain an initial quantization value when the initial weight value of the current network layer does not reach any weight value boundary, wherein the current network layer is the network layer of the weight adjustment value to be quantized;

[0144] A setting module is used to adjust the initial quantization value to a quantization value in the opposite direction to obtain the target quantization value when the initial weight value of the current network layer reaches any weight value boundary; when the initial weight value of the current network layer does not reach any weight value boundary, directly use the initial quantization value as the target quantization value.

[0145] In a possible implementation manner, the quantization module is used to:

[0146] When the weight adjustment value of the current network layer is greater than 0, quantizing the weight adjustment value of the current network layer to a first integer value as the initial quantization value, wherein the first integer value is a preset positive integer, and the first integer value indicates that the weight adjustment value is adjusted in a positive increment direction;

[0147] When the weight adjustment value of the current network layer is equal to 0, 0 is used as the initial quantization value, and when the initial quantization value is 0, it indicates that the weight adjustment value is not adjusted;

[0148] When the weight adjustment value of the current network layer is less than 0, the weight adjustment value of the current network layer is quantized to a second integer value as the initial quantization value, wherein the second integer value is a preset negative integer, and the second integer value indicates that the weight adjustment value is adjusted in the reverse incremental direction.

[0149] In a possible implementation, each training sample in the training set includes an input vector;

[0150] The prediction unit comprises:

[0151] An input layer calculation module, used to input the training set into the input layer of the initial neural network model, and perform matrix multiplication operation on the input vector in the training set and the matrix composed of multiple initial weight values ​​of the input layer to obtain the output result of the input layer;

[0152] A hidden layer calculation module, used for inputting the output result of the input layer into the hidden layer of the initial neural network model, performing nonlinear transformation on the output result of the input layer through a V-type relu activation function, performing matrix multiplication operation on the output result after the nonlinear transformation and a matrix composed of multiple initial weight values ​​of the hidden layer, to obtain the output result of the hidden layer;

[0153] The output layer calculation module is used to input the output result of the hidden layer into the output layer of the initial neural network model, and normalize the output result of the hidden layer to obtain the probability distribution of each category, and select the category with the largest probability as the prediction value of the initial neural network model for the input vector.

[0154] The device for training a neural network model based on a storage-computing integrated chip provided in an embodiment of the present application can train a neural network model on a storage-computing integrated chip. Specifically, the storage-computing integrated chip can initialize the weight parameters of the initial neural network model, obtain the initial weight value of each network layer of the initial neural network model, adjust the conductance value of the storage-computing integrated chip according to the initial weight value of each network layer, and input the training set into the initial neural network model, output the predicted value of each training sample, and determine the current loss value according to the difference between the predicted value and the true value of each training sample, calculate the error gradient of the current loss value to the weight parameters of each network layer, and determine the error gradient of each network layer according to the error gradient of each network layer. After the weight adjustment value is obtained, the weight adjustment value can be quantized to a preset integer range for each network layer according to the relationship between the initial weight value and the weight value boundary of the network layer, and the target quantization value corresponding to the weight adjustment value can be obtained. Then, the weight update interface is called to adjust the initial weight value along the adjustment direction represented by the target quantization value corresponding to the initial weight value to obtain the adjusted weight value, and the conductance value of the storage-computing integrated chip is adjusted according to the adjusted weight value of each network layer to obtain the updated neural network model. Iterative training is performed through the above method until the current loss value meets the preset convergence condition or reaches the preset iteration number threshold, and the final target neural network model required can be obtained. It can be seen from this that in the model training process, similar to adjusting the weights of the neural network by gradient descent, the conductance value of the storage-computing integrated array will also be adjusted according to the error feedback. By applying an appropriate voltage (such as a pulse), the conductance of the storage-computing integrated chip can be changed toward the desired value, thereby optimizing the performance of the neural network. Moreover, by quantizing the weight adjustment value to a preset integer range that only represents the adjustment direction, not only the integer characteristics of the storage and computing integrated chip are met, but also because the conductance value only needs to be updated according to the adjustment direction and does not require an accurate update value, the time for updating the weight during subsequent on-chip training can be significantly reduced.

[0155] In addition, when quantizing the weight adjustment value to a preset integer range, the embodiment of the present application does not directly quantize it, but first determines whether the initial weight value of each network layer reaches the weight value boundary. If the initial weight value of a network layer reaches any weight value boundary, the weight adjustment value of the network layer will be quantized to a preset integer range to obtain the initial quantization value, and then the initial quantization value will be adjusted in the opposite direction to obtain the final required target quantization value, thereby realizing the weight constraint method of extreme value rebound, avoiding the problem of nonlinear uncontrollable weight data, and thus reducing the risk of chip damage. When performing weight training for neural network training, in the process of forward propagation, when calculating the output part of each layer unit in the hidden layer, by adopting a new neuron activation function, namely the V-type relu function, it can be ensured that the weight will not be offset in the negative direction as a whole after multiple updates, and the representation range of the storage and computing integrated chip is fully utilized.

[0156] Based on the above method embodiments, another embodiment of the present application provides a storage-computing integrated chip, which is at least used to implement the method described in any one of the embodiments.

[0157] Based on the above method embodiment, another embodiment of the present application provides an electronic device, which includes at least one storage and computing integrated chip as described above.

[0158] The above device embodiment corresponds to the method embodiment and has the same technical effect as the method embodiment. For specific description, please refer to the method embodiment. The device embodiment is obtained based on the method embodiment. For specific description, please refer to the method embodiment part, which will not be repeated here. It can be understood by those skilled in the art that the accompanying drawings are only schematic diagrams of one embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present application.

[0159] Those skilled in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be changed accordingly and located in one or more devices different from the embodiment. The modules in the above embodiment can be combined into one module, or can be further divided into multiple sub-modules.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for training a neural network model based on a storage-computing integrated chip, characterized in that: The method is applied to a storage-computing integrated chip, and the method includes: Initializing the weight parameters of the initial neural network model, obtaining the initial weight value of each network layer of the initial neural network model, and adjusting the conductance value of the storage-computing integrated chip according to the initial weight value of each network layer, wherein the weight value corresponding to the same weight parameter is positively correlated with the corresponding conductance value; Inputting a training set into the initial neural network model, outputting a predicted value of each training sample, and determining a current loss value according to a difference between the predicted value and the true value of each training sample, wherein the training set includes a plurality of the training samples and a true value corresponding to each of the training samples; Calculating the error gradient of the current loss value to the weight parameter of each network layer, and determining the weight adjustment value of each network layer according to the error gradient of each network layer, wherein the weight adjustment value represents the weight change to be adjusted; For each network layer, according to the relationship between the initial weight value and the weight value boundary of the network layer, the weight adjustment value is quantized into a preset integer range to obtain a target quantization value corresponding to the weight adjustment value, wherein different quantization values ​​within the preset integer range represent different adjustment directions respectively; For the initial weight value of each network layer, a weight update interface is called to adjust the initial weight value along the adjustment direction represented by the target quantization value corresponding to the initial weight value to obtain the adjusted weight value, and the conductance value of the storage-computing integrated chip is adjusted according to the adjusted weight value of each network layer to obtain an updated neural network model; The training set is continuously trained based on the updated neural network model until the current loss value satisfies a preset convergence condition or reaches a preset threshold of iteration times, and the obtained neural network model is used as the target neural network model.

2. The method according to claim 1, characterized in that For each network layer, according to the relationship between the initial weight value and the weight value boundary of the network layer, the weight adjustment value is quantized into a preset integer range to obtain a target quantization value corresponding to the weight adjustment value, including: When the initial weight value of the current network layer reaches any weight value boundary, the weight adjustment value of the current network layer is quantized into the preset integer range to obtain an initial quantization value, and the initial quantization value is adjusted to a quantization value in the opposite direction to obtain the target quantization value, wherein the current network layer is the network layer of the weight adjustment value to be quantized; When the initial weight value of the current network layer does not reach any weight value boundary, the weight adjustment value of the current network layer is quantized to the preset integer range to obtain an initial quantization value, and the initial quantization value is directly used as the target quantization value.

3. The method according to claim 2, characterized in that Quantizing the weight adjustment value of the current network layer into the preset integer range to obtain an initial quantization value, including: When the weight adjustment value of the current network layer is greater than 0, quantizing the weight adjustment value of the current network layer to a first integer value as the initial quantization value, wherein the first integer value is a preset positive integer, and the first integer value indicates that the weight adjustment value is adjusted in a positive increment direction; When the weight adjustment value of the current network layer is equal to 0, 0 is used as the initial quantization value, and when the initial quantization value is 0, it indicates that the weight adjustment value is not adjusted; When the weight adjustment value of the current network layer is less than 0, the weight adjustment value of the current network layer is quantized to a second integer value as the initial quantization value, wherein the second integer value is a preset negative integer, and the second integer value indicates that the weight adjustment value is adjusted in the reverse incremental direction.

4. The method according to any one of claims 1 to 3, characterized in that Each training sample in the training set includes an input vector; The training set is input into the initial neural network model, and the predicted value of each training sample is output, including: Inputting the training set into the input layer of the initial neural network model, and performing a matrix multiplication operation on the input vector in the training set and a matrix consisting of a plurality of initial weight values ​​of the input layer to obtain an output result of the input layer; Inputting the output result of the input layer into the hidden layer of the initial neural network model, performing nonlinear transformation on the output result of the input layer through a V-type relu activation function, performing matrix multiplication operation on the output result after the nonlinear transformation and a matrix composed of multiple initial weight values ​​of the hidden layer, and obtaining the output result of the hidden layer; The output result of the hidden layer is input into the output layer of the initial neural network model, and the output result of the hidden layer is normalized to obtain the probability distribution of each category, and the category with the largest probability is selected as the prediction value of the initial neural network model for the input vector.

5. A device for training a neural network model based on a storage and computing integrated chip, characterized in that: The device is applied to a storage-computing integrated chip, and the device includes: An initialization unit, used to initialize the weight parameters of the initial neural network model and obtain the initial weight value of each network layer of the initial neural network model; A conductance adjustment unit, used to adjust the conductance value of the storage-computing integrated chip according to the initial weight value of each network layer, wherein the weight value corresponding to the same weight parameter is positively correlated with the corresponding conductance value; A prediction unit, used to input a training set into the initial neural network model and output a predicted value of each training sample, wherein the training set includes a plurality of the training samples and a true value corresponding to each of the training samples; A determination unit, configured to determine a current loss value according to a difference between a predicted value and a true value of each of the training samples; A calculation unit, used to calculate the error gradient of the current loss value to the weight parameter of each network layer, and determine the weight adjustment value of each network layer according to the error gradient of each network layer, wherein the weight adjustment value represents the weight change to be adjusted; a quantization unit, configured to quantize the weight adjustment value into a preset integer range for each network layer according to the relationship between the initial weight value of the network layer and the weight value boundary, and obtain a target quantization value corresponding to the weight adjustment value, wherein different quantization values ​​within the preset integer range represent different adjustment directions; A weight adjustment unit, configured to call a weight update interface for the initial weight value of each network layer, adjust the initial weight value along an adjustment direction represented by the target quantization value corresponding to the initial weight value, and obtain an adjusted weight value; The conductance adjustment unit is used to adjust the conductance value of the storage-computation integrated chip according to the adjusted weight value of each network layer to obtain an updated neural network model; An iterative training unit is used to continue training the training set based on the updated neural network model until the current loss value meets a preset convergence condition or reaches a preset threshold of iteration times, and the obtained neural network model is used as a target neural network model.

6. The device according to claim 5, characterized in that The quantization unit comprises: A quantization module, configured to quantize the weight adjustment value of the current network layer within the preset integer range to obtain an initial quantization value when the initial weight value of the current network layer reaches any weight value boundary, and quantize the weight adjustment value of the current network layer within the preset integer range to obtain an initial quantization value when the initial weight value of the current network layer does not reach any weight value boundary, wherein the current network layer is the network layer of the weight adjustment value to be quantized; A setting module is used to adjust the initial quantization value to a quantization value in the opposite direction to obtain the target quantization value when the initial weight value of the current network layer reaches any weight value boundary; when the initial weight value of the current network layer does not reach any weight value boundary, directly use the initial quantization value as the target quantization value.

7. The device according to claim 6, characterized in that The quantization module is used for: When the weight adjustment value of the current network layer is greater than 0, quantizing the weight adjustment value of the current network layer to a first integer value as the initial quantization value, wherein the first integer value is a preset positive integer, and the first integer value indicates that the weight adjustment value is adjusted in a positive increment direction; When the weight adjustment value of the current network layer is equal to 0, 0 is used as the initial quantization value, and when the initial quantization value is 0, it indicates that the weight adjustment value is not adjusted; When the weight adjustment value of the current network layer is less than 0, the weight adjustment value of the current network layer is quantized to a second integer value as the initial quantization value, wherein the second integer value is a preset negative integer, and the second integer value indicates that the weight adjustment value is adjusted in the reverse incremental direction.

8. The device according to any one of claims 5 to 7, characterized in that: Each training sample in the training set includes an input vector; The prediction unit comprises: An input layer calculation module, used to input the training set into the input layer of the initial neural network model, and perform matrix multiplication operation on the input vector in the training set and the matrix composed of multiple initial weight values ​​of the input layer to obtain the output result of the input layer; A hidden layer calculation module, used for inputting the output result of the input layer into the hidden layer of the initial neural network model, performing nonlinear transformation on the output result of the input layer through a V-type relu activation function, performing matrix multiplication operation on the output result after the nonlinear transformation and a matrix composed of multiple initial weight values ​​of the hidden layer, to obtain the output result of the hidden layer; The output layer calculation module is used to input the output result of the hidden layer into the output layer of the initial neural network model, and normalize the output result of the hidden layer to obtain the probability distribution of each category, and select the category with the largest probability as the prediction value of the initial neural network model for the input vector.

9. A storage and computing integrated chip, characterized in that: The integrated storage and computing chip is at least used to implement the method as described in any one of claims 1-4.

10. An electronic device, characterized in that: The electronic device comprises at least one storage and computing integrated chip as described in claim 9.

Images