CN117725974A - Modularized neural network training platform based on multipath NPU - Google Patents

Abstract

The invention discloses a modularized neural network training platform based on multiple NPUs, which relates to the technical field of neural network training, and trains a neural network model to be trained by adopting a global optimal network parameter guiding algorithm, a parallel data optimizing algorithm and a parallel weight optimizing algorithm based on the multiple NPUs, so that the training effect of the neural network model is improved, the training time of the neural network is effectively shortened, and the training effect of the neural network is improved.

Description

A modular neural network training platform based on multi-channel NPU

Technical field

The invention relates to the technical fields of image recognition and neural networks, and in particular to a modular neural network training platform based on multi-channel NPU.

Background technique

With the emergence of massive data, artificial intelligence technology is developing rapidly. Machine learning is an inevitable product of the development of artificial intelligence to a certain stage. It is committed to mining valuable potential information from large amounts of data through computing methods.

In the existing technology, image recognition has been widely used, such as building image recognition neural networks for parts classification, defect identification, portrait recognition, etc. Before using these neural networks for image recognition, it is often necessary to train the neural networks. Neural network training refers to the training of artificial neural networks. Input enough samples into the neural network, and adjust the structure of the network (mainly adjusting the weights) through a certain algorithm so that the output of the network matches the expected value. This process is neural network training. The training process of the existing image recognition neural network mainly involves gradient descent training of the neural network multiple times to obtain a trained neural network. However, this process takes a long time and the training effect is poor, resulting in image The recognition effect is not good.

Contents of the invention

The purpose of the present invention is to provide a modular neural network training platform based on multi-channel NPU, which solves the problems existing in the existing technology.

The present invention is realized through the following technical solutions:

A modular neural network training platform based on multi-channel NPU, including training data and NPU data acquisition module, division module, first core initialization module, first update module, second core initialization module, second update module, and loop module and output modules;

The training data and NPU data acquisition module is used to obtain the neural network model to be trained, and determine the idle cores in the multi-channel NPU to obtain multiple target cores;

The dividing module is used to divide the plurality of target cores into two parts to obtain a first core group and a second core group. The first core group is used for first parallel training, and the second core group For second parallel training;

The first core initialization module is used to deploy a plurality of neural network models to be trained on each first core in the first core group, obtain a first neural network model to be trained, and provide a plurality of first neural network models to be trained respectively. The network model generates initial network parameter values, obtains the first network parameter vector corresponding to the first neural network model to be trained, and completes the initialization of the first core;

The first update module is used to run all the first cores simultaneously in a parallel operation manner, and for each first neural network model to be trained in the first core, adopt a global optimal network parameter guidance algorithm and a parallel data optimization algorithm. The first network parameter vector of each first to-be-trained neural network model is first updated until the number of first updates reaches the upper limit, and the local optimal network parameter vector in each first core is determined based on the first network parameter vector. ;

The second core initialization module is used to deploy multiple neural network models to be trained on the second core in the second core group to obtain a second neural network model to be trained, based on the local optimal network parameter vector in the first core , use the reduction method to determine the global optimal network parameter vector, and use the global optimal network parameter vector as the initial network parameters of the second neural network model to be trained on the second core to obtain the second neural network model corresponding to the second to be trained. Network parameter vector to complete the initialization of the second core;

The second update module is used to run all second cores simultaneously in a parallel operation manner, and for each second neural network model to be trained in the second core, use a parallel weight optimization algorithm to update each second neural network model to be trained. The second network parameter vector of the network model is updated for the second time until the number of second updates reaches the upper limit or the iteration end requirements are met, and the second update is ended;

The loop module is configured to, when the number of second updates reaches the upper limit, use the second network parameter vector in the second core as the global optimal network parameter vector, and return to the first update step;

The output module is used to, when iteration end requirements are met, use the second network parameter vector in the second core as the final network parameters of the neural network to be trained, and complete the training of the neural network model to be trained.

In a possible implementation, generating initial network parameter values for multiple first neural network models to be trained includes: generating the first to be trained using a method of randomly generating network parameters between the upper limit of the network parameter and the lower limit of the network parameter. Network parameter values of the neural network model, or a mixed sequence strategy is used to generate network parameter values of the first neural network model to be trained.

In a possible implementation, a globally optimal network parameter guidance algorithm and a parallel data optimization algorithm are used to perform a first update on the first network parameter vector of each first neural network model to be trained until the number of first updates is reached. The upper limit is to determine the local optimal network parameter vector in each first core according to the first network parameter vector, including:

A1. Obtain the first fitness corresponding to the first network parameter vector of the first neural network model to be trained in parallel;

A2. Determine the first historical optimal value of each first network parameter vector according to the first fitness, and obtain the first historical optimal value and the first fitness corresponding to the first historical optimal value;

A3. According to the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, use the reduction method to determine the local optimal network parameter vector in the first core;

A4. Determine whether the first maximum update times threshold has been reached. If so, output the local optimal network parameter vector in the first core. Otherwise, use the global optimal network parameter guidance algorithm to update the first network parameter vector and return to step A1. .

In a possible implementation, a reduction method is used to determine the local optimal network parameter vector in the first core based on the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, including:

According to the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, n first fitnesses are obtained;

Compare the i-th first fitness with the n-i-th first fitness, select the larger first fitness to enter the next round of comparison, until the number of selected first fitnesses is only one, and the first core is obtained The local optimal network parameter vector in ; among them, when n is an odd number, the middle number directly enters the next round of comparison.

In a possible implementation, using the global optimal network parameter guidance algorithm to update the first network parameter vector includes:

Performing the first guided update on the i-th first network parameter vector is:

Among them, f(t) represents the random number between [0,1] generated during the t-th local training, and all f(t) are uniformly distributed, and X _i (t) represents the i-th random number during the t-th local training. The first network parameter vector, i=1,2,...,L, L represents the total number of the first network parameter vector, X _i (t+1) represents the updated X _i (t); X _{i, best} (t) Indicates the state of maximum fitness corresponding to the i-th first network parameter vector in the previous t local training processes, that is, the historical optimal value of the i-th first network parameter vector; represents the average value corresponding to the historical optimal values of all first network parameter vectors; X _best represents the global optimal value. During the first cycle, each first core runs independently at this time and does not know the global optimal value. Therefore, the local optimal network parameter vector in the first core is regarded as the global optimal value; ω(t) represents the inertia weight during the t-th local training;

Divide all the first network parameter vectors into two parts, one part is K better network parameter vectors, and the other part is K worse network parameter vectors; the K better network parameter vectors are the first half with the largest fitness Network parameter vector, when the first network parameter vector is an odd number, the first network parameter vector with the largest fitness does not participate in segmentation;

According to K optimal network parameter vectors, the network parameter center vector is obtained as:

Among them, C(t) represents the network parameter center vector, X _k (t) represents the kth better network parameter, and F(X _k (t) represents the fitness of the kth better network parameter;

According to the network parameter center vector C(t), the second guided update is performed on the inferior network parameter vector as:

X _k' (t+1)＝X _k' (t)+rand*(C(t)-X _k' (t))

Among them, X _k' (t) represents the k'th worse network parameter vector, X _k' (t+1) represents the updated X _k' (t), and rand represents a random number between (0,1) ;

Determine whether the fitness of the inferior network parameter vector after the second guidance update is greater than the fitness before the second guidance update. If so, accept the second guidance update of the inferior network parameter vector X _k' (t) to complete the first An update of the network parameter vector, otherwise the second guided update of the inferior network parameter vector X _k' (t) is rejected to complete the update of the first network parameter vector.

In a possible implementation, the inertia weight is:

Among them, ω _max represents the maximum value of the inertia weight, ω _min represents the minimum value of the inertia weight, iter represents the maximum number of times to update the first network parameter vector, and t represents the current number of times to update the first network parameter vector.

In a possible implementation, a parallel weight optimization algorithm is used to perform a second update on the second network parameter vector of each second neural network model to be trained until the number of second updates reaches the upper limit or the iteration end requirements are met, including :

B1. Build a task node for each second neural network model to be trained on the second core, build a main task node, and divide private memory and a public memory for each task node;

B2. Initialize the iteration counter t'=1, initialize the scale matrix B _t' =I;

B3. For all the second neural network models to be trained, obtain the error function value E(w _t' ) corresponding to the second neural network model to be trained, and put the error function value E(w _t' ) into the common memory, To be called by each task node; among them, the error function value E(w _t' ) is obtained through the second neural network model to be trained on any task node, w _t' represents the second neural network to be trained during the t'th training. second network parameter vector;

B4. Obtain the gradient matrix of the second neural network model to be trained. And put the scale matrix B _t' and the gradient matrix g _t' into the public memory for call by each task node; among them, each column in the gradient matrix g _t' is a network parameter of the second neural network model to be trained. gradient;

Among them, I represents the unit scale matrix, represents the gradient operator, E(w _t' ) represents the error function, w _t' represents the weight vector corresponding to the second neural network model to be trained, B _t' is an M-order square matrix, and M represents the second neural network model to be trained. weight dimension;

B5. Each task node runs in parallel, and calls the scale matrix B _t' and gradient matrix g _t' from the common memory for each task node to determine the search direction d _m of the second to-be-trained neural network on the m-th task node as the scale. The multiplication and accumulation data of the m-th row data in the matrix B _t' and the m-th column data in the gradient matrix g _t' ; where, m=1,2,...,M, the search direction d _m is stored in the m-th task node corresponding to in private memory;

B6. Each task node runs in parallel, and the search direction d _m is retrieved from its private memory for each task node. According to the search direction d _m and the golden section method is used, the mth in the second neural network model to be trained on the task node is determined. The update step size λ _m corresponding to each network _parameter is stored in the private memory corresponding to the mth task node;

B7. Each task node runs in parallel, and the update step size λ _m is retrieved from its private memory for each task node, and the m-th network parameter in the second neural network model to be trained on the task node is updated according to the update step size λ _m . Update; obtain the m-th gradient component based on the m-th updated network parameter of the second to-be-trained neural network model on the m-th task node;

B9. Each task node runs in parallel. According to the m-th gradient component, the gradient matrix g _m corresponding to the m-th task node is re-determined, and the scale matrix B _t' corresponding to the m-th task node is calculated based on the gradient matrix g _m . renew;

B10. Output the m-th column element in the scale matrix B _t' and the m-th network parameter corresponding to the second neural network model to be trained through each task node to the main task node;

B11. Use the main task node to construct the m network parameters into an updated second network parameter vector, and determine whether the fitness corresponding to the updated second network parameter vector is greater than the set threshold. If so, the iteration end requirements are met. And end the second update process, otherwise enter step B12;

B12. Determine whether the count value of the iteration counter t' is greater than the upper limit. If so, determine that the number of second updates has reached the upper limit, and end the second update process. Otherwise, increase the count value of the iteration counter t' by one, and add each task node to The elements in the mth column of the output scale matrix B _m form the scale matrix B _t' , use the updated second network parameter vector as the network parameters of the second neural network model to be trained on all task nodes, and return to step B3.

In a possible implementation, obtaining the error function value E(w _t' ) corresponding to the second neural network model to be trained is:

Among them, s=1,2,...,S, S represents the total number of training samples, n=1,2,...,N, N represents the total number of output neurons corresponding to the second neural network model to be trained; y _n (x _s , w _t' ) represents the actual output of the nth output neuron of the second neural network model to be trained when it takes the training sample x _s as input and w _t' as the network parameter; d _sn represents The expected output corresponding to y _n (x _s ,w _t' ); d _max,n represents the maximum value of the nth expected output corresponding to all training samples, and each training sample corresponds to an expected output vector; d _min,n represents The minimum value of the nth expected output corresponding to all training samples.

In a possible implementation, the update step size λ _m corresponding to the m-th network parameter in the second neural network model to be trained on the task node is determined based on the search direction d _m and using the golden section method, including:

Use the search direction d _m to construct a search matrix D with only one row. The m-th element in the search matrix is the search direction d _m , and the other elements are 0. The total number of elements of the search matrix is the same as the total number of network parameters of the second neural network model to be trained. same;

Construct a step matrix λ' with only one column. The m-th element in the step matrix is the update step λ _m to be determined, and the other elements are 0. The total number of elements of the step matrix is the same as that of the second neural network model to be trained. The total number of network parameters is the same

Construct solution conditions based on search matrix D and search matrix D And solve the solution condition to determine the update step size λ _m ;

According to the update step size λ _m , the m-th network parameter in the second neural network model to be trained on the task node is updated as:

w _m '=w _m +λ _m d _m

Among them, w _m represents the m-th network parameter in the second neural network model to be trained, and w _m ' represents the updated w _m .

In a possible implementation, the gradient matrix g _m corresponding to the m -th task node is re-determined based on the m -th gradient component, and the scale matrix B _t' corresponding to the m -th task node is calculated based on the gradient matrix g _m Make updates including:

Based on the gradient matrix g _t' , use the m-th gradient component to update the m-th element in the gradient matrix g _t' , and obtain the gradient matrix g _m corresponding to the m-th task node;

According to the gradient matrix g _m corresponding to the m-th task node, the scale matrix B _t' corresponding to the m-th task node is updated as:

Among them, B _t' ' represents the updated B _t' , W represents the weight difference matrix, W=w _t' '-w _t' , w _t'' represents the second to be trained updated according to the update step size λ _m The weight vector of the neural network model, w _t' represents the weight vector of the second neural network model to be trained before updating, T represents the transpose, Z=g _m -g _t' , and Z represents the gradient difference matrix.

The invention provides a modular neural network training platform based on multi-channel NPU, which is based on multi-channel NPU and adopts the global optimal network parameter guidance algorithm, parallel data optimization algorithm and parallel weight optimization algorithm to train the neural network model to be trained. , not only improves the training effect of the neural network model, but also effectively reduces the training time of the neural network, improves the training effect of the neural network, thereby improving the accuracy of image recognition.

Description of the drawings

In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention. Therefore, it should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic structural diagram of a multi-channel NPU-based modular neural network training platform provided by an embodiment of the present invention.

Marks and corresponding parts names in the attached drawings:

1-Training data and NPU data acquisition module, 2-Division module, 3-First core initialization module, 4-First update module, 5-Second core initialization module, 6-Second update module, 7-Cycle module, 8-Output module.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with the examples and drawings. The schematic embodiments of the present invention and their descriptions are only used to explain the present invention and do not as a limitation of the invention.

As shown in Figure 1, a modular neural network training platform based on multi-channel NPU (Neural Processor Unit, neural network processor) includes training data and NPU data acquisition module 1, division module 2, and first core initialization module 3 , the first update module 4, the second core initialization module 5, the second update module 6, the loop module 7 and the output module 8.

The training data and NPU data acquisition module 1 is used to obtain the neural network model to be trained, determine the idle cores in the multi-channel NPU, and obtain multiple target cores.

Optionally, the training data and NPU data acquisition module 1 is also used to acquire training data, which is used to train the neural network model to be trained.

The division module 2 is used to divide multiple target cores into two parts to obtain a first core group and a second core group. The first core group is used for first parallel training, and the second core group is used for second parallel training. .

Since the data calculation amount of the second core group is less than the data calculation amount of the first core group, the cores can be divided according to the actual situation, for example, the first number of cores in the first core group and the second number of cores in the second core group. The ratio of the number of cores can be 7:3 or 8:2, etc.

Optionally, when dividing core groups, you can try to avoid multiple cores on one NPU being located in different core groups.

The first core initialization module 3 is configured to deploy a plurality of neural network models to be trained on each first core in the first core group, obtain a first neural network model to be trained, and provide a plurality of first neural networks to be trained respectively. The model generates initial network parameter values, obtains the first network parameter vector corresponding to the first neural network model to be trained, and completes the initialization of the first core.

Optionally, the upper limit of the network parameter threshold and the lower limit of the network parameter threshold can be set. Between the upper limit of the network parameter threshold and the lower limit of the network parameter threshold, a first network parameter vector corresponding to the first neural network model to be trained is generated in a randomly generated manner, The vector here can be understood as a single-row matrix including all network parameters of the first neural network model to be trained. An initialization method similar to chaos sequence may also be used to initialize the first network parameter vector corresponding to the first neural network model to be trained.

The first update module 4 is configured to run all first cores simultaneously in a parallel operation manner, and for each first neural network model to be trained in the first core, use a global optimal network parameter guidance algorithm and a parallel data optimization algorithm to The first network parameter vector of each first neural network model to be trained is first updated until the number of first updates reaches the upper limit, and the local optimal network parameter vector in each first core is determined according to the first network parameter vector.

In the existing technology, when training image recognition neural networks (such as when training convolutional neural networks for product classification or portrait recognition), traditional optimization algorithms not only need to process data serially, but also have the problem of easily falling into local optima. This leads to poor training effect of the neural network and poor final image recognition effect. Therefore, this embodiment uses the global optimal network parameter guidance algorithm and the parallel data optimization algorithm to perform the first update of the first network parameter vector of each first neural network model to be trained, which not only optimizes the training effect of the neural network, but also effectively Improved training efficiency.

The second core initialization module 5 is used to deploy multiple neural network models to be trained on the second core in the second core group to obtain the second neural network model to be trained. According to the local optimal network parameter vector in the first core, Use the reduction method to determine the global optimal network parameter vector, and use the global optimal network parameter vector as the initial network parameters of the second neural network model to be trained on the second core to obtain the second network corresponding to the second neural network model to be trained. Parameter vector to complete the initialization of the second core.

The local optimal network parameter vector in the first core is the network parameter vector obtained by running multiple first cores in parallel. It can be regarded as having been trained multiple times, but has not reached the optimal position. Therefore, the local optimal network parameter can be The optimal value in the vector is used as the second network parameter vector, and a parallel weight optimization algorithm is used to update the second network parameter vector of each second neural network model to be trained to further optimize the network parameters and achieve the expected training effect while maintaining a high training efficiency.

The second update module 6 is configured to run all second cores simultaneously in a parallel operation manner, and for each second neural network model to be trained in the second core, use a parallel weight optimization algorithm to update each second neural network model to be trained. The second network parameter vector of the model is updated for the second time until the number of second updates reaches the upper limit or the iteration end requirement is met, and the second update is ended.

In the process of training the image recognition network, traditional algorithms often construct kernel functions to calculate data, and repeatedly schedule a kernel function to perform operations through serial operations to achieve data updates. This method is simple and easy to implement, but The efficiency is low, so this embodiment uses a parallel weight optimization algorithm to perform a second update on the second network parameter vector of each second neural network model to be trained.

The loop module 7 is configured to, when the number of second updates reaches the upper limit, use the second network parameter vector in the second core as the global optimal network parameter vector, and return to the first update step.

The output module 8 is used to, when the iteration end requirements are met, use the second network parameter vector in the second core as the final network parameters of the neural network to be trained, and complete the training of the neural network model to be trained.

In a possible implementation, generating initial network parameter values for multiple first neural network models to be trained includes: generating the first to be trained using a method of randomly generating network parameters between the upper limit of the network parameter and the lower limit of the network parameter. Network parameter values of the neural network model, or a mixed sequence strategy is used to generate network parameter values of the first neural network model to be trained.

First, the global optimal network parameter guidance algorithm is used to search, and then a local optimal network parameter vector is determined from each first core, and then the global optimal in this overall cycle is determined based on all local optimal network parameter vectors. Network parameter vector, and then use a parallel weight optimization algorithm to conduct a fine search for the global optimal network parameter vector to obtain the network parameters with target accuracy. If the target accuracy is not reached after a certain number of fine searches, the next overall cycle will be entered.

In a possible implementation, a globally optimal network parameter guidance algorithm and a parallel data optimization algorithm are used to perform a first update on the first network parameter vector of each first neural network model to be trained until the number of first updates is reached. The upper limit is to determine the local optimal network parameter vector in each first core according to the first network parameter vector, including:

A1. Obtain the first fitness corresponding to the first network parameter vector of the first neural network model to be trained in parallel.

Optionally, the fitness function of the first neural network model to be trained can be the negative value of its error function. For example, when the first neural network model to be trained does not have a BP neural network, its error function can be: E represents the error, p=1,2,…,P, P represents the total number of input data, q=1,2,…,Q, Q represents the total output of the BP neural network;/> represents the actual output corresponding to the q-th output neuron of the BP neural network when the p-th data is input, y' _pq represents the expected output corresponding to the q-th output neuron of the BP neural network when the p-th data is input; then the fitness The function can be/>

A2. Determine the first historical optimal value of each first network parameter vector according to the first fitness, and obtain the first historical optimal value and the first fitness corresponding to the first historical optimal value.

The first historical optimal value also participates in subsequent updates. When the first network parameter vector is updated for the first time, its historical optimal value is itself.

A3. According to the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, use the reduction method to determine the local optimal network parameter vector in the first core.

A4. Determine whether the first maximum update times threshold has been reached. If so, output the local optimal network parameter vector in the first core. Otherwise, use the global optimal network parameter guidance algorithm to update the first network parameter vector and return to step A1. .

In a possible implementation, a reduction method is used to determine the local optimal network parameter vector in the first core based on the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, including:

According to the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, n first fitnesses are obtained.

Compare the i-th first fitness with the n-i-th first fitness, select the larger first fitness to enter the next round of comparison, until the number of selected first fitnesses is only one, and the first core is obtained The local optimal network parameter vector in . Among them, when n is an odd number, the middle number directly enters the next round of comparison.

Determining the local optimal network parameter vector in the first core through the reduction method can effectively reduce the amount of data comparison and save computing resources.

In a possible implementation, using the global optimal network parameter guidance algorithm to update the first network parameter vector includes:

Performing the first guided update on the i-th first network parameter vector is:

Among them, f(t) represents the random number between [0,1] generated during the t-th local training, and all f(t) are uniformly distributed, and X _i (t) represents the i-th random number during the t-th local training. The first network parameter vector, i=1,2,...,L, L represents the total number of the first network parameter vector, and X _i (t+1) represents the updated X _i (t). X _i,best (t) represents the state of maximum fitness corresponding to the i-th first network parameter vector in the previous t local training processes, that is, the historical optimal value of the i-th first network parameter vector. Represents the average value corresponding to the historical optimal values of all first network parameter vectors. X _best represents the global optimal value. During the first cycle, each first core runs independently and does not know the global optimal value. Therefore, the local optimal network parameter vector in the first core is regarded as the global optimal value. value. ω(t) represents the inertia weight during the t-th local training.

The first guidance update provided by this embodiment improves the local development capability in the late iteration period and prevents the algorithm from prematurely converging and falling into a local optimum.

Divide all first network parameter vectors into two parts, one part is K better network parameter vectors, and the other part is K worse network parameter vectors. The K better network parameter vectors are half of the first network parameter vectors with the greatest fitness. When the first network parameter vector is an odd number, the first network parameter vector with the greatest fitness does not participate in the segmentation.

According to K optimal network parameter vectors, the network parameter center vector is obtained as:

Among them, C(t) represents the network parameter center vector, X _k (t) represents the kth better network parameter, and F(X _k (t) represents the fitness of the kth better network parameter.

According to the network parameter center vector C(t), the second guided update is performed on the inferior network parameter vector as:

X _k' (t+1)＝X _k' (t)+rand*(C(t)-X _k' (t))

Among them, X _k' (t) represents the k'th worse network parameter vector, X _k' (t+1) represents the updated X _k' (t), and rand represents a random number between (0,1) .

Determine whether the fitness of the inferior network parameter vector after the second guidance update is greater than the fitness before the second guidance update. If so, accept the second guidance update of the inferior network parameter vector X _k' (t) to complete the first An update of the network parameter vector, otherwise the second guided update of the inferior network parameter vector X _k' (t) is rejected to complete the update of the first network parameter vector.

High fitness can prove that the position of the network parameter vector is better, and it has certain reference value for other network parameter vectors with low fitness. Therefore, the K better network parameter vectors can be used as a reference, and the K worse network parameter vectors can be used as a reference. Update to speed up the update of the entire network parameter population, and use the survival of the fittest mechanism to update the inferior network parameter vector, which can avoid the development of inferior networks.

Optionally, after each boot update, the network parameters can be processed out of bounds to avoid abnormal data.

In a possible implementation, the inertia weight is:

Among them, ω _max represents the maximum value of the inertia weight, ω _min represents the minimum value of the inertia weight, iter represents the maximum number of times to update the first network parameter vector, and t represents the current number of times to update the first network parameter vector.

By setting the inertia weight related to the number of updates, the local search ability and global search ability are balanced, and the overall efficiency of the algorithm is improved. Although the first update performs local search and global search, its accuracy may not be enough, so the second update can be performed to further improve the search accuracy and ensure the overall training effect.

In a possible implementation, a parallel weight optimization algorithm is used to perform a second update on the second network parameter vector of each second neural network model to be trained until the number of second updates reaches the upper limit or the iteration end requirements are met, including :

B1. Build a task node for each second neural network model to be trained on the second core, build a main task node, and divide private memory and a public memory for each task node.

B2. Initialize the iteration counter t'=1, and initialize the scale matrix B _t' =I.

B3. For all the second neural network models to be trained, obtain the error function value E(w _t' ) corresponding to the second neural network model to be trained, and put the error function value E(w _t' ) into the common memory, To be called by each task node. Among them, the error function value E(w _t' ) is obtained through the second neural network model to be trained on any task node, and w _t' represents the second network parameter vector applied by the second neural network to be trained during the t'th training.

B4. Obtain the gradient matrix of the second neural network model to be trained. And put the scale matrix B _t' and the gradient matrix g _t' into the common memory for call by each task node. Among them, each column in the gradient matrix g _t' is the gradient of a network parameter in the second neural network model to be trained.

Among them, I represents the unit scale matrix, represents the gradient operator, E(w _t' ) represents the error function, w _t' represents the weight vector corresponding to the second neural network model to be trained, B _t' is an M-order square matrix, and M represents the second neural network model to be trained. Weight dimension.

B5. Each task node runs in parallel, and calls the scale matrix B _t' and gradient matrix g _t' from the common memory for each task node to determine the search direction d _m of the second to-be-trained neural network on the m-th task node as the scale. The multiplication and accumulation data of the m-th row data in matrix B _t' and the m-th column data in gradient matrix g _t' . Among them, m=1,2,...,M, and the search direction _dm is stored in the private memory corresponding to the mth task node.

B6. Each task node runs in parallel, and the search direction d _m is retrieved from its private memory for each task node. According to the search direction d _m and the golden section method is used, the mth in the second neural network model to be trained on the task node is determined. The update step size λ _m corresponding _to each network parameter is stored in the private memory corresponding to the mth task node.

B7. Each task node runs in parallel, and the update step size λ _m is retrieved from its private memory for each task node, and the m-th network parameter in the second neural network model to be trained on the task node is updated according to the update step size λ _m . Make an update. Obtain the mth gradient component according to the mth updated network parameter of the second to-be-trained neural network model on the mth task node.

B9. Each task node runs in parallel. According to the m-th gradient component, the gradient matrix g _m corresponding to the m-th task node is re-determined, and the scale matrix B _t' corresponding to the m-th task node is calculated based on the gradient matrix g _m . renew.

B10. Output the m-th column element in the scale matrix B _t' and the m-th network parameter corresponding to the second to-be-trained neural network model to the main task node through each task node.

B11. Use the main task node to construct the m network parameters into an updated second network parameter vector, and determine whether the fitness corresponding to the updated second network parameter vector is greater than the set threshold. If so, the iteration end requirements are met. And end the second update process, otherwise enter step B12.

B12. Determine whether the count value of the iteration counter t' is greater than the upper limit. If so, determine that the number of second updates has reached the upper limit, and end the second update process. Otherwise, increase the count value of the iteration counter t' by one, and add each task node to The elements in the mth column of the output scale matrix B _m form the scale matrix B _t' , use the updated second network parameter vector as the network parameters of the second neural network model to be trained on all task nodes, and return to step B3.

In a possible implementation, obtaining the error function value E(w _t' ) corresponding to the second neural network model to be trained is:

Among them, s=1,2,…,S, S represents the total number of training samples, n=1,2,…,N, and N represents the total number of output neurons corresponding to the second neural network model to be trained. y _n (x _s ,w _t' ) means that when the second neural network model to be trained takes the training sample x _s as the input and w _t' as the network parameter, the nth output neuron of the second neural network model to be trained is actual output. d _sn represents the expected output corresponding to y _n (x _s ,w _t' ). d _max,n represents the maximum value of the nth expected output corresponding to all training samples, and each training sample corresponds to an expected output vector. d _min,n represents the minimum value of the nth expected output corresponding to all training samples.

In a possible implementation, the update step size λ _m corresponding to the m-th network parameter in the second neural network model to be trained on the task node is determined based on the search direction d _m and using the golden section method, including:

Use the search direction d _m to construct a search matrix D with only one row. The m-th element in the search matrix is the search direction d _m , and the other elements are 0. The total number of elements of the search matrix is the same as the total number of network parameters of the second neural network model to be trained. same.

Construct a step matrix λ' with only one column. The m-th element in the step matrix is the update step λ _m to be determined, and the other elements are 0. The total number of elements of the step matrix is the same as that of the second neural network model to be trained. The total number of network parameters is the same

Construct solution conditions based on search matrix D and search matrix D And solve the solution condition to determine the update step size λ _m .

According to the update step size λ _m , the m-th network parameter in the second neural network model to be trained on the task node is updated as:

w _m '=w _m +λ _m d _m

Among them, w _m represents the m-th network parameter in the second neural network model to be trained, and w _m ' represents the updated w _m .

In a possible implementation, the gradient matrix g _m corresponding to the m -th task node is re-determined based on the m -th gradient component, and the scale matrix B _t' corresponding to the m -th task node is calculated based on the gradient matrix g _m Make updates including:

Based on the gradient matrix g _t' , the m-th gradient component is used to update the m-th element in the gradient matrix g _t' to obtain the gradient matrix g _m corresponding to the m-th task node.

Calculating gradients using traditional methods will consume a lot of computing resources and cannot guarantee high accuracy. For example, if you want to use V sets of training data to train a neural network, then V work items are needed to calculate the objective function. The weight of the neural network structure has C elements. If you use the traditional gradient definition to calculate the gradient of the weight, you need to call C times. The objective function calculates the partial derivative of each element of the weight, and the computing efficiency is greatly reduced. Therefore, this embodiment adopts another gradient matrix updating method, which includes: dividing the degree of parallelism according to the training data, each task node calculating the partial derivative of the weight based on a set of training data, and finally obtaining the partial derivative of the objective function. Assume that there are V training data, and the second neural network structure weight has C elements. When the m-th network parameter in the second neural network model to be trained is updated, the updated network parameters will be synchronized to all nodes, and then each The second neural network to be trained uses only one training data to obtain the partial derivatives of the weight, and then reduces and sums the partial derivatives of the weight to obtain the gradient matrix g _m .

The method of reduction and summation is: for M partial derivatives corresponding to the same network parameter, add the m-th partial derivative and the M-m-th partial derivative, and put the addition result into the next round of addition data. There is only one data left in the next round of addition data, and the final partial derivative corresponding to a network parameter is obtained. Among them, when n is an odd number, the intermediate partial derivatives directly enter the next round of addition.

When data integration and data communication are required, the main task node sends and receives data to other task nodes. Although this application adds a process of data sending and receiving, it reduces the serial calls of functions and effectively improves training efficiency.

According to the gradient matrix g _m corresponding to the m-th task node, the scale matrix B _t' corresponding to the m-th task node is updated as:

Among them, B _t' ' represents the updated B _t' , W represents the weight difference matrix, W=w _t' '-w _t' , w _t'' represents the second to be trained updated according to the update step size λ _m The weight vector of the neural network model, w _t' represents the weight vector of the second neural network model to be trained before updating, T represents the transpose, Z=g _m -g _t' , and Z represents the gradient difference matrix.

The above specific embodiments further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Within the spirit and principles of the present invention, any modifications, equivalent substitutions, improvements, etc. shall be included in the protection scope of the present invention.

Claims (10)

1. A modular neural network training platform based on multi-channel NPU, which is characterized in that it includes a training data and NPU data acquisition module, a dividing module, a first core initialization module, a first update module, a second core initialization module, and a third core initialization module. 2. Update module, cycle module and output module; The training data and NPU data acquisition module is used to obtain the neural network model to be trained, and determine the idle cores in the multi-channel NPU to obtain multiple target cores; The dividing module is used to divide the plurality of target cores into two parts to obtain a first core group and a second core group. The first core group is used for first parallel training, and the second core group For second parallel training; The first core initialization module is used to deploy a plurality of neural network models to be trained on each first core in the first core group, obtain a first neural network model to be trained, and provide a plurality of first neural network models to be trained respectively. The network model generates initial network parameter values, obtains the first network parameter vector corresponding to the first neural network model to be trained, and completes the initialization of the first core; The first update module is used to run all the first cores simultaneously in a parallel operation manner, and for each first neural network model to be trained in the first core, adopt a global optimal network parameter guidance algorithm and a parallel data optimization algorithm. The first network parameter vector of each first to-be-trained neural network model is first updated until the number of first updates reaches the upper limit, and the local optimal network parameter vector in each first core is determined based on the first network parameter vector. ; The second core initialization module is used to deploy multiple neural network models to be trained on the second core in the second core group to obtain a second neural network model to be trained, based on the local optimal network parameter vector in the first core , use the reduction method to determine the global optimal network parameter vector, and use the global optimal network parameter vector as the initial network parameters of the second neural network model to be trained on the second core to obtain the second neural network model corresponding to the second to be trained. Network parameter vector to complete the initialization of the second core; The second update module is used to run all second cores simultaneously in a parallel operation manner, and for each second neural network model to be trained in the second core, use a parallel weight optimization algorithm to update each second neural network model to be trained. The second network parameter vector of the network model is updated for the second time until the number of second updates reaches the upper limit or the iteration end requirements are met, and the second update is ended; The loop module is configured to, when the number of second updates reaches the upper limit, use the second network parameter vector in the second core as the global optimal network parameter vector, and return to the first update step; The output module is used to, when iteration end requirements are met, use the second network parameter vector in the second core as the final network parameters of the neural network to be trained, and complete the training of the neural network model to be trained. 2. The modular neural network training platform based on multi-channel NPU according to claim 1, characterized in that generating initial network parameter values for a plurality of first neural network models to be trained includes: using the upper limit of the network parameter and The method of randomly generating network parameters between the lower limits of the network parameters generates the network parameter values of the first neural network model to be trained, or using a mixed sequence strategy to generate the network parameter values of the first neural network model to be trained. 3. The modular neural network training platform based on multi-channel NPU according to claim 1, characterized in that, a global optimal network parameter guidance algorithm and a parallel data optimization algorithm are used to train the first neural network model of each first neural network model to be trained. A network parameter vector is first updated until the number of first updates reaches the upper limit, and the local optimal network parameter vector in each first core is determined based on the first network parameter vector, including: A1. Obtain the first fitness corresponding to the first network parameter vector of the first neural network model to be trained in parallel; A2. Determine the first historical optimal value of each first network parameter vector according to the first fitness, and obtain the first historical optimal value and the first fitness corresponding to the first historical optimal value; A3. According to the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, use the reduction method to determine the local optimal network parameter vector in the first core; A4. Determine whether the first maximum update times threshold has been reached. If so, output the local optimal network parameter vector in the first core. Otherwise, use the global optimal network parameter guidance algorithm to update the first network parameter vector and return to step A1. . 4. The modular neural network training platform based on multi-channel NPU according to claim 3, characterized in that, according to the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, the reduction method is adopted. Determine the local optimal network parameter vector in the first core, including: According to the first fitness corresponding to the first historical optimal value corresponding to the first network parameter vector, n first fitnesses are obtained; Compare the i-th first fitness with the n-i-th first fitness, select the larger first fitness to enter the next round of comparison, until the number of selected first fitnesses is only one, and the first core is obtained The local optimal network parameter vector in ; among them, when n is an odd number, the middle number directly enters the next round of comparison. 5. The modular neural network training platform based on multi-channel NPU according to claim 3, characterized in that using the global optimal network parameter guidance algorithm to update the first network parameter vector includes: Performing the first guided update on the i-th first network parameter vector is: Among them, f(t) represents the random number between [0,1] generated during the t-th local training, and all f(t) are uniformly distributed, and X _i (t) represents the i-th random number during the t-th local training. The first network parameter vector, i=1,2,...,L, L represents the total number of the first network parameter vector, X _i (t+1) represents the updated X _i (t); X _{i, best} (t) Indicates the state of maximum fitness corresponding to the i-th first network parameter vector in the previous t local training processes, that is, the historical optimal value of the i-th first network parameter vector; represents the average value corresponding to the historical optimal values of all first network parameter vectors; X _best represents the global optimal value. During the first cycle, each first core runs independently at this time and does not know the global optimal value. Therefore, the local optimal network parameter vector in the first core is regarded as the global optimal value; ω(t) represents the inertia weight during the t-th local training; Divide all the first network parameter vectors into two parts, one part is K better network parameter vectors, and the other part is K worse network parameter vectors; the K better network parameter vectors are the first half with the largest fitness Network parameter vector, when the first network parameter vector is an odd number, the first network parameter vector with the largest fitness does not participate in segmentation; According to K optimal network parameter vectors, the network parameter center vector is obtained as: Among them, C(t) represents the network parameter center vector, X _k (t) represents the kth better network parameter, and F(X _k (t) represents the fitness of the kth better network parameter; According to the network parameter center vector C(t), the second guided update is performed on the inferior network parameter vector as: X _k' (t+1)＝X _k' (t)+rand*(C(t)-X _k' (t)) Among them, X _k' (t) represents the k'th worse network parameter vector, X _k' (t+1) represents the updated X _k' (t), and rand represents a random number between (0,1) ; Determine whether the fitness of the inferior network parameter vector after the second guidance update is greater than the fitness before the second guidance update. If so, accept the second guidance update of the inferior network parameter vector X _k' (t) to complete the first An update of the network parameter vector, otherwise the second guided update of the inferior network parameter vector X _k' (t) is rejected to complete the update of the first network parameter vector. 6. The modular neural network training platform based on multi-channel NPU according to claim 5, characterized in that the inertia weight is: Among them, ω _max represents the maximum value of the inertia weight, ω _min represents the minimum value of the inertia weight, iter represents the maximum number of times to update the first network parameter vector, and t represents the current number of times to update the first network parameter vector. 7. The modular neural network training platform based on multi-channel NPU according to claim 5, characterized in that a parallel weight optimization algorithm is used to perform a second update on the second network parameter vector of each second neural network model to be trained. , until the number of second updates reaches the upper limit or meets the iteration end requirements, including: B1. Build a task node for each second neural network model to be trained on the second core, build a main task node, and divide private memory and a public memory for each task node; B2. Initialize the iteration counter t'=1, initialize the scale matrix B _t' =I; B3. For all the second neural network models to be trained, obtain the error function value E(w _t' ) corresponding to the second neural network model to be trained, and put the error function value E(w _t' ) into the common memory, To be called by each task node; among them, the error function value E(w _t' ) is obtained through the second neural network model to be trained on any task node, w _t' represents the second neural network to be trained during the t'th training. second network parameter vector; B4. Obtain the gradient matrix of the second neural network model to be trained. And put the scale matrix B _t' and the gradient matrix g _t' into the public memory for call by each task node; among them, each column in the gradient matrix g _t' is a network parameter of the second neural network model to be trained. gradient; Among them, I represents the unit scale matrix, represents the gradient operator, E(w _t' ) represents the error function, w _t' represents the weight vector corresponding to the second neural network model to be trained, B _t' is an M-order square matrix, and M represents the second neural network model to be trained. weight dimension; B5. Each task node runs in parallel, and calls the scale matrix B _t' and gradient matrix g _t' from the common memory for each task node to determine the search direction d _m of the second to-be-trained neural network on the m-th task node as the scale. The multiplication and accumulation data of the m-th row data in the matrix B _t' and the m-th column data in the gradient matrix g _t' ; where, m=1,2,...,M, the search direction d _m is stored in the m-th task node corresponding to in private memory; B6. Each task node runs in parallel, and the search direction d _m is retrieved from its private memory for each task node. According to the search direction d _m and the golden section method is used, the mth in the second neural network model to be trained on the task node is determined. The update step size λ _m corresponding to each network _parameter is stored in the private memory corresponding to the mth task node; B7. Each task node runs in parallel, and the update step size λ _m is retrieved from its private memory for each task node, and the m-th network parameter in the second neural network model to be trained on the task node is updated according to the update step size λ _m . Update; obtain the m-th gradient component based on the m-th updated network parameter of the second to-be-trained neural network model on the m-th task node; B9. Each task node runs in parallel. According to the m-th gradient component, the gradient matrix g _m corresponding to the m-th task node is re-determined, and the scale matrix B _t' corresponding to the m-th task node is calculated based on the gradient matrix g _m . renew; B10. Output the m-th column element in the scale matrix B _t' and the m-th network parameter corresponding to the second neural network model to be trained through each task node to the main task node; B11. Use the main task node to construct the m network parameters into an updated second network parameter vector, and determine whether the fitness corresponding to the updated second network parameter vector is greater than the set threshold. If so, the iteration end requirements are met. And end the second update process, otherwise enter step B12; B12. Determine whether the count value of the iteration counter t' is greater than the upper limit. If so, determine that the number of second updates has reached the upper limit, and end the second update process. Otherwise, increase the count value of the iteration counter t' by one, and add each task node to The elements in the mth column of the output scale matrix B _m form the scale matrix B _t' , use the updated second network parameter vector as the network parameters of the second neural network model to be trained on all task nodes, and return to step B3. 8. The modular neural network training platform based on multi-channel NPU according to claim 7, characterized in that, obtaining the error function value E(w _t' ) corresponding to the second neural network model to be trained is: Among them, s=1,2,...,S, S represents the total number of training samples, n=1,2,...,N, N represents the total number of output neurons corresponding to the second neural network model to be trained; y _n (x _s , w _t' ) represents the actual output of the nth output neuron of the second neural network model to be trained when it takes the training sample x _s as input and w _t' as the network parameter; d _sn represents The expected output corresponding to y _n (x _s ,w _t' ); d _max,n represents the maximum value of the nth expected output corresponding to all training samples, and each training sample corresponds to an expected output vector; d _min,n represents The minimum value of the nth expected output corresponding to all training samples. 9. The modular neural network training platform based on multi-channel NPU according to claim 7, characterized in that, according to the search direction _dm and using the golden section method, the mth in the second neural network model to be trained on the task node is determined The update step size λ _m corresponding to each network parameter includes: Use the search direction d _m to construct a search matrix D with only one row. The m-th element in the search matrix is the search direction d _m , and the other elements are 0. The total number of elements of the search matrix is the same as the total number of network parameters of the second neural network model to be trained. same; Construct a step matrix λ' with only one column. The m-th element in the step matrix is the update step λ _m to be determined, and the other elements are 0. The total number of elements of the step matrix is the same as that of the second neural network model to be trained. The total number of network parameters is the same Construct solution conditions based on search matrix D and search matrix D And solve the solution condition to determine the update step size λ _m ; According to the update step size λ _m , the m-th network parameter in the second neural network model to be trained on the task node is updated as: w _m '=w _m +λ _m d _m Among them, w _m represents the m-th network parameter in the second neural network model to be trained, and w _m ' represents the updated w _m . 10. The modular neural network training platform based on multi-channel NPU according to claim 7, characterized in that, according to the m-th gradient component, the gradient matrix g _m corresponding to the m-th task node is re-determined, and according to the gradient The matrix g _m updates the scale matrix B _t' corresponding to the m-th task node, including: Based on the gradient matrix g _t' , use the m-th gradient component to update the m-th element in the gradient matrix g _t' , and obtain the gradient matrix g _m corresponding to the m-th task node; According to the gradient matrix g _m corresponding to the m-th task node, the scale matrix B _t' corresponding to the m-th task node is updated as: Among them, B _t′ ′ represents the updated B _t′ , W represents the weight difference matrix, W=w _t′ ′-w _t′ , w _{t′ ′} represents the second to be trained updated according to the update step size λ _m The weight vector of the neural network model, w _t' represents the weight vector of the second neural network model to be trained before updating, T represents the transpose, Z=g _m -g _t' , and Z represents the gradient difference matrix.