Neural network model construction method based on heterogeneous chip energy consumption and related device

Abstract

The application discloses a neural network model construction method based on heterogeneous chip energy consumption and related devices, and the method comprises the following steps: testing the operator energy consumption of model operators of different algorithm models on different heterogeneous chips by using preset test cases to obtain a model operator energy consumption table, wherein the model operator energy consumption table comprises an operator name, a chip model, an execution time and an energy consumption amount of the model operator; constructing a reinforcement learning data set according to the model operators in the model operator energy consumption table; searching for optimal energy consumption operators in the reinforcement learning data set based on a preset strategy network and a preset reward function according to a reinforcement learning strategy, and constructing an optimal energy consumption neural network model. The neural network model constructed according to the optimal energy consumption operators can fundamentally reduce operation energy consumption, and the model accuracy can also be ensured by optimizing and combining the operators. The application can solve the technical problem that the existing model compression and light-weight processing technologies cannot fundamentally reduce operation energy consumption while ensuring model accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural networks, in particular to a neural network model construction method based on energy consumption of heterogeneous chips and related devices. BACKGROUND

[0002] The energy consumption of intelligent AI chips is limited by the rated power, especially the lightweight intelligent chips deployed at the edge, which have very limited energy consumption resources for model operation. In this case, the algorithm usually needs to reduce the operation energy consumption as much as possible while meeting the accuracy requirement. There are two common practices: 1) compressing the model to reduce the operation consumption generated by one inference; 2) using a relatively mature small lightweight model, since the small model has smaller energy consumption, it is suitable for deployment on chips with limited energy consumption resources.

[0003] However, model compression technology cannot sufficiently reduce energy consumption. In order to meet the excessive pruning of energy consumption requirements, under the premise of energy consumption resource limitation, more network structure damage is brought, and the model accuracy will also decrease; while the lightweight model, although the model size is small, when the model runs on a heterogeneous chip, it may still cause the instantaneous floating point operation per second to be too high due to the underlying operation mode, causing the instantaneous power to exceed the chip power consumption limit, thereby causing the chip to overload and overheat, and even causing a crash. SUMMARY

[0004] The present application provides a neural network model construction method based on energy consumption of heterogeneous chips and related devices, to solve the technical problem that the existing model compression and lightweight processing technology cannot fundamentally reduce the operation energy consumption while ensuring the model accuracy.

[0005] Therefore, the first aspect of the present application provides a neural network model construction method based on energy consumption of heterogeneous chips, comprising:

[0006] testing the operator energy consumption of the model operators of different algorithm models on different heterogeneous chips by using preset test cases, to obtain a model operator energy consumption table, the model operator energy consumption table including the operator name, chip model, execution time and energy consumption of the model operator;

[0007] constructing a reinforcement learning data set according to the model operators in the model operator energy consumption table;

[0008] based on a reinforcement learning strategy, searching for an optimal energy consumption operator in the reinforcement learning data set according to a preset policy network and a preset reward function, and constructing an optimal energy consumption neural network model.

[0009] Preferably, the testing of the operator energy consumption of the model operators of different algorithm models on different heterogeneous chips by using preset test cases, to obtain a model operator energy consumption table, further comprises:

[0010] extracting structural hyperparameters of different algorithm models, the structural hyperparameters including the number of network layers, the number of neurons, and the size of convolution kernels.

[0011] Preferably, the optimal energy consumption algorithm is searched in the reinforcement learning dataset based on a preset policy network and a preset reward function, and the optimal energy consumption neural network model is constructed based on the reinforcement learning strategy.

[0012] The model operator is defined as a state based on the reinforcement learning strategy, and a state space is formed.

[0013] A set of executable operation actions is defined at each state to form an action space.

[0014] The energy consumption algorithm is iteratively searched in the reinforcement learning dataset based on the state space, the action space, and a preset reward function by the preset policy network with the optimal energy consumption as a constraint condition until a stop condition is reached, and the optimal energy consumption algorithm is obtained.

[0015] The optimal energy consumption neural network model is constructed according to the optimal energy consumption algorithm.

[0016] Preferably, the optimal energy consumption algorithm is searched in the reinforcement learning dataset based on a preset policy network and a preset reward function, and the optimal energy consumption neural network model is constructed based on the reinforcement learning strategy, and further includes:

[0017] The policy network is constructed based on a multilayer perceptron to obtain a preset policy network.

[0018] The second aspect of the present application provides a neural network model construction device based on energy consumption of heterogeneous chips, comprising:

[0019] The energy consumption test unit is used to test the operator energy consumption of the model operator of different algorithm models on different heterogeneous chips by using a preset test case, and a model operator energy consumption table is obtained, the model operator energy consumption table including the operator name, chip model, execution time, and energy consumption of the model operator.

[0020] The dataset construction unit is used to construct a reinforcement learning dataset according to the model operator in the model operator energy consumption table.

[0021] The optimal searching unit is used to search the optimal energy consumption algorithm in the reinforcement learning dataset based on a preset policy network and a preset reward function based on a reinforcement learning strategy, and construct an optimal energy consumption neural network model.

[0022] Preferably, it further includes:

[0023] The operator extraction unit is configured to extract structural hyperparameters of different algorithm models, and generate model operators, wherein the structural hyperparameters include the number of network layers, the number of neurons, and the size of convolution kernels.

[0024] Preferably, the optimal search unit is specifically configured to:

[0025] The model operator is defined as a state based on a reinforcement learning strategy, and a state space is formed.

[0026] A set of executable operation actions is defined at each state, and an action space is formed.

[0027] The energy consumption operator is iteratively searched in the reinforcement learning data set based on the state space, the action space, and a preset reward function by using a preset policy network as a constraint condition of energy consumption optimization, until a stop condition is reached, and an optimal energy consumption operator is obtained.

[0028] An optimal energy consumption neural network model is constructed according to the optimal energy consumption operator.

[0029] Preferably, the method further comprises:

[0030] The network construction unit is configured to construct a policy network based on a multi-layer perceptron, and obtain a preset policy network.

[0031] The third aspect of the present application provides a neural network model construction device based on energy consumption of a heterogeneous chip, the device comprising a processor and a memory;

[0032] The memory is configured to store program code and transmit the program code to the processor.

[0033] The processor is configured to execute the neural network model construction method based on energy consumption of a heterogeneous chip according to instructions in the program code.

[0034] The fourth aspect of the present application provides a computer readable storage medium for storing program code, wherein the program code is used to execute the neural network model construction method based on energy consumption of a heterogeneous chip.

[0035] As can be seen from the above technical solutions, the embodiments of the present application have the following advantages:

[0036] In the present application, a neural network model construction method based on heterogeneous chip energy consumption is provided, which includes: testing the operator energy consumption of model operators of different algorithm models on different heterogeneous chips by using preset test cases to obtain a model operator energy consumption table, the model operator energy consumption table including the operator name, chip model, execution time and energy consumption of the model operator; constructing a reinforcement learning data set according to the model operators in the model operator energy consumption table; searching for an optimal energy consumption operator in the reinforcement learning data set based on a preset strategy network and a preset reward function based on a reinforcement learning strategy, and constructing an optimal energy consumption neural network model.

[0037] The neural network model construction method based on heterogeneous chip energy consumption provided in the present application splits a model into multiple model operators, tests the energy consumption of different model operators on different heterogeneous chips by using preset test cases, searches for an optimal energy consumption operator based on an energy consumption optimization strategy, and constructs a neural network model based on the optimal energy consumption operator, which can fundamentally reduce operation energy consumption and ensure model accuracy through optimized combination of operators. Therefore, the present application can solve the technical problem that existing model compression and lightweight processing technologies cannot fundamentally reduce operation energy consumption while ensuring model accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 A flowchart of the neural network model construction method based on heterogeneous chip energy consumption provided for the embodiments of the present application is shown in the figure.

[0039] Figure 2 A structure diagram of the neural network model construction device based on heterogeneous chip energy consumption provided for the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0040] In order to enable personnel in the technical field to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0041] In order to facilitate understanding, please refer to Figure 1 The embodiments of the neural network model construction method based on heterogeneous chip energy consumption provided in the present application include:

[0042] Step 101, test the operator energy consumption of model operators of different algorithm models on different heterogeneous chips by using preset test cases to obtain a model operator energy consumption table, the model operator energy consumption table including the operator name, chip model, execution time and energy consumption of the model operator.

[0043] Further, step 101 further includes:

[0044] The structural hyperparameters of different algorithm models are extracted to generate model operators, and the structural hyperparameters include the number of network layers, the number of neurons, and the size of convolution kernels.

[0045] It should be noted that, in order to compare the energy consumption of different model operators on the heterogeneous chips, a set of standard test cases, i.e., preset test cases, are defined in this embodiment. The specific test cases include image data to be processed by the model operators and various codes for implementing image data processing. This set of test cases can cover the test requirements of all types of model operators and meet the test needs of different input data.

[0046] The structural parameters and network hyperparameters of different algorithm models can all be used as model operators for testing. In addition to the number of network layers, the number of neurons, and the size of convolution kernels mentioned in this embodiment, other algorithm model-related parameters can also be extracted to generate model operators, and the specifics are not limited. Specifically, to implement energy consumption testing, each model operator needs to implement corresponding codes on different heterogeneous chips, and the energy consumption of each model operator is recorded. In addition, since the heterogeneous chips have different computing characteristics, the energy consumption of different algorithm model operators on different heterogeneous chips is different, so the energy consumption of the model operators with obvious differences in statistics is very effective for modeling.

[0047] The operator energy consumption test mainly measures the execution time of the model operator during the implementation of the preset test case using a timer, measures the energy consumption of each model operator on the heterogeneous chip using an energy consumption measurement tool, and then constructs a model operator energy consumption table based on the obtained information. It can be known from the table that the execution time and energy consumption of each model operator on a specific type of heterogeneous chip; therefore, the model operator energy consumption table can be used to select and combine different model operators, select more efficient, less energy-consuming, and more accurate operators for a specific task, and construct an optimal neural network model. In addition, if the energy consumption measurement tool is a software tool, the workload, frequency change, and the like of the CPU and GPU can be directly monitored to estimate the energy consumption.

[0048] Step 102, constructing a reinforcement learning data set according to the model operators in the model operator energy consumption table.

[0049] It can be understood that the reinforcement learning data set is used in the subsequent reinforcement learning optimization process, and the purpose is to find the model operator with the optimal energy consumption and the most suitable performance in the reinforcement learning data set, and to construct a model based on this. The chip model, execution time, and energy consumption of a model operator in the model operator energy consumption table are a group of data, and multiple groups of data form a reinforcement learning data set.

[0050] Step 103, based on the reinforcement learning strategy, searching for the optimal energy consumption operator in the reinforcement learning data set according to the preset policy network and the preset reward function, and constructing an optimal energy consumption neural network model.

[0051] Further, step 103 comprises:

[0052] Defining the model operator as a state based on the reinforcement learning strategy, forming a state space;

[0053] Defining a set of executable operation actions in each state, forming an action space;

[0054] Performing iterative search for the energy consumption operator in the reinforcement learning data set based on the state space, the action space and the preset reward function through the preset policy network with the optimal energy consumption as a constraint condition, until a stop condition is reached, to obtain the optimal energy consumption operator;

[0055] Constructing an optimal energy consumption neural network model according to the optimal energy consumption operator.

[0056] Further, step 103 further comprises:

[0057] Constructing the policy network based on the multilayer perceptron to obtain the preset policy network.

[0058] It should be noted that the executable operation actions in the action space refer to the selection operations such as increasing or decreasing the number of layers and the number of neurons in the model operator. The optimal energy consumption as a constraint condition can also consider the degree of adaptation of the operator performance to the target task, which is a point that can be concerned in specific application scenarios. The preset policy network takes the current state as input and outputs a probability distribution for expressing the probability of selecting each action under the current state; and the preset policy network can be constructed based on the multilayer perceptron, or other types of network structures can be selected, which are only examples and are not limited. The preset reward function is negatively related to the energy consumption of the neural network structure, so it is generally defined as a negative index value.

[0059] In addition, the iterative optimization learning task formulated based on the reinforcement learning strategy can be optimized and trained in advance based on the loss function. First, data related to energy consumption optimization can be collected, including energy consumption data, performance data and information related to operator selection of different neural network structures, such as operator parameter settings. Then, prepare the training data set, including input features and target labels, the input features include heterogeneous chip types, operator types, test cases, etc., and the target labels include energy consumption data and performance data, etc. Then, according to the defined policy network, the input features are received and the probability distribution of selecting each action (i.e. selecting operator parameters) under the current state is output; the loss function is constructed based on the preset reward function and the energy consumption penalty term, and by setting appropriate weight parameters λ, the optimization objectives of energy consumption and performance are balanced. The loss function is defined as:

[0060] L = -R + l max(0, E - E limint )

[0061] Wherein, R is a reward term, that is, a preset reward function, used to measure the performance of the neural network structure, and the value can be the accuracy of the model on the validation set or other task-related evaluation indicators, and is also the target to be maximized in the embodiment; λ is a weight parameter for balancing the reward term and the energy consumption, and a larger λ will emphasize more on reducing energy consumption, and a smaller λ will emphasize more on improving performance; E and E limint are the total energy consumption and the energy consumption upper limit of the neural network respectively; the total energy consumption is the target to be minimized, and the energy consumption upper limit is the limited value of the energy consumption resource. max(0, E - E limint ) represents an energy consumption penalty term, which will add a positive value to the loss function if the total energy consumption exceeds the energy consumption upper limit, so as to ensure that the energy consumption is limited within an acceptable range during the optimization process. The loss function can update the parameters of the policy network during the reinforcement learning training process, so that the policy network can find an operator with good performance and low energy consumption under the constraint of energy consumption optimization, and construct an optimal energy consumption neural network model.

[0062] During the training process, the parameters of the policy network are constantly updated through interaction with the environment, so that the policy network can find a neural network structure with good performance and energy consumption not exceeding the limit under the constraint of energy consumption optimization. In each training step, the operator parameters are selected according to the output probability distribution of the policy network based on the current parameters of the policy network. The selected operator parameters are applied to the neural network structure, and the energy consumption is measured on the actual device or simulator. The energy consumption data and performance data are transmitted to the policy network as rewards and returns to update the parameters of the policy network to optimize energy consumption and performance. In addition, in order to balance exploration and utilization during the training process, an exploration strategy such as ε-greedy can be used to balance randomness and selection according to the policy network. Moreover, the training process can be terminated according to the preset stopping condition, for example, reaching a certain number of training rounds or reaching a certain performance level.

[0063] The neural network model construction method based on energy consumption of heterogeneous chips provided in the embodiments of the present application splits the model into multiple model operators, tests the energy consumption of different model operators on different heterogeneous chips through preset test cases, performs reinforcement learning search based on the energy consumption optimization strategy, and finds the optimal energy consumption operator. The neural network model constructed according to the optimal energy consumption operator can fundamentally reduce the operation energy consumption, and the optimized combination of operators can also ensure the model accuracy. Therefore, the embodiments of the present application can solve the technical problem that the existing model compression and lightweight processing technologies cannot fundamentally reduce the operation energy consumption while ensuring the model accuracy.

[0064] For ease of understanding, please refer to Figure 2The application provides an embodiment of a neural network model construction device based on energy consumption of a heterogeneous chip, which comprises the following steps of:

[0065] An energy consumption test unit 201 is configured to test operator energy consumption of model operators of different algorithm models on different heterogeneous chips by using preset test cases, and obtain a model operator energy consumption table, wherein the model operator energy consumption table comprises an operator name, a chip model, an execution time and an energy consumption amount of the model operator;

[0066] A data set construction unit 202 is configured to construct a reinforcement learning data set according to the model operator in the model operator energy consumption table;

[0067] An optimal search unit 203 is configured to search for an optimal energy consumption operator in the reinforcement learning data set based on a preset strategy network and a preset reward function according to a reinforcement learning strategy, and construct an optimal energy consumption neural network model.

[0068] Further, the application further comprises the following steps of:

[0069] An operator extraction unit 204 is configured to extract structural hyperparameters of different algorithm models, and generate model operators, wherein the structural hyperparameters comprise a network layer number, a neuron number and a convolution kernel size.

[0070] Further, the optimal search unit 203 is specifically configured to:

[0071] define the model operator as a state based on the reinforcement learning strategy, and form a state space;

[0072] define a set of executable operation actions in each state, and form an action space;

[0073] perform iterative search for an energy consumption operator in the reinforcement learning data set based on the state space, the action space and the preset reward function by using the preset strategy network as a constraint condition of optimal energy consumption, until a stop condition is reached, and obtain the optimal energy consumption operator;

[0074] construct the optimal energy consumption neural network model according to the optimal energy consumption operator.

[0075] Further, the application further comprises the following steps of:

[0076] A network construction unit 205 is configured to construct a strategy network based on a multilayer perceptron, and obtain the preset strategy network.

[0077] The application further provides a neural network model construction device based on energy consumption of a heterogeneous chip, which comprises a processor and a memory.

[0078] The memory is configured to store program code, and transmit the program code to the processor.

[0079] The processor is configured to execute the method of constructing a neural network model based on energy consumption of a heterogeneous chip according to instructions in the program code.

[0080] The application further provides a computer-readable storage medium for storing program code, the program code being used to execute the method of constructing a neural network model based on energy consumption of a heterogeneous chip.

[0081] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. For example, the described apparatus embodiments are merely schematic, and the division of the units is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0082] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0083] In addition, each functional unit in the embodiments of the present application can be integrated in a processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be in the form of hardware, or in the form of a software functional unit.

[0084] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, includes a plurality of instructions for executing all or part of the steps of the method described in various embodiments of the present application by a computer device (which can be a personal computer, a server, or a network device, etc.). The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (English full name: Read-Only Memory, English abbreviation: ROM), a random access memory (English full name: Random Access Memory, English abbreviation: RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0085] The above-described and the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A neural network model construction method based on heterogeneous chip energy consumption, characterized in that, The method comprises the following steps: obtaining a model operator energy consumption table by testing the energy consumption of model operators of different algorithm models on different heterogeneous chips using preset test cases, wherein the model operator energy consumption table comprises the operator name, chip model, execution time and energy consumption of the model operator; constructing a reinforcement learning data set according to the model operators in the model operator energy consumption table; constructing a policy network based on a multilayer perceptron to obtain a preset policy network; searching for an optimal energy consumption operator in the reinforcement learning data set based on a reinforcement learning strategy, the preset policy network and a preset reward function, and constructing an optimal energy consumption neural network model, wherein the searching process comprises the following steps: defining the model operators as states based on the reinforcement learning strategy to form a state space; defining a set of executable operation actions at each state to form an action space; performing iterative search for an energy consumption operator in the reinforcement learning data set based on the state space, the action space and a preset reward function by taking the optimal energy consumption as a constraint condition through the preset policy network until a stop condition is reached to obtain the optimal energy consumption operator; constructing an optimal energy consumption neural network model according to the optimal energy consumption operator. 2.The method of claim 1, wherein, The method further comprises the following steps before obtaining the model operator energy consumption table by testing the energy consumption of model operators of different algorithm models on different heterogeneous chips using preset test cases: extracting structural hyperparameters of different algorithm models to generate model operators, wherein the structural hyperparameters comprise the number of network layers, the number of neurons and the size of convolution kernels.

3. The apparatus for constructing a neural network model based on energy consumption of a heterogeneous chip, characterized by, The method comprises the following steps: an energy consumption testing unit is configured to obtain a model operator energy consumption table by testing the energy consumption of model operators of different algorithm models on different heterogeneous chips using preset test cases, wherein the model operator energy consumption table comprises the operator name, chip model, execution time and energy consumption of the model operator; a data set construction unit is configured to construct a reinforcement learning data set according to the model operators in the model operator energy consumption table; a network construction unit is configured to construct a policy network based on a multilayer perceptron to obtain a preset policy network; an optimal searching unit is configured to search for an optimal energy consumption operator in the reinforcement learning data set based on a reinforcement learning strategy, the preset policy network and a preset reward function, and construct an optimal energy consumption neural network model, wherein the optimal searching unit is specifically configured to: define the model operators as states based on the reinforcement learning strategy to form a state space; define a set of executable operation actions at each state to form an action space; perform iterative search for an energy consumption operator in the reinforcement learning data set based on the state space, the action space and a preset reward function by taking the optimal energy consumption as a constraint condition through the preset policy network until a stop condition is reached to obtain the optimal energy consumption operator; construct an optimal energy consumption neural network model according to the optimal energy consumption operator.

4. The neural network model construction apparatus based on heterogeneous chip energy consumption according to claim 3, characterized in that, The method further comprises the following steps: an operator extraction unit is configured to extract structural hyperparameters of different algorithm models to generate model operators, wherein the structural hyperparameters comprise the number of network layers, the number of neurons and the size of convolution kernels.

5. A neural network model construction device based on energy consumption of a heterogeneous chip, characterized by, The device comprises a processor and a memory; the memory is configured to store program code and transmit the program code to the processor; The processor is configured to execute the method for constructing a neural network model based on energy consumption of a heterogeneous chip according to instructions in the program code.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium is configured to store program code for executing the method for constructing a neural network model based on energy consumption of a heterogeneous chip according to instructions in the program code.

Images