CN114897098A

CN114897098A - Automatic mixing precision quantification method and device

Info

Publication number: CN114897098A
Application number: CN202210634034.0A
Authority: CN
Inventors: 张川; 葛荧萌; 冀贞昊; 张在琛; 黄永明; 尤肖虎
Original assignee: Network Communication and Security Zijinshan Laboratory
Current assignee: Network Communication and Security Zijinshan Laboratory
Priority date: 2022-06-06
Filing date: 2022-06-06
Publication date: 2022-08-12
Also published as: WO2023236609A1

Abstract

The invention provides an automatic mixing precision quantification method and device, wherein the method comprises the following steps: acquiring intermediate variables to be quantized, which are generated by a multi-input multi-output MIMO detector in the process of signal detection of an MIMO system; training the agent through a deep reinforcement learning algorithm, and quantizing the decimal bit width of each intermediate variable based on a strategy stored in the trained agent; and quantizing the integer bit width of each intermediate variable based on a probability density function. The method and the device can automatically realize the distribution of the quantization bit width of different intermediate variables in the MIMO detector, avoid a large amount of quantization bit width redundancy and save hardware resources.

Description

Automatic mixing precision quantification method and device

Technical Field

The invention relates to the technical field of machine learning, in particular to an automatic mixing precision quantification method and device.

Background

As the demand for throughput and stability of communication systems continues to increase, multiple-input multiple-output (MIMO) systems are receiving widespread attention due to their potential for high spectral efficiency. Since the computational complexity of optimal detection algorithms in massive MIMO scenarios is hard to bear in hardware implementations, many hardware-friendly detectors have been implemented today and present significant advantages in terms of throughput, energy efficiency and area efficiency. However, the conventional hardware-friendly detector mainly focuses on arithmetic implementation rather than quantization optimization, and in order to save design workload, a unified quantization scheme that uses the same quantization bit width for all variables is mostly adopted, which causes a large amount of quantization bit width redundancy, resulting in hardware resource waste.

Disclosure of Invention

The invention provides an automatic mixed precision quantization method and device, which are used for solving the defect of hardware resource waste caused by a large amount of quantization bit width redundancy caused by the fact that an MIMO detector uses the same quantization scheme for all variables in the prior art, can automatically realize the allocation of the quantization bit width of different intermediate variables in the MIMO detector, avoid a large amount of quantization bit width redundancy and save hardware resources.

The invention provides an automatic mixing precision quantification method, which comprises the following steps:

acquiring intermediate variables to be quantized, which are generated by an MIMO detector in the process of signal detection of an MIMO system;

training the agent through a deep reinforcement learning algorithm, and quantizing the decimal bit width of each intermediate variable based on a strategy stored in the trained agent;

and quantizing the integer bit width of each intermediate variable.

According to the automatic mixing precision quantification method provided by the invention, the training of the intelligent agent through the deep reinforcement learning algorithm comprises the following steps:

in each round, initializing an environment, interacting the intelligent agent with the environment based on a Markov decision process, and storing interaction data;

and controlling the intelligent agent to update the strategy stored in the intelligent agent based on the currently stored interaction data until the maximum number of rounds is reached.

According to an automatic blending precision quantification method provided by the invention, the initialization environment comprises the following steps:

randomly extracting a plurality of intermediate variables;

initializing the decimal bit width of all intermediate variables to be a preset maximum decimal bit width, and initializing the integer bit width of all intermediate variables to be a preset maximum integer bit width;

determining a plurality of states, wherein the plurality of states comprise serial numbers and decimal bit widths corresponding to the plurality of intermediate variables;

randomly selecting and returning a state from the plurality of states.

According to the automatic mixed precision quantification method provided by the invention, the interaction between the intelligent agent and the environment based on the Markov decision process and the storage of interaction data comprises the following steps:

executing the following steps at each moment until the current moment reaches the maximum moment, and storing interaction data generated at each moment, wherein the interaction data comprises a state, an action value and a reward value:

determining an action value according to the current state and the current strategy; wherein the action value is used for representing the variation of the decimal bit width;

modifying the current quantization scheme according to the action value to obtain a modified quantization scheme;

and evaluating the modified quantization scheme according to a reward function to obtain a reward value, and simultaneously selecting and returning to the next state.

According to the automatic mixed precision quantization method provided by the invention, the quantizing the decimal bit width of each intermediate variable based on the trained strategy stored in the agent comprises the following steps:

converting the trained strategies stored in the intelligent agent into the statistical distribution of the decimal bit width of each intermediate variable through a Monte Carlo simulation algorithm;

quantifying the decimal bit width of the intermediate variable based on the statistical distribution of the decimal bit width of the intermediate variable;

the strategy is used for representing a mapping relation between a state and an action value, the state comprises a sequence number corresponding to an intermediate variable and a decimal bit width, and the action value is used for representing a variation of the decimal bit width.

According to the automatic mixed precision quantification method provided by the invention, the method for converting the trained strategy stored in the intelligent agent into the statistic distribution of the decimal bit width of each intermediate variable through the Monte Carlo simulation algorithm comprises the following steps:

executing the following steps during each test until the maximum test times are reached, and obtaining the frequency of each intermediate variable measuring different decimal bit widths so as to obtain the statistical distribution of the decimal bit widths of each intermediate variable:

for each intermediate variable, determining an action value based on the current state of the intermediate variable and a strategy stored in the trained agent;

altering a decimal bit width of the intermediate variable based on the action value.

According to the automatic mixed precision quantization method provided by the invention, the quantizing the decimal bit width of the intermediate variable based on the statistical distribution of the decimal bit width of the intermediate variable comprises the following steps:

and determining the average value of the statistical distribution of the decimal point width of the intermediate variable as the decimal point width of the intermediate variable based on the statistical distribution of the decimal point width of the intermediate variable.

According to the automatic mixed precision quantization method provided by the invention, the quantizing the integer bit width of each intermediate variable comprises the following steps:

generating a plurality of data of each intermediate variable through a Monte Carlo simulation algorithm to obtain a first data set;

extracting data in the first data set which is not in the amplitude range of the quantization scheme to obtain a second data set; the amplitude of the quantization scheme is the difference of an expression which takes 2 as a base number and takes an integer bit width as an argument and an expression which takes 2 as a base number and takes a negative decimal bit width as an argument;

determining the minimum integer bit width meeting the preset condition as the integer bit width of the intermediate variable; wherein the preset conditions are as follows: the ratio of the number of data in the second data set to the number of data in the first data set is less than or equal to a preset threshold.

The present invention also provides an automatic blending precision quantizing device, including:

the variable acquisition module is used for acquiring intermediate variables to be quantized, which are generated by the MIMO detector in the process of signal detection of the MIMO system;

the first quantization module is used for training the agent through a deep reinforcement learning algorithm and quantizing the decimal bit width of each intermediate variable based on a strategy stored in the trained agent;

and the second quantization module is used for quantizing the integer bit width of each intermediate variable based on a probability density function.

The present invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the automatic blending precision quantization method as described in any of the above.

The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements an automatic blending accuracy quantification method as described in any of the above.

The invention also provides a computer program product comprising a computer program which, when executed by a processor, implements an automatic blending accuracy quantification method as described in any one of the above.

According to the automatic mixed precision quantification method and device, the intelligent agent is trained through a deep reinforcement learning algorithm, the decimal bit width of each intermediate variable to be quantified is quantified based on a strategy stored in the trained intelligent agent, and the distribution of the decimal bit widths of different intermediate variables can be automatically realized; then, the integer bit width of each intermediate variable is quantized, and the distribution of the integer bit widths of different intermediate variables can be automatically realized. Therefore, the method and the device can automatically realize the distribution of the quantization bit width of different intermediate variables in the MIMO detector, avoid a large amount of quantization bit width redundancy and save hardware resources.

Drawings

In order to more clearly illustrate the technical solutions of the present invention or the prior art, the drawings needed for the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and those skilled in the art can also obtain other drawings according to the drawings without creative efforts.

FIG. 1 is a flow chart of an automatic blending accuracy quantification method provided by the present invention;

FIG. 2 shows the number of different decimated variables N provided by the present invention _ext Under the condition, a curve graph of the change of the reward value along with the network updating times is obtained;

FIG. 3 is a floating point AMP detector provided by the present invention and at a different N _ext A comparison graph of BER performance of the fractional quantized AMP detector below;

FIG. 4 shows BER 10 provided by the present invention ^-3 When is different from N _ext A plot of the average fractional bit width of the values versus SNR loss;

FIG. 5 shows the difference L provided by the present invention _a Under the condition, a curve graph of the change of the reward value along with the network updating times is obtained;

FIG. 6 is a graph comparing BER performance of a floating-point AMP detector, a fractional quantized AMP detector, and a fractional integer quantized AMP detector provided by the present invention;

FIG. 7 is a graph comparing BER performance of AMP detectors provided by the present invention and unified quantized AMP detectors using different fractional bit widths;

FIG. 8 is a schematic structural diagram of an automatic blending precision quantizing device provided by the present invention;

fig. 9 is a schematic structural diagram of an electronic device provided by the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Referring to fig. 1, fig. 1 is a flow chart of an automatic blending precision quantization method according to the present invention. As shown in fig. 1, the method for quantizing the accuracy of automatic mixing for a MIMO detector provided by the present invention mainly comprises the following steps:

step 101, acquiring intermediate variables to be quantized, which are generated by an MIMO detector in the process of signal detection of an MIMO system;

102, training the agent through a deep reinforcement learning algorithm, and quantifying the decimal bit width of each intermediate variable based on a strategy stored in the trained agent;

and 103, quantizing the integer bit width of each intermediate variable.

In step 101, regarding the MIMO system, taking uplink as an example, assume that one ue configures N _t Root transmitting antenna, base station terminal configuration N _r MIMO system of root receiving antenna, where N _t ＜N _r . In general, the MIMO system can be simplified to the following real-number domain mathematical model:

y＝Hx+n (1)

wherein,

representing a received signal vector;

represents a transmitted signal vector; h represents 2N _r ×2N _t Of dimensionsThe channel matrix, which is assumed to be an independent and equally distributed (i.i.d.) rayleigh channel in this embodiment, has a mean value of 0 and a variance of 1/2N _r And the receiving end knows the channel condition;

representing additive white Gaussian noise, with a noise mean of 0 and a variance of

The MIMO detector is used for detecting signals of the MIMO system. In this embodiment, the MIMO detector may be a Bayesian Message Passing (BMP) detector, but the embodiment is not limited thereto, and the MIMO detector may also be another detector.

In this step, intermediate variables to be quantized, which are generated by the MIMO detector in the process of performing signal detection on the MIMO system, are obtained, and the intermediate variables generated by different MIMO detectors are different.

In step 102, the agent is trained through a deep reinforcement learning algorithm, and the decimal bit width of each intermediate variable to be quantized is quantized based on a strategy stored in the trained agent, so that the distribution of the decimal bit widths of different intermediate variables can be automatically realized.

Optionally, in step 102, training the agent through a deep reinforcement learning algorithm includes the following sub-steps:

step 1021, in each round, initializing an environment, interacting the agent with the environment based on a Markov decision process, and storing interaction data;

step 1022, judging whether the number of times of the preset round is reached, if so, turning to step 1023, and if not, turning to step 1021;

step 1023, controlling the agent to update the strategy stored in the agent based on the currently stored interactive data;

step 1024, judging whether the maximum number of rounds is reached, if so, finishing the training, and if not, turning to step 1021.

In step 1021, a Deep Reinforcement Learning (DRL) algorithm is used as a branch of Machine Learning (ML) to focus on interacting with the environment to make a correct decision. In the present embodiment, it is assumed that the process of interaction of the agent with the environment is a Markov Decision Process (MDP).

In each round, the environment is initialized, the process of interaction between the agent and the environment is assumed to be a Markov decision process, the agent and the environment can be interacted based on the Markov decision process, and interaction data is stored during the interaction.

In step 1023, the agent updates the policy stored in the agent once every preset number of rounds is reached based on the currently stored interaction data.

In the embodiment, the agent and the environment are continuously interacted based on a Markov decision process, and the agent is trained through a deep reinforcement learning algorithm, so that the agent can obtain a strategy through learning.

Optionally, in step 1021, initializing the environment may include:

step 10211, randomly extracting a plurality of intermediate variables;

step 10212, initializing the decimal bit width of all intermediate variables to be a preset maximum decimal bit width, and initializing the integer bit width of all intermediate variables to be a preset maximum integer bit width;

step 10213, determining a plurality of states, wherein the plurality of states include serial numbers and decimal bit widths corresponding to a plurality of intermediate variables;

step 10214, randomly select and return to a state from a plurality of states.

In step 10211, considering the strict requirement of bit-error rate (BER) performance in MIMO detection and the large correlation between different intermediate variables, considering all the intermediate variables in one round may introduce serious reward misjudgment, resulting in poor quantization result. Thus, in one round, N is randomly drawn _ext (1＜N _ext ≤N _all ) An intermediate variable to obtain a variable set S _ext (ii) a It is composed ofIn, N _ext Indicating a predetermined number of randomly extracted intermediate variables, N _all Representing the number of all intermediate variables, S _ext Representing a set of randomly drawn intermediate variables. Although only N is considered in a single round _ext Correlations between the extracted variables are taken into account, but as long as the number of rounds is sufficient.

In step 10213, the state is defined as a vector (oh (k), oh (q) _k ) K ∈ {1, 2., N) } _all K represents the sequence number of the kth intermediate variable to be quantized; q. q.s _k ∈{0,1,...,q _max }，q _k Decimal place width, q, representing the k intermediate variable _max Denotes q _k The preset maximum value of (a), that is, the preset maximum decimal bit width of the kth intermediate variable; operation oh () represents the return of one-hot coded values.

In this step, the decimal bit width of the plurality of randomly extracted intermediate variables is initialized to a preset maximum decimal bit width q _max Therefore, a plurality of states, i.e., (oh (k'), oh (q), to which the plurality of intermediate variables correspond, respectively, can be obtained _max ))，k'∈S _ext 。

Step 10214, randomly selecting and returning a state (oh (k'), oh (q) from a plurality of states _k' ))，k'∈S _ext 。

In this embodiment, the plurality of intermediate variables are randomly extracted first, then the decimal bit width of all the intermediate variables is initialized to the preset maximum decimal bit width, the integer bit width of all the intermediate variables is initialized to the preset maximum integer bit width, and finally one state is randomly selected from the plurality of states respectively corresponding to the plurality of intermediate variables and returned, so that the initialization of the environment can be realized.

Optionally, step 1021, interacting the agent with the environment based on the markov decision process, and storing interaction data, including:

step 10215, determining an action value according to the current state and the current strategy at each moment;

step 10216, modifying the current quantization scheme according to the action value to obtain a modified quantization scheme;

step 10217, evaluating the modified quantization scheme according to the reward function to obtain a reward value, selecting and returning to the next state, and turning to step 10218; the method comprises the steps that interactive data generated at each moment are stored simultaneously, and the interactive data comprise states, action values and reward values;

step 10218, judging whether the current time reaches the maximum time, if yes, going to step 1022, if no, going to step 10215.

In order to understand the interaction process of the agent and the environment more clearly, the quantization scheme, the action and the reward function are first described separately as follows:

(1) quantization scheme

Since the linear quantization can be implemented efficiently in hardware, the present embodiments all employ linear quantization. The specific quantization scheme is as follows: the symbol part, the integer part and the decimal part take 1, p and q respectively, and are abbreviated as 1-p-q. For intermediate variables of value v, the quantized value v _Q Can be expressed as:

v _Q ＝round(clip(v,-B,B)/C)×C (2)

wherein B is 2 ^p -2 ^-q And C is 2 ^-q 。

(2) Movement of

The action represents the amount of change in the decimal place width. Bit width after action (denoted as q' _k ) Is not in the set {0, 1.,. q., } _max In the method, q 'can be cut' _k Into the range [0, q ] _max ]In (1). However, this would increase to take 0 or q _max Probability (less than 0 or greater than q) _max Is forced to 0 or q _max ) Resulting in a greater likelihood that the agent will stay at both extreme points. Therefore, if q is _k Greater than q _max It can be calculated starting from 0; if q is _k Less than 0, may be selected from q _max The calculation is started. Thus, it can be expressed by the following expression of fractional bit width variation:

q′ _k ＝q _k +a _t mod(q _max +1) (3)

taking action at agent a _t Fraction of the last, k-th variableBit width is changed to q' _k 。

Defining the action influence range as the maximum absolute value of the decimal bit width change and recording as L _a . Then, the motion space

Is a collection

When L is _a When the decimal point is 1, the decimal point bit width can only be changed by 1 bit at most, which may cause low learning efficiency and easily fall into a local optimal solution. When in use

The agent may flexibly change the current fractional bit width to any other fractional bit width, but this may lead to instability of the learning process and degradation of convergence performance.

(3) Reward function

The reward depends mainly on BER performance and fractional bit width. Since MIMO detection has strict performance requirements, the reward function aims to reduce the decimal bit width while ensuring BER performance.

Evaluation on BER performance: monte carlo simulations are commonly used to test the BER performance of a particular detection algorithm. Considering that accurate BER requires a large number of samples to perform monte carlo simulation, it can be pre-calculated before the agent starts learning, denoted as P _b . In evaluating the BER performance of the quantization detector, the environment uses a small number of samples to simultaneously emulate the BER performance of the floating point detector and the quantization detector to save computation time. The corresponding BER of the floating point detector and the quantization detector are written as

And

the BER relative error of the quantization detector is defined as

But since the number of samples simulated in evaluating the BER performance of the quantization detector is small,

may be equal to 0. Therefore, use instead

As a relative error. If the relative error is greater than a threshold value epsilon ₂ Then the performance of the quantitative detector cannot be guaranteed and the environment returns a reward r _t Is-1. Otherwise, the environment will consider the decimal point width.

Representing the average decimal place width of all variables of the current quantization scheme as

Relative error less than epsilon ₂ The reward function of the time is defined as

Wherein theta is ₁ And theta ₂ Are empirical parameters. The exponential function may allow the environment to return a larger reward value with a smaller bit width than a linear function, resulting in a greater tendency for the agent to reduce the bit width. The overall reward function is as follows:

then, the environment selects the next intermediate variable, which is recorded as the kth 'intermediate variable, and the state is updated to (oh (k'), oh (q) _k' ))。

Based on the above description, the following steps 10215-10218 are described in detail:

in step 10215, at time t, according to the current state

And current policy

Determining an action value

Wherein,

and

respectively representing a state space and an action space, and the strategy is used for representing the mapping relation between the state and the action values.

In step 10216, the action value is determined

And modifying the current quantization scheme to obtain a modified quantization scheme.

The current quantization scheme: v. of _Q Round (clip (v, -B, B)/C) × C, wherein,

and

modified quantization scheme: v. of _Q (clip (v, -B ', B ')/C '). times.C ', where q ' _k ＝q _k +a _t ，

And

in step 10217, the modified quantization scheme is evaluated according to a reward function to obtain a reward value

Representing a reward functionSimultaneously selecting and returning to the next state (oh (k'), oh (q) _k' ))。

In step 10218, the agent interacts with the environment once at each time, i.e. step 10215 and 10217 are executed once, and the current time t refers to: moment when the agent is interacting with the environment, maximum moment t _max Means that: and the preset maximum moment when the intelligent agent interacts with the environment. Repeatedly executing the above steps 10215-10217 until the current time t reaches the maximum time t _max And simultaneously storing the interaction data generated at each moment, wherein the interaction data comprises a state, an action value and a reward value.

Optionally, in step 102, quantizing the decimal bit width of each intermediate variable based on the policy stored in the trained agent includes:

step 1025, converting the strategy stored in the trained agent into the statistical distribution of the decimal bit width of each intermediate variable through a Monte Carlo simulation algorithm;

and step 1026, determining the decimal bit width of the intermediate variable based on the statistical distribution of the decimal bit width of the intermediate variable.

In step 1025, the following sub-steps may be included:

step 10251, in each test, determining an action value for each intermediate variable based on the current state of the intermediate variable and a strategy stored in the trained agent;

step 10252, changing the decimal bit width of the intermediate variable based on the action value;

step 10253, judging whether the maximum test times is reached, if so, turning to step 10254, and if not, turning to step 10251;

step 10254, obtaining the frequency of each intermediate variable measuring with different decimal bit widths;

step 10255, obtain the decimal fraction bit width statistical distribution of each intermediate variable.

In step 10251, at each test, the intermediate variable selected by the fixed environment each time is the kth intermediate variable, according to the current state (oh (k), oh (q) _k ) And current policy

Determining an action value

In step 10252, based on the action value

Changing decimal bit width q 'of k-th intermediate variable' _k ＝q _k +a _t 。

In step 10254, after the maximum number of tests is reached, the frequency with which the kth intermediate variable takes on different decimal places is obtained.

In step 10255, the statistical distribution of decimal places of the k-th variable learned by the agent can be approximated after the maximum test time is sufficiently large.

In this embodiment, the policy stored in the trained agent may be converted into the statistical distribution of the decimal bit width of the intermediate variable through a monte carlo simulation algorithm.

In step 1026, based on the statistical distribution of the decimal bit width of the intermediate variable, the average of the statistical distribution of the decimal bit width of the intermediate variable is determined as the decimal bit width of the intermediate variable.

In the present embodiment, the mean value of statistical distribution of decimal place widths of the kth variable is calculated as q _k An estimate of the expected value.

Optionally, step 103 may comprise the following sub-steps:

step 1031, generating a plurality of data of each intermediate variable through a Monte Carlo simulation algorithm to obtain a first data set;

step 1032, extracting data in the first data set, which is not in the amplitude range of the quantization scheme, to obtain a second data set; the amplitude of the quantization scheme is the difference of an expression which takes 2 as a base number and takes an integer bit width as an argument and an expression which takes 2 as a base number and takes a negative decimal bit width as an argument;

step 1033, determining the minimum integer bit width meeting the preset condition as the integer bit width of the intermediate variable; wherein the preset conditions are as follows: the ratio of the number of data in the second data set to the number of data in the first data set is less than or equal to a preset threshold.

In step 1031, a plurality of data of the k-th variable intermediate variable are generated by a monte carlo simulation algorithm, so as to obtain a first data set S.

In step 1032, for the 1-p-q quantization scheme, the quantization scheme magnitude is B-2 ^p -2 ^-q The quantization scheme amplitude range is [ -B, B [ -B [ ]]Extracting the data not in [ -B, B ] in the first data set S]The data in the range constitutes a second data set S ', S' ({ v | | v | > | B |, v ∈ S }.

In step 1033, it will suffice

Is determined as the integer bit width of the intermediate variable, i.e. of the intermediate variable

Wherein epsilon ₁ Representing a preset threshold. When epsilon ₁ When 0, then card (S') is 0, meaning that the quantization scheme can cover all values of the variable.

In this embodiment, the integer bit width of each intermediate variable is quantized, and the allocation of the integer bit widths of different intermediate variables can be automatically realized.

In summary, according to the automatic hybrid precision quantization method for the MIMO detector provided by the present invention, the agent is trained through a deep reinforcement learning algorithm, and the decimal bit width of each intermediate variable to be quantized is quantized based on a policy stored in the trained agent, so that the distribution of the decimal bit widths of different intermediate variables can be automatically realized; then, the integer bit width of each intermediate variable is quantized, and the distribution of the integer bit widths of different intermediate variables can be automatically realized. Therefore, the method and the device can automatically realize the distribution of the quantization bit width of different intermediate variables in the MIMO detector, avoid a large amount of quantization bit width redundancy and save hardware resources.

The method provided in the present embodiment is verified by the following specific examples.

In the present embodiment, the configuration is taken as an example, the transmission signal vector is uniformly and randomly generated from 16QAM modulation, the channel matrix is randomly generated from a rayleigh channel model with the number of transmission antennas being 8 and the number of reception antennas being 128, and the MIMO detector is selected as an Approximate Message Passing (AMP) detector, and the number of iterations is set to 4. Through certain simulation experiments, epsilon is found ₁ ＝10 ^-4 ,ε ₂ ＝0.8,θ ₁ 10 and θ ₂ A good tradeoff between BER performance and quantization bit width is achieved at 4. Therefore, these hyper-parameters are fixed during the following simulation experiments.

The strategy network and the value network are both composed of full-connection DNN with 6 hidden layers, the strategy network is used for storing the strategy learned by the intelligent agent, and the value network is used for evaluating the value of the current state. The dimensions of the 6 hidden layers are 64, 128, 256, 128, 64, respectively. The dimensions of the input layers of both the policy network and the value network are equal to the dimensions of the state space. The dimension of the output layer of the policy network is equal to the dimension of the action space, and the dimension of the value network is 1. The learning rate is set to 0.002, the agent adopts a near-end Policy Optimization (PPO) algorithm, attenuation factors in the PPO algorithm are set to 0.99, an element clipping value is set to 0.2, and decimal bit width of each variable is learned until the maximum number of rounds is reached.

The simulation results were analyzed as follows:

1)N _ext influence of size

The reward returned by the environment is related to the number of iterations as shown in figure 2. Different N _ext In this case, as the number of iterations increases, the reward value increases first and then fluctuates around a certain value, illustrating how the agent learns to better quantify the intermediate variables in the detection algorithm. When N is present _ext When 1, the agent may reduce the decimal point width of the variable to a very low value, i.e., a larger prize value as shown in fig. 2, since no correlation between different intermediate variables is considered. With N _ext Increase, on convergenceThe prize value becomes smaller. This is primarily because the reward false positive problem makes the agent more challenging in reducing the decimal point width.

Floating-point AMP detector and fractional quantization AMP detector in this embodiment are at different N _ext The BER performance comparison below is shown in FIG. 3, where the legend entry "N _ext 1 represents N _ext The performance curve of the fractional quantized AMP detector in this embodiment when 1, and others are similar. Although the agent is in N _ext The decimal bit width allocated is the least when the decimal bit width is 1, but the performance is seriously reduced. N is a radical of _ext ＝5，N _ext 15 and N _ext The performance of the fractional quantized AMP detector at 21 is substantially the same as the performance of a floating AMP detector.

In FIG. 4, the BER is 10 ^-3 When is different from N _ext Average fractional bit width of values versus SNR loss. The horizontal axis represents the difference N in the present embodiment _ext The fractional quantization AMP detector of (1) reduces SNR loss compared to a floating AMP detector as a reference. The vertical axis represents the average decimal place width. As shown in FIG. 4, the average fractional bit width is N _ext The lowest, but greater performance loss, is 1. When N is present _ext When 5, the fractional quantization AMP detector can allocate fewer fractional bit widths while maintaining the performance of the original algorithm. And N is _ext 15 and N _ext Loss of performance and N of 21 _ext The difference is small at 5, but the average fractional bit width is large. Thus in the following analysis, N will be _ext And fixing to 5.

2)L _a Influence of size

Different L _a In this case, the change of the prize value with the number of network updates is shown in fig. 5. And L _a In comparison with the case of 1, L _a Convergence performance is slightly superior when 2. At L _a 1 and L _a In both cases 2, the prize value eventually converges to approximately the same value. When L is _a When the convergence rate is 5, the convergence rate becomes very slow, and the reward value at the time of convergence becomes smaller than those in the other two cases. Therefore, when L is _a Too large, the convergence speed and the final decimal width are unsatisfactory. In the following discussion, we will refer to L _a Is fixed to2。

3) Integer quantization

After the decimal bit width is obtained, the integer bit width can be calculated according to the above expression (5). Fig. 6 shows a BER performance comparison of a floating-point AMP detector, a fractional quantized AMP detector, and a fractional integer quantized AMP detector. It can be seen that PDF-based integer quantization can perfectly preserve the BER performance of the floating-point AMP detector.

4) Comparison with a unified quantization scheme

The automatic mixed precision quantization scheme proposed by the present embodiment is compared with the unified quantization scheme. The uniformly quantized integer bit width can be obtained by PDF-based integer quantization, i.e., 6 bits. With respect to fractional bit widths, the BER performance of a floating point AMP detector and a unified quantized AMP detector using different fractional bit widths are compared as shown in fig. 7. Unified quantized AMPs with 4-bit and 5-bit fractional bit widths suffer from severe performance degradation. A unified quantized AMP detector with a 6bit fractional bit width can almost perfectly restore the performance of a floating AMP detector. Thus, the quantization scheme for the unified quantization AMP detector is taken to be 1-6-6. The average bit width comparison of the AMP detector based on automatic mix precision quantization and the unified quantized AMP detector is shown in table 1. The AMP detector based on automatic mixed precision quantization reduces the quantization bit width by 57.2% and 58% in integer bit width and fractional bit width, respectively, compared to the unified quantization AMP detector.

The average integer and fractional bit widths of the AMP detector based on automatic mix precision quantization and the unified quantization AMP detector can be compared by table 1.

TABLE 1

As can be seen from the above analysis, the automatic hybrid precision quantization method according to this embodiment automatically implements allocation of quantization bit widths of different intermediate variables in the MIMO detector, thereby avoiding a large amount of quantization bit width redundancy.

The following describes the automatic blending precision quantization apparatus provided by the present invention, and the automatic blending precision quantization apparatus described below and the automatic blending precision quantization method described above may be referred to in correspondence with each other.

Referring to fig. 8, fig. 8 is a schematic structural diagram of an automatic blending precision quantizing device provided by the present invention. As shown in fig. 8, the automatic blending precision quantizing device provided by the present invention may include:

a variable obtaining module 10, configured to obtain an intermediate variable to be quantized, which is generated by the MIMO detector in a process of performing signal detection on the MIMO system;

the first quantization module 20 is configured to train the agent through a deep reinforcement learning algorithm, and quantize the decimal bit width of each intermediate variable based on a policy stored in the trained agent;

a second quantization module 30, configured to quantize the integer bit width of each intermediate variable based on a probability density function.

Optionally, the first quantization module 20 comprises:

the interaction unit is used for initializing the environment in each round, interacting the intelligent agent with the environment based on a Markov decision process and storing interaction data;

and the updating unit is used for controlling the intelligent agent to update the strategy stored in the intelligent agent based on the currently stored interactive data until the maximum round times are reached.

Optionally, the interaction unit is specifically configured to:

randomly extracting a plurality of intermediate variables;

randomly selecting and returning a state from the plurality of states.

Optionally, the interaction unit is specifically configured to:

Optionally, the first quantization module 20 comprises:

the statistical distribution unit is used for converting the trained strategies stored in the intelligent agent into the statistical distribution of the decimal bit width of each intermediate variable through a Monte Carlo simulation algorithm;

a decimal bit width quantization unit, configured to quantize the decimal bit width of the intermediate variable based on statistical distribution of the decimal bit width of the intermediate variable;

Optionally, the statistical distribution unit is specifically configured to:

Optionally, the fractional-bit-width quantization unit is specifically configured to:

and determining the average value of the statistical distribution of the decimal bit width of the intermediate variable as the decimal bit width of the intermediate variable based on the statistical distribution of the decimal bit width of the intermediate variable.

Optionally, the second quantization module 30 is specifically configured to:

Fig. 9 illustrates a physical structure diagram of an electronic device, which may include, as shown in fig. 8: a processor (processor)810, a communication Interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication Interface 820 and the memory 830 communicate with each other via the communication bus 840. The processor 810 may invoke logic instructions in the memory 830 to perform an automatic blending precision quantization method comprising:

and quantizing the integer bit width of each intermediate variable.

In addition, the logic instructions in the memory 830 may be implemented in software functional units and stored in a computer readable storage medium when the logic instructions are sold or used as independent products. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

In another aspect, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is capable of executing the automatic blending precision quantification method provided by the above methods, and the method includes:

and quantizing the integer bit width of each intermediate variable.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements a method for automatic blending accuracy quantification provided by the above methods, the method comprising:

and quantizing the integer bit width of each intermediate variable.

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. An automatic blending precision quantization method, comprising:

acquiring intermediate variables to be quantized, which are generated by a multi-input multi-output MIMO detector in the process of signal detection of an MIMO system;

and quantizing the integer bit width of each intermediate variable.

2. The method of claim 1, wherein training an agent through a deep reinforcement learning algorithm comprises:

3. The method of claim 2, wherein the initializing the environment comprises:

randomly extracting a plurality of intermediate variables;

randomly selecting and returning a state from the plurality of states.

4. The method of claim 3, wherein the Markov decision-based process of interacting an agent with an environment and storing interaction data comprises:

5. The MIMO detector-oriented automatic hybrid-precision quantization method according to any one of claims 1 to 4, wherein the quantizing the decimal bit width of each of the intermediate variables based on the trained stored policy in the agent comprises:

6. The method of claim 5, wherein the transforming the trained stored strategies in the agent into a fractional bit wide statistical distribution of each intermediate variable by a Monte Carlo simulation algorithm comprises:

7. The MIMO detector-oriented automatic hybrid precision quantization method of claim 5, wherein the quantizing the fractional bit width of the intermediate variable based on the statistical distribution of the fractional bit width of the intermediate variable comprises:

8. The MIMO detector-oriented automatic hybrid precision quantization method of claim 1, wherein the quantizing the integer bit width of each of the intermediate variables comprises:

9. An automatic blending precision quantization apparatus, comprising:

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor, when executing the program, implements the automatic blending accuracy quantification method of any one of claims 1 to 8.

11. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, implements the automatic blending precision quantization method of any of claims 1 to 8.