CN114897098A - Automatic mixing precision quantification method and device - Google Patents

Abstract

The present invention provides an automatic mixed precision quantization method and device, wherein the method includes: acquiring intermediate variables to be quantized generated by a multiple-input multiple-output MIMO detector in the process of performing signal detection on a MIMO system; The algorithm trains the agent, and quantifies the decimal bit width of each intermediate variable based on the strategy stored in the trained agent; based on the probability density function, quantifies the integer bit width of each intermediate variable quantify. The invention can automatically realize the allocation of quantization bit widths of different intermediate variables in the MIMO detector, avoid a large number of quantization bit width redundancy, and save hardware resources.

Description

A kind of automatic mixed precision quantization method and device

technical field

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

Background technique

With the ever-increasing demands on the throughput and stability of communication systems, multiple-input multiple-output (MIMO, multiple-input multiple-output) systems have received extensive attention due to their potential for high spectral efficiency. Since the computational complexity of optimal detection algorithms in massive MIMO scenarios is unbearable in hardware implementation, many hardware-friendly detectors have been implemented and exhibit obvious advantages in throughput, energy efficiency, and area efficiency. However, the existing hardware-friendly detectors mainly focus on arithmetic implementation rather than quantization optimization, and in order to save design work, most of them adopt a unified quantization scheme that uses the same quantization bit width for all variables, resulting in a large number of redundant quantization bit widths. excess, resulting in a waste of hardware resources.

SUMMARY OF THE INVENTION

The present invention provides an automatic mixed-precision quantization method and device, which are used to solve the defect of a large amount of quantization bit width redundancy caused by the use of the same quantization scheme for all variables in the prior art, resulting in waste of hardware resources. The allocation of quantization bit widths of different intermediate variables in the MIMO detector is realized, a large amount of quantization bit width redundancy is avoided, and hardware resources are saved.

The present invention provides an automatic mixed precision quantization method, comprising:

Obtaining intermediate variables to be quantized generated by the MIMO detector in the process of performing signal detection on the MIMO system;

The agent is trained through a deep reinforcement learning algorithm, and the decimal width of each intermediate variable is quantified based on the strategy stored in the trained agent;

The integer bit width of each of the intermediate variables is quantized.

According to an automatic mixed-precision quantization method provided by the present invention, the training of an agent through a deep reinforcement learning algorithm includes:

In each round, initialize the environment, interact with the environment based on Markov decision process, and store the interaction data;

Every time the preset number of rounds is reached, the agent is controlled to update the strategy stored in the agent based on the currently stored interaction data until the maximum number of rounds is reached.

According to an automatic mixed-precision quantization method provided by the present invention, the initialization environment includes:

Randomly select multiple intermediate variables;

Initialize the decimal width of all intermediate variables to the preset maximum decimal width, and initialize the integer width of all intermediate variables to the preset maximum integer width;

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

A state is randomly selected from the plurality of states and returned.

According to an automatic mixed-precision quantification method provided by the present invention, the agent interacts with the environment based on the Markov decision process, and stores the interaction data, including:

The following steps are performed at each moment until the current moment reaches the maximum moment, and the interaction data generated at each moment is stored, and the interaction data includes state, action value and reward value:

Determine the action value according to the current state and the current strategy; wherein, the action value is used to represent the amount of change in the width of the decimal place;

Modify the current quantization scheme according to the action value to obtain the modified quantization scheme;

The modified quantization scheme is evaluated according to the reward function, the reward value is obtained, and the next state is selected and returned.

According to an automatic mixed-precision quantization method provided by the present invention, the quantization of the decimal width of each intermediate variable based on the strategy stored in the trained agent includes:

Convert the strategy stored in the trained agent into a statistical distribution of the decimal width of each intermediate variable through a Monte Carlo simulation algorithm;

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

The strategy is used to represent the mapping relationship between a state and an action value, the state includes a sequence number and a decimal width corresponding to an intermediate variable, and the action value is used to represent a change in the decimal width.

According to an automatic mixed-precision quantization method provided by the present invention, the strategy stored in the trained agent is transformed into a statistical distribution of the decimal width of each intermediate variable by using a Monte Carlo simulation algorithm, including: :

The following steps are performed in each test until the maximum number of tests is reached, and the frequency of taking different decimal places for each of the intermediate variables is obtained, so as to obtain the statistical distribution of the decimal places of each of the intermediate variables:

For each of the intermediate variables, determine an action value based on the current state of the intermediate variables and the strategy stored in the trained agent;

The decimal width of the intermediate variable is changed based on the action value.

According to an automatic mixed-precision quantization method provided by the present invention, the quantization of the decimal width of the intermediate variable based on the statistical distribution of the decimal width of the intermediate variable includes:

A mean value of the statistical distribution of the decimal width of the intermediate variable is determined as the decimal width of the intermediate variable based on the statistical distribution of the decimal width of the intermediate variable.

According to an automatic mixed-precision quantization method provided by the present invention, the quantization of the integer bit width of each of the intermediate variables includes:

Generate a number of data of each of the intermediate variables through a Monte Carlo simulation algorithm to obtain a first data set;

Extracting the data in the first data set that is not within the range of the quantization scheme amplitude, to obtain a second data set; wherein, the quantization scheme amplitude is an expression with a base of 2 and an integer bit width as an independent variable minus 2 as an expression Base, the difference of expressions with negative decimal places as arguments;

Determining the minimum integer bit width that satisfies the preset condition as the integer bit width of the intermediate variable; wherein, the preset condition is: the number of data in the second data set is equal to the number of data in the first data set. The ratio is less than or equal to the preset threshold.

The present invention also provides an automatic mixed-precision quantization device, comprising:

A variable acquisition module, configured to acquire intermediate variables to be quantified generated by the MIMO detector in the process of performing signal detection on the MIMO system;

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

The second quantization module is configured to quantize the integer bit width of each of the intermediate variables based on the probability density function.

The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, when the processor executes the program, the automatic mixing as described in any of the above is realized Accuracy quantization method.

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the automatic mixed-precision quantization method as described above.

The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the automatic mixed-precision quantization method described above.

The present invention provides an automatic mixed precision quantization method and device, which trains an agent through a deep reinforcement learning algorithm, and quantifies the decimal bit width of each intermediate variable to be quantized based on the strategy stored in the trained agent. , can automatically realize the allocation of the decimal bit width of different intermediate variables; then, quantify the integer bit width of each intermediate variable, can automatically realize the allocation of the integer bit width of different intermediate variables. Therefore, the present invention can automatically realize the allocation of quantization bit widths of different intermediate variables in the MIMO detector, avoid a large amount of quantization bit width redundancy, and save hardware resources.

Description of drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are the For some embodiments of the invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

1 is a schematic flowchart of an automatic mixed-precision quantization method provided by the present invention;

Fig. 2 is the change curve diagram of reward value along with network update times under different extraction variable number N _ext situation provided by the present invention;

Fig. 3 is the BER performance comparison diagram of the floating-point AMP detector provided by the present invention and the small-quantization _AMP detector under different Next;

4 is a comparison diagram of the average decimal bit width and SNR loss of different _Next values when BER=10 ⁻³ provided by the present invention;

5 is _a graph of the change of reward value with the number of network updates under different La situations provided by the present invention;

Fig. 6 is the BER performance comparison diagram of floating-point AMP detector, fractional quantization AMP detector and fractional integer quantization AMP detector provided by the present invention;

Fig. 7 is the BER performance comparison diagram of the AMP detector provided by the present invention and the unified quantization AMP detector using different decimal bit widths;

8 is a schematic structural diagram of an automatic mixed-precision quantization device provided by the present invention;

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

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Please refer to FIG. 1 . FIG. 1 is one of the schematic flow charts of the automatic mixed precision quantization method provided by the present invention. As shown in FIG. 1 , the automatic mixed-precision quantization method for a multiple-input multiple-output MIMO detector provided by the present invention mainly includes the following steps:

Step 101: Obtain an intermediate variable to be quantized generated by the MIMO detector in the process of performing signal detection on the MIMO system;

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

Step 103: Quantize the integer bit width of each intermediate variable.

In step 101, regarding the MIMO system, taking the uplink as an example, assume a MIMO system in which N _t transmit antennas are configured at the user end and N _r receive antennas are configured at the base station end, where N _t <N _r . In general, this MIMO system can be simplified to the following real-domain mathematical model:

y=Hx+n (1)

表示接收信号矢量；

表示发送信号矢量；H表示2N_r×2N_t维度的信道矩阵，本实施例中假设该信道为独立同分布(i.i.d.，independent andidentically distributed)的瑞利信道，其均值为0，方差为1/2N_r，并且接收端已知信道状况；

表示加性高斯白噪声，噪声均值为0，方差为

in,

Represents the received signal vector;

represents the transmitted signal vector; H represents the channel matrix of dimension 2N _r × 2N _t . In this embodiment, it is assumed that the channel is an independent and identically distributed (iid) Rayleigh channel with a mean value of 0 and a variance of 1/2N. _r , and the receiver knows the channel condition;

represents additive white Gaussian noise, the noise mean is 0, and the variance is

The MIMO detector is used for signal detection of the MIMO system. In this embodiment, the MIMO detector may be a Bayesian message passing (BMP, Bayesian message passing) detector, but this embodiment is not limited thereto, and the MIMO detector may also be other detectors.

In this step, the intermediate variables to be quantized generated by the MIMO detector during the signal detection process of the MIMO system are acquired, and the intermediate variables generated by different MIMO detectors are different.

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

Optionally, in step 102, the agent is trained through a deep reinforcement learning algorithm, including the following sub-steps:

Step 1021. In each round, initialize the environment, interact with the agent based on the Markov decision process, and store the interaction data;

Step 1022, determine whether the preset number of rounds is reached, if so, go to step 1023, if not, go to step 1021;

Step 1023, controlling the agent to update the strategy stored in the agent based on the currently stored interaction data;

Step 1024 , determine whether the maximum number of rounds has been reached, if so, the training ends; if not, go to step 1021 .

In step 1021, a deep reinforcement learning (DRL, deep reinforcement learning) algorithm, as a branch of machine learning (ML, machine learning), focuses on interacting with the environment to make correct decisions. In this embodiment, it is assumed that the process of interaction between the agent and the environment is a Markov decision process (MDP, Markov decision process).

In each round, the environment is initialized, and it is assumed that the interaction process between the agent and the environment is a Markov decision process. Based on the Markov decision process, the agent can interact with the environment, and the interaction data is stored during the interaction.

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

In this embodiment, the agent is constantly interacting with the environment based on the 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 extract multiple intermediate variables;

Step 10212: Initialize the decimal width of all intermediate variables to the preset maximum decimal width, and initialize the integer width of all intermediate variables to the preset maximum integer width;

Step 10213, determine a plurality of states, and the plurality of states include the serial numbers and decimal bit widths corresponding to a plurality of intermediate variables;

Step 10214: Randomly select and return a state from multiple states.

In step 10211, considering the strict requirements for bit-error-rate (BER, bit-error-rate) performance in MIMO detection and the large correlation between different intermediate variables, considering all intermediate variables in one round will introduce Serious reward misjudgment, resulting in poor quantitative results. Therefore, in a round, N _ext (1<N _ext ≤N _all ) intermediate variables are randomly selected to obtain a variable set S _ext ; where N _ext represents the preset number of randomly selected intermediate variables, and N _all represents The number of all intermediate variables, _Sext represents the set of randomly drawn intermediate variables. Although only the correlations between the _Next drawn variables are considered in an episode, as long as the number of episodes is sufficient, the correlations between all variables will be considered.

In step 10213, the state is defined as a vector (oh(k), oh(q _k )), where k∈{1,2,...,N _all }, k represents the kth intermediate variable to be quantized q _k ∈{0,1,...,q _max }, q _k represents the decimal width of the kth intermediate variable, q _max represents the preset maximum value of q _k , that is, the kth intermediate variable The preset maximum decimal width; the operation oh() means to return the one-hot encoded value.

In this step, the decimal width of multiple intermediate variables randomly selected is initialized to the preset maximum decimal width q _max , so multiple states corresponding to multiple intermediate variables can be obtained, namely (oh(k' ), oh(q _max )), k'∈S _ext .

Step 10214: Randomly select and return a state (oh(k'), oh(q _k' )), k'∈S _ext , from multiple states.

In this embodiment, multiple intermediate variables are randomly selected first, then the decimal width of all intermediate variables is initialized to the preset maximum decimal width, and the integer width of all intermediate variables is initialized to the preset maximum integer width, Finally, a state is randomly selected from multiple states corresponding to multiple intermediate variables and a state is returned, which can realize the initialization of the environment.

Optionally, in step 1021, the agent interacts with the environment based on the Markov decision process, and the interaction data is stored, including:

Step 10215, at each moment, determine the action value according to the current state and the current strategy;

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

Step 10217: Evaluate the modified quantification scheme according to the reward function, obtain the reward value, select and return to the next state at the same time, and go to step 10218; wherein, the interaction data generated at each moment is stored at the same time, and the interaction data includes state, action value and reward value;

Step 10218: Determine whether the current time reaches the maximum time, if yes, go to step 1022, if not, go to step 10215.

In order to understand the interaction process between the agent and the environment more clearly, the quantification scheme, action and reward function are firstly explained, as follows:

(1) Quantization scheme

Since linear quantization can be efficiently implemented in hardware, linear quantization is adopted in this embodiment. The specific quantization scheme is as follows: the sign part, the integer part and the fractional part take 1, p and q respectively, abbreviated as 1-pq. For an intermediate variable with value v, the quantized value v _Q can be expressed as:

v _Q = round(clip(v,-B,B)/C)×C (2)

where B= ^2p -2- ^q and C=2- ^q .

(2) Action

The action represents the amount of change in the width of the decimal places. If the bit width (denoted as q' _k ) after taking the action is not in the set {0,1,...,q _max }, q' _k can be clipped into the range [0,q _max ]. However, this increases the probability of taking 0 or _qmax (values less than 0 or greater than _qmax are forced to 0 or _qmax ), causing the agent to be more likely to stay at the two extreme points. Therefore, if q _k is greater than q _max , the calculation can be started from 0; if q _k is less than 0, the calculation can be started from q _max . Therefore, it can be represented by the following expression for the change in decimal width:

q′ _k = _{q k} +at mod(q _max +1) (3)

After the agent takes action a _t , the fractional width of the kth variable becomes q' _k .

就是集合

当L_a＝1时，小数位宽最多只能改变1位，可能导致学习效率低，容易陷入局部最优解。当

时，智能体可以灵活地将当前小数位宽更改为任何其他小数位宽，但是这可能会导致学习过程的不稳定和收敛性能的下降。The influence range of the action is defined as the maximum absolute value of the change in the width of decimal places, denoted as L _a . So, the action space

is a collection

When La ₌ 1, the width of the decimal place can only be changed by 1 digit at most, which may lead to low learning efficiency and easy to fall into the local optimal solution. when

, the agent has the flexibility to change the current fractional width to any other fractional width, but this may lead to instability in the learning process and degradation of convergence performance.

(3) Reward function

The reward mainly depends on the BER performance and the fractional width. Since MIMO detection has strict performance requirements, the reward function is designed to reduce the fractional bit width under the premise of guaranteeing BER performance.

和

量化检测器的BER相对误差定义为

但是由于在评估量化检测器的BER性能时仿真的样本数量较少，

可能会等于0。因此，改为使用

作为相对误差。如果相对误差大于阈值ε₂，则无法保证量化检测器的性能，环境返回奖励r_t＝-1。否则，环境将考虑小数位宽。Regarding the evaluation of BER performance: Monte Carlo simulations are often used to test the BER performance of a specific detection algorithm. Considering that accurate BER requires a large number of samples for Monte Carlo simulation, it can be pre-computed before the agent starts learning, denoted as P _b . When evaluating the BER performance of the quantized detector, the environment simulates the BER performance of both the floating-point detector and the quantized detector using a small number of samples to save computation time. The corresponding BERs of the floating-point detector and the quantized detector are written as

and

The BER relative error of the quantization detector is defined as

However, due to the small number of simulated samples when evaluating the BER performance of the quantized detector,

May be equal to 0. So instead use

as a relative error. If the relative error is greater than the threshold ε ₂ , the performance of the quantized detector cannot be guaranteed, and the environment returns a reward r _t =-1. Otherwise, the environment will consider the decimal width.

相对误差小于ε₂时的奖励函数定义为

其中θ₁和θ₂是经验参数。与线性函数相比，指数函数可以让环境在小数位宽较小的情况下返回更大的奖励值，从而导致智能体更倾向于减小小数位宽。综上奖励函数如下：Express the average decimal width of all variables in the current quantization scheme as

The reward function when the relative error is less than _ε2 is defined as

where θ ₁ and θ ₂ are empirical parameters. Compared to linear functions, exponential functions allow the environment to return larger reward values with smaller fractional widths, causing the agent to be more inclined to reduce the fractional widths. In summary, the reward function is as follows:

Then, the environment selects the next intermediate variable, denoted as the k'th intermediate variable, and the state is updated to (oh(k'), oh(q _k' )).

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

和当前策略

确定动作值

其中，

和

分别表示状态空间和动作空间，策略用于表征状态和动作值之间的映射关系。In step 10215, at time t, according to the current state

and current strategy

determine action value

in,

and

represent the state space and action space, respectively, and the policy is used to represent the mapping relationship between the state and action values.

修改当前量化方案，得到修改后的量化方案。In step 10216, according to the action value

Modify the current quantization scheme to obtain a modified quantization scheme.

和

Current quantization scheme: v _Q =round(clip(v,-B,B)/C)×C, where,

and

和

Modified quantization scheme: v _Q =round(clip(v,-B',B')/C')×C', where q' _k = _{q k} +at ,

and

表示奖励函数，同时选择并返回下一状态(oh(k'),oh(q_k'))。In step 10217, the modified quantification scheme is evaluated according to the reward function, and the reward value is obtained

Represents the reward function while choosing and returning the next state (oh(k'), oh(q _k' )).

In step 10218, at each moment, the agent interacts with the environment once, that is, steps 10215-10217 are executed once. The current moment t refers to the moment when the agent is interacting with the environment, and the maximum moment t _max refers to: The preset maximum moment when the agent interacts with the environment. The above steps 10215-10217 are repeatedly performed until the current time t reaches the maximum time t _max , and the interaction data generated at each time is stored at the same time, and the interaction data includes the state, the action value and the reward value.

Optionally, in step 102, the decimal width of each intermediate variable is quantified based on the strategy stored in the trained agent, including:

Step 1025: Convert the strategy stored in the trained agent into a statistical distribution of the decimal width of each intermediate variable through a Monte Carlo simulation algorithm;

Step 1026: Determine the decimal width of the intermediate variable based on the statistical distribution of the decimal width of the intermediate variable.

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

Step 10251, during each test, for each intermediate variable, determine the action value based on the current state of the intermediate variable and the strategy stored in the trained agent;

Step 10252, change the decimal place width of the intermediate variable based on the action value;

Step 10253, determine whether the maximum number of tests has been reached, if so, go to step 10254, if not, go to step 10251;

Step 10254: Obtain the frequency that each intermediate variable takes different decimal place widths;

Step 10255: Obtain the statistical distribution of the decimal place width of each intermediate variable.

确定动作值

In step 10251, in 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 the current strategy

determine action value

更改第k个中间变量的小数位宽q′_k＝q_k+a_t。In step 10252, based on the action value

Change the decimal width of the k-th intermediate variable q′ _k = _{q k} +at .

In step 10254, after the maximum number of tests is reached, the frequency of the kth intermediate variable taking different decimal place widths is obtained.

In step 10255, after the maximum test time is large enough, the statistical distribution of the decimal width of the kth variable learned by the agent can be approximated.

In this embodiment, a Monte Carlo simulation algorithm can be used to convert the strategy stored in the trained agent into a decimal-wide statistical distribution of intermediate variables.

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

In this embodiment, the mean value of the statistical distribution of the decimal place width of the kth variable is calculated as an estimate of the expected value of _qk .

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

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

Step 1032: Extract the data in the first data set that is not within the range of the quantization scheme amplitude value to obtain a second data set; wherein, the quantization scheme amplitude value is an expression with a base of 2 and an integer bit width as an independent variable minus 2. difference of expressions in base and with negative decimal width as argument;

Step 1033: Determine the minimum integer bit width that satisfies the preset condition as the integer bit width of the intermediate variable; wherein, the preset condition is: 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 the preset threshold.

In step 1031, several data of the k-th variable intermediate variable are generated by the Monte Carlo simulation algorithm to obtain the first data set S.

In step 1032, for the 1-pq quantization scheme, the amplitude of the quantization scheme is B=2 ^p -2 ^-q , the amplitude range of the quantization scheme is [-B, B], and the first data set S is extracted that is not in [-B ,B], form the second data set S′, S′={v||v|>|B|,v∈S}.

的最小整数位宽确定为中间变量的整数位宽，即In step 1033, it will be satisfied

The minimum integer bit width of is determined as the integer bit width of the intermediate variable, that is

Among them, ε ₁ represents a preset threshold. When ε ₁ =0, then card(S')=0, indicating 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 width of different intermediate variables can be automatically realized.

To sum up, the present invention provides an automatic mixed-precision quantization method for a multiple-input multiple-output MIMO detector, which trains the agent through a deep reinforcement learning algorithm, and quantifies each agent based on the strategy stored in the trained agent. Quantizing the decimal bit width of each intermediate variable to be quantized can automatically realize the allocation of the decimal bit width of different intermediate variables; then, quantizing the integer bit width of each intermediate variable can automatically realize the integer bit width of different intermediate variables. allocation. Therefore, the present invention can automatically realize the allocation of quantization bit widths 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 this embodiment is verified below with a specific example.

In this embodiment, the following configuration is used as an example. The transmit signal vector is uniformly and randomly generated from 16QAM modulation, the channel matrix is randomly generated from the Rayleigh channel model with 8 transmit antennas and 128 receive antennas, and the MIMO detector is selected as an approximate message The AMP (approximate message passing) detector is set to 4 iterative times. After certain simulation experiments, it is found that ε ₁ =10 ^-4 , ε ₂ =0.8, θ ₁ =10 and θ ₂ =4 can achieve a good trade-off between BER performance and quantization bit width. Therefore, these hyperparameters were fixed during the following simulation experiments.

Both the policy network and the value network consist of a fully connected DNN with 6 hidden layers. The policy network is used to store the policy learned by the agent, and the value network is used to evaluate the value of the current state. The dimensions of the 6 hidden layers are 64, 128, 256, 256, 128, and 64, respectively. The dimension of the input layer of both the policy network and the value network is equal to the dimension 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 the Proximal Policy Optimization (PPO) algorithm, the attenuation factor in the PPO algorithm is set to 0.99, the ∈ clipping value is set to 0.2, and the decimal width of each variable is learned until the maximum round is reached. number.

The simulation results are analyzed as follows:

1) The effect of N _ext size

The relationship between the reward returned by the environment and the number of iterations is shown in Figure 2. In different _Next situations, as the number of iterations increases, the reward value first increases and then fluctuates around a certain value, indicating that the agent is learning how to better quantify the intermediate variables in the detection algorithm. When N _ext = 1, since the correlation between different intermediate variables is not considered, the agent can reduce the decimal width of variables to a very low value, that is, a larger reward value as shown in Figure 2. As N _ext increases, the reward value at convergence becomes smaller. This is mainly because the reward misjudgment problem makes the agent more challenging to reduce the decimal width.

The BER performance comparison between the floating-point AMP detector and the small- _quantization AMP detector in this embodiment under different Next conditions is shown in Figure 3, where the legend entry " _Next = 1" indicates this embodiment when _Next = 1 The performance curves of a small number of quantified AMP detectors in , others are similar. Although the agent allocates the least fractional bit width when _Next = 1, the performance is severely degraded. The performance of the small-quantization AMP detector for N _ext =5, N _ext =15 and N _ext =21 is basically the same as that of the floating-point AMP detector.

Figure 4 shows a plot of the average fractional width versus SNR loss for different _Next values at BER = 10 ⁻³ . The horizontal axis represents the SNR loss of the small number of quantized AMP detectors of different _Next in this example compared to the floating-point AMP detector as a benchmark. The vertical axis represents the average decimal place width. As shown in Figure 4, the average fractional bit width is the lowest when _Next = 1, but its performance loss is relatively large. When N _ext =5, the fractional quantized AMP detector can allocate less fractional bit width while maintaining the performance of the original algorithm. On the other hand, the performance loss of N _ext =15 and _Next =21 is smaller than that of N _ext =5, but the average decimal bit width is larger. Therefore, in the following analysis, N _ext is fixed to 5.

2) The influence of the size of L _a

Figure 5 shows the change of the reward value with the number of network updates under different _La conditions. Compared with the case of La ₌ 1, the convergence performance with La ₌ 2 is slightly better. In both cases _{La =} 1 and La ₌ 2, the reward value eventually converges to approximately the same value. When La ₌ 5, the convergence speed becomes very slow, and the reward value at convergence is smaller than the other two cases. Therefore, when _La is too large, neither the convergence speed nor the final fractional bit width is satisfactory. In the following discussion, we fix _La to 2.

3) Integer quantization

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

4) Comparison with the unified quantification scheme

The automatic mixed-precision quantization scheme proposed in this embodiment is compared with the unified quantization scheme. The integer bit width of uniform quantization can be obtained by PDF-based integer quantization, that is, 6 bits. As for the fractional bit width, the BER performance comparison of the floating-point AMP detector and the uniformly quantized AMP detector using different fractional bit widths is shown in Figure 7. Uniform quantized AMPs with 4-bit and 5-bit fractional bit widths suffer severe performance degradation. A uniformly quantized AMP detector with a 6-bit fractional bit width can almost perfectly restore the performance of a floating-point AMP detector. Therefore, the quantization scheme of the unified quantization AMP detector is taken as 1-6-6. The average bit width comparison between the AMP detector based on automatic mixed-precision quantization and the unified quantization AMP detector is shown in Table 1. Compared with the unified quantization AMP detector, 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.

Table 1 compares the average integer and fractional bit widths of the automatic mixed-precision quantization based AMP detector and the unified quantization AMP detector.

Table 1

It can be seen from the above analysis that the automatic mixed-precision quantization method of this embodiment avoids a large amount of quantization bit width redundancy due to the automatic realization of the allocation of quantization bit widths of different intermediate variables in the MIMO detector.

The automatic mixed-precision quantization device provided by the present invention is described below, and the automatic mixed-precision quantization device described below and the automatic mixed-precision quantization method described above can be referred to each other correspondingly.

Please refer to FIG. 8 , which is a schematic structural diagram of an automatic mixed-precision quantization apparatus provided by the present invention. As shown in Figure 8, the automatic mixed precision quantization device provided by the present invention may include:

A variable acquisition module 10, configured to acquire intermediate variables to be quantified generated by the MIMO detector in the process of performing signal detection on the MIMO system;

The first quantization module 20 is used to train the agent through a deep reinforcement learning algorithm, and quantify the decimal width of each of the intermediate variables based on the strategy stored in the trained agent;

The second quantization module 30 is configured to quantize the integer bit width of each of the intermediate variables based on the probability density function.

Optionally, the first quantization module 20 includes:

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

The updating unit is configured to control the agent to update the strategy stored in the agent based on the currently stored interaction data every time a preset number of rounds is reached, until the maximum number of rounds is reached.

Optionally, the interaction unit is specifically used for:

Randomly select multiple intermediate variables;

Initialize the decimal width of all intermediate variables to the preset maximum decimal width, and initialize the integer width of all intermediate variables to the preset maximum integer width;

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

A state is randomly selected from the plurality of states and returned.

Optionally, the interaction unit is specifically used for:

The following steps are performed at each moment until the current moment reaches the maximum moment, and the interaction data generated at each moment is stored, and the interaction data includes state, action value and reward value:

Determine the action value according to the current state and the current strategy; wherein, the action value is used to represent the amount of change in the width of the decimal place;

Modify the current quantization scheme according to the action value to obtain the modified quantization scheme;

The modified quantization scheme is evaluated according to the reward function, the reward value is obtained, and the next state is selected and returned.

Optionally, the first quantization module 20 includes:

A statistical distribution unit, used for converting the strategy stored in the trained agent into a statistical distribution of the decimal width of each intermediate variable through a Monte Carlo simulation algorithm;

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

The strategy is used to represent the mapping relationship between a state and an action value, the state includes a sequence number and a decimal width corresponding to an intermediate variable, and the action value is used to represent a change in the decimal width.

Optionally, the statistical distribution unit is specifically used for:

The following steps are performed in each test until the maximum number of tests is reached, and the frequency of taking different decimal places for each of the intermediate variables is obtained, so as to obtain the statistical distribution of the decimal places of each of the intermediate variables:

For each of the intermediate variables, determine an action value based on the current state of the intermediate variables and the strategy stored in the trained agent;

The decimal width of the intermediate variable is changed based on the action value.

Optionally, the fractional width quantization unit is specifically used for:

Based on the statistical distribution of the decimal width of the intermediate variable, the mean value of the statistical distribution of the decimal width of the intermediate variable is determined as the decimal width of the intermediate variable.

Optionally, the second quantization module 30 is specifically used for:

Generate a number of data of each of the intermediate variables through a Monte Carlo simulation algorithm to obtain a first data set;

Extracting the data in the first data set that is not within the range of the quantization scheme amplitude, to obtain a second data set; wherein, the quantization scheme amplitude is an expression with a base of 2 and an integer bit width as an independent variable minus 2 as an expression Base, the difference of expressions with negative decimal places as arguments;

Determining the minimum integer bit width that satisfies the preset condition as the integer bit width of the intermediate variable; wherein, the preset condition is: the number of data in the second data set is equal to the number of data in the first data set. The ratio is less than or equal to the preset threshold.

FIG. 9 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 8 , the electronic device may include: a processor (processor) 810, a communication interface (Communications Interface) 820, a memory (memory) 830, and a communication bus 840, The processor 810 , the communication interface 820 , and the memory 830 communicate with each other through the communication bus 840 . The processor 810 may invoke logic instructions in the memory 830 to perform an automatic mixed-precision quantization method including:

Obtaining intermediate variables to be quantized generated by the MIMO detector in the process of performing signal detection on the MIMO system;

The agent is trained through a deep reinforcement learning algorithm, and the decimal width of each intermediate variable is quantified based on the strategy stored in the trained agent;

The integer bit width of each of the intermediate variables is quantized.

In addition, the above-mentioned logic instructions in the memory 830 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

In another aspect, the present invention also provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can Execute the automatic mixed-precision quantization method provided by the above methods, the method includes:

Obtaining intermediate variables to be quantized generated by the MIMO detector in the process of performing signal detection on the MIMO system;

The agent is trained through a deep reinforcement learning algorithm, and the decimal width of each intermediate variable is quantified based on the strategy stored in the trained agent;

The integer bit width of each of the intermediate variables is quantized.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, is implemented to execute the automatic mixed-precision quantization method provided by the above methods, the method include:

Obtaining intermediate variables to be quantized generated by the MIMO detector in the process of performing signal detection on the MIMO system;

The agent is trained through a deep reinforcement learning algorithm, and the decimal width of each intermediate variable is quantified based on the strategy stored in the trained agent;

The integer bit width of each of the intermediate variables is quantized.

The device embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

From the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and certainly can also be implemented by hardware. Based on this understanding, the above-mentioned technical solutions can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic A disc, an optical disc, etc., includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; 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 invention.

Claims (11)

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;

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.

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

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.

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

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.

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

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.

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:

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.

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:

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.

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:

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.

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:

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.

9. An automatic blending precision quantization apparatus, comprising:

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.

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.

Images