Menu

文档

DB4AI-Query:模型训练和推断

openGauss当前版本支持了原生DB4AI能力,通过引入原生AI算子,简化操作流程,充分利用数据库优化器、执行器的优化与执行能力,获得高性能的数据库内模型训练能力。更简化的模型训练与预测流程、更高的性能表现,让开发者在更短时间内能更专注于模型的调优与数据分析上,而避免了碎片化的技术栈与冗余的代码实现。

关键字解析

表 1 DB4AI语法及关键字

  

名称

描述

语句

CREATE MODEL

创建模型并进行训练,同时保存模型。

PREDICT BY

利用已有模型进行推断。

关键字

TARGET

训练/推断任务的目标列名。

FEATURES

训练/推断任务的数据特征列名。

MODEL

训练任务的模型名称。

使用指导

  1. 本版本支持的算法概述。

    当前版本的DB4AI支持基于SGD算子的逻辑回归(目前支持二分类任务)、线性回归和支持向量机算法(分类任务),以及基于K-Means算子的Kmeans聚类算法。

  2. 模型训练语法说明。

    • CREATE MODEL

      使用“CREATE MODEL”语句可以进行模型的创建和训练。模型训练SQL语句,现有一个数据集为kmeans_2d,该表的数据内容如下:

      openGauss=# select * from kmeans_2d;
       id |              position
      ----+-------------------------------------
        1 | {74.5268815685995,88.2141939294524}
        2 | {70.9565760521218,98.8114827475511}
        3 | {76.2756086327136,23.8387574302033}
        4 | {17.8495847294107,81.8449544720352}
        5 | {81.2175785354339,57.1677675866522}
        6 | {53.97752255667,49.3158342130482}
        7 | {93.2475341879763,86.934042100329}
        8 | {72.7659293473698,19.7020415100269}
        9 | {16.5800288529135,75.7475957670249}
       10 | {81.8520747194998,40.3476078575477}
       11 | {76.796671198681,86.3827232690528}
       12 | {59.9231450678781,90.9907738864422}
       13 | {70.161884885747,19.7427458665334}
       14 | {11.1269539105706,70.9988166182302}
       15 | {80.5005071521737,65.2822235273197}
       16 | {54.7030725912191,52.151339428965}
       17 | {103.059707058128,80.8419883321039}
       18 | {85.3574452036992,14.9910179991275}
       19 | {28.6501615960151,76.6922890325077}
       20 | {69.7285806713626,49.5416352967732}
      (20 rows)
      

      该表的字段position的数据类型为 double precision[].

    • 以Kmeans为例,训练一个模型。从kmeans_2d训练集中指定position为特征列,使用kmeans算法,创建并保存模型point_kmeans。

      openGauss=# CREATE MODEL point_kmeans USING kmeans FEATURES position FROM kmeans_2d WITH num_centroids=3;
      NOTICE:  Hyperparameter max_iterations takes value DEFAULT (10)
      NOTICE:  Hyperparameter num_centroids takes value 3
      NOTICE:  Hyperparameter tolerance takes value DEFAULT (0.000010)
      NOTICE:  Hyperparameter batch_size takes value DEFAULT (10)
      NOTICE:  Hyperparameter num_features takes value DEFAULT (2)
      NOTICE:  Hyperparameter distance_function takes value DEFAULT (L2_Squared)
      NOTICE:  Hyperparameter seeding_function takes value DEFAULT (Random++)
      NOTICE:  Hyperparameter verbose takes value DEFAULT (0)
      NOTICE:  Hyperparameter seed takes value DEFAULT (0)
      MODEL CREATED. PROCESSED 1
      

      上述命令中:

      • “CREATE MODEL”语句用于模型的训练和保存。
      • USING关键字指定算法名称。
      • FEATURES用于指定训练模模型的特征,需根据训练数据表的列名添加。
      • TARGET指定模型的训练目标,它可以是训练所需数据表的列名,也可以是一个表达式,例如: price > 10000。
      • WITH用于指定训练模型时的超参数。当超参未被用户进行设置的时候,框架会使用默认数值。

        针对不同的算子,框架支持不同的超参组合:

        表 2 算子支持的超参

        算子

        超参

        GD

        (logistic_regression、linear_regression、svm_classification)

        optimizer(char); verbose(bool); max_iterations(int); max_seconds(double); batch_size(int); learning_rate(double); decay(double); tolerance(double)

        其中,SVM限定超参lambda(double)

        Kmeans

        max_iterations(int); num_centroids(int); tolerance(double); batch_size(int); num_features(int); distance_function(char); seeding_function(char*); verbose(int);seed(int)

        当前各个超参数设置的默认值和取值范围如下:

        表 3 超参的默认值以及取值范围

        算子

        超参(默认值)

        取值范围

        超参描述

        GD (logistic_regression、linear_regression、svm_classification)

        optimizer = gd(梯度下降法)

        gd/ngd(自然梯度下降)

        优化器

        verbose = false

        T/F

        日志显示

        max_iterations = 100

        (0, INT_MAX_VALUE]

        最大迭代次数

        max_seconds = 0 (不对运行时长设限制)

        [0,INT_MAX_VALUE]

        运行时长

        batch_size = 1000

        (0, MAX_MEMORY_LIMIT]

        一次训练所选取的样本数

        learning_rate = 0.8

        (0, DOUBLE_MAX_VALUE]

        学习率

        decay = 0.95

        (0, DOUBLE_MAX_VALUE]

        权值衰减率

        tolerance = 0.0005

        (0, DOUBLE_MAX_VALUE]

        公差

        seed = 0(对seed取随机值)

        [0, INT_MAX_VALUE]

        种子

        just for SVM:lambda = 0.01

        (0, DOUBLE_MAX_VALUE)

        正则化参数

        Kmeans

        max_iterations = 10

        [1, INT_MAX_VALUE]

        最大迭代次数

        num_centroids = 10

        [1, MAX_MEMORY_LIMIT]

        簇的数目

        tolerance = 0.00001

        (0,1)

        中心点误差

        batch_size = 10

        [1, MAX_MEMORY_LIMIT]

        一次训练所选取的样本数

        num_features = 2

        [1, GS_MAX_COLS]

        输入样本特征数

        distance_function = “L2_Squared”

        L1\L2\L2_Squared\Linf

        正则化方法

        seeding_function = “Random++”

        “Random++”\“KMeans||”

        初始化种子点方法

        verbose = 0U

        { 0, 1, 2 }

        冗长模式

        seed = 0U

        [0, INT_MAX_VALUE]

        种子

        MAX_MEMORY_LIMIT = 最大内存加载的元组数量

        GS_MAX_COLS = 数据库单表最大属性数量

    • 模型保存成功,则返回创建成功信息:

      MODEL CREATED. PROCESSED x
      
  3. 查看模型信息。

    当训练完成后模型会被存储到系统表gs_model_warehouse中。系统表gs_model_warehouse可以查看到关于模型本身和训练过程的相关信息。

    用户可以通过查看系统表的方式查看模型,例如查看模型名为“point_kmeans”的SQL语句如下:

    openGauss=# select * from gs_model_warehouse where modelname='point_kmeans';
    -[ RECORD 1 ]---------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
    modelname             | point_kmeans
    modelowner            | 10
    createtime            | 2021-04-30 17:30:39.59044
    processedtuples       | 20
    discardedtuples       | 0
    pre_process_time      | 6.2001e-05
    exec_time             | .000185272
    iterations            | 5
    outputtype            | 23
    modeltype             | kmeans
    query                 | CREATE MODEL point_kmeans USING kmeans FEATURES position FROM kmeans_2d WITH num_centroids=3;
    modeldata             |
    weight                |
    hyperparametersnames  | {max_iterations,num_centroids,tolerance,batch_size,num_features,distance_function,seeding_function,verbose,seed}
    hyperparametersvalues | {10,3,1e-05,10,2,L2_Squared,Random++,0,0}
    hyperparametersoids   | {23,23,701,23,23,1043,1043,23,23}
    coefnames             | {original_num_centroids,actual_num_centroids,dimension,distance_function_id,seed,coordinates}
    coefvalues            | {3,3,2,2,572368998,"(77.282589,23.724434)(74.421616,73.239455)(18.551682,76.320914)"}
    coefoids              |
    trainingscoresname    |
    trainingscoresvalue   |
    modeldescribe         | {"id:1,objective_function:542.851169,avg_distance_to_centroid:108.570234,min_distance_to_centroid:1.027078,max_distance_to_centroid:297.210108,std_dev_distance_to_centroid:105.053257,cluster_size:5","id:2,objective_function:5825.982139,avg_distance_to_centroid:529.634740,min_distance_to_centroid:100.270449,max_distance_to_centroid:990.300588,std_dev_distance_to_centroid:285.915094,cluster_size:11","id:3,objective_function:220.792591,avg_distance_to_centroid:55.198148,min_distance_to_centroid:4.216111,max_distance_to_centroid:102.117204,std_dev_distance_to_centroid:39.319118,cluster_size:4"}
    
  4. 利用已存在的模型做推断任务。

    使用“SELECT”和“PREDICT BY”关键字利用已有模型完成推断任务。

    查询语法:SELECT…PREDICT BY…(FEATURES…)…FROM…;

    openGauss=# SELECT id, PREDICT BY point_kmeans (FEATURES position) as pos FROM (select * from kmeans_2d limit 10);
     id | pos
    ----+-----
      1 |   2
      2 |   2
      3 |   1
      4 |   3
      5 |   2
      6 |   2
      7 |   2
      8 |   1
      9 |   3
     10 |   1
    (10 rows)
    

    针对相同的推断任务,同一个模型的结果是稳定的。且基于相同的超参数和训练集训练的模型也具有稳定性,同时AI模型训练存在随机成分(每个batch的数据分布、随机梯度下降),所以不同的模型间的计算表现、结果允许存在小的差别。

  5. 查看执行计划。

    使用explain语句可对“CREATE MODEL”和“PREDICT BY”的模型训练或预测过程中的执行计划进行分析。Explain关键字后可直接拼接CREATE MODEL/ PREDICT BY语句(子句),也可接可选的参数,支持的参数如下:

    表 4 EXPLAIN支持的参数

    参数名

    描述

    ANALYZE

    布尔型变量,追加运行时间、循环次数等描述信息

    VERBOSE

    布尔型变量,控制训练的运行信息是否输出到客户端

    COSTS

    布尔型变量

    CPU

    布尔型变量

    DETAIL

    布尔型变量,不可用。

    NODES

    布尔型变量,不可用

    NUM_NODES

    布尔型变量,不可用

    BUFFERS

    布尔型变量

    TIMING

    布尔型变量

    PLAN

    布尔型变量

    FORMAT

    可选格式类型:TEXT / XML / JSON / YAML

    示例:

    openGauss=# Explain CREATE MODEL patient_logisitic_regression USING logistic_regression FEATURES second_attack, treatment TARGET trait_anxiety > 50 FROM patients WITH batch_size=10, learning_rate = 0.05;
    NOTICE:  Hyperparameter batch_size takes value 10
    NOTICE:  Hyperparameter decay takes value DEFAULT (0.950000)
    NOTICE:  Hyperparameter learning_rate takes value 0.050000
    NOTICE:  Hyperparameter max_iterations takes value DEFAULT (100)
    NOTICE:  Hyperparameter max_seconds takes value DEFAULT (0)
    NOTICE:  Hyperparameter optimizer takes value DEFAULT (gd)
    NOTICE:  Hyperparameter tolerance takes value DEFAULT (0.000500)
    NOTICE:  Hyperparameter seed takes value DEFAULT (0)
    NOTICE:  Hyperparameter verbose takes value DEFAULT (FALSE)
    NOTICE:  GD shuffle cache size 212369
                                QUERY PLAN
    -------------------------------------------------------------------
     Gradient Descent  (cost=0.00..0.00 rows=0 width=0)
       ->  Seq Scan on patients  (cost=0.00..32.20 rows=1776 width=12)
    (2 rows)
    
  6. 异常场景。

    • 训练阶段。

      • 场景一:当超参数的设置超出取值范围,模型训练失败,返回ERROR,并提示错误,例如:

        openGauss=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment TARGET trait_anxiety  FROM patients WITH optimizer='aa';
        NOTICE:  Hyperparameter batch_size takes value DEFAULT (1000)
        NOTICE:  Hyperparameter decay takes value DEFAULT (0.950000)
        NOTICE:  Hyperparameter learning_rate takes value DEFAULT (0.800000)
        NOTICE:  Hyperparameter max_iterations takes value DEFAULT (100)
        NOTICE:  Hyperparameter max_seconds takes value DEFAULT (0)
        NOTICE:  Hyperparameter optimizer takes value aa
        ERROR:  Invalid hyperparameter value for optimizer. Valid values are: gd, ngd. (default is gd)
        
      • 场景二:当模型名称已存在,模型保存失败,返回ERROR,并提示错误原因:

        openGauss=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment TARGET trait_anxiety  FROM patients;
        NOTICE:  Hyperparameter batch_size takes value DEFAULT (1000)
        NOTICE:  Hyperparameter decay takes value DEFAULT (0.950000)
        NOTICE:  Hyperparameter learning_rate takes value DEFAULT (0.800000)
        NOTICE:  Hyperparameter max_iterations takes value DEFAULT (100)
        NOTICE:  Hyperparameter max_seconds takes value DEFAULT (0)
        NOTICE:  Hyperparameter optimizer takes value DEFAULT (gd)
        NOTICE:  Hyperparameter tolerance takes value DEFAULT (0.000500)
        NOTICE:  Hyperparameter seed takes value DEFAULT (0)
        NOTICE:  Hyperparameter verbose takes value DEFAULT (FALSE)
        NOTICE:  GD shuffle cache size 5502
        ERROR:  The model name "patient_linear_regression" already exists in gs_model_warehouse.
        
      • 场景三:FEATURE或者TARGETS列是*,返回ERROR,并提示错误原因:

        openGauss=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES *  TARGET trait_anxiety  FROM
        patients;
        ERROR:  FEATURES clause cannot be *
        -----------------------------------------------------------------------------------------------------------------------、
        openGauss=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment TARGET *  FROM patients;
        ERROR:  TARGET clause cannot be *
        
      • 场景四:对于无监督学习方法使用TARGET关键字,或者在监督学习方法中不适用TARGET关键字,均会返回ERROR,并提示错误原因:

        openGauss=# CREATE MODEL patient_linear_regression USING linear_regression FEATURES second_attack,treatment FROM patients;
        ERROR:  Supervised ML algorithms require TARGET clause
        -----------------------------------------------------------------------------------------------------------------------------
        CREATE MODEL patient_linear_regression USING linear_regression TARGET trait_anxiety  FROM patients;   ERROR:  Supervised ML algorithms require FEATURES clause
        
      • 场景五:当GUC参数statement_timeout设置了时长,训练超时执行的语句将被终止:执行CREATE MODEL语句。训练集的大小、训练轮数(iteration)、提前终止条件(tolerance、max_seconds)、并行线程数(nthread)等参数都会影响训练时长。当时长超过数据库限制,语句被终止模型训练失败。

    • 推断阶段。

      • 场景六:当模型名在系统表中查找不到,数据库会报ERROR:

        openGauss=# select id, PREDICT BY patient_logistic_regression (FEATURES second_attack,treatment) FROM patients;
        ERROR:  There is no model called "patient_logistic_regression".
        
      • 场景七:当做推断任务FEATURES的数据维度和数据类型与训练集存在不一致,将报ERROR,并提示错误原因,例如:

        openGauss=# select id, PREDICT BY patient_linear_regression (FEATURES second_attack) FROM patients;
        ERROR:  Invalid number of features for prediction, provided 1, expected 2
        CONTEXT:  referenced column: patient_linear_regression_pred
        -------------------------------------------------------------------------------------------------------------------------------------
        openGauss=# select id, PREDICT BY patient_linear_regression (FEATURES 1,second_attack,treatment) FROM patients;
        ERROR:  Invalid number of features for prediction, provided 3, expected 2
        CONTEXT:  referenced column: patient_linear_regression_pre
        

本文档遵循知识共享许可协议CC 4.0 (http://creativecommons.org/Licenses/by/4.0/)。