# RegressionLinearCoderConfigurer

Coder configurer for linear regression model with high-dimensional data

## Description

A RegressionLinearCoderConfigurer object is a coder configurer of a linear regression model (RegressionLinear) with high-dimensional data.

A coder configurer offers convenient features to configure code generation options, generate C/C++ code, and update model parameters in the generated code.

• Configure code generation options and specify the coder attributes for linear model parameters by using object properties.

• Generate C/C++ code for the predict and update functions of the linear regression model by using generateCode. Generating C/C++ code requires MATLAB® Coder™.

• Update model parameters in the generated C/C++ code without having to regenerate the code. This feature reduces the effort required to regenerate, redeploy, and reverify C/C++ code when you retrain the linear model with new data or settings. Before updating model parameters, use validatedUpdateInputs to validate and extract the model parameters to update.

This flow chart shows the code generation workflow using a coder configurer.

For the code generation usage notes and limitations of a linear regression model, see the Code Generation sections of RegressionLinear, predict, and update.

## Creation

After training a linear regression model by using fitrlinear, create a coder configurer for the model by using learnerCoderConfigurer. Use the properties of a coder configurer to specify the coder attributes of the predict and update arguments. Then, use generateCode to generate C/C++ code based on the specified coder attributes.

## Properties

expand all

### predict Arguments

The properties listed in this section specify the coder attributes of the predict function arguments in the generated code.

Coder attributes of the predictor data to pass to the generated C/C++ code for the predict function of the linear regression model, specified as a LearnerCoderInput object.

When you create a coder configurer by using the learnerCoderConfigurer function, the input argument X determines the default values of the LearnerCoderInput coder attributes:

• SizeVector — The default value is the array size of the input X.

• If the Value attribute of the ObservationsIn property for the RegressionLinearCoderConfigurer is 'rows', then this SizeVector value is [n p], where n corresponds to the number of observations and p corresponds to the number of predictors.

• If the Value attribute of the ObservationsIn property for the RegressionLinearCoderConfigurer is 'columns', then this SizeVector value is [p n].

To switch the elements of SizeVector (for example, to change [n p] to [p n]), modify the Value attribute of the ObservationsIn property for the RegressionLinearCoderConfigurer accordingly. You cannot modify the SizeVector value directly.

• VariableDimensions — The default value is [0 0], which indicates that the array size is fixed as specified in SizeVector.

You can set this value to [1 0] if the SizeVector value is [n p] or to [0 1] if it is [p n], which indicates that the array has variable-size rows and fixed-size columns. For example, [1 0] specifies that the first value of SizeVector (n) is the upper bound for the number of rows, and the second value of SizeVector (p) is the number of columns.

• DataType — This value is single or double. The default data type depends on the data type of the input X.

• Tunability — This value must be true, meaning that predict in the generated C/C++ code always includes predictor data as an input.

You can modify the coder attributes by using dot notation. For example, to generate C/C++ code that accepts predictor data with 100 observations (in rows) of three predictor variables (in columns), specify these coder attributes of X for the coder configurer configurer:

configurer.X.SizeVector = [100 3];
configurer.X.DataType = 'double';
configurer.X.VariableDimensions = [0 0];
[0 0] indicates that the first and second dimensions of X (number of observations and number of predictor variables, respectively) have fixed sizes.

To allow the generated C/C++ code to accept predictor data with up to 100 observations, specify these coder attributes of X:

configurer.X.SizeVector = [100 3];
configurer.X.DataType = 'double';
configurer.X.VariableDimensions = [1 0];
[1 0] indicates that the first dimension of X (number of observations) has a variable size and the second dimension of X (number of predictor variables) has a fixed size. The specified number of observations, 100 in this example, becomes the maximum allowed number of observations in the generated C/C++ code. To allow any number of observations, specify the bound as Inf.

Coder attributes of the predictor data observation dimension ('ObservationsIn' name-value pair argument of predict), specified as an EnumeratedInput object.

When you create a coder configurer by using the learnerCoderConfigurer function, the 'ObservationsIn' name-value pair argument determines the default values of the EnumeratedInput coder attributes:

• Value — The default value is the predictor data observation dimension you use when creating the coder configurer, specified as 'rows' or 'columns'. If you do not specify 'ObservationsIn' when creating the coder configurer, the default value is 'rows'.

• SelectedOption — This value is always 'Built-in'. This attribute is read-only.

• BuiltInOptions — Cell array of 'rows' and 'columns'. This attribute is read-only.

• IsConstant — This value must be true.

• TunabilityThe default value is false if you specify 'ObservationsIn','rows' when creating the coder configurer, and true if you specify 'ObservationsIn','columns'. If you set Tunability to false, the software sets Value to 'rows'. If you specify other attribute values when Tunability is false, the software sets Tunability to true.

Number of output arguments to return from the generated C/C++ code for the predict function of the linear regression model, specified as 1. predict returns YHat (predicted responses) in the generated C/C++ code.

The NumOutputs property is equivalent to the '-nargout' compiler option of codegen (MATLAB Coder). This option specifies the number of output arguments in the entry-point function of code generation. The object function generateCode generates two entry-point functions—predict.m and update.m for the predict and update functions of a linear regression model, respectively—and generates C/C++ code for the two entry-point functions. The specified value for the NumOutputs property corresponds to the number of output arguments in the entry-point function predict.m.

Data Types: double

### update Arguments

The properties listed in this section specify the coder attributes of the update function arguments in the generated code. The update function takes a trained model and new model parameters as input arguments, and returns an updated version of the model that contains the new parameters. To enable updating the parameters in the generated code, you need to specify the coder attributes of the parameters before generating code. Use a LearnerCoderInput object to specify the coder attributes of each parameter. The default attribute values are based on the model parameters in the input argument Mdl of learnerCoderConfigurer.

Coder attributes of the linear predictor coefficients (Beta of a linear regression model), specified as a LearnerCoderInput object.

The default attribute values of the LearnerCoderInput object are based on the input argument Mdl of learnerCoderConfigurer:

• SizeVector — This value must be [p 1], where p is the number of predictors in Mdl.

• VariableDimensions — This value must be [0 0], indicating that the array size is fixed as specified in SizeVector.

• DataType — This value is 'single' or 'double'. The default data type is consistent with the data type of the training data you use to train Mdl.

• Tunability — This value must be true.

Coder attributes of the bias term (Bias of a linear regression model), specified as a LearnerCoderInput object.

The default attribute values of the LearnerCoderInput object are based on the input argument Mdl of learnerCoderConfigurer:

• SizeVector — This value must be [1 1].

• VariableDimensions — This value must be [0 0], indicating that the array size is fixed as specified in SizeVector.

• DataType — This value is 'single' or 'double'. The default data type is consistent with the data type of the training data you use to train Mdl.

• Tunability — This value must be true.

### Other Configurer Options

File name of the generated C/C++ code, specified as a character vector.

The object function generateCode of RegressionLinearCoderConfigurer generates C/C++ code using this file name.

The file name must not contain spaces because they can lead to code generation failures in certain operating system configurations. Also, the name must be a valid MATLAB function name.

After creating the coder configurer configurer, you can specify the file name by using dot notation.

configurer.OutputFileName = 'myModel';

Data Types: char

Verbosity level, specified as true (logical 1) or false (logical 0). The verbosity level controls the display of notification messages at the command line.

ValueDescription
true (logical 1)The software displays notification messages when your changes to the coder attributes of a parameter result in changes for other dependent parameters.
false (logical 0)The software does not display notification messages.

To enable updating machine learning model parameters in the generated code, you need to configure the coder attributes of the parameters before generating code. The coder attributes of parameters are dependent on each other, so the software stores the dependencies as configuration constraints. If you modify the coder attributes of a parameter by using a coder configurer, and the modification requires subsequent changes to other dependent parameters to satisfy configuration constraints, then the software changes the coder attributes of the dependent parameters. The verbosity level determines whether or not the software displays notification messages for these subsequent changes.

After creating the coder configurer configurer, you can modify the verbosity level by using dot notation.

configurer.Verbose = false;

Data Types: logical

### Options for Code Generation Customization

To customize the code generation workflow, use the generateFiles function and the following three properties with codegen (MATLAB Coder), instead of using the generateCode function.

After generating the two entry-point function files (predict.m and update.m) by using the generateFiles function, you can modify these files according to your code generation workflow. For example, you can modify the predict.m file to include data preprocessing, or you can add these entry-point functions to another code generation project. Then, you can generate C/C++ code by using the codegen (MATLAB Coder) function and the codegen arguments appropriate for the modified entry-point functions or code generation project. Use the three properties described in this section as a starting point to set the codegen arguments.

codegen (MATLAB Coder) arguments, specified as a cell array.

This property enables you to customize the code generation workflow. Use the generateCode function if you do not need to customize your workflow.

Instead of using generateCode with the coder configurer configurer, you can generate C/C++ code as follows:

generateFiles(configurer)
cgArgs = configurer.CodeGenerationArguments;
codegen(cgArgs{:})
If you customize the code generation workflow, modify cgArgs accordingly before calling codegen.

If you modify other properties of configurer, the software updates the CodeGenerationArguments property accordingly.

Data Types: cell

List of tunable input arguments of the entry-point function predict.m for code generation, specified as a cell array. The cell array contains another cell array that includes coder.PrimitiveType (MATLAB Coder) objects and coder.Constant (MATLAB Coder) objects.

If you modify the coder attributes of predict arguments, then the software updates the corresponding objects accordingly. If you specify the Tunability attribute as false, then the software removes the corresponding objects from the PredictInputs list.

The cell array in PredictInputs is equivalent to configurer.CodeGenerationArguments{6} for the coder configurer configurer.

Data Types: cell

List of the tunable input arguments of the entry-point function update.m for code generation, specified as a cell array of a structure including coder.PrimitiveType (MATLAB Coder) objects. Each coder.PrimitiveType object includes the coder attributes of a tunable machine learning model parameter.

If you modify the coder attributes of a model parameter by using the coder configurer properties (update Arguments properties), then the software updates the corresponding coder.PrimitiveType object accordingly. If you specify the Tunability attribute of a machine learning model parameter as false, then the software removes the corresponding coder.PrimitiveType object from the UpdateInputs list.

The structure in UpdateInputs is equivalent to configurer.CodeGenerationArguments{3} for the coder configurer configurer.

Data Types: cell

## Object Functions

 generateCode Generate C/C++ code using coder configurer generateFiles Generate MATLAB files for code generation using coder configurer validatedUpdateInputs Validate and extract machine learning model parameters to update

## Examples

collapse all

Train a machine learning model, and then generate code for the predict and update functions of the model by using a coder configurer.

Simulate 10,000 observations from the model

$y={x}_{100}+2{x}_{200}+e.$

• $X={x}_{1},...,{x}_{1000}$ is a 10,000-by-1000 numeric matrix with standard normal elements.

• $e$ is a random normal error with mean 0 and standard deviation 0.3.

rng('default') % For reproducibility
n = 10000;
p = 1000;
X = randn(n,p);
Y = X(:,100) + 2*X(:,200) + 0.3*randn(n,1);

Train a linear regression model using the simulated data. Pass the transposed predictor matrix Xnew to fitrlinear, and use the 'ObservationsIn' name-value pair argument to specify that the columns of Xnew correspond to observations.

Xnew = X';
Mdl = fitrlinear(Xnew,Y,'ObservationsIn','columns');

Mdl is a RegressionLinear object.

Create a coder configurer for the RegressionLinear model by using learnerCoderConfigurer. Specify the predictor data Xnew, and use the 'ObservationsIn' name-value pair argument to specify the observation dimension of Xnew. The learnerCoderConfigurer function uses these input arguments to configure the coder attributes of the corresponding input arguments of predict.

configurer = learnerCoderConfigurer(Mdl,Xnew,'ObservationsIn','columns')
configurer =
RegressionLinearCoderConfigurer with properties:

Update Inputs:
Beta: [1x1 LearnerCoderInput]
Bias: [1x1 LearnerCoderInput]

Predict Inputs:
X: [1x1 LearnerCoderInput]
ObservationsIn: [1x1 EnumeratedInput]

Code Generation Parameters:
NumOutputs: 1
OutputFileName: 'RegressionLinearModel'

Properties, Methods

configurer is a RegressionLinearCoderConfigurer object, which is a coder configurer of a RegressionLinear object.

To generate C/C++ code, you must have access to a C/C++ compiler that is configured properly. MATLAB Coder locates and uses a supported, installed compiler. You can use mex -setup to view and change the default compiler. For more details, see Change Default Compiler.

Generate code for the predict and update functions of the linear regression model (Mdl) with default settings.

generateCode(configurer)
generateCode creates these files in output folder:
'initialize.m', 'predict.m', 'update.m', 'RegressionLinearModel.mat'

The generateCode function completes these actions:

• Generate the MATLAB files required to generate code, including the two entry-point functions predict.m and update.m for the predict and update functions of Mdl, respectively.

• Create a MEX function named RegressionLinearModel for the two entry-point functions.

• Create the code for the MEX function in the codegen\mex\RegressionLinearModel folder.

• Copy the MEX function to the current folder.

Display the contents of the predict.m, update.m, and initialize.m files by using the type function.

type predict.m
function varargout = predict(X,varargin) %#codegen
% Autogenerated by MATLAB, 03-Oct-2020 19:22:45
[varargout{1:nargout}] = initialize('predict',X,varargin{:});
end
type update.m
function update(varargin) %#codegen
% Autogenerated by MATLAB, 03-Oct-2020 19:22:45
initialize('update',varargin{:});
end
type initialize.m
function [varargout] = initialize(command,varargin) %#codegen
% Autogenerated by MATLAB, 03-Oct-2020 19:22:45
coder.inline('always')
persistent model
if isempty(model)
end
switch(command)
case 'update'
% Update struct fields: Beta
%                       Bias
model = update(model,varargin{:});
case 'predict'
% Predict Inputs: X, ObservationsIn
X = varargin{1};
if nargin == 2
[varargout{1:nargout}] = predict(model,X);
else
PVPairs = cell(1,nargin-2);
for i = 1:nargin-2
PVPairs{1,i} = varargin{i+1};
end
[varargout{1:nargout}] = predict(model,X,PVPairs{:});
end
end
end

Train a linear regression model using a partial data set, and create a coder configurer for the model. Use the properties of the coder configurer to specify coder attributes of the linear regression model parameters. Use the object function of the coder configurer to generate C code that predicts responses for new predictor data. Then retrain the model using the entire data set, and update parameters in the generated code without regenerating the code.

Train Model

Simulate 10,000 observations from the model

$y={x}_{100}+2{x}_{200}+e.$

• $X={x}_{1},...,{x}_{1000}$ is a 10,000-by-1000 numeric matrix with standard normal elements.

• $e$ is a random normal error with mean 0 and standard deviation 0.3.

rng('default') % For reproducibility
n = 10000;
p = 1000;
X = randn(n,p);
Y = X(:,100) + 2*X(:,200) + 0.3*randn(n,1);

Train a linear regression model using the first 500 observations. Transpose the predictor data, and use the 'ObservationsIn' name-value pair argument to specify that the columns of XTrain correspond to observations.

XTrain = X(1:500,:)';
YTrain = Y(1:500);
Mdl = fitrlinear(XTrain,YTrain,'ObservationsIn','columns');

Mdl is a RegressionLinear object.

Create Coder Configurer

Create a coder configurer for the RegressionLinear model by using learnerCoderConfigurer. Specify the predictor data XTrain, and use the 'ObservationsIn' name-value pair argument to specify the observation dimension of XTrain. The learnerCoderConfigurer function uses these input arguments to configure the coder attributes of the corresponding input arguments of predict.

configurer = learnerCoderConfigurer(Mdl,XTrain,'ObservationsIn','columns');

configurer is a RegressionLinearCoderConfigurer object, which is a coder configurer of a RegressionLinear object.

Specify Coder Attributes of Parameters

Specify the coder attributes of the linear regression model parameters so that you can update the parameters in the generated code after retraining the model. This example specifies the coder attributes of the predictor data that you want to pass to the generated code.

Specify the coder attributes of the X property of configurer so that the generated code accepts any number of observations. Modify the SizeVector and VariableDimensions attributes. The SizeVector attribute specifies the upper bound of the predictor data size, and the VariableDimensions attribute specifies whether each dimension of the predictor data has a variable size or fixed size.

configurer.X.SizeVector = [1000 Inf];
configurer.X.VariableDimensions
ans = 1x2 logical array

0   1

The size of the first dimension is the number of predictor variables. This value must be fixed for a machine learning model. Because the predictor data contains 1000 predictors, the value of the SizeVector attribute must be 1000 and the value of the VariableDimensions attribute must be 0.

The size of the second dimension is the number of observations. Setting the value of the SizeVector attribute to Inf causes the software to change the value of the VariableDimensions attribute to 1. In other words, the upper bound of the size is Inf and the size is variable, meaning that the predictor data can have any number of observations. This specification is convenient if you do not know the number of observations when generating code.

The order of the dimensions in SizeVector and VariableDimensions depends on the coder attributes of ObservationsIn.

configurer.ObservationsIn
ans =
EnumeratedInput with properties:

Value: 'columns'
SelectedOption: 'Built-in'
BuiltInOptions: {'rows'  'columns'}
IsConstant: 1
Tunability: 1

When the Value attribute of the ObservationsIn property is 'columns', the first dimension of the SizeVector and VariableDimensions attributes of X corresponds to the number of predictors, and the second dimension corresponds to the number of observations. When the Value attribute of ObservationsIn is 'rows', the order of the dimensions is switched.

Generate Code

To generate C/C++ code, you must have access to a C/C++ compiler that is configured properly. MATLAB Coder locates and uses a supported, installed compiler. You can use mex -setup to view and change the default compiler. For more details, see Change Default Compiler.

Generate code for the predict and update functions of the linear regression model (Mdl).

generateCode(configurer)
generateCode creates these files in output folder:
'initialize.m', 'predict.m', 'update.m', 'RegressionLinearModel.mat'

The generateCode function completes these actions:

• Generate the MATLAB files required to generate code, including the two entry-point functions predict.m and update.m for the predict and update functions of Mdl, respectively.

• Create a MEX function named RegressionLinearModel for the two entry-point functions.

• Create the code for the MEX function in the codegen\mex\RegressionLinearModel folder.

• Copy the MEX function to the current folder.

Verify Generated Code

Pass some predictor data to verify whether the predict function of Mdl and the predict function in the MEX function return the same predicted responses. To call an entry-point function in a MEX function that has more than one entry point, specify the function name as the first input argument.

YHat = predict(Mdl,XTrain,'ObservationsIn','columns');
YHat_mex = RegressionLinearModel('predict',XTrain,'ObservationsIn','columns');

Compare YHat and YHat_mex.

max(abs(YHat-YHat_mex))
ans = 0

In general, YHat_mex might include round-off differences compared to YHat. In this case, the comparison confirms that YHat and YHat_mex are equal.

Retrain Model and Update Parameters in Generated Code

Retrain the model using the entire data set.

retrainedMdl = fitrlinear(X',Y,'ObservationsIn','columns');

Extract parameters to update by using validatedUpdateInputs. This function detects the modified model parameters in retrainedMdl and validates whether the modified parameter values satisfy the coder attributes of the parameters.

params = validatedUpdateInputs(configurer,retrainedMdl);

Update parameters in the generated code.

RegressionLinearModel('update',params)

Verify Generated Code

Compare the outputs from the predict function of retrainedMdl and the predict function in the updated MEX function.

YHat = predict(retrainedMdl,X','ObservationsIn','columns');
YHat_mex = RegressionLinearModel('predict',X','ObservationsIn','columns');
max(abs(YHat-YHat_mex))
ans = 0

The comparison confirms that YHat and YHat_mex are equal.

expand all