Prepare Data

This example builds models for the acoustic scene classification (ASC) task, which classifies environments from the sounds they produce. ASC is a generic multiclass classification problem that is foundational for context awareness in devices, robots, and other applications [2].

Download Data Set

To run the example, you must first download the data set [2]. The size of the full data set is approximately 15.5 GB. Depending on your machine and internet connection, downloading the data can take about four hours.

The tempdir command creates a temporary folder, which might be autodeleted. To retain the downloaded data set, change the temporary folder (downloadFolder) to a local directory.

downloadFolder = tempdir; 
dataset_asc = fullfile(downloadFolder,"TUT-acoustic-scenes-2017");

if ~datasetExists(dataset_asc)
    disp("Downloading TUT-acoustic-scenes-2017 (15.5 GB) ...")
    HelperDownload_TUT_acoustic_scenes_2017(dataset_asc);
end

Load Metadata

Create a table from the training set metadata. Name the table variables FileName, AcousticScene, and SpecificLocation.

trainMetaData = readtable(fullfile(dataset_asc,"TUT-acoustic-scenes-2017-development","meta"), ...
    Delimiter={'\t'}, ...
    ReadVariableNames=false);
trainMetaData.Properties.VariableNames = ["FileName","AcousticScene","SpecificLocation"];
head(trainMetaData)

             FileName             AcousticScene    SpecificLocation
    __________________________    _____________    ________________

    {'audio/b020_90_100.wav' }      {'beach'}          {'b020'}    
    {'audio/b020_110_120.wav'}      {'beach'}          {'b020'}    
    {'audio/b020_100_110.wav'}      {'beach'}          {'b020'}    
    {'audio/b020_40_50.wav'  }      {'beach'}          {'b020'}    
    {'audio/b020_50_60.wav'  }      {'beach'}          {'b020'}    
    {'audio/b020_30_40.wav'  }      {'beach'}          {'b020'}    
    {'audio/b020_160_170.wav'}      {'beach'}          {'b020'}    
    {'audio/b020_170_180.wav'}      {'beach'}          {'b020'}    

Determine a table from the test set metadata. Name the table variables FileName, AcousticScene, and SpecificLocation.

testMetaData = readtable(fullfile(dataset_asc,"TUT-acoustic-scenes-2017-evaluation","meta"), ...
    Delimiter={'\t'}, ...
    ReadVariableNames=false);
testMetaData.Properties.VariableNames = ["FileName","AcousticScene","SpecificLocation"];
head(testMetaData)

         FileName         AcousticScene    SpecificLocation
    __________________    _____________    ________________

    {'audio/1245.wav'}      {'beach'}          {'b174'}    
    {'audio/1456.wav'}      {'beach'}          {'b174'}    
    {'audio/1318.wav'}      {'beach'}          {'b174'}    
    {'audio/967.wav' }      {'beach'}          {'b174'}    
    {'audio/203.wav' }      {'beach'}          {'b174'}    
    {'audio/777.wav' }      {'beach'}          {'b174'}    
    {'audio/231.wav' }      {'beach'}          {'b174'}    
    {'audio/768.wav' }      {'beach'}          {'b174'}    

Check if the training and test sets have common recording locations.

sharedRecordingLocations = intersect(testMetaData.SpecificLocation,trainMetaData.SpecificLocation);
disp("Number of specific recording locations in both training and test sets = " + numel(sharedRecordingLocations))

Number of specific recording locations in both training and test sets = 0

The specific recording locations in the test set do not intersect with the specific recording locations in the training set. This setup makes it easier to validate that the trained models can generalize to real-world scenarios.

The first variable in each metadata table contains the filenames of the data. Concatenate the filenames with the file paths.

trainFilePaths = fullfile(dataset_asc,"TUT-acoustic-scenes-2017-development",trainMetaData.FileName);
testFilePaths = fullfile(dataset_asc,"TUT-acoustic-scenes-2017-evaluation",testMetaData.FileName);

Create Audio Datastores

The metadata might contain files that are not included in the data set. Remove the file paths and acoustic scene labels that correspond to the missing files.

ads = audioDatastore(dataset_asc,IncludeSubfolders=true);
allFiles = ads.Files;

trainIdxToRemove = ~ismember(trainFilePaths,allFiles);
trainFilePaths(trainIdxToRemove) = [];
trainLabels = categorical(trainMetaData.AcousticScene);
trainLabels(trainIdxToRemove) = [];

testIdxToRemove = ~ismember(testFilePaths,allFiles);
testFilePaths(testIdxToRemove) = [];
testLabels = categorical(testMetaData.AcousticScene);
testLabels(testIdxToRemove) = [];

Create audio datastores for the training and test sets. Set the Labels property of the audioDatastore (Audio Toolbox) to the acoustic scene. Call countEachLabel (Audio Toolbox) to verify an even distribution of labels in both the training and test sets.

adsTrain = audioDatastore(trainFilePaths, ...
    Labels=trainLabels, ...
    IncludeSubfolders=true);
display(countEachLabel(adsTrain))

  15×2 table

         Label          Count
    ________________    _____

    beach                312 
    bus                  312 
    cafe/restaurant      312 
    car                  312 
    city_center          312 
    forest_path          312 
    grocery_store        312 
    home                 312 
    library              312 
    metro_station        312 
    office               312 
    park                 312 
    residential_area     312 
    train                312 
    tram                 312 

adsTest = audioDatastore(testFilePaths, ...
    Labels=categorical(testMetaData.AcousticScene), ...
    IncludeSubfolders=true);
display(countEachLabel(adsTest))

  15×2 table

         Label          Count
    ________________    _____

    beach                108 
    bus                  108 
    cafe/restaurant      108 
    car                  108 
    city_center          108 
    forest_path          108 
    grocery_store        108 
    home                 108 
    library              108 
    metro_station        108 
    office               108 
    park                 108 
    residential_area     108 
    train                108 
    tram                 108 

You can reduce the size of the data set to speed up the run time, which can affect accuracy. In general, reducing the size of the data set is a good practice for development and debugging. To reduce the size of the data set, you can set speedupExample to true. This example uses the entire data set.

speedupExample = false;
if speedupExample
    adsTrain = splitEachLabel(adsTrain,20);
    adsTest = splitEachLabel(adsTest,10);
end

Extract Features

You must extract features from the audio data to train machine learning models using the fitting functions in Statistics and Machine Learning Toolbox™. In this example, you extract features by using wavelet scattering. As shown in [5], wavelet scattering can provide a good representation of acoustic scenes.

Call read (Audio Toolbox) to get the data and sample rate of an audio file from the training set. Each audio file in the data set has a consistent sample rate and duration (10 seconds). Normalize the audio and listen to it. Display the corresponding label.

[data,adsInfo] = read(adsTrain);
data = data./max(data,[],"all");

fs = adsInfo.SampleRate;
sound(data,fs)

disp("Acoustic scene = " + string(adsTrain.Labels(1)))

Acoustic scene = beach

Call reset (Audio Toolbox) to return the datastore to its initial condition.

reset(adsTrain)

Define a waveletScattering (Wavelet Toolbox) object. The invariance scale and quality factors were determined through trial and error.

sf = waveletScattering(SignalLength=size(data,1), ...
    SamplingFrequency=fs,InvarianceScale=0.75,QualityFactors=[4 1]);

Convert the audio signal to mono, and then call featureMatrix (Wavelet Toolbox) to return the scattering coefficients for the scattering decomposition framework sf.

dataMono = mean(data,2);
scatteringCoeffients = featureMatrix(sf,dataMono,Transform="log");

Average the scattering coefficients over the 10-second audio clip.

featureVector = mean(scatteringCoeffients,2);
disp("Number of wavelet features per 10-second clip = " + numel(featureVector));

Number of wavelet features per 10-second clip = 286

The helper function HelperWaveletFeatureVector performs these steps for a single audio clip. Use a tall array with cellfun and HelperWaveletFeatureVector to parallelize the feature extraction steps for all the data.

When you perform calculations on tall arrays, MATLAB® uses either a parallel pool (default if you have Parallel Computing Toolbox™) or the local MATLAB session. This example uses a parallel pool. If you want to run the example using the local MATLAB session, change the global execution environment by using the mapreducer function.

Extract wavelet feature vectors for all the data in the training and test sets.

train_set_tall = tall(adsTrain);

Starting parallel pool (parpool) using the 'Processes' profile ...
Connected to parallel pool with 6 workers.

scatteringTrain = cellfun(@(x)HelperWaveletFeatureVector(x,sf),train_set_tall,UniformOutput=false);
xTrain = gather(scatteringTrain);

Evaluating tall expression using the Parallel Pool 'Processes':
- Pass 1 of 1: Completed in 24 min 59 sec
Evaluation completed in 24 min 59 sec

xTrain = cell2mat(xTrain')';

test_set_tall = tall(adsTest);
scatteringTest = cellfun(@(x)HelperWaveletFeatureVector(x,sf),test_set_tall,UniformOutput=false);
xTest = gather(scatteringTest);

Evaluating tall expression using the Parallel Pool 'Processes':
- Pass 1 of 1: Completed in 8 min 40 sec
Evaluation completed in 8 min 40 sec

xTest = cell2mat(xTest')';

Train Machine Learning Models

When training machine learning models, you can train several models and then select the best model based on its accuracy on the test data. However, when the target environment is an embedded system, or you have limited resources (such as memory and hardware speed), a good practice is to keep several models and compare them further using Model-Based Design tools. You can perform verification and validation to narrow your choice and select the model that provides acceptable accuracy, is best suited for the target environment, and meets application-specific requirements.

This example involves training these seven model types:

Train the seven models and save them in the cell array multiMdls.

MdlNames = ["Bilayered NN","Linear Discriminant", "Subspace Discriminant", ...
    "Bag of Trees", "Linear SVM","Logistic Regression","Decision Tree"]';
yTrain = grp2idx(adsTrain.Labels);
yTest = grp2idx(adsTest.Labels);

rng("default") % For reproducibility
multiMdls = cell(7,1);

% Bilayered NN
multiMdls{1} = fitcnet(xTrain,yTrain,LayerSizes=25,Standardize=true);

% Linear Discriminant
multiMdls{2} = fitcdiscr(xTrain,yTrain);

% Subspace Discriminant
multiMdls{3} = fitcensemble(xTrain,yTrain, ...
    Method="Subspace",Learners="discriminant", ...
    NumLearningCycles=30,NPredToSample=25);

% Bag of Trees
multiMdls{4} = fitcensemble(xTrain,yTrain, ...
    Method="Bag",Learners="tree", ...
    NumLearningCycles=50);

% Linear SVM
multiMdls{5} = fitcecoc(xTrain,yTrain);

% Logistic Regression
tLinear = templateLinear(Learner="logistic");
multiMdls{6} = fitcecoc(xTrain,yTrain,Learners=tLinear);

% Decision Tree
multiMdls{7} = fitctree(xTrain,yTrain);

Measure Model Accuracy and Size

Test the models with the test data by using the helper function HelperMdlMetrics. This function returns a table of model metrics, including the model accuracy as a percentage and the model size in KB.

tbl = HelperMdlMetrics(multiMdls,xTest,yTest);
tbl.Properties.RowNames = MdlNames;

Specify the output display format as bank so the numbers appear aligned when you display the values in the table tbl.

format("bank")

Display the model accuracy and the model size.

disp(tbl)

                             Accuracy    Model Size
                             ________    __________

    Bilayered NN              57.22       10612.82 
    Linear Discriminant       74.69       12514.67 
    Subspace Discriminant     69.20       22060.23 
    Bag of Trees              50.37       37585.22 
    Linear SVM                58.40       11435.21 
    Logistic Regression       21.23        1934.36 
    Decision Tree             40.56       10728.79 

Restore the default display format.

format("default")

Import Models to Simulink for Prediction

You can use the accuracy and model size metrics to select the best machine learning model for deployment in Simulink. By using more than one trained model in Simulink, you can compare model performance using additional metrics, such as speed and memory footprint, to further narrow your choice.

This example provides the Simulink model slexVandVPredictExample.slx, which includes prediction blocks for all the trained machine learning models in multiMdls.

Prepare Data

Create an input signal (input) in the form of an array for the Simulink model. The first column of the array contains the timeVector variable, which includes the points in time at which the observations enter the model. The other columns contain the wavelet feature vectors.

timeVector = (0:length(xTest)-1)';
input = [timeVector,xTest];

Save Models for MATLAB Function Blocks

Statistics and Machine Learning Toolbox provides prediction blocks for neural network, ensemble (bag of trees), SVM, logistic regression, and decision tree models:

For linear discriminant and subspace discriminant models, you must implement the prediction functionality of the models using MATLAB Function blocks. Save the model in the MATLAB formatted binary file (MAT-file) by using the saveLearnerForCoder function, and then load the file in the MATLAB Function block by using the loadLearnerForCoder function.

saveLearnerForCoder(multiMdls{2},'linDiscMdl.mat')
saveLearnerForCoder(multiMdls{3},'subDiscMdl.mat')

Simulate Simulink Model

Open the Simulink model slexVandVPredictExample.slx.

open_system("slexVandVPredictExample.slx")

Simulate the model and export the simulation output to the workspace.

simOut = sim("slexVandVPredictExample.slx");

Close the model.

close_system("slexVandVPredictExample.slx")

Examine Performance Using Simulink Profiler and SIL/PIL Manager

You can examine model performance using the execution time in both the Simulink model and the generated code.

Execute Time Profile in Simulink Using Simulink Profiler

Simulink Profiler allows you to examine the model and block execution and to identify issues that can contribute to poor simulation performance. This process helps you compare the prediction blocks for the trained machine learning models in terms of simulation speed.

Open the Simulink model slexVandVPredictExample.slx.

open_system("slexVandVPredictExample.slx")

Start Simulink Profiler.

When the simulation is complete, the Profiler Report pane opens and displays the simulation profile for the model. In the Path column, expand the Prediction subsystem. The pane lists the performance times of the prediction blocks in descending order, from slowest to fastest.

In this example, the Bag of Trees model is the slowest, and the Neural Network model is the fastest, followed by the Decision Tree and Linear Discriminant models.

Execute Time Profile in Generated Code Using SIL/PIL Manager

During simulation, the prediction blocks that use MATLAB Function blocks typically run faster than blocks that use masked subsystem implementation. This difference occurs because the software converts MATLAB Function blocks into C code and compiles the code into mex-files before running the block for simulation. Suppose you are interested in comparing the performance of the blocks for embedded applications. After enabling the software-in-the-loop (SIL) profiling, you can generate code from all the blocks and compare their performance using SIL simulation.

To configure code execution profiling for SIL mode simulation:

For more information, see Create Execution-Time Profile for Generated Code (Embedded Coder).

To generate function execution data, you must insert measurement probes into the generated code. The software can insert measurement probes for an atomic subsystem only if you set the Function packaging field (on the Code Generation tab of the Block Parameters dialog box) to either Nonreusable function or Reusable function. To enable granular execution profiling for each predict block, place the prediction blocks inside a subsystem. For each prediction block that is not a MATLAB Function block:

You do not need to place the MATLAB Function Blocks (linear discriminant and subspace discriminant models) inside a subsystem. For each subsystem block and MATLAB Function block:

The prediction implementations that use MATLAB Function blocks require support for variable signals to be enabled for code generation. To enable code generation support:

This example provides the Simulink model slexVandVPredictSILProfilingExample.slx, which is configured for SIL profiling. Open the Simulink model.

open_system("slexVandVPredictSILProfilingExample.slx")

Run the model in SIL mode and collect profiling information. On the Apps tab, in the Code Verification, Validation, and Test group, open the SIL/PIL Manager app. On the SIL/PIL tab, ensure that the SIL/PIL Mode option under the Prepare section is set to Software-in-the-loop (SIL). On the SIL/PIL tab, click Run Verification to run the simulation and log profiling metrics.

After the SIL verification is complete, in the Results section of the SIL/PIL tab, click Compare Runs and select Code Profile Analyzer. In the Code Profile Analyzer window, under the Analysis section, select Function Execution. Navigate to the end of the list of blocks to locate the prediction blocks, and compare their execution times.

As shown in the results table, the Bag of Trees model is not the slowest implementation, as it was in the Simulink Profiler results. The fastest implementation is now Decision Tree, followed by Neural Network. Linear Discriminant and Subspace Discriminant are now the slowest.

Display the total execution time for each block. In this example, the prediction blocks are sections 8 through 14 in the code execution profiling object.

executionProfileLog = out.executionProfile;
ep = executionProfileLog.Sections(8:end);
blockNames = {ep.Name}'

blockNames = 7×1 cell
    {'Neural Network'       }
    {'Bag of Trees'         }
    {'Linear SVM'           }
    {'Logistic Regression'  }
    {'Decision Tree'        }
    {'Linear Discriminant'  }
    {'Subspace Discriminant'}

totalExeTimes = [ep.TotalExecutionTimeInTicks]

totalExeTimes = 1×7 uint64 row vector

     10179112     68217005    211051307    206157401      1622284   1174648581    354784700

[totalExeTimesSorted,blockIndSorted] = sort(totalExeTimes);
exeTimeTable = table(blockNames(blockIndSorted),totalExeTimesSorted');
disp(exeTimeTable)

              Var1                  Var2   
    _________________________    __________

    {'Decision Tree'        }       1622284
    {'Neural Network'       }      10179112
    {'Bag of Trees'         }      68217005
    {'Logistic Regression'  }     206157401
    {'Linear SVM'           }     211051307
    {'Subspace Discriminant'}     354784700
    {'Linear Discriminant'  }    1174648581

Close the models.

close_system("slexVandVPredictExample.slx")
close_system("slexVandVPredictSILProfilingExample.slx")

Measure Runtime Memory Usage Using Static Code Metrics Report

Another important measure for an embedded application deployed to low-power hardware is the amount of runtime memory required for running the application on target hardware. Embedded Coder® provides features for measuring the runtime memory footprint of the application in the form of a metric report when you generate embedded C/C++ code.

The code generator performs static analysis of the generated C/C++ code and provides the metrics in the Static Code Metrics Report section of the HTML Code Generation Report. To enable this metric for each prediction block inside the Prediction subsystem, you must place each prediction block inside an atomic subsystem and change Function packaging to Nonreusable function, as described in Execute Time Profile in Generated Code Using SIL/PIL Manager. Use the Simulink model slexVandVPredictSILProfilingExample.slx, which is configured for profiling generated code. Open the Simulink model.

open_system("slexVandVPredictSILProfilingExample.slx")

To generate the Static Code Metrics Report:

After the code is generated, the code generation report opens. In the Content section, navigate to Static Code Metrics Report. Go to the Function Information section at the end of the report, and inspect the Accumulated Stack Size column in the table. The stack size is sorted in descending order, by default.

The Linear Discriminant block uses the highest amount of stack memory, followed by the Neural Network and Subspace Discriminant blocks. The Decision Tree block uses the lowest amount of runtime memory, followed closely by the Bag of Trees block.

The stack memory reported by static code metrics is the amount of memory the application uses while running. The application might include additional data, such as machine learning model parameters that are not changed at runtime. This type of data is usually stored in read-only memory, such as flash memory or ROM, and is not included in the report.

On a typical embedded system, read-only memory is usually more abundant and cheaper than runtime memory. If the machine learning model includes many parameters with large arrays, deploying the application might be problematic. The best way to find out if an application can fit the hardware is to compile, build, and deploy the application to the embedded system. The read-only application data is usually defined in the source file with the suffix "_data.c" in the code generation report. You can access this data in the Code section of the report under Data Files.

An indirect way of comparing data memory usage in machine learning models is to measure the size of the machine learning model parameters. See Measure Model Accuracy and Size for the size of the models in this example. Sort the models in ascending order by size.

sortrows(tbl,'Model Size')

ans=7×2 table
                             Accuracy    Model Size
                             ________    __________

    Logistic Regression       21.235       1934.4  
    Bilayered NN              57.222        10613  
    Decision Tree             40.556        10729  
    Linear SVM                58.395        11435  
    Linear Discriminant       74.691        12515  
    Subspace Discriminant     69.198        22060  
    Bag of Trees               50.37        37585  

The Logistic Regression model is the smallest, and the Bag of Trees model is the largest. When making decisions about memory usage, look at both the stack and read-only memory footprints. Also, consider the amount of RAM (for stack) and ROM (flash or read-only memory) available on the target hardware to make the most informed decision.

Close the model.

close_system("slexVandVPredictSILProfilingExample.slx")

Perform Code Coverage Analysis Using Simulink Coverage

Simulink® Coverage™ performs model and code coverage analysis that measures testing completeness in models and generated code. Use Simulink Coverage to gain insight into which parts of your model are exercised during simulation, and to uncover bugs in the model or gaps in your testing.

Enable Coverage and Select Metrics

Open the Simulink model slexVandVPredictExample.slx.

open_system("slexVandVPredictExample.slx")

Before you analyze coverage results for your model, you need to enable coverage and select the coverage metrics you want to see:

Analyze Coverage and View Results

Next, you need to simulate the model to collect the coverage results. First, simulate the model by clicking the Analyze Coverage button on the Coverage tab. When the simulation is complete, the app highlights the model objects based on their coverage completeness.

The blocks highlighted in green are executed during the simulation. If the block is a subsystem or masked subsystem, the green highlight means all blocks inside the subsystem are executed. In this case, four out of seven blocks have full coverage.

Masked subsystems highlighted in red are partially executed, which means some blocks under the mask are not executed. In this case, the Decision Tree block is partially executed.

Coverage information is not available for the Linear Discriminant and Subspace Discriminant MATLAB Function blocks.

If a subsystem is highlighted in red, you can navigate to the subsystem to find the specific blocks that are not executed. To find out what parts of the prediction algorithm are not executed, click the arrow on the lower left of the Decision Tree block to see the internal blocks under the block mask.

In this case, the findnode subsystem has a condition that is not executed. This condition is executed if the Decision Tree block has been pruned, which is not the case for this decision tree model. Therefore, you can ignore the red flag because the rest of the prediction algorithm is executed.

Simulink Coverage can measure code coverage and indicate the untested elements of your design. You can collect coverage for the C/C++ code in supported model elements, MATLAB Function blocks, or the code generated from models in SIL mode or processor-in-the-loop (PIL) mode. For more information, see Code Coverage for Models in Software-in-the-Loop (SIL) Mode and Processor-in-the-Loop (PIL) Mode (Simulink Coverage).

Close the model.

close_system("slexVandVPredictExample.slx")

Run Back-to-Back Test Using Simulink Test

Simulink® Test™ provides tools for authoring, managing, and executing systematic, simulation-based tests of models, generated code, and simulated or physical hardware. Simulink Test includes simulation, baseline, and equivalence test templates that let you perform functional, unit, regression, and back-to-back testing using SIL, PIL, and real-time hardware-in-the-loop (HIL) modes.

You can compare results between your model and generated code by running equivalence tests in different environments, such as model simulation and SIL, PIL, or HIL execution. Equivalence tests are sometimes referred to as back-to-back tests.

Create test cases and test harnesses automatically by using the Test for Component wizard.

Expand the Results hierarchy in the Results and Artifacts panel. Select label:1 under Equivalence Criteria Result. The upper plot shows that the output signals align, and the lower plot shows no difference between the output signals.

More About Compliance with Industry Standards

IEC Certification Kit provides tool qualification artifacts, certificates, and test suites, and generates traceability matrices. The kit helps you qualify bat365® code generation and verification tools, and streamline certification of your embedded systems to ISO® 26262, IEC 61508, EN 50128, ISO 25119, and related functional-safety standards such as IEC 62304 and EN 50657.

IEC 62304 (Medical device software – Software life cycle processes) is an international standard that describes the software development and maintenance processes required for medical device software. For example, if you require FDA‌ approval for your hearing aid software you can use IEC Certification Kit to validate your design.

For FDA approval, medical device manufacturers must validate the software tools they use in the development of a medical device by demonstrating that the tools have an acceptably low risk of harm, even in the presence of an incorrect output. IEC Certification Kit allows you to perform relevant verification and validation tasks and document the results. For more information, see FDA Software Validation.

For more information on how to use IEC Certification Kit, see Get Started (IEC Certification Kit).

Helper Functions

HelperWaveletFeatureExtractor

function features = HelperWaveletFeatureVector(x,sf)
% Copyright 2019-2021 The bat365, Inc.
x = mean(x,2);
features = featureMatrix(sf,x,Transform="log");
features = mean(features,2);
end

HelperMdlMetrics

The HelperMdlMetrics function takes a cell array of trained models (Mdls) and test data sets (X and Y) and returns a table of model metrics that includes the model accuracy as a percentage and the model size in KB. The helper function uses the whos function to estimate the model size. However, the size returned by the whos function can be larger than the actual model size required in the generated code for deployment. For example, the generated code does not include information that is not needed for prediction. Consider a CompactClassificationECOC model that uses logistic regression learners. The binary learners in a CompactClassificationECOC model object in the MATLAB workspace contain the ModelParameters property. However, the model prepared for deployment in the generated code does not contain this property.

function tbl = HelperMdlMetrics(Mdls,X,Y)
numMdl = length(Mdls);
metrics = NaN(numMdl,2);
for i = 1 : numMdl
    Mdl = Mdls{i};
    MdlInfo = whos("Mdl");
    metrics(i,:) = [(1-loss(Mdl,X,Y))*100 MdlInfo.bytes/1024];
end
tbl = array2table(metrics, ...
    VariableNames=["Accuracy","Model Size"]);
end

References

[1] Seshia, Sanjit A., et al. "Toward Verified Artificial Intelligence." Communications of the ACM, Vol. 65, No. 7, July 2022, pp. 46–55.

[2] A. Mesaros, T. Heittola, and T. Virtanen. "Acoustic Scene Classification: an Overview of DCASE 2017 Challenge Entries." In proc. International Workshop on Acoustic Signal Enhancement, 2018.

[3] Huszar, Ferenc. "Mixup: Data-Dependent Data Augmentation." InFERENCe. November 03, 2017. Accessed January 15, 2019. https://www.inference.vc/mixup-data-dependent-data-augmentation/.

[4] Han, Yoonchang, Jeongsoo Park, and Kyogu Lee. "Convolutional neural networks with binaural representations and background subtraction for acoustic scene classification." The Detection and Classification of Acoustic Scenes and Events (DCASE) (2017): pp. 1–5.

[5] Lostanlen, Vincent, and Joakim Anden. "Binaural scene classification with wavelet scattering." Technical Report, DCASE2016 Challenge, 2016.

[6] A. J. Eronen, V. T. Peltonen, J. T. Tuomi, A. P. Klapuri, S. Fagerlund, T. Sorsa, G. Lorho, and J. Huopaniemi, "Audio-based context recognition." IEEE Trans. on Audio, Speech, and Language Processing, Vol 14, No. 1, Jan. 2006, pp. 321–329.

[7] TUT Acoustic scenes 2017, Development dataset

[8] TUT Acoustic scenes 2017, Evaluation dataset