Script implementing the terrain neural net

You must have created the images with CreateTerrainImages before running this script. Saves the trained network in the file LunarNet.mat

See also FindDirectory

Contents

%--------------------------------------------------------------------------
%   Copyright (c) 2021 Princeton Satellite Systems, Inc.
%   All rights reserved.
%--------------------------------------------------------------------------
%   Since version 2021.1
%--------------------------------------------------------------------------

v = ver;
any(strcmp('Deep Learning Toolbox', {v.Name}))

ans =

  logical

   1

Get the images

pImages = FindDirectory('LROImages');
c0 = cd;
cd(pImages);
label = load('Label');
cd(c0);

t          = categorical(label.t);
nClasses   = max(label.t);
imds       = imageDatastore(pImages,'labels',t);
labelCount = countEachLabel(imds);

% Display a few snapshots
NewFig('Lunar Snapshots');
n = 4;
m = 5;
ks = sort(randi(length(label.t),1,n*m)); % random selection
for i = 1:n*m
	subplot(n,m,i);
	imshow(imds.Files{ks(i)});
  title(sprintf('Image %d: %d',ks(i),label.t(ks(i))))
end

% We need the size of the images for the input layer
img = readimage(imds,1);

% Split into training and testing sets
fracTraining = 0.8;
[imdsTrain,imdsTest] = splitEachLabel(imds,fracTraining,'randomized');

Error using imageDatastore (line 138)
Input folders or files contain non-standard file extensions.

Use FileExtensions Name-Value pair to include the non-standard file extensions.

Error in TerrainNeuralNet (line 26)
imds       = imageDatastore(pImages,'labels',t);

Training

This gives the structure of the convolutional neural net

layers = [
    imageInputLayer(size(img))

    convolution2dLayer(3,8,'Padding','same')
    batchNormalizationLayer
    reluLayer

    maxPooling2dLayer(2,'Stride',2)

    convolution2dLayer(3,32,'Padding','same')
    batchNormalizationLayer
    reluLayer

    maxPooling2dLayer(2,'Stride',2)

    fullyConnectedLayer(nClasses)
    softmaxLayer
    classificationLayer
        ];
disp(layers)

options = trainingOptions('sgdm', ...
    'InitialLearnRate',0.01, ...
    'MaxEpochs',6, ...
    'MiniBatchSize',100,...
    'ValidationData',imdsTest, ...
    'ValidationFrequency',10, ...
    'ValidationPatience',inf,...
    'Shuffle','every-epoch', ...
    'Verbose',false, ...
    'Plots','training-progress');
disp(options)
fprintf('Fraction for training %8.2f%%\n',fracTraining*100);

terrainNet = trainNetwork(imdsTrain,layers,options);

Test the neural net

predLabels  = classify(terrainNet,imdsTest);
testLabels  = imdsTest.Labels;

accuracy = sum(predLabels == testLabels)/numel(testLabels);
fprintf('Accuracy is %8.2f%%\n',accuracy*100)

save('LunarNet','terrainNet')


%--------------------------------------
% $Date$
% $Revision$

Published with MATLAB® R2021a