Документация

url = "https://github.com/brendenlake/omniglot/raw/master/python/images_background.zip";
downloadFolder = tempdir;
filename = fullfile(downloadFolder,"images_background.zip");

trainingSetFolder = fullfile(downloadFolder,'images_background');
if ~exist(trainingSetFolder,"dir")
    disp("Downloading Omniglot training data (4.5 MB)...")
    websave(filename,url);
    unzip(filename,downloadFolder);
end
disp("Training data downloaded.")

imds = omniglotImageDatastore(trainingSetFolder);

idxs = randperm(numel(imds.Files),8);

for i = 1:numel(idxs)
    subplot(4,2,i)
    imshow(readimage(imds,idxs(i)))
    title(imds.Labels(idxs(i)), "Interpreter","none");
end

batchSize = 10;
[pairImage1,pairImage2,pairLabel] = getSiameseBatch(imds,batchSize);

for i = 1:batchSize    
    if pairLabel(i) == 1
        s = "similar";
    else
        s = "dissimilar";
    end
    subplot(2,5,i)   
    imshow([pairImage1(:,:,:,i) pairImage2(:,:,:,i)]);
    title(s)
end

layers = [
    imageInputLayer([105 105 1],'Name','input1','Normalization','none')
    convolution2dLayer(10,64,'Name','conv1','WeightsInitializer','narrow-normal','BiasInitializer','narrow-normal')
    reluLayer('Name','relu1')
    maxPooling2dLayer(2,'Stride',2,'Name','maxpool1')
    convolution2dLayer(7,128,'Name','conv2','WeightsInitializer','narrow-normal','BiasInitializer','narrow-normal')
    reluLayer('Name','relu2')
    maxPooling2dLayer(2,'Stride',2,'Name','maxpool2')
    convolution2dLayer(4,128,'Name','conv3','WeightsInitializer','narrow-normal','BiasInitializer','narrow-normal')
    reluLayer('Name','relu3')
    maxPooling2dLayer(2,'Stride',2,'Name','maxpool3')
    convolution2dLayer(5,256,'Name','conv4','WeightsInitializer','narrow-normal','BiasInitializer','narrow-normal')
    reluLayer('Name','relu4')
    fullyConnectedLayer(4096,'Name','fc1','WeightsInitializer','narrow-normal','BiasInitializer','narrow-normal')];

lgraph = layerGraph(layers);

dlnet = dlnetwork(lgraph);

fcWeights = dlarray(0.01*randn(1,4096));
fcBias = dlarray(0.01*randn(1,1));

fcParams = struct(...
    "FcWeights",fcWeights,...
    "FcBias",fcBias);

numIterations = 10000;
miniBatchSize = 180;

learningRate = 6e-5;
trailingAvgSubnet = [];
trailingAvgSqSubnet = [];
trailingAvgParams = [];
trailingAvgSqParams = [];
gradDecay = 0.9;
gradDecaySq = 0.99;

executionEnvironment = "auto";

plots = "training-progress";

plotRatio = 16/9;

if plots == "training-progress"
    trainingPlot = figure;
    trainingPlot.Position(3) = plotRatio*trainingPlot.Position(4);
    trainingPlot.Visible = 'on';
    
    trainingPlotAxes = gca;
    
    lineLossTrain = animatedline(trainingPlotAxes);
    xlabel(trainingPlotAxes,"Iteration")
    ylabel(trainingPlotAxes,"Loss")
    title(trainingPlotAxes,"Loss During Training")
end

% Loop over mini-batches.
for iteration = 1:numIterations
    
    % Extract mini-batch of image pairs and pair labels
    [X1,X2,pairLabels] = getSiameseBatch(imds,miniBatchSize);
    
    % Convert mini-batch of data to dlarray. Specify the dimension labels
    % 'SSCB' (spatial, spatial, channel, batch) for image data
    dlX1 = dlarray(single(X1),'SSCB');
    dlX2 = dlarray(single(X2),'SSCB');
    
    % If training on a GPU, then convert data to gpuArray.
    if (executionEnvironment == "auto" && canUseGPU) || executionEnvironment == "gpu"
        dlX1 = gpuArray(dlX1);
        dlX2 = gpuArray(dlX2);
    end  
    
    % Evaluate the model gradients and the generator state using
    % dlfeval and the modelGradients function listed at the end of the
    % example.
    [gradientsSubnet, gradientsParams,loss] = dlfeval(@modelGradients,dlnet,fcParams,dlX1,dlX2,pairLabels);
    lossValue = double(gather(extractdata(loss)));
    
    % Update the Siamese subnetwork parameters.
    [dlnet,trailingAvgSubnet,trailingAvgSqSubnet] = ...
        adamupdate(dlnet,gradientsSubnet, ...
        trailingAvgSubnet,trailingAvgSqSubnet,iteration,learningRate,gradDecay,gradDecaySq);
    
    % Update the fullyconnect parameters.
    [fcParams,trailingAvgParams,trailingAvgSqParams] = ...
        adamupdate(fcParams,gradientsParams, ...
        trailingAvgParams,trailingAvgSqParams,iteration,learningRate,gradDecay,gradDecaySq);
      
    % Update the training loss progress plot.
    if plots == "training-progress"
        addpoints(lineLossTrain,iteration,lossValue);
    end
    drawnow;
end

url = 'https://github.com/brendenlake/omniglot/raw/master/python/images_evaluation.zip';
downloadFolder = tempdir;
filename = fullfile(downloadFolder,'images_evaluation.zip');

testingSetFolder = fullfile(downloadFolder,'images_evaluation');
if ~exist(testingSetFolder,'dir')
    disp('Downloading Omniglot test data (3.2 MB)...')
    websave(filename,url);
    unzip(filename,downloadFolder);
end
disp("Test data downloaded.")

imdsTest = omniglotImageDatastore(testingSetFolder);

numClasses = numel(unique(imdsTest.Labels))

accuracy = zeros(1,5);
accuracyBatchSize = 150;

for i = 1:5
    
    % Extract mini-batch of image pairs and pair labels
    [XAcc1,XAcc2,pairLabelsAcc] = getSiameseBatch(imdsTest,accuracyBatchSize);
    
    % Convert mini-batch of data to dlarray. Specify the dimension labels
    % 'SSCB' (spatial, spatial, channel, batch) for image data.
    dlXAcc1 = dlarray(single(XAcc1),'SSCB');
    dlXAcc2 = dlarray(single(XAcc2),'SSCB');
    
    % If using a GPU, then convert data to gpuArray.
    if (executionEnvironment == "auto" && canUseGPU) || executionEnvironment == "gpu"
       dlXAcc1 = gpuArray(dlXAcc1);
       dlXAcc2 = gpuArray(dlXAcc2);
    end    
    
    % Evaluate predictions using trained network
    dlY = predictSiamese(dlnet,fcParams,dlXAcc1,dlXAcc2);
   
    % Convert predictions to binary 0 or 1
    Y = gather(extractdata(dlY));
    Y = round(Y);
    
    % Compute average accuracye for the minibatch
    accuracy(i) = sum(Y == pairLabelsAcc)/accuracyBatchSize;
end

% Compute accuracy over all minibatches
averageAccuracy = mean(accuracy)*100

testBatchSize = 10;

[XTest1,XTest2,pairLabelsTest] = getSiameseBatch(imdsTest,testBatchSize);
    
% Convert test batch of data to dlarray. Specify the dimension labels
% 'SSCB' (spatial, spatial, channel, batch) for image data and 'CB' 
% (channel, batch) for labels
dlXTest1 = dlarray(single(XTest1),'SSCB');
dlXTest2 = dlarray(single(XTest2),'SSCB');

% If using a GPU, then convert data to gpuArray
if (executionEnvironment == "auto" && canUseGPU) || executionEnvironment == "gpu"
   dlXTest1 = gpuArray(dlXTest1);
   dlXTest2 = gpuArray(dlXTest2);
end

% Calculate the predicted probability
dlYScore = predictSiamese(dlnet,fcParams,dlXTest1,dlXTest2);
YScore = gather(extractdata(dlYScore));

% Convert predictions to binary 0 or 1
YPred = round(YScore);    

% Extract data to plot
XTest1 = extractdata(dlXTest1);
XTest2 = extractdata(dlXTest2);

% Plot images with predicted label and predicted score
testingPlot = figure;
testingPlot.Position(3) = plotRatio*testingPlot.Position(4);
testingPlot.Visible = 'on';
    
for i = 1:numel(pairLabelsTest)
     
    if YPred(i) == 1
        predLabel = "similar";
    else
        predLabel = "dissimilar" ;
    end
    
    if pairLabelsTest(i) == YPred(i)
        testStr = "\bf\color{darkgreen}Correct\rm\newline";
        
    else
        testStr = "\bf\color{red}Incorrect\rm\newline";
    end
    
    subplot(2,5,i)        
    imshow([XTest1(:,:,:,i) XTest2(:,:,:,i)]);        
    
    title(testStr + "\color{black}Predicted: " + predLabel + "\newlineScore: " + YScore(i)); 
end

function Y = forwardSiamese(dlnet,fcParams,dlX1,dlX2)
% forwardSiamese accepts the network and pair of training images, and returns a
% prediction of the probability of the pair being similar (closer to 1) or 
% dissimilar (closer to 0). Use forwardSiamese during training.

    % Pass the first image through the twin subnetwork
    F1 = forward(dlnet,dlX1);
    F1 = sigmoid(F1);
    
    % Pass the second image through the twin subnetwork
    F2 = forward(dlnet,dlX2);
    F2 = sigmoid(F2);
    
    % Subtract the feature vectors
    Y = abs(F1 - F2);
    
    % Pass the result through a fullyconnect operation
    Y = fullyconnect(Y,fcParams.FcWeights,fcParams.FcBias);
    
    % Convert to probability between 0 and 1.
    Y = sigmoid(Y);
end

function Y = predictSiamese(dlnet,fcParams,dlX1,dlX2)
% predictSiamese accepts the network and pair of images, and returns a
% prediction of the probability of the pair being similar (closer to 1)
% or dissimilar (closer to 0). Use predictSiamese during prediction.

    % Pass the first image through the twin subnetwork
    F1 = predict(dlnet,dlX1);
    F1 = sigmoid(F1);
    
    % Pass the second image through the twin subnetwork
    F2 = predict(dlnet,dlX2);
    F2 = sigmoid(F2);
    
    % Subtract the feature vectors
    Y = abs(F1 - F2);
    
    % Pass the result through a fullyconnect operation
    Y = fullyconnect(Y,fcParams.FcWeights,fcParams.FcBias);
    
    % Convert to probability between 0 and 1.
    Y = sigmoid(Y);
end

function [gradientsSubnet,gradientsParams,loss] = modelGradients(dlnet,fcParams,dlX1,dlX2,pairLabels)
% The modelGradients function calculates the binary cross-entropy loss between the
% paired images and returns the loss and the gradients of the loss with respect to
% the network learnable parameters

    % Pass the image pair through the network 
    Y = forwardSiamese(dlnet,fcParams,dlX1,dlX2);
    
    % Calculate binary cross-entropy loss
    loss = binarycrossentropy(Y,pairLabels);
       
    % Calculate gradients of the loss with respect to the network learnable
    % parameters
    [gradientsSubnet,gradientsParams] = dlgradient(loss,dlnet.Learnables,fcParams);
end

function loss = binarycrossentropy(Y,pairLabels)
    % binarycrossentropy accepts the network's prediction, Y, the true
    % label, pairLabels, and returns the binary cross-entropy loss value.
    
    % Get the precision of the prediction to prevent errors due to floating
    % point precision    
    y = extractdata(Y);
    if(isa(y,'gpuArray'))
        precision = classUnderlying(y);
    else
        precision = class(y);
    end
      
    % Convert values less than floating point precision to eps.
    Y(Y < eps(precision)) = eps(precision);
    %convert values between 1-eps and 1 to 1-eps.
    Y(Y > 1 - eps(precision)) = 1 - eps(precision);
    
    % Calculate binary cross-entropy loss for each pair
    loss = -pairLabels.*log(Y) - (1 - pairLabels).*log(1 - Y);
    
    % Sum over all pairs in minibatch and normalize.
    loss = sum(loss)/numel(pairLabels);
end

function [X1,X2,pairLabels] = getSiameseBatch(imds,miniBatchSize)
% getSiameseBatch returns a randomly selected batch or paired images. On
% average, this function produces a balanced set of similar and dissimilar
% pairs.

    pairLabels = zeros(1,miniBatchSize);
    imgSize = size(readimage(imds,1));
    X1 = zeros([imgSize 1 miniBatchSize]);
    X2 = zeros([imgSize 1 miniBatchSize]);

    for i = 1:miniBatchSize
        choice = rand(1);
        if choice < 0.5
            [pairIdx1,pairIdx2,pairLabels(i)] = getSimilarPair(imds.Labels);
        else
            [pairIdx1,pairIdx2,pairLabels(i)] = getDissimilarPair(imds.Labels);
        end
        X1(:,:,:,i) = imds.readimage(pairIdx1);
        X2(:,:,:,i) = imds.readimage(pairIdx2);
    end
end

function [pairIdx1,pairIdx2,pairLabel] = getSimilarPair(classLabel)
% getSimilarSiamesePair returns a random pair of indices for images
% that are in the same class and the similar pair label = 1.

    % Find all unique classes.
    classes = unique(classLabel);
    
    % Choose a class randomly which will be used to get a similar pair.
    classChoice = randi(numel(classes));
    
    % Find the indices of all the observations from the chosen class.
    idxs = find(classLabel==classes(classChoice));
    
    % Randomly choose two different images from the chosen class.
    pairIdxChoice = randperm(numel(idxs),2);
    pairIdx1 = idxs(pairIdxChoice(1));
    pairIdx2 = idxs(pairIdxChoice(2));
    pairLabel = 1;
end

function  [pairIdx1,pairIdx2,label] = getDissimilarPair(classLabel)
% getDissimilarSiamesePair returns a random pair of indices for images
% that are in different classes and the dissimilar pair label = 0.

    % Find all unique classes.
    classes = unique(classLabel);
    
    % Choose two different classes randomly which will be used to get a dissimilar pair.
    classesChoice = randperm(numel(classes),2);
    
    % Find the indices of all the observations from the first and second classes.
    idxs1 = find(classLabel==classes(classesChoice(1)));
    idxs2 = find(classLabel==classes(classesChoice(2)));
    
    % Randomly choose one image from each class.
    pairIdx1Choice = randi(numel(idxs1));
    pairIdx2Choice = randi(numel(idxs2));
    pairIdx1 = idxs1(pairIdx1Choice);
    pairIdx2 = idxs2(pairIdx2Choice);
    label = 0;
end

function imds = omniglotImageDatastore(path)
% omniglotImageDatastore accepts the folder path to the Omniglot
% dataset, loads the data into MATLAB as an ImageDatastore object and
% labels the images based on the file paths.

    % Load the data as an ImageDatastore object
    imds = imageDatastore(path,'IncludeSubfolders',true,'LabelSource','none');
    
    % Generate labels for each image from the names of the last two
    % directories for that image location
    files = string(imds.Files);
    parts = split(files,filesep);
    labels = join(parts(:,(end-2):(end-1)),'_');
    imds.Labels = categorical(labels);
end