Load of Transfer Learning on GPU

Hi Guys,

I've been trying the transfer learning examples provided by mathworks with my own datasete (105 different labels) on the pretrained models matlab provided.

When training, I noticed that the Training Progress window dows say 'Hardware Resourse : Single GPU' and I've seen the 'Cuda library needs to be recompiled...' warning which I guess that indicates matlab does see a gpu existence.

My problem is while transfer learning using alexnet or currently training using vgg-16

• takes 3 hours and beyond (Alexnet spent 4 hours)
• GPU load in task manager shows 5% maximum and sometimes even 0% (CPU load is around 20%)...

Is this normal for a single 2080Ti card or is it not trained on the gpu for some reasons?

Thanks in advance !

-----------------------------------------------------------------------

imds = imageDatastore('Train_single','IncludeSubfolders',true,'LabelSource','foldernames');
inputSize = [224 224];
imds.ReadFcn = @(loc)imresize(imread(loc),inputSize);
[imdsTrain,imdsValidation] = splitEachLabel(imds,0.7,'randomized');
net = vgg16();
layersTransfer = net.Layers(1:end-3);
numClasses = numel(categories(imdsTrain.Labels));
layers = [
    layersTransfer
    fullyConnectedLayer(numClasses,'WeightLearnRateFactor',10,'BiasLearnRateFactor',10)
    softmaxLayer
    classificationLayer];
options = trainingOptions('sgdm', ...
    'MiniBatchSize',10, ...
    'MaxEpochs',6, ...
    'InitialLearnRate',1e-4, ...
    'ValidationData',imdsValidation, ...
    'ValidationFrequency',3, ...
    'Verbose',false, ...
    'Plots','training-progress');
netTransfer = trainNetwork(imdsTrain,layers,options'Hㄍ);
YPred = classify(netTransfer,imdsValidation);
accuracy = mean(YPred == imdsValidation.Labels);

Answer 1

[Joss Knight]

Your problem is this line:

imds.ReadFcn = @(loc)imresize(imread(loc),inputSize);

You should remove it and instead use an augmentedImageDatastore

imds = imageDatastore('Train_single','IncludeSubfolders',true,'LabelSource','foldernames');
inputSize = [224 224];
[imdsTrain,imdsValidation] = splitEachLabel(imds,0.7,'randomized');
augdsTrain = augmentedImageDatastore(inputSize, imdsTrain);
augdsValidation = augmentedImageDatastore(inputSize, imdsValidation);
net = vgg16();
layersTransfer = net.Layers(1:end-3);
numClasses = numel(categories(imdsTrain.Labels));
layers = [
    layersTransfer
    fullyConnectedLayer(numClasses,'WeightLearnRateFactor',10,'BiasLearnRateFactor',10)
    softmaxLayer
    classificationLayer];
options = trainingOptions('sgdm', ...
    'MiniBatchSize',10, ...
    'MaxEpochs',6, ...
    'InitialLearnRate',1e-4, ...
    'ValidationData',augdsValidation, ...
    'ValidationFrequency',3, ...
    'Verbose',false, ...
    'Plots','training-progress');
netTransfer = trainNetwork(augdsTrain,layers,options);
YPred = classify(netTransfer,augdsValidation);
accuracy = mean(YPred == imdsValidation.Labels);

Using a ReadFcn forces ImageDatastore to load images sequentially and therefore file I/O becomes a big bottleneck.