Convert Image Classification Network into Regression Network using ResNet18

Tried above task following this example but have error

I will appreciate it if anyone can help, and thanks in advance. The images are in augmented image datastore

My code:

net = resnet18;

layers = net.Layers;

numResponses = 1;

layers = [

layers(1:68)

fullyConnectedLayer(numResponses)

regressionLayer];

options = trainingOptions('sgdm',...

'InitialLearnRate',0.001, ...

'ValidationData',{augimdsImagesP11Test,P11Test_betaRepBSN8},...

'Plots','training-progress',...

'Verbose',false);

net2 = trainNetwork(augimdsImagesP11Train,P11Train2_beta,layers,options);

ERROR =

Error using trainNetwork (line 170)

Layers argument must be an array of layers or a layer graph.

Answer 1

[Madhav Thakker]

Hi Kenneth,

I understand that you want to convert the pretrained resent-18 to regression network.

While training the network, we must pass the Layer graph or an array of layers to our network. I am attaching a code snippet that does the same with resnet-18 network and modifies the network to be used as a regression network.

net = resnet18;
lgraph = layerGraph(net);
 
numClasses = 1;
newLearnableLayer = fullyConnectedLayer(numClasses, ...
'Name','new_fc', ...
'WeightLearnRateFactor',10, ...
'BiasLearnRateFactor',10);
regress = regressionLayer('Name', 'new_reg');
 
lgraph = replaceLayer(lgraph,"fc1000",newLearnableLayer);
lgraph = removeLayers(lgraph, "ClassificationLayer_predictions");
lgraph = replaceLayer(lgraph,"prob",regress);

Note that, the input to the network is same as the pretrained resnet-18, i.e., images with shape (224, 224, 3).

For more information, see https://www.mathworks.com/help/deeplearning/ug/convert-classification-network-into-regression-network.html

Hope this helps.

[Kenneth Afebu]

Thanks @Madhav Thakker. It worked.

Also the images must be in 4-D arrays.