Can I get output data from CNN convolution layer without training?

I create a CNN and assigned to a variable named layers. I want to get output from a layer without training the network such as from convolution2dLayer. For example I want to feed layers with one image and I want to get output data from "pool4" layer. Is it possible? I tried activations function but I get errors. I tried both with augmented image datastore and normal datastore. As a simple explanation all i wanted to do is make mathematical processes like convolution and pooling for images.

layers = [

imageInputLayer(inputSize,"Name","data")

convolution2dLayer([11 11],96,"Name","conv1","BiasLearnRateFactor",2,"Stride",[4 4])

reluLayer("Name","relu1")

maxPooling2dLayer([3 3],"Name","pool1","Stride",[2 2])

convolution2dLayer([5 5],128,"Name","conv2","BiasLearnRateFactor",2,"Padding",[2 2 2 2])

reluLayer("Name","relu2")

maxPooling2dLayer([3 3],"Name","pool2","Stride",[2 2])

convolution2dLayer([3 3],256,"Name","conv3","BiasLearnRateFactor",2,"Padding",[1 1 1 1])

reluLayer("Name","relu3")

maxPooling2dLayer([3 3],"Name","pool3","Stride",[2 2])

convolution2dLayer([3 3],384,"Name","conv4","BiasLearnRateFactor",2,"Padding",[1 1 1 1])

reluLayer("Name","relu4")

maxPooling2dLayer([3 3],"Name","pool4","Stride",[2 2])

fullyConnectedLayer(4096,"Name","fc1","BiasLearnRateFactor",2)

reluLayer("Name","relu6")

dropoutLayer(0.5,"Name","drop1")

fullyConnectedLayer(4096,"Name","fc2","BiasLearnRateFactor",2)

reluLayer("Name","relu7")

dropoutLayer(0.5,"Name","drop2")

fullyConnectedLayer(2,"Name","fc3","BiasLearnRateFactor",2)

softmaxLayer("Name","prob")

classificationLayer("Name","output")];

>> featuresTrain = activations(layers,augimdsTrain,'pool4','OutputAs','rows');

Check for incorrect argument data type or missing argument in call to

function 'activations'.

[MFK]

I think i found the answer. Is below code correct for this purpose?

layers = [
imageInputLayer([512 512 3],Normalization="none")
convolution2dLayer([11 11],96,"Name","conv1","BiasLearnRateFactor",2,"Stride",[4 4])
reluLayer("Name","relu1")
maxPooling2dLayer([3 3],"Name","pool1","Stride",[2 2])
convolution2dLayer([5 5],128,"Name","conv2","BiasLearnRateFactor",2,"Padding",[2 2 2 2])
reluLayer("Name","relu2")
maxPooling2dLayer([3 3],"Name","pool2","Stride",[2 2])
convolution2dLayer([3 3],256,"Name","conv3","BiasLearnRateFactor",2,"Padding",[1 1 1 1])
reluLayer("Name","relu3")
maxPooling2dLayer([3 3],"Name","pool3","Stride",[2 2])
convolution2dLayer([3 3],384,"Name","conv4","BiasLearnRateFactor",2,"Padding",[1 1 1 1])
reluLayer("Name","relu4")
maxPooling2dLayer([3 3],"Name","pool4","Stride",[2 2])
reluLayer("Name","relu6")
flattenLayer('Name','flatten1')
]

net=dlnetwork(layers);
dlX = dlarray(double((img)),'SSC');
feature= forward(net,dlX);

Answer 1

[Matt J]

When not training, it is better to use predict(), rather than forward().

layers = [
imageInputLayer([512 512 3],Normalization="none")
convolution2dLayer([11 11],96,"Name","conv1","BiasLearnRateFactor",2,"Stride",[4 4])
reluLayer("Name","relu1")
maxPooling2dLayer([3 3],"Name","pool1","Stride",[2 2])
convolution2dLayer([5 5],128,"Name","conv2","BiasLearnRateFactor",2,"Padding",[2 2 2 2])
reluLayer("Name","relu2")
maxPooling2dLayer([3 3],"Name","pool2","Stride",[2 2])
convolution2dLayer([3 3],256,"Name","conv3","BiasLearnRateFactor",2,"Padding",[1 1 1 1])
reluLayer("Name","relu3")
maxPooling2dLayer([3 3],"Name","pool3","Stride",[2 2])
convolution2dLayer([3 3],384,"Name","conv4","BiasLearnRateFactor",2,"Padding",[1 1 1 1])
reluLayer("Name","relu4")
maxPooling2dLayer([3 3],"Name","pool4","Stride",[2 2])
reluLayer("Name","relu6")
flattenLayer('Name','flatten1')
];

net=dlnetwork(layers);
dlX = dlarray(rand(512,512,3),'SSC');
feature= predict(net,dlX);

whos feature

  Name             Size            Bytes  Class      Attributes

  feature      13824x1             55300  dlarray