How to Increase Alexnet Image Input layer Image size?

How to Increase Alexnet Image Input layer Image size?

[Karthiga Mahalingam]

The best option would be to adjust your input image size according to the input layer using preprocessing techniques like imresize because otherwise you would have to perform transfer learning and create your own inputImageLayer that involves changing convolutional layers and fc layer parameters- but as you can see, the latter is a lot more involved than the former.

Answer 1

[amir ali aljarrah]

can you change the input by use this code

net = alexnet;
layersTransfer = net.Layers(2:end-3);
layers = [
    imageInputLayer([100,100,3]);
    layersTransfer
    fullyConnectedLayer(numClasses)
    softmaxLayer
    classificationLayer];

[Susama Bagchi]

Hello Mr. Aljarrah,

I tried your procedure. but during the transfer learning, it is throwing error that pool5 layer size is not matching with its input image size.

Can you please help in this regard?

[Daniel Vieira]

It won't be that simple. If you change the input layer size, the convolutional layers deal with the new size just fine, they will just operate on the new size and pass on to the next layers. But as soon as you reach the first fullyConected layer you will have mismatching sizes, because the fully connected layers expect fixed sizes.

So, to change the input size you will also have to change everything from the first fully connected layer to the end. Something like this:

net = alexnet;

newSize=[100 100 3];
numClasses=10;
layersTransfer = net.Layers(2:16);

layers = [
    imageInputLayer(newSize)
    layersTransfer      % layers preserved
    fullyConnectedLayer(1000)   % new 'middle' fc
    reluLayer()                 %
    dropoutLayer(0.10)          %
    fullyConnectedLayer(numClasses) % last fc
    softmaxLayer                    %
    classificationLayer];

analyzeNetwork(layers)

I just noticed I answered a 5 year old question... I hope it helps somebody else :p

[fa der]

thhx