Is it possible to change the kernel size of ResNet50 pretrained networks?

Below is my code to change the kernel size in the pretrained network ResNet50, however it turns out these errors

• "Unable to set the 'FilterSize' property of class 'Convolution2DLayer' because it is read-only." and
• "Error in (line 41) net.Layers(convLayerIndex).FilterSize = newKernelSize;"

May I know the problem of this code? and how can I solve this?

% Load the pre-trained ResNet50 model

net = resnet50;

% Display network layers

disp(net.Layers);

% Get input layer size

Input_Layer_Size = net.Layers(1).InputSize(1:2);

% Resize training and validation images

Resized_Training_Image = augmentedImageDatastore(Input_Layer_Size, DatasetTrain);

Resized_Validation_Image = augmentedImageDatastore(Input_Layer_Size, DatasetValidation);

% Define new layers

Number_of_Classes = numel(categories(DatasetTrain.Labels));

New_Feature_Learner = fullyConnectedLayer(Number_of_Classes, ...

'Name', 'Disease_Feature_Learner', ...

'WeightLearnRateFactor', 10, ...

'BiasLearnRateFactor', 10);

New_Classifier_Layer = classificationLayer('Name', 'Disease_Classifier');

% Replace layers in the network

net = layerGraph(net);

net = replaceLayer(net, 'fc1000', New_Feature_Learner);

net = replaceLayer(net, 'ClassificationLayer_fc1000', New_Classifier_Layer);

% Find the convolutional layer whose kernel size you want to modify

convLayerName = 'conv1'; % Change this to the name of the specific layer you want to modify

convLayerIndex = arrayfun(@(l) strcmp(l.Name, convLayerName), net.Layers);

if any(convLayerIndex)

% Modify the kernel size of the convolutional layer

newKernelSize = [5, 5]; % Change this to the desired new kernel size [height, width]

net.Layers(convLayerIndex).FilterSize = newKernelSize;

else

disp('Convolutional layer not found in the network.');

end

% Set up training options

Size_of_Minibatch = 4;

Validation_Frequency = floor(numel(Resized_Training_Image.Files) / Size_of_Minibatch);

Training_Options = trainingOptions('sgdm', ...

'MiniBatchSize', Size_of_Minibatch, ...

'MaxEpochs', 6, ...

'InitialLearnRate', 3e-4, ...

'Shuffle', 'every-epoch', ...

'ValidationData', Resized_Validation_Image, ...

'ValidationFrequency', Validation_Frequency, ...

'Verbose', false, ...

'Plots', 'training-progress');

% Train the network

net = trainNetwork(Resized_Training_Image, net, Training_Options);

Answer 1

[Binaya]

Hi Amiera

I understand that you would like to change the filter size of one of the convolution layers of the "ResNet50" pre-trained model.

The task you are trying to accomplish is not possible for the following reasons:

1. The "ResNet50" model is already pre-trained in MATLAB.
2. When this model is loaded, the trained model weights are already assigned into the network.
3. Trying to modify the filter size of a convolution layer of this network would lead to change in number of weights of this new model which does leads to non-existent weights.
4. To avoid this, the convolution layer hyperparameters are "read-only" properties.

If you want the hyperparameters to be changed, you can try the following steps:

1. Load the "net" into a layer graph(say "lgraph") using "layerGraph".
2. Create a new convolution layer using "convolution2dLayer" function.
3. Replace the layer corresponding to "net.Layers(convLayerIndex)" in "lgraph" with the new convolution layer using "replaceLayer" function.
4. Convert the layer graph "lgraph" into a new network using "assembleNetwork" function.

Please refer to the following documentation for more details:

1. "layerGraph": https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.html
2. "replaceLayer": https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.replacelayer.html
3. "convolution2dLayer": https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.convolution2dlayer.html
4. "assembleNetwork": https://www.mathworks.com/help/deeplearning/ref/assemblenetwork.html

I hope this helps

Regards

Binaya