You can import the architecture of the pretrained models to MATLAB without weights for the keras, caffe and ONNX networks. You can refer to importCaffeLayers, importKerasLayers, importONNXLayers.
For the pretrained network in MATLAB we’ll get both the Layers and the trained Weights. You can remove the weights if required as follows,
net = alexnet;
layers = net.Layers;
layers(2).Weights=[];
You can also use the isprop function to check whether the property Weights or any other related trainable property exists for a layer so that you can remove them.
if isprop(layers(2),'Weights') == true
layers(2).Weights = [];
end
