Modifying inceptionv3 (pre-trained deep learning network)

Hello,

I am trying to modify inceptionv3 (this is built-in function of 'Deep Learning Toolbox') to make a new network for deep learning, such as different feature map and connections of layers, etc. Of course, I can simply modify the network as below,

-----------------------------------------------------------------------------------------------------------------------------------------------------------------------

net = inceptionv3;

[learnableLayer, classLayer] = findLayersToReplace(lgraph);

new_layer = fullyConnectedLayer(numClasses, 'Name', 'flower_fc', 'WeightLearnRateFactor',10, 'BiasLearnRateFactor',10);

lgraph = replaceLayer(lgraph,learnableLayer.Name,new_layer);

-----------------------------------------------------------------------------------------------------------------------------------------------------------------------

But, I can not edit the very detail structure of this inceptionv3 to customize...

My question is how to look into the details of inceptionv3, the built-in function, to modify its structure and parameters.

I tried to modify 'inceptionv3' with this command 'edit inceptionv3' and look into the code as below

[SL: Removed the text of the inceptionv3 function that is part of Deep Learning Toolbox. Please do not post the text of MathWorks functions.]

As you see the code is too simple to represent the inceptionv3 network. How can I look into the details of inceptionv3

Thank you for help!

Answer 1

[Sean de Wolski]

You can import the layers from the base workspace of MATLAB into the Deep Network Designer and modify it there.

[JUNGUK KO]

Thank you so much for your kind answer :)

Answer 2

[Steven Lord]

I believe the network is stored in a MAT-file. Don't modify the contents of that MAT-file, as doing so could negatively impact other functions or examples in Deep Learning Toolbox that use this network. I would prefer the approach you've described, where you modify the network as returned by the inceptionv3 function.

[JUNGUK KO]

yes... I think so

thank you :)