How to use removeLayers without layerNames (How to remove layers if the layers don't have names)?

Question

Somdip Dey 2019-10-14

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/485322-how-to-use-removelayers-without-layernames-how-to-remove-layers-if-the-layers-don-t-have-names

评论： Somdip Dey 2019-10-24

When I am using transfer learning with ResNet50 I am removing the last 3 layers of ResNet as follows:

net = resnet50;
lgraph = layerGraph(net);
lgraph = removeLayers(lgraph, {'fc1000','fc1000_softmax','ClassificationLayer_fc1000'});

However, syntax for removeLayers is

newlgraph = removeLayers(lgraph,layerNames)

Right now I am trying to use transfer learning with MobileNetv2, but the last 3 layers are not named and doesn't have layerNames specified after checking the layers with the following code:

net = mobilenetv2;
lgraph = layerGraph(net);
alllayers = lgraph.Layers

How can I remove the last 3 layers or n number of layers of MobileNetv2 in order to use transfer learning with my own classifier?

2 个评论
显示无隐藏无

Jalaj Gambhir 2019-10-17

Can you post the output of your 'ver' command?

Somdip Dey 2019-10-24

在 MATLAB Online 中打开

The output of 'ver' command as follows:

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

MATLAB Version: 9.7.0.1216025 (R2019b) Update 1

MATLAB License Number: 955201

Operating System: Mac OS X Version: 10.14.4 Build: 18E226

Java Version: Java 1.8.0_202-b08 with Oracle Corporation Java HotSpot(TM) 64-Bit Server VM mixed mode

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

MATLAB Version 9.7 (R2019b)

Simulink Version 10.0 (R2019b)

Computer Vision Toolbox Version 9.1 (R2019b)

Deep Learning Toolbox Version 13.0 (R2019b)

Image Acquisition Toolbox Version 6.1 (R2019b)

Image Processing Toolbox Version 11.0 (R2019b)

Statistics and Machine Learning Toolbox Version 11.6 (R2019b)

Symbolic Math Toolbox Version 8.4 (R2019b)

My code is as follows:

%%Load pre-trained network model
    net = mobilenetv2;
    lgraph = layerGraph(net);
    %%Load DataSet
    imds = imageDatastore('/SOMEPATH/DS', ...
        'IncludeSubfolders',true, ...
        'LabelSource','foldernames');
%Split DataSet into two sets: Training & Validation. Here Split is done on
    %9:1 ratio.
    [imdsTrain,imdsValidation] = splitEachLabel(imds,0.75,'randomized');
    %Get inputSize of images
    inputSize = net.Layers(1).InputSize;
 
    %Remove last layer (3 layers from array)
    lgraph = removeLayers(lgraph, {'fc1000','fc1000_softmax','ClassificationLayer_fc1000'});
 
    %Fetch num of classes
    numClasses = numel(categories(imdsTrain.Labels));
    %Add a fully connected layer, a dropout layer, a softmax layer
    newlayers = [
        fullyConnectedLayer(numClasses,'WeightLearnRateFactor',20,'BiasLearnRateFactor',20,'Name','lastFCCustom')
        dropoutLayer(0.5,'Name','DROPOUT')
        softmaxLayer('Name','softmax')
        classificationLayer('Name','classoutput')];
    
    lgraph = addLayers(lgraph,newlayers);
    lgraph = connectLayers(lgraph,'avg_pool','lastFCCustom');
    
    layers = lgraph.Layers;
    
    connections = lgraph.Connections;
 
    layers(1:111) = freezeWeights(layers(1:111));
    lgraph = createLgraphUsingConnections(layers,connections);
....
....
....

The output in the console is as follow:

    Error using nnet.cnn.LayerGraph>iValidateLayerName (line 663)
    Layer 'ClassificationLayer_fc1000' does not exist.
    
    Error in nnet.cnn.LayerGraph/removeLayers (line 234)
                iValidateLayerName( ...
    
    Error in mobilenetv2_custom (line 32)
    lgraph = removeLayers(lgraph,
    {'fc1000','fc1000_softmax','ClassificationLayer_fc1000'});

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Jalaj Gambhir 2019-10-18

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/485322-how-to-use-removelayers-without-layernames-how-to-remove-layers-if-the-layers-don-t-have-names#answer_396953

在 MATLAB Online 中打开

Hi,

I cannot reproduce the issue at my end. The same snippet of code that you mentioned, gives the layer details including the Layer Names of all 155 Layers of MobileNetv2. And, I am able to remove layers with the procedure you mentioned with resnet50.

Another method of creating a new layer graph with the required layer is to use addLayers as mentioned below:

newlgraph = layerGraph; 
net = mobilenetv2; 
lgraph = layerGraph(net); 
alllayers = lgraph.Layers; 
newlgraph = addLayers(newlgraph, alllayers(1:end-3)) 