hello everyone, I have faced this error on my CNN: (Layer 5 is expected to have a different size)

I'm trying to implement the CNN algorithm that used on paper (A Deep-Network Solution Towards Model-less Obstacle Avoidance)for Lei Tai, Shaohua Li, and Ming Liu; and when I put their specification of CNN layers; I got the following error: using nnet.cnn.layer.Layer>iInferSize (line 266) Layer 5 is expected to have a different size. if anyone has an idea what is going on, which size they mean? and why I got this error? plese, let me Know.

    layers = [imageInputLayer([120 160 1],'Normalization','none');

          convolution2dLayer(5,32,'NumChannels',1);
          reluLayer();
          maxPooling2dLayer(2,'Stride',2);

         convolution2dLayer(5,32,'NumChannels',1);
         reluLayer();
          maxPooling2dLayer(2,'Stride',2);

          convolution2dLayer(5,64)
          reluLayer();
          maxPooling2dLayer(2,'Stride',2);

          fullyConnectedLayer(5);
          softmaxLayer
          classificationLayer()];

[KALYAN ACHARJYA]

The above link is not available, pls provide the detail code, so that members can help specifically.

[Khadija Al Jabri]

if true
  % code
% Load the sample data as an |ImageDatastore| object.
digitDatasetPath = fullfile(matlabroot,'toolbox','nnet','nndemos', ...
      'nndatasets','DigitDataset');
digitData = imageDatastore(digitDatasetPath, ...
      'IncludeSubfolders',true,'LabelSource','foldernames');

%% Specify Training and Test Sets trainingNumFiles = 50; rng(1) % For reproducibility [trainDigitData,testDigitData] = splitEachLabel(digitData, ... trainingNumFiles,'randomize');

%% Define the Network Layers layers = [imageInputLayer([120 160 1],'Normalization','none');

          convolution2dLayer(5,32,'NumChannels',1);
          reluLayer();
          maxPooling2dLayer(2,'Stride',2);

         convolution2dLayer(5,32,'NumChannels',1);
         reluLayer();
          maxPooling2dLayer(2,'Stride',2);

          convolution2dLayer(5,64)
          reluLayer();
          maxPooling2dLayer(2,'Stride',2);

          fullyConnectedLayer(5);
          softmaxLayer
          classificationLayer()];

%% Specify the Training Options options = trainingOptions('sgdm','MaxEpochs',15, ... 'InitialLearnRate',0.0001);

%% Train the Network Using Training Data convnet = trainNetwork(trainDigitData,layers,options);

%% Classify the Images in the Test Data and Compute Accuracy YTest = classify(convnet,testDigitData); TTest = testDigitData.Labels;

% Calculate the accuracy. accuracy = sum(YTest == TTest)/numel(TTest)

end

Answer 1

[Javier Pinzón]

Hello Khadija,

The error is located in the "NumChannels", it must have the same amout of channels of the filters used in the poir Convolution layer, so, the correct way to write it is:

          convolution2dLayer(5,32,'NumChannels',1);
          reluLayer();
          maxPooling2dLayer(2,'Stride',2);

         convolution2dLayer(5,32,'NumChannels', 32);
         reluLayer();
          maxPooling2dLayer(2,'Stride',2);

          convolution2dLayer(5,64)
          reluLayer();
          maxPooling2dLayer(2,'Stride',2);

          fullyConnectedLayer(5);
          softmaxLayer
          classificationLayer()];

In other cases, it may be no necessary to specify the number of channels, and let it be automaticaly get.

Hop it helps and thanks if the answer is accepted.

Any question, feel free to ask.

Regards,

Javier

[Khadija Al Jabri]

Its work finally, thank you so much. I appreciate your reply. and one more thing, I have 1104 depth images for 5 categories where the five categories are distributed as the following:

 Label    Count
    _____    _____

    0         68  
    1        324  
    2        386  
    3        268  
    4         58

I would like to choose randomly 750 images for training and 354 images for testing. How can I do that on Matlab?

[Javier Pinzón]

Hello Khadija,

First of all, having such small amount of images in 2 categories, may affect the recognition of them, for that reason, I highly recommend you to increase at least up to 260 the images in the category "0" and "4", unless those categories were not really important for your application.

After having a similar quantity of images, Normally you use 70~75% for training, and the rest for validation (or test). Choose this amount ramdonly in each category and the lets play =).

About if you can do that in matlab, yes, there are different ways to do, for example, saving the path in an array, and choose them randomly and separate them. there are many examples in the matlab "help".

Hope it helps.

Regards, Javier

[Khadija Al Jabri]

Hello Mr.Javier

Thank you for your cooperation.

Those data are a benchmark dataset. I got them from this paper: <https://ram-lab.com/papers/2016/iros_2016.pdf> Actually, I try to implement what they did in this paper in order to make sure that its work, then I will collect my own dataset.

Regards

Khadija

[Javier Pinzón]

I do understand, so what you can do is to usa a "randperm" comand, to choose randomly data from each category, goind thoroughtly the structure of the category, without choosing twice the same image.

Regards

[Khadija Al Jabri]

Dear Mr.Javier

could you please help me in this regard? How can I use "randperm" for each category with Matlab? I tried to do it but I can't find the appropriate command!

I looking forward to hearing from you

Regards

[Javier Pinzón]

Hello Khadija,

you can check this link:

Randperm

With that examples, you can create an array of an amount of numbers. After that, you can play with the datastore. At this moment I cannot show you an example, unfortunately.

Regards

[Khadija Al Jabri]

Mr.Javier

Many thank for your assistant. I do appreciate it.

Regards