Error using reshape

[snez...@gmail.com]

Hi Hagay,,

Thanks for your code.

My input data is 2000 sample hyperspectral images  as matrix array 25*25*37 (25 is the size of images and 37 is the number of bands) and 1000 test data for tree type classification( output 3 class). I tried to run Mnist.m on my data but it took a long time (more than one day) and I didn't get result. Then I resized my images to 5*5*37 with this config setting :

 %%%%%%%%%%%%%%%%%%%%%  input settings  %%%%%%%%%%%%%%%%%%

 

  net.hyperParam.sizeFmInput = [5 5]; % size of input feature map

  net.hyperParam.numFmInput  = 925;        % number of feature map of the input

 

 %%%%%%%%%%%%%%%%%%%%%  Layers specification %%%%%%%%%%%%%%%%%%

 

  net.layers{end+1}.properties = struct('type',2,'numFm',12  ,'kernel',5,'pad',2);

  net.layers{end+1}.properties = struct('type',2,'numFm',24 ,'kernel',13);

  net.layers{end+1}.properties = struct('type',1,'numFm',128);

  net.layers{end+1}.properties = struct('type',1,'numFm',3);

 % net.layers{end+1}.properties = struct('type',0,'numFm',3); %softmax

 

 %%%%%%%%%%%%%%%%%%%%%  Hyper params - training %%%%%%%%%%%%%%%%%%

 

  net.hyperParam.trainLoopCount = 1000; %on how many images to train before evaluating the network

  net.hyperParam.testImageNum   = 2000;    % after each loop, on how many images to evaluate network performance

  net.hyperParam.ni_initial     = 0.05; % ni to start training process

  net.hyperParam.ni_final       = 0.025;         % final ni to stop the training process

  net.hyperParam.normalizeNetworkInput=0; % default is '1'

  

Now I is 2000*1 that each row in I cell array has 925 (5*5*37) values. 

  

Unfortunately, I've got these errors ;

Verifying backProp..Network is OK. Verification time=221.90

Error using reshape

To RESHAPE the number of elements must not

change.

Error in GetNetworkInputs (line 134)

    image = reshape(image , inputDim);

Error in Train (line 54)

 assert(ndims((zeros([net.hyperParam.sizeFmInput

 net.hyperParam.numFmInput])))==ndims(GetNetworkInputs(dataset.I{1},net,1)),

 'Error - input does not match network

 configuration (input size + num FM)');

Error in demoMnist (line 15)

net   =  Train(MNIST,net, 15000);

Could you please tell me what's the problem? and How much time is required to process this amount of data normally?

Thanks in advance!

Somayeh

 

[Hagay Garty]

You should set numFmInput as the number of bands, ie

net.hyperParam.numFmInput  = 37

I suggest starting without scaling first, keep your data as 25x25x37 and set

net.hyperParam.sizeFmInput = [5 5]; % size of input feature map

net.hyperParam.numFmInput  = 37;        % number of feature map of the input

 

Also to get quicker updates, reduce trainLoopCount to 100 and set testImageNum to 200

If you still have issues converging, you can try with/without the normalizeNetworkInput flag

[snez...@gmail.com]

Thanks so much for your reply!

I did what you suggested but the scesRate was completely constant=67.50% and all data were classified in first class. and I've got these errors :

Error using

images.internal.imageDisplayValidateParams>validateCData

(line 115)

Multi-plane image inputs must be RGB images of

size MxNx3.

Error in

images.internal.imageDisplayValidateParams (line

27)

common_args.CData =

validateCData(common_args.CData,image_type);

Error in images.internal.imageDisplayParseInputs

(line 78)

common_args =

images.internal.imageDisplayValidateParams(common_args);

Error in imshow (line 245)

    images.internal.imageDisplayParseInputs({'Parent','Border','Reduce'},preparsed_varargin{:});

    

Error in checkNetwork (line 168)

        imshow(im,'Border','loose');

Error in demoMnist (line 18)

checkNetwork(net,Inf,MNIST,1);

BR

Somayeh

[Hagay Garty]

The error is caused since imshow cannot show an image with more than 3 colors. so you can simply remark this call.

As for the single class classification, make sure the labels vector and training data are ordered randomly. If it's not then you should get some warning when training starts.

[snez...@gmail.com]

Hi again,

Thanks for your answers!

 I've checked the labels and training data. they've ordered randomly and I didn't get any warning when training started. This is first 100 values of my labels vector for instance:

labels =

     0     2     2     0     0     1     0     0     0     0     1     0     1     2     0     0     0     0      0     0     0     0     0     2     0     0     2      0     0     0     1     0     2     1     0     1

     0     2     0     0     2     1     0     1     0      0     2     0     0     2     0     0     0     0      0     1     0     0     0     2     0     0     0      0     0     0     1     0     0     0     0     0

     2     1     0     2     1     1     0     0     1      2     0     2     0     0     0     0     0     1      0     0     0     0     1     0     2     0     2      1

 

As you see they are in random order but the number of class 0 is more. Could you please tell me if you think there is another reason for this wrong classification? Should I modify some layers or kernels?

BR

Somayeh

[snez...@gmail.com]

Hi again,

I've tested some things to change the result but unfortunately, nothing happened. Do you think is it possible to change the result with some modification in optimization algorithm such as smaller steps? In case, how can I do that?

BR

Somayeh

[Hagay Garty]

Not sure what is your issue. You can try without normalization, did you try to set 'normalizeNetworkInput ' to 0?

You can control the learning rate in the config file. Set 'net.hyperParam.ni_initial' to any value you wish