How to change Batchsize during training

Hi,

I have data with size of (224,224,3,4) in 'SSCB' format. During traing, data size is required to change to ( 7,7,3,4*1024). data is divided in smaller chuncks ( with help of a 7*7 window size ) and added to Batchsize dimension.

I have tested resize2dLayer and also designed a custom layer for reshaping the data but it seems that MATLAB layers don't include Batchsize dimension as dimension of data

i.e, size of input data X in DimChangeLayer is (224,224,3) not (224,224,3,4). how can I solve it and have a output of size ( 7,7,3,4*1024). Thanks

my code:

Input_name = 'MSA_input';
input_size= [224 , 224 , 3];
C = input_size(3) ;
H = input_size(1);
W = input_size(2);
win_s=7;
n_win1 = round(H/win_s);
n_win2 = round(W/win_s);
num_Win = n_win1*n_win2;


 l = [imageInputLayer(input_size, 'Name',Input_name, 'Normalization', 'none', 'NormalizationDimension', 'auto', 'DataAugmentation', 'none')];
 l = [l DimChangeLayer( 'dimchanger4' , [win_s,win_s , C,num_Win )]; % 

net =dlnetwork(l);

and this is my DimChangeLayer code :

classdef DimChangeLayer < nnet.layer.Layer

    properties
        Output_size
    end
    
    methods
        
        function layer = DimChangeLayer(name , out_size)
            % layer = DimChangeLayer
            
            % Set layer name.
            layer.Name = name;
          
            % Set layer description.
            layer.Description = "change dim";
            
            % layer otputsize
            layer.Output_size = out_size;
            
        end
        
        
           function [Z] = predict(layer,X)
                Z = reshape( X , layer.Output_size);
               
           end

           function [Z] = forward(layer,X)
                Z = reshape( X , layer.Output_size);
           end 
    end
end

[yanqi liu]

yes，sir，why not use reshape to change data dimension in prepare step，may be upload your data mat file to do some analysis

[sheyda Ghanbaralizadeh]

thanks @yanqi liu but this is dlnetwork and you can't use reshape in the middle of layers. you must use a layer that performs reshape function . as you can see I designed such layer but number of batches are not cosidered as data size so they won't be altered by reshape in DimChangeLayer.

Input_name = 'MSA_input';
input_size= [224 , 224 , 3];
C = input_size(3) ;
H = input_size(1);
W = input_size(2);
win_s=7;
n_win1 = round(H/win_s);
n_win2 = round(W/win_s);
num_Win = n_win1*n_win2;


l = [imageInputLayer(input_size, 'Name',Input_name, 'Normalization', 'none', 'NormalizationDimension', 'auto', 'DataAugmentation', 'none')];
% %  l = [l resize2dLayer('OutputSize',[H*W , 1])];
%  l = [l DimChangeLayer( 'dimchanger1' , [H*W , C] )];
l = [l DimChangeLayer( 'dimchanger4' , [win_s, win_s , C, 4*num_Win] )];

net =dlnetwork(l);

%% input and train
inp = randn(224,224,3);
X = zeros(224,224,3,4); 
X(:,:,:,1) = inp;
X(:,:,:,2) = inp;
X(:,:,:,3) = inp;
X(:,:,:,4) = inp;


dlX = dlarray(single(X),'SSCB');
 size(dlX)


outt = randn(224,224,3);
Y = zeros(224,224,3,4); 
Y(:,:,:,1) = outt;
Y(:,:,:,2) = outt;
Y(:,:,:,3) = outt;
Y(:,:,:,4) = outt;

dlY= dlarray(single(Y),'SSCB');
  size(dlY)

loss = dlfeval(@modelGradients,net,dlX,dlY);


function loss = modelGradients(net,dlX,dlY)
  dlYPred = forward(net,dlX);
  size(dlYPred)
  loss = My_mse(dlYPred,dlY);

end

function my_m = My_mse(x,z)
  my_m = mean((x(:)-z(:)).^2)/2;
end

this is my whole code with DimChangeLayer and input data

Answer 1

[Srivardhan Gadila]

When we call dlnetwork to create a dlnetwork object, it validates if all the layers in the layers array are valid or not and during this process some sample inputs are passed which include, few random inputs with single batch size and few inputs with batch size greater than one. Hence your custom layer fails this check as you have one constant (4 in this case) multiplied with a variable and this should be changed to as follows:

classdef DimChangeLayer < nnet.layer.Layer
    
    properties
        Output_size
    end
    
    methods
        
        function layer = DimChangeLayer(name , out_size)
            % layer = DimChangeLayer
            
            % Set layer name.
            layer.Name = name;
            
            % Set layer description.
            layer.Description = "change dim";
            
            % layer otputsize
            layer.Output_size = out_size;
            
        end
        
        
        function [Z] = predict(layer,X)
            sz = layer.Output_size;
            Z = reshape( X ,sz(1), sz(2), sz(3), []);
        end
    end
end

You can refer to this example more information: Define Custom Deep Learning Layer with Formatted Inputs. Although in the example they also inherit from nnet.layer.Formattable super class, it is not needed in this case.