how to change softmaxlayer with regression in matconvnet

Question

I am trying to train MNIST data set with single output. It means when i give an 28*28 input (image) the model gives us a just number. For example i give '5', the model give me as a result 4.9,5, 5.002 or close to 5. So I have red some documents. People tells softmaxlayer have to be changed with regression layer. For doing do this. I am using matconvnet library and its mnist example. I have changed my network and written regression layer loss function. these are my codes:

net.layers = {} ;
net.layers{end+1} = struct('type', 'conv', ...
                           'weights', {{f*randn(5,5,1,20, 'single'), zeros(1, 20, 'single')}}, ...
                           'stride', 1, ...
                           'pad', 0) ;
net.layers{end+1} = struct('type', 'pool', ...
                           'method', 'max', ...
                           'pool', [2 2], ...
                           'stride', 2, ...
                           'pad', 0) ;
net.layers{end+1} = struct('type', 'conv', ...
                           'weights', {{f*randn(5,5,20,50, 'single'),zeros(1,50,'single')}}, ...
                           'stride', 1, ...
                           'pad', 0) ;
net.layers{end+1} = struct('type', 'pool', ...
                           'method', 'max', ...
                           'pool', [2 2], ...
                           'stride', 2, ...
                           'pad', 0) ;
net.layers{end+1} = struct('type', 'conv', ...
                           'weights', {{f*randn(4,4,50,500, 'single'),  zeros(1,500,'single')}}, ...
                           'stride', 1, ...
                           'pad', 0) ;
net.layers{end+1} = struct('type', 'relu') ;
net.layers{end+1} = struct('type', 'conv', ...
                           'weights', {{f*randn(1,1,500,1, 'single'), zeros(1,1,'single')}}, ...
                           'stride', 1, ...
                           'pad', 0) ;                         
 net.layers{end+1} = struct('type', 'normloss');

this is regression loss function:

function Y = vl_normloss(X,c,dzdy)
size(X)%1 1 1 100
size(c)%1 100

if nargin <= 2

Y = 0.5*sum((squeeze(X)'-c).^2);
size(Y)%1 1
Y      % 1.7361e+03
else
size(Y)
Y = +((squeeze(X)'-c))*dzdy;
Y = reshape(Y,size(X));
end

I changed opts.errorFunction = 'multiclass' ; to 'none' Also i add

case 'normloss'
      res(i+1).x = vl_normloss(res(i).x,l.class) ;

to vl_simplenn script

But when i run train this error occurs

Error using vl_nnconv DEROUTPUT dimensions are incompatible with X and FILTERS.

Error in vl_simplenn (line 415) [res(i).dzdx, dzdw{1}, dzdw{2}] = ...

what i have to do for solving this problem? thank you

Answer 1

I found the solution. I have made a mistake. There are 2 different line have to be changed in vl_simplenn script but i changed only one line. This code works.

Answer 2

I've a question.

net.layers{end+1} = struct('type', 'conv', ...
                           'weights', {{f*randn(1,1,500,1, 'single'), zeros(1,1,'single')}}, ...
                           'stride', 1, ...
                           'pad', 0) ;                         
 net.layers{end+1} = struct('type', 'normloss');

In conv layer the output [1 1 500 1] shows a descriptor of 1x500? In the loss layer how you using that value? Shouldn't the loss layer softmax predict the class prob and then you find the highest prob corresponding class? Or in this case the output of [1 1 500 1] is the probability?