Activations differ with increased number of input.

I input the first frame of an image sequence to a pre-trained network and get activation from a certain layer, no problem here. But when I test it in this loop:

% workdat is HxWx3xS image sequence
% extract_from_here is the layer I want to extract activation/feature map from
for i = 1: size(workdat, 4)
    tmp = activations(net, workdat(:,:,:,1:i), extract_from_here, 'OutputAs', 'columns');
    ac(:,i) = tmp(:,1);
end

Then, the interesting case happens when I do the comparison below:

for k = 1:size(ac,2)
    prc(:,k) = 100*abs((ac(:,1)-ac(:,k))./ac(:,1));
end

So what I do is just comparing activation of first frame in the sequence between activations of the same frame when the activations are extracted by giving HxWx3xN to the network where N is running from 1 to S.

In prc array, I've seen errors of 3.4 %. Why is this the case? Why MATLAB extracts activation of the same image different when I input it to CNN alone and when I input it to CNN with more images?

Note: When I input the same image to the network multiple times alone, I get no difference:

ac1 = activations(net, workdat(:,:,:,1), extract_from_here, 'OutputAs', 'columns');
ac2 = activations(net, workdat(:,:,:,1), extract_from_here, 'OutputAs', 'columns');
nnz(ac1-ac2) % always gives 0.

Answer 1

[Raunak Gupta]

Hi,

I tried to reproduce the same with Example mentioned here which uses Alexnet and MerchData. The percentage error calculated for me was ranging from very low of 10^-7 to 0.1%. This issue is occurring due to the ‘MiniBatch’ Name-Value pair in activations function. If you keep the MiniBatchSize as 1, there is no difference in output with the above approach. Below is the code that I tried.

unzip('MerchData.zip'); 
imds = imageDatastore('MerchData', ... 
'IncludeSubfolders',true, ... 
'LabelSource','foldernames'); 
img = readall(imds); 
net = alexnet;
bigImage = zeros(227,227,3,75); 
for i=1:75 
    bigImage(:,:,:,i) = img{i}; 
end 
for i = 1: size(bigImage, 4) 
    tmp = activations(net, bigImage(:,:,:,1:i), 'fc7', 'OutputAs', 'columns', 'MiniBatchSize',1); 
    ac(:,i) = tmp(:,1); 
end 
for k = 1:size(ac,2) 
    prc(:,k) = 100*abs((ac(:,1)-ac(:,k))./ac(:,1)); 
end 
max(prc(:)) % Gives output 0

I have brought this issue of MiniBatchSize to the notice of our developers. They will investigate the matter further.