onnx export issue with the average image

hi i export alexnet from matlab with onnx , but i dont have same features at layers so i checked the average image.

as i see there is issue with the average image.

i tried to import the network to c++(it passed some process use python cntk to dnn) and tested with all inputs zeros but i see with subtruction layer we have not same average image !

% convnet is alexnet

AverageImage=convnet.Layers(1).AverageImage;

figure,imagesc(AverageImage(:,:,1))

while in c++ output of layer data_Sub (first subtruction layer) . when all inputs are zero i suppose it should be the negative of the average .

i also check the channels and the values if there was some mix in rgb or some rotation but there is nothing clear here.

also as i see with other frameworks( tensorflow ..) and other netwroks vgg,googlnet

there is diffrence in implementation of the deep learning net . so i dont see way to have same featurs in matlab and python and c++ . only to retrain on the images in python or c++ .