Use of mean_file

[Frederic Commandeur]

Dear Caffe users,

I have an issue with the use of the mean_file flag of the data layer.
I specified a binaryproto as the mean file in my train and test prototxt and it definitely improves the performances of my network when training.

However, when i want to classify a new case with the trained network, it seems that the mean file is not used.
Despite specifying the same binaryproto in my deployment prototxt file, the performances are low compared to the loss and accuracy obtained on the test set during the training.
So I explicitly remove the mean file to my new case before giving it to the net, and it appears to work very well.

I wonder what is the interest to specify the mean file in the deployment prototxt if it is not used?
Is it the correct way to do it? Or is there any other way?

Thank you very much
Frederic