Extracting Network Model From Weights

[Alex Ensafi]

Hello,

If the ".caffemodel" file contains a serialized NetParameter with all the relevant network layers and trained weights, why do we need to specify both "-model" (the ".prototxt" file) and "-weights" (the ".caffemodel" file) when testing a network?  Shouldn't you be able to extract the ".prototxt" information from the ".caffemodel" file?  Unless I am missing something, it seems redundant to have to specify both.

Thank you,

Alex

[Przemek D]

Trained weights do not fully specify convolutional layers. You can have an arbitrarily shaped data blob and convolve it with a 3x3 filter bank.
Then there is order of layers, splits and concats, activation functions... none of that stuff is put in the caffemodel file (and probably some more that I don't know of).