how does caffe processes deploy.prototxt?

[Ali Mousavi]

I'm still confused on how caffe processes deploy.

If I'm using pycaffe to do "Net Surgery" i need to load a deploy prototxt and an init caffemodel. If i turn a train.prototxt into deploy.prototxt, and the train.prototxt has gaussian/Xavier weight initialization will the weights in the net surgery be effected?

for example, if I have the following in deploy.prototxt:

 weight_filler {
      type: "gaussian"
      std: 0.001
    }

and I have the following in python code

net = caffe.Net("deploy.prototxt", "init.caffemodel)

print deeplab_net.params[layer][0].data

will this be effected by gaussian initialization or not(yet)?

of course I can remove that from the deploy.prototxt but I'm trying to understand the order of operation of caffe.

Thank You

[bingca...@gmail.com]

When deploy a trained model, you need the *.caffemodel as well as the deploy.prototxt, which defines *.caffemodel and the network architecture. Deploy does not change the *.caffemodel. Therefore, the weight_filter is not involved. The weight_filter is only used for training.

cheers,

Bingcai Zhang

[Ali Mousavi]

Thank You