use pre-trained layers

[Farhad Dalirani]

I want to use the first five layers plus the RoI layer of caffenet in fast R-CNN(https://github.com/rbgirshick/fast-rcnn). And then add my new neural branch to it.
But I don't know how to do that.

[Corvus Corax]

You need to make sure the layers you want to keep are named and sized identically in each network definition.

Then you load both networks in python, and copy the weights over from one to the other, something like this:

rcnn = caffe.Net('fastrcnn.prototxt',caffe.TEST,weights='fastrcnn.caffemodel')
mine = caffe.Net('mynetwork.prototxt',caffe.TRAIN)

def copy_weights(from_net,to_net):
    for key in from_net.params.keys():
        try:            
            for i in range(len(from_net.params[key])):
                #print i,"from", np.array(from_net.params[key][i].shape), "to", np.array(to_net.params[key][i].shape)
                to_net.params[key][i].data.flat = from_net.params[key][i].data.flat
        except Exception as e:
            print("copying data for layer ",key, " failed with ",e)
            if key in to_net.params.keys():
                if (len(from_net.params[key])>len(to_net.params[key])):
                    print("target layer has no bias weights")
                else:
                    print("copying from shape",np.array(from_net.params[key][i].shape), "to", np.array(to_net.params[key][i].shape))
            else:
                print "layer not defined in destination network"


copy_weights(rcnn,mine)