Loading from caffemodel File - Python

[David Mascharka]

I'm having some trouble loading a caffemodel in python. I'm just trying to get access to the network right now so I've compiled the protobuf file to caffe_pb2. I'm using the caffemodel available from here for this. Here's what I've got:

import caffe_pb2

model = open('yu-koltun-net.caffemodel', 'rb')
binary_content = model.read()

protobuf = caffe_pb2.NetParameter()
protobuf.ParseFromString(binary_content)

layers = protobuf.layers

# now see if there are any layers
print len(layers) # outputs 0

How are there no layers? Is this the wrong way to read a caffemodel file? I know I can do

import caffe

net = caffe.Net('yu-koltun-net.prototxt', 'yu-koltun-net.caffemodel', caffe.TEST)

but I just want to access some information about the layers.

[David Mascharka]

Ah, I was using probobuf.layers which is deprecated. I should be using protobuf.layer and now it works.

[sk06]

Hi,

I want to read convolution_param specs like kernel_size, pad, stride, axis info of a convolution layer directly from caffemodel file. Please help me with some suggestions.

Thanks

SK06

[Ankit Singh]

@sk06, You could use the convolution_param from layer to get the specs (similarly how you would do with a prototxt file),

from caffe.proto.caffe_pb2 import NetParameter

with open('path_to_caffemodel.caffemodel', 'rb') as fp:
    net = NetParameter()
    net.ParseFromString(fp.read())

for idx, layer in enumerate(net.layer):
    if(layer.type == 'Convolution'):
        conv_param = layer.convolution_param
        print(idx, layer.name)
        print(conv_param.kernel_size)
        print(conv_param.pad)
        print(conv_param.stride)
        print(conv_param.axis)