Initialize a net with random weights

[Rafael Soares Padilha]

Hi all,

I'm new to Caffe and I'm trying to figure out how to initialize an untrained net and perform the classification without any previous training.

Usually, I would init my net in python with:

net = caffe.Net(model_def,      # defines the structure of the model
                model_weights,  # contains the trained weights
                caffe.TEST)     # use test mode (e.g., don't perform dropout)

Is there a way of initializing it without the model_weights, for example?

Another way would be to create a .caffemodel file with random weights. If this is possible, how would I do that? 

Thank you guys for your attention!

[Jan]

That is possible. There is another constructor for caffe.Net in python, that only takes the net.prototxt and the phase flag (e.g. caffe.TEST) as parameters: See https://github.com/BVLC/caffe/blob/master/python/caffe/_caffe.cpp#L89-L97 and https://github.com/BVLC/caffe/blob/master/python/caffe/_caffe.cpp#L240. This constructor just scaffolds the net and inits the layer params according to the config options (usually some kind of randomness).

Jan

[Rafael Soares Padilha]

Thanks, Jan! That was exactly what I was looking for!

[Rishabh Shah]

Hey Rafael, could you tell me how you managed to solve this issue?