Creating custom connectivity in PyBrain neural net

问题:

I want to create an artificial neural network (in PyBrain) that follows the following layout:

However, I cannot find the proper way to achieve this. The only option that I see in the documentation is the way to create fully connected layers, which is not what I want: I want some of my input nodes to be connected to the second hidden layer and not to the first one.

回答1:

The solution is to use the connection type of your choice, but with slicing parameters: inSliceFrom, inSliceTo, outSliceFrom and outSliceTo. I agree the documentation should mention this, so far it's only in the Connection class' comments.

Here is example code for your case:

#create network and modules
net = FeedForwardNetwork()
inp = LinearLayer(9)
h1 = SigmoidLayer(2)
h2 = TanhLayer(2)
outp = LinearLayer(1)
# add modules
net.addOutputModule(outp)
net.addInputModule(inp)
net.addModule(h1)
net.addModule(h2)
# create connections
net.addConnection(FullConnection(inp, h1, inSliceTo=6))
net.addConnection(FullConnection(inp, h2, inSliceFrom=6))
net.addConnection(FullConnection(h1, h2))
net.addConnection(FullConnection(h2, outp))
# finish up
net.sortModules()

回答2:

An alternative way to the one suggested by schaul is to use multiple input layers.

#create network
net = FeedForwardNetwork()

# create and add modules
input_1 = LinearLayer(6)
net.addInputModule(input_1)
input_2 = LinearLayer(3)
net.addInputModule(input_2)
h1 = SigmoidLayer(2)
net.addModule(h1)
h2 = SigmoidLayer(2)
net.addModule(h2)
outp = SigmoidLayer(1)
net.addOutputModule(outp)

# create connections
net.addConnection(FullConnection(input_1, h1))
net.addConnection(FullConnection(input_2, h2))
net.addConnection(FullConnection(h1, h2))
net.addConnection(FullConnection(h2, outp))

net.sortModules()