.. _netmodcon: Building Networks with Modules and Connections ============================================== This chapter will guide you to use PyBrain's most basic structural elements: the :class:`FeedForwardNetwork` and :class:`RecurrentNetwork` classes and with them the :class:`Module` class and the :class:`Connection` class. We have already seen how to create networks with the ``buildNetwork`` shortcut - but since this technique is limited in some ways, we will now explore how to create networks from the ground up. Feed Forward Networks --------------------- We will start with a simple example, building a multi layer perceptron. First we make a new :class:`FeedForwardNetwork` object:: >>> from pybrain.structure import FeedForwardNetwork >>> n = FeedForwardNetwork() Next, we're constructing the input, hidden and output layers:: >>> from pybrain.structure import LinearLayer, SigmoidLayer >>> inLayer = LinearLayer(2) >>> hiddenLayer = SigmoidLayer(3) >>> outLayer = LinearLayer(1) There are a couple of different classes of layers. For a complete list check out the :mod:`modules` package. In order to use them, we have to add them to the network:: >>> n.addInputModule(inLayer) >>> n.addModule(hiddenLayer) >>> n.addOutputModule(outLayer) We can actually add multiple input and output modules. The net has to know which of its modules are input and output modules, in order to forward propagate input and to back propagate errors. It still needs to be explicitly determined how they should be connected. For this we use the most common connection type, which produces a full connectivity between layers, by connecting each neuron of one layer with each neuron of the other. This is implemented by the :class:`FullConnection` class:: >>> from pybrain.structure import FullConnection >>> in_to_hidden = FullConnection(inLayer, hiddenLayer) >>> hidden_to_out = FullConnection(hiddenLayer, outLayer) As with modules, we have to explicitly add them to the network:: >>> n.addConnection(in_to_hidden) >>> n.addConnection(hidden_to_out) All the elements are in place now, so we can do the final step that makes our MLP usable, which is to call the ``.sortModules()`` method:: >>> n.sortModules() This call does some internal initialization which is necessary before the net can finally be used: for example, the modules are sorted topologically. Examining a Network ------------------- We can actually print networks and examine their structure:: >>> print n FeedForwardNetwork-6 Modules: [, , ] Connections: [ 'SigmoidLayer-7'>, 'LinearLayer-8'>] Note that the output on your machine will not necessarily be the same. One way of using the network is to call its 'activate()' method with an input to be transformed:: >>> n.activate([1, 2]) array([-0.11302355]) Again, this might look different on your machine - the weights of the connections have already been initialized randomly. To have a look at those parameters, just check the ``.params`` field of the connections: We can access the trainable parameters (weights) of a connection directly, or read all weights of the network at once:: >>> in_to_hidden.params array([ 1.37751406, 1.39320901, -0.24052686, -0.67970042, -0.5999425 , -1.27774679]) >>> hidden_to_out.params array([-0.32156782, 1.09338421, 0.48784924]) The network encapsulating the modules actually holds the parameters too. You can check them out here:: >>> n.params array([ 1.37751406, 1.39320901, -0.24052686, -0.67970042, -0.5999425 , -1.27774679, -0.32156782, 1.09338421, 0.48784924]) As you can see, the last three parameters of the network equal the parameters of the second connection. Naming your Networks structure ------------------------------ In some settings it makes sense to give the parts of a network explicit identifiers. The structural components are derive from the :class:`Named` class, which means that they have an attribute `.name` by which you can identify it by. If no name is given, a new name will be generated automatically. Subclasses can also be named by passing the `name` argument on initialization:: >>> LinearLayer(2) >>> LinearLayer(2, name="foo") .. note: Although names should be unique in your PyBrain environment, this is not enforced by the library. By using names for your networks, printouts look more concise and readable. They also ensure that your network components are named in the same way every time you run your program. Using Recurrent Networks ------------------------ In order to allow recurrency, networks have to be able to "look back in time". Due to this, the :class:`RecurrentNetwork` class is different from the :class:`FeedForwardNetwork` class in the substantial way, that the complete history is saved. This is actually memory consuming, but necessary for some learning algorithms. To create a recurrent network, just do as with feedforward networks but use the appropriate class:: >>> from pybrain.structure import RecurrentNetwork >>> n = RecurrentNetwork() We will quickly build up a network that is the same as in the example above: >>> n.addInputModule(LinearLayer(2, name='in')) >>> n.addModule(SigmoidLayer(3, name='hidden')) >>> n.addOutputModule(LinearLayer(1, name='out')) >>> n.addConnection(FullConnection(n['in'], n['hidden'], name='c1')) >>> n.addConnection(FullConnection(n['hidden'], n['out'], name='c2')) The :class:`RecurrentNetwork` class has one additional method, ``.addRecurrentConnection()``, which looks back in time one timestep. We can add one from the hidden to the hidden layer:: >>> n.addRecurrentConnection(FullConnection(n['hidden'], n['hidden'], name='c3')) If we now activate the network, we will get different outputs each time:: >>> n.sortModules() >>> n.activate((2, 2)) array([-0.1959887]) >>> n.activate((2, 2)) array([-0.19623716]) >>> n.activate((2, 2)) array([-0.19675801]) Of course, we can clear the history of the network. This can be done by calling the `reset` method:: >>> n.reset() >>> n.activate((2, 2)) array([-0.1959887]) >>> n.activate((2, 2)) array([-0.19623716]) >>> n.activate((2, 2)) array([-0.19675801]) After the call to ``.reset()``, we are getting the same outputs as just after the objects creation.