Persist the neural network built with PyBrain

A story when you want to create a neural network with the machine learning library PyBrain and make it persistent with S3.

Use pickle

Using python's standard serialization module pickle, the neural network seems to be saved and restored, and you can also ʻactivate () for objects that are properly pickle.load () `.

However, I was addicted to the problem that learning by BackpropTrainer.train () etc. was not restarted. The detailed behavior of pickle is unknown, and the cause has not been investigated.

Use NetworkWriter

Using NetworkWriter, which is provided as a utility of PyBrain, solves the above problem and resumes learning. It's easy to use, so I think this is fine.

import

python

from pybrain.tools.customxml import NetworkWriter, NetworkReader

Export

python

NetworkWriter.writeToFile(network, filename_local)

Read

python

network = NetworkReader.readFrom(filename_local)

The file is saved in XML, and I can understand that it is doing neural network, and it is somewhat more secure than the pickle format.

python

<?xml version="1.0" ?>
<PyBrain>
        <Network class="pybrain.structure.networks.feedforward.FeedForwardNetwork" name="FeedForwardNetwork-8">
                <name val="u'FeedForwardNetwork-8'"/>
                <Modules>
                        <LinearLayer class="pybrain.structure.modules.linearlayer.LinearLayer" inmodule="True" name="in">
                                <dim val="8"/>
                                <name val="'in'"/>
                        </LinearLayer>
                        <LinearLayer class="pybrain.structure.modules.linearlayer.LinearLayer" name="out" outmodule="True">
                                <dim val="1"/>
                                <name val="'out'"/>
                        </LinearLayer>
                        <BiasUnit class="pybrain.structure.modules.biasunit.BiasUnit" name="bias">
                                <name val="'bias'"/>
                        </BiasUnit>
                        <SigmoidLayer class="pybrain.structure.modules.sigmoidlayer.SigmoidLayer" name="hidden0">
                                <dim val="3"/>
                                <name val="'hidden0'"/>
                        </SigmoidLayer>
                </Modules>
                <Connections>
                        <FullConnection class="pybrain.structure.connections.full.FullConnection" name="FullConnection-6">
                                <inmod val="bias"/>
                                <outmod val="out"/>
                                <Parameters>[0.6554487520957738]</Parameters>
                        </FullConnection>
                        <FullConnection class="pybrain.structure.connections.full.FullConnection" name="FullConnection-7">
                                <inmod val="bias"/>
                                <outmod val="hidden0"/>
                                <Parameters>[0.8141201069100833, -1.9519540651889176, 0.3483014480876096]</Parameters>
                        </FullConnection>
                        <FullConnection class="pybrain.structure.connections.full.FullConnection" name="FullConnection-5">
                                <inmod val="in"/>
                                <outmod val="hidden0"/>
                                <Parameters>[0.32489279837910084, 0.34480786433574551, 0.75045803824057666, -0.58411948692771176, -0.12327324616272992, -0.41228675759787226, -0.85553671683893218, -1.3320521945223582, -1.0531422952418676, -0.40839301403900452, -2.7565756871565674, -1.6723188687051469, -1.3597994054921079, 0.24852450267525059, -0.40924881241151689, 0.54037857219934371, 1.0960673042273468, 1.3324258379470664, 0.29047259837334116, -0.022417631256966383, 0.44571376571760984, 0.6492450404233816, -0.29105564158278247, 1.2243353023571237]</Parameters>
                        </FullConnection>
                        <FullConnection class="pybrain.structure.connections.full.FullConnection" name="FullConnection-4">
                                <inmod val="hidden0"/>
                                <outmod val="out"/>
                                <Parameters>[0.25616738836523284, -2.2028123481048487, -0.11026881677981226]</Parameters>
                        </FullConnection>
                </Connections>
        </Network>
</PyBrain>