kaynewest / cybrain

Neural Networks in Cython, inspired by PyBrain.

CyBrain

Neural Networks in Cython, inspired by PyBrain, but 70x faster.

Check Comparison.py for a first approach. In this example we train a CyBrain network to solve the XOR problem. At the end we use PyBrain to solve the same problem. A speed comparison indicates that CyBrain is 70x faster.

XOR Example

from cybrain import Neuron, LogisticNeuron, Connection, Layer, Network, BiasUnit, Trainer, SoftMaxLayer
import numpy as np
from time import time
nnet = Network()

#TRUTH TABLE (DATA)
X =     [[0.0,0.0]];     Y = [[0.0]]
X.append([1.0,0.0]); Y.append([1.0])
X.append([0.0,1.0]); Y.append([1.0])
X.append([1.0,1.0]); Y.append([0.0])


#CONVERT DATA TO NUMPY ARRAY
X, Y = np.array(X), np.array(Y)

#CREATE LAYERS
Lin = Layer( 2, names= ['a','b'] )
Lhidden = Layer( 2, LogisticNeuron , names= ['c','d'] )
Lout = Layer( 1, LogisticNeuron , names= ['e'] )
bias = Layer( 1, BiasUnit, names= ['bias'] )

#ADD LAYERS TO NETWORK
nnet.addInputLayer(Lin)
nnet.addLayer(Lhidden)
nnet.addOutputLayer(Lout)
nnet.addAutoInputLayer(bias)

#CONNECT LAYERS
Lin.connectTo(Lhidden)
Lhidden.connectTo(Lout)
bias.connectTo(Lhidden)
bias.connectTo(Lout)

#CREATE BATCH TRAINER
rate = 0.1
batch = Trainer( nnet, X, Y, rate )

#TRAIN
t1 = time()
batch.epochs(2000)
print "Time CyBrain {}".format(time()-t1)

#PRINT RESULTS
for x in X:
    print "{} ==> {}".format( x, nnet.activateWith(x, return_value= True) )

Same Example in PyBrain

from pybrain.tools.shortcuts import buildNetwork
from pybrain import LinearLayer, SigmoidLayer, FeedForwardNetwork, FullConnection, BiasUnit, SoftmaxLayer
from pybrain.supervised.trainers.backprop import BackpropTrainer
from pybrain.structure.modules.tanhlayer import TanhLayer
from pybrain.datasets import SupervisedDataSet
from time import time


ds = SupervisedDataSet(2,1 )

ds.addSample((0, 0), (0,))
ds.addSample((0, 1), (1,))
ds.addSample((1, 0), (1,))
ds.addSample((1, 1), (0,))


net = buildNetwork(2, 2, 1, bias=True, outputbias= True, hiddenclass=SigmoidLayer)
trainer = BackpropTrainer(net, ds, learningrate= 0.1)

t1 = time()
trainer.trainEpochs(2000)
print "Time PyBrain {}".format(time()-t1)

#PRINT RESULTS
for x in X:
    print "{} ==> {}".format( x, net.activate(x) )

Outputs

Time CyBrain 0.111654996872
[ 0.  0.] ==> [ '0.0365560102866' ]
[ 1.  0.] ==> [ '0.951081842587'  ]
[ 0.  1.] ==> [ '0.951928021684'  ]
[ 1.  1.] ==> [ '0.0332036251855' ]

Time PyBrain 7.03572702408
[ 0.  0.] ==> [  1.67662906e-08]
[ 1.  0.] ==> [ 0.99999998]
[ 0.  1.] ==> [ 0.99999998]
[ 1.  1.] ==> [  7.30255101e-09]