CyBrain

Neural Networks in Cython, inspired by PyBrain, but focused on speed.

Check examples/logistic_XOR_example.py for a first approach.

XOR Example

import cybrain as cb
import numpy as np
from time import time

#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 NETWORK
nnet = cb.Network()

#CREATE LAYERS
Lin = cb.LinearLayer(2)
Lhidden = cb.LogisticLayer(2)
Lout = cb.LogisticLayer(1)
bias = cb.BiasUnit()

#ADD LAYERS TO NETWORK
nnet.inputLayers = [Lin]
nnet.hiddenLayers = [Lhidden]
nnet.outputLayers = [Lout]
nnet.autoInputLayers = [bias]

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

#CREATE BATCH TRAINER
rate = 0.1
nnet.setup()
batch = cb.FullBatchTrainer(nnet, X, Y, rate)


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

#PRINT RESULTS
for i in range(len(X)):
    print "{} ==> {}".format(X[i], nnet.activate(X[i:i+1,:]))

Same Example in PyBrain

import cybrain as cb
import numpy as np
from time import time

#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 NETWORK
nnet = cb.Network()

#CREATE LAYERS
Lin = cb.LinearLayer(2)
Lhidden = cb.LogisticLayer(2)
Lout = cb.LogisticLayer(1)
bias = cb.BiasUnit()

#ADD LAYERS TO NETWORK
nnet.inputLayers = [Lin]
nnet.hiddenLayers = [Lhidden]
nnet.outputLayers = [Lout]
nnet.autoInputLayers = [bias]

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

#CREATE BATCH TRAINER
rate = 0.1
nnet.setup()
batch = cb.FullBatchTrainer(nnet, X, Y, rate)


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

#PRINT RESULTS
for i in range(len(X)):
    print "{} ==> {}".format(X[i], np.array(nnet.activate(X[i:i+1,:])))

Outputs

Time CyBrain 0.211654996872
[ 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]

Roadmap

• Implement other common Layers: Tanh, Softmax, etc [High].
• Refactor variable and methods according to python style guide [Medium].
• Cython memoryview operations optimization [High]
• Auto-encoders and CNNs [Requires research]
• Try to organize the folders [Cython research]
• Solutions to Andrew Ng's ML class with CyBrain [Soon]

About the Author

My name is Cristian Garcia, I am a Mathematical Engineer and Software developer, so I claim to be a Mathematical Developer. I live in Colombia but plan to travel abroad as soon as I get my collage degree to do a Master in Artificial Intelligence or get a job as a Mathematical Developer. If you happen to need a Mathematical Developer at your company and find my code reasonable, contact me.