def testAnd():
examples = [[[0, 0, 1], [0.1]], [[0, 1, 1], [.1]], [[1, 0, 1], [.1]],
[[1, 1, 1], [.9]]]
net = Network.Network([3, 1])
t = BackProp()
t.TrainOnLine(examples, net)
return net
def testXor():
examples = [[[0, 0, 1], [.1]], [[0, 1, 1], [.9]], [[1, 0, 1], [.9]], [[1, 1, 1], [.1]]]
net = Network.Network([3, 3, 1])
t = BackProp(speed=.8)
t.TrainOnLine(examples, net, errTol=0.2)
return net
def _trainExamples(self, ex, arch=[3, 1], useAvgErr=False):
net = Network.Network(arch)
t = Trainers.BackProp()
t.TrainOnLine(ex,
net,
errTol=self.trainTol,
useAvgErr=useAvgErr,
silent=True)
errs = [abs(x[-1] - net.ClassifyExample(x)) for x in ex]
return net, errs
def testOr():
examples = [[[0, 0, 1], [0.1]], [[0, 1, 1], [.9]], [[1, 0, 1], [.9]], [[1, 1, 1], [.9]]]
net = Network.Network([3, 1])
t = BackProp()
t.TrainOnLine(examples, net, maxIts=1000, useAvgErr=0)
print('classifications:')
for example in examples:
res = net.ClassifyExample(example[0])
print('%f -> %f' % (example[1][0], res))
return net
def testLinear():
examples = [
[.1, .1],
[.2, .2],
[.3, .3],
[.4, .4],
[.8, .8],
]
net = Network.Network([1, 2, 1])
t = BackProp(speed=.8)
t.TrainOnLine(examples, net, errTol=0.1, useAvgErr=0)
print('classifications:')
for example in examples:
res = net.ClassifyExample(example[:-1])
print('%f -> %f' % (example[-1], res))
return net
def _trainExamples(self, ex, arch=[3, 1]):
net = Network.Network(arch)
t = Trainers.BackProp()
t.TrainOnLine(ex, net, errTol=self.trainTol, useAvgErr=0, silent=1)
errs = map(lambda x, y=net: abs(x[-1] - y.ClassifyExample(x)), ex)
return errs
def CrossValidationDriver(examples,
attrs=[],
nPossibleVals=[],
holdOutFrac=.3,
silent=0,
tolerance=0.3,
calcTotalError=0,
hiddenSizes=None,
**kwargs):
"""
**Arguments**
- examples: the full set of examples
- attrs: a list of attributes to consider in the tree building
*This argument is ignored*
- nPossibleVals: a list of the number of possible values each variable can adopt
*This argument is ignored*
- holdOutFrac: the fraction of the data which should be reserved for the hold-out set
(used to calculate the error)
- silent: a toggle used to control how much visual noise this makes as it goes.
- tolerance: the tolerance for convergence of the net
- calcTotalError: if this is true the entire data set is used to calculate
accuracy of the net
- hiddenSizes: a list containing the size(s) of the hidden layers in the network.
if _hiddenSizes_ is None, one hidden layer containing the same number of nodes
as the input layer will be used
**Returns**
a 2-tuple containing:
1) the net
2) the cross-validation error of the net
**Note**
At the moment, this is specific to nets with only one output
"""
nTot = len(examples)
if not kwargs.get('replacementSelection', 0):
testIndices, trainIndices = SplitData.SplitIndices(nTot,
holdOutFrac,
silent=1,
legacy=1,
replacement=0)
else:
testIndices, trainIndices = SplitData.SplitIndices(nTot,
holdOutFrac,
silent=1,
legacy=0,
replacement=1)
trainExamples = [examples[x] for x in trainIndices]
testExamples = [examples[x] for x in testIndices]
nTrain = len(trainExamples)
if not silent:
print('Training with %d examples' % (nTrain))
nInput = len(examples[0]) - 1
nOutput = 1
if hiddenSizes is None:
nHidden = nInput
netSize = [nInput, nHidden, nOutput]
else:
netSize = [nInput] + hiddenSizes + [nOutput]
net = Network.Network(netSize)
t = Trainers.BackProp()
t.TrainOnLine(trainExamples,
net,
errTol=tolerance,
useAvgErr=0,
silent=silent)
nTest = len(testExamples)
if not silent:
print('Testing with %d examples' % nTest)
if not calcTotalError:
xValError, _ = CrossValidate(net, testExamples, tolerance)
else:
xValError, _ = CrossValidate(net, examples, tolerance)
if not silent:
print('Validation error was %%%4.2f' % (100 * xValError))
net._trainIndices = trainIndices
return net, xValError