pybrain - ClassificationDataSet - how to understand the output when using SoftmaxLayer

Question

I am trying to build first classifier with Pybrain neural network along with specialized ClassificationDataSet and I am not sure I fully understand it works.

So I have a pandas dataframe with 6 feature columns and 1 column for class label (Survived, just 0 or 1).

I build a dataset out of it:

ds = ClassificationDataSet(6, 1, nb_classes=2)
for i in df[['Gender', 'Pclass', 'AgeFill', 'FamilySize', 'FarePerPerson', 'Deck','Survived']].values:
    ds.addSample(tuple(i[:-1]), i[-1])
ds._convertToOneOfMany()
return ds

Ok, I check how dataset looks like:

for i, m in ds:
    i, m


(array([ 1.,  3.,  2.,  2.,  1.,  8.]), array([1, 0]))
(array([ 0.,  1.,  1.,  2.,  0.,  2.]), array([0, 1]))

And I already have a problem. What means [1,0] or [0,1]? Is it just '0' or '1' of original 'survived' column? How to get back to original values?

Later, when I finish with training of my network:

net = buildNetwork(6, 6, 2, hiddenclass=TanhLayer, bias=True,  outclass=SoftmaxLayer)
trainer = BackpropTrainer(net, ds)
trainer.trainEpochs(10)

I will try to activate it on my another dataset (for which I want to do actual classification) and I will get a pairs of activation results for each of 2 output neurons, but how to understand which output neuron corresponds to which original class? Probably this is something obvious, but I am not able to understand it from the docs, unfortunately.

Answer 1

Ok, looks like pybrain uses position to determine which class it means by (0,1) or (1,0).

To go back to original 0 or 1 mark you need to use argmax() function. So for example if I already have a trained network and I want to validate it on the same data as I used for training I could do this:

for inProp, num in ds:
    out = net.activate(inProp).argmax()
    if out == num.argmax():
        true+=1
    total+=1
res = true/total

inProp will look like a tuple of my input values for activation, num - a tuple of expected two-neuron output (either (0,1) or (1,0)) and num.argmax() will translate it into just 0 or 1 - real output.

I might be wrong since this is a pure heuristic, but it works in my example.