names = ('ridge regression', 'IRLS', 'gradient descent', 'conjugate gradient',
'BGFS')
# This threshold node will be appended to the trained classifier node, to transform the probabilities output by the logistic regression to the set {0, 1}
threshold_node = ThresholdNode(threshold=.5, output_values=[0, 1])
# Loop over the different training algorithms and train and test them
for node, name in zip(nodes, names):
# Construct the flow
flow = mdp.Flow([node, threshold_node])
# Train the regression_node
flow.train([[[xtrain, ytrain]], []])
# Apply to the test set
y = flow.execute(xtest)
# Compute the discriminant line
y1 = (node.w[0] * r[0] + node.b - 0.5) / -node.w[1]
y2 = (node.w[0] * r[1] + node.b - 0.5) / -node.w[1]
# Plot the discriminant line
pylab.plot(r, [y1, y2], label=name)
# Calculate the error
print "error rate %s: %.3f" % (name, loss_01(y, ytest))
pylab.legend()
pylab.show()
threshold_node = ThresholdNode(threshold=.5, output_values=[0, 1])
# Loop over the different training algorithms and train and test them
for node, name in zip(nodes, names):
# Construct the flow
flow = mdp.Flow([node, threshold_node])
# Train the regression_node
flow.train([[[xtrain, ytrain]], []])
# Apply to the test set
y = flow.execute(xtest)
# Compute the discriminant line
if isinstance(node, RidgeRegressionNode):
y1 = (node.beta[1] * r[0] + node.beta[0] - .5) / -node.beta[2]
y2 = (node.beta[1] * r[1] + node.beta[0] - .5) / -node.beta[2]
else:
y1 = (node.w[0] * r[0] + node.b) / -node.w[1]
y2 = (node.w[0] * r[1] + node.b) / -node.w[1]
# Plot the discriminant line
pylab.plot(r, [y1, y2], label=name)
# Calculate the error
print "error rate %s: %.3f" % (name, loss_01(y, ytest))
pylab.legend()
pylab.show()
def test_loss_01():
''' Test zero one loss on simple case '''
assert utils.loss_01(np.array([[1], [2], [3]]), np.array([[1], [2],
[4]])) == 1. / 3
def test_loss_01():
''' Test zero one loss on simple case '''
assert utils.loss_01(np.array([[1, 2, 3]]).T,
np.array([[1, 2, 4]]).T) == 1. / 3