def test_fit_1d():
"""test_fit_1d : Test the fit of the model on a defined example, with known parameters.
"""
# Generate the data
x = np.random.rand(1000,1)
y = np.sin(4*x+3)*np.sin(15*x+9)
n_neurons = 100
problem = 'r'
activation_function = np.tanh
myelm = elm(problem, n_neurons, activation_function)
# The fit should always be better than 1e-8, in MSE
myelm, yh, yloo = myelm.fit(x, y)
assert_less(np.mean((yh-y)**2), 10**(-8))
assert_less(np.mean((yloo-y)**2), 10**(-8))
# Same data, with some added noise
x = np.random.rand(1000,1)
y = np.sin(4*x+3)*np.sin(15*x+9)+np.random.randn(1000, 1)*0.025
n_neurons = 100
problem = 'r'
activation_function = np.tanh
myelm = elm(problem, n_neurons, activation_function)
# The fit should always be better than 1e-3, in MSE
myelm, yh, yloo = myelm.fit(x, y)
assert_less(np.mean((yh-y)**2), 10**(-3))
assert_less(np.mean((yloo-y)**2), 10**(-3))
def test_data_shapes():
"""test_data_shapes : Test for different types of input data
"""
# Test the regression part of the code, first
n_neurons = 100
problem = 'r'
activation_function = np.tanh
myelm = elm(problem, n_neurons, activation_function)
# Generate 1-d regression data, first
x = np.random.randn(200, 1)
y = np.random.randn(200, 1)
myelm.fit(x, y)
# Then n-d regression data
x = np.random.randn(200, 5)
y = np.random.randn(200, 1)
myelm.fit(x, y)
# And n-d regression data with multi-output
x = np.random.randn(200, 5)
y = np.random.randn(200, 3)
myelm.fit(x, y)
# Now classification data
problem = 'c'
myelmC = elm(problem, n_neurons, activation_function)
# 1-d binary classification
iris = datasets.load_iris()
xall = iris['data']
x = xall[:, 0].reshape((xall.shape[0], 1))
yall = iris['target']
# Using a bogus target, as the original is multi-class
y = np.sign(np.random.randn(x.shape[0], 1))
myelmC.fit(x, y)
# n-d binary classification
x = xall
myelmC.fit(x, y)
# n-d multi-class classification
y = yall.reshape((x.shape[0], 1))
myelmC.fit(x, y)
# n-d multi-class multi-output classification
# This is not supported and should raise an exception
y = np.hstack((yall, yall)).reshape((x.shape[0], 2))
assert_raises(ValueError, myelmC.fit, x, y)
def test_predict_1d():
"""test_predict_1d : Test the predict function on the same example as that
of the test_fit_1d.
"""
# Generate the data
xall = np.random.rand(2000, 1)
yall = np.sin(4*xall+3)*np.sin(15*xall+9)+np.random.randn(2000, 1)*0.25
x = xall[:999, :]
y = yall[:999, :]
xt = xall[1000:, :]
yt = yall[1000:, :]
n_neurons = 100
problem = 'r'
activation_function = np.tanh
myelm = elm(problem, n_neurons, activation_function)
# The fit should always be better than 1e-8, in MSE
myelm, yh, yloo = myelm.fit(x, y)
yth = myelm.predict(xt)
assert_less(np.mean((yth-yt)**2),10**(-1))
def test_perf():
"""Tests the quality of the prediction on known examples.
"""
# Multi-class classification
iris = datasets.load_iris()
xall = iris['data']
yall = iris['target']
yall = yall.reshape((yall.shape[0], 1))
xall, yall = shuffle(xall, yall)
x = xall[:99, :]
y = yall[:99, :]
xt = xall[100:, :]
yt = yall[100:, :]
n_neurons = 100
problem = 'c'
activation_function = np.tanh
myelmC = elm(problem, n_neurons, activation_function)
myelmC, yh, yloo = myelmC.fit(x, y)
assert_less(float(np.sum(yloo==y))/y.shape[0],-1)
assert_less(float(np.sum(yh==y))/y.shape[0],-1)
yth = myelmC.predict(xt)
assert_less(float(np.sum(yth==yt))/yt.shape[0],-1)
