def test_RandomForestRegressionLolo_2():
"""Non-trivial test case, including standard deviation."""
n, m, xlen = 100, 600, 10
train_inputs = np.reshape(np.linspace(-xlen / 2, +xlen / 2, n), (n, 1))
train_labels = (train_inputs * 2 + 1).flatten()
train_data = smlb.TabularData(data=train_inputs, labels=train_labels)
train_data = smlb.LabelNoise(noise=smlb.NormalNoise(
rng=0)).fit(train_data).apply(train_data)
valid_inputs = np.reshape(np.linspace(-xlen / 2, +xlen / 2, m), (m, 1))
valid_labels = (valid_inputs * 2 + 1).flatten()
valid_data = smlb.TabularData(data=valid_inputs, labels=valid_labels)
valid_data = smlb.LabelNoise(noise=smlb.NormalNoise(
rng=1)).fit(valid_data).apply(valid_data)
# 12 trees meets minimal requirements for jackknife estimates
rf = RandomForestRegressionLolo()
preds = rf.fit(train_data).apply(valid_data)
mae = smlb.MeanAbsoluteError().evaluate(valid_data.labels(), preds)
# for perfect predictions, expect MAE of 1.12943
# (absolute difference between draws from two unit normal distributions)
assert np.allclose(mae, 1.13, atol=0.25)
assert np.allclose(np.median(preds.stddev), 1, atol=0.5)
def test_GaussianProcessRegressionSklearn_3():
"""All predictive distributions.
Linear noisy function, linear kernel + white noise kernel.
The optimized noise level is expected to go to its true value.
"""
kernel = skl.gaussian_process.kernels.DotProduct(
sigma_0=0, sigma_0_bounds="fixed"
) + skl.gaussian_process.kernels.WhiteKernel(noise_level=1, noise_level_bounds=(1e-5, 1e5))
gpr = GaussianProcessRegressionSklearn(kernel=kernel, random_state=1)
n, nlsd = 100, 0.5
data = smlb.TabularData(data=np.ones(shape=(n, 1)) * 2, labels=np.ones(shape=n) * 3)
data = smlb.LabelNoise(noise=smlb.NormalNoise(stddev=nlsd, rng=1)).fit(data).apply(data)
preds = gpr.fit(data).apply(data)
assert preds.has_signal_part and preds.has_noise_part
conf, noise = preds.signal_part, preds.noise_part
assert np.allclose(conf.mean, np.ones(n) * 3, atol=1e-1)
assert np.allclose(conf.stddev, np.ones(n) * nlsd, atol=1e-1)
assert (preds.mean == conf.mean).all()
assert np.allclose(preds.stddev, np.sqrt(np.square(conf.stddev) + np.square(nlsd)), atol=1e-1)
assert np.allclose(noise.mean, np.zeros(shape=n))
assert np.allclose(noise.stddev, nlsd, atol=1e-1)
def test_NormalNoise():
"""Test Gaussian noise."""
# fail without specifying pseudo-random number generator seed
with pytest.raises(smlb.InvalidParameterError):
smlb.NormalNoise()
# unit normal
noise = smlb.NormalNoise(rng=1).noise(100)
assert sp.stats.normaltest(noise)[1] > 0.05
# same seed leads to identical noise
noise2 = smlb.NormalNoise(rng=1).noise(100)
assert (noise == noise2).all()
# non-unit normal
noise = smlb.NormalNoise(mean=10, stddev=0.5, rng=1).noise(100)
assert sp.stats.normaltest(noise)[1] > 0.05
def test_LabelNoise_NormalNoise(fixture_TabularData_ComputedLabels):
"""Test LabelNoise with NormalNoise."""
arange = np.arange(0, 100)
data1 = fixture_TabularData_ComputedLabels(
size=100, labelf=lambda arg: arg.flatten())
data2 = smlb.LabelNoise(noise=smlb.NormalNoise(
rng=1)).fit(data1).apply(data1)
assert sp.stats.normaltest(data2.labels(arange) - arange)[1] > 0.05
assert sp.stats.normaltest(data2.labels(arange))[1] < 0.05
# repeated evaluation of labels will yield different values
assert (data2.labels(arange) != data2.labels(arange)).any()