def test_scalar_quantity_distribution():
# Regression test for gh-12336
angles = Distribution([90., 30., 0.] * u.deg)
sin_angles = np.sin(angles) # This failed in 4.3.
assert isinstance(sin_angles, Distribution)
assert isinstance(sin_angles, u.Quantity)
assert_array_equal(sin_angles, Distribution(np.sin([90., 30., 0.]*u.deg)))
def test_histogram():
arr = np.random.randn(2, 3, 1000)
distr = Distribution(arr)
hist, bins = distr.pdf_histogram(bins=10)
assert hist.shape == (2, 3, 10)
assert bins.shape == (2, 3, 11)
def test_histogram():
arr = np.random.randn(2, 3, 1000)
distr = Distribution(arr)
hist, bins = distr.pdf_histogram(bins=10)
assert hist.shape == (2, 3, 10)
assert bins.shape == (2, 3, 11)
def setup_class(self):
with NumpyRNGContext(12345):
self.data = np.random.normal(np.array([1, 2, 3, 4])[:, np.newaxis],
np.array([3, 2, 4, 5])[:, np.newaxis],
(4, 10000))
self.distr = Distribution(self.data * u.kpc)
def test_reprs():
darr = np.arange(30).reshape(3, 10)
distr = Distribution(darr * u.kpc)
assert 'n_samples=10' in repr(distr)
assert 'n_samples=10' in str(distr)
assert r'n_{\rm samp}=10' in distr._repr_latex_()
def test_reprs():
darr = np.arange(30).reshape(3, 10)
distr = Distribution(darr * u.kpc)
assert 'n_samples=10' in repr(distr)
assert 'n_samples=10' in str(distr)
assert r'n_{\rm samp}=10' in distr._repr_latex_()
def test_array_repr_latex():
# as of this writing ndarray does not have a _repr_latex_, and this test
# ensure distributions account for that. However, if in the future ndarray
# gets a _repr_latex_, we can skip this.
arr = np.random.randn(4, 1000)
if hasattr(arr, '_repr_latex_'):
pytest.skip('in this version of numpy, ndarray has a _repr_latex_')
distr = Distribution(arr)
assert distr._repr_latex_() is None
def test_array_repr_latex():
# as of this writing ndarray does not have a _repr_latex_, and this test
# ensure distributions account for that. However, if in the future ndarray
# gets a _repr_latex_, we can skip this.
arr = np.random.randn(4, 1000)
if hasattr(arr, '_repr_latex_'):
pytest.skip('in this version of numpy, ndarray has a _repr_latex_')
distr = Distribution(arr)
assert distr._repr_latex_() is None
def test_distr_cannot_view_new_dtype():
# A Distribution has a very specific structured dtype with just one
# element that holds the array of samples. As it is not clear what
# to do with a view as a new dtype, we just error on it.
# TODO: with a lot of thought, this restriction can likely be relaxed.
distr = Distribution([2., 3., 4.])
with pytest.raises(ValueError, match='with a new dtype'):
distr.view(np.dtype('f8'))
# Check subclass just in case.
ad = Angle(distr, 'deg')
with pytest.raises(ValueError, match='with a new dtype'):
ad.view(np.dtype('f8'))
with pytest.raises(ValueError, match='with a new dtype'):
ad.view(np.dtype('f8'), Distribution)
def setup_class(self):
with NumpyRNGContext(12345):
self.data = np.random.normal(np.array([1, 2, 3, 4])[:, np.newaxis],
np.array([3, 2, 4, 5])[:, np.newaxis],
(4, 10000))
self.distr = Distribution(self.data * u.kpc)
def test_init_scalar():
parr = np.random.poisson(
np.array([1, 5, 30, 400])[:, np.newaxis], (4, 1000))
with pytest.raises(TypeError) as exc:
Distribution(parr.ravel()[0])
assert exc.value.args[
0] == "Attempted to initialize a Distribution with a scalar"
def test_quantity_init(self):
# Test that we can initialize directly from a Quantity
pq = self.parr << u.ct
pqd = Distribution(pq)
assert isinstance(pqd, u.Quantity)
assert isinstance(pqd, Distribution)
assert isinstance(pqd.value, Distribution)
assert_array_equal(pqd.value.distribution, self.parr)
def test_quantity_init_with_distribution(self):
# Test that we can initialize a Quantity from a Distribution.
pd = Distribution(self.parr)
qpd = pd << u.ct
assert isinstance(qpd, u.Quantity)
assert isinstance(qpd, Distribution)
assert qpd.unit == u.ct
assert_array_equal(qpd.value.distribution, pd.distribution.astype(float))
def test_index_assignment_quantity():
arr = np.random.randn(2, 1000)
distr = Distribution(arr*u.kpc)
d1q, d2q = distr
assert isinstance(d1q, Distribution)
assert isinstance(d2q, Distribution)
ndistr = ds.normal(center=[1, 2]*u.kpc, std=[3, 4]*u.kpc, n_samples=1000)
n1, n2 = ndistr
assert isinstance(n1, ds.Distribution)
assert isinstance(n2, ds.Distribution)
def test_index_assignment_array():
arr = np.random.randn(2, 1000)
distr = Distribution(arr)
d1a, d2a = distr
assert isinstance(d1a, Distribution)
assert isinstance(d2a, Distribution)
ndistr = ds.normal(center=[1, 2], std=[3, 4], n_samples=1000)
n1, n2 = ndistr
assert isinstance(n1, ds.Distribution)
assert isinstance(n2, ds.Distribution)
def test_add_distribution(self):
another_data = (np.random.randn(4, 10000)
* np.array([1000, .01, 80, 10])[:, np.newaxis]
+ np.array([2000, 0, 0, 500])[:, np.newaxis])
# another_data is in pc, but main distr is in kpc
another_distr = Distribution(another_data * u.pc)
combined_distr = self.distr + another_distr
expected = np.median(self.data + another_data/1000,
axis=-1) * self.distr.unit
assert_quantity_allclose(combined_distr.pdf_median, expected)
expected = np.var(self.data + another_data/1000, axis=-1) * self.distr.unit**2
assert_quantity_allclose(combined_distr.pdf_var, expected)
def test_distr_angle_view_as_quantity():
# Check that Quantity subclasses decay to Quantity appropriately.
distr = Distribution([2., 3., 4.])
ad = Angle(distr, 'deg')
qd = ad.view(u.Quantity)
assert not isinstance(qd, Angle)
assert isinstance(qd, u.Quantity)
assert isinstance(qd, Distribution)
# View directly as DistributionQuantity class.
qd2 = ad.view(qd.__class__)
assert not isinstance(qd2, Angle)
assert isinstance(qd2, u.Quantity)
assert isinstance(qd2, Distribution)
assert_array_equal(qd2.distribution, qd.distribution)
qd3 = ad.view(qd.dtype, qd.__class__)
assert not isinstance(qd3, Angle)
assert isinstance(qd3, u.Quantity)
assert isinstance(qd3, Distribution)
assert_array_equal(qd3.distribution, qd.distribution)
def test_distr_angle():
# Check that Quantity subclasses decay to Quantity appropriately.
distr = Distribution([2., 3., 4.])
ad = Angle(distr, 'deg')
ad_plus_ad = ad + ad
assert isinstance(ad_plus_ad, Angle)
assert isinstance(ad_plus_ad, Distribution)
ad_times_ad = ad * ad
assert not isinstance(ad_times_ad, Angle)
assert isinstance(ad_times_ad, u.Quantity)
assert isinstance(ad_times_ad, Distribution)
ad += ad
assert isinstance(ad, Angle)
assert isinstance(ad, Distribution)
assert_array_equal(ad.distribution, ad_plus_ad.distribution)
with pytest.raises(u.UnitTypeError):
ad *= ad
def test_numpy_init(self):
# Test that we can initialize directly from a Numpy array
Distribution(self.parr)
class TestDistributionStatistics():
def setup_class(self):
with NumpyRNGContext(12345):
self.data = np.random.normal(np.array([1, 2, 3, 4])[:, np.newaxis],
np.array([3, 2, 4, 5])[:, np.newaxis],
(4, 10000))
self.distr = Distribution(self.data * u.kpc)
def test_shape(self):
# Distribution shape
assert self.distr.shape == (4, )
assert self.distr.distribution.shape == (4, 10000)
def test_size(self):
# Total number of values
assert self.distr.size == 4
assert self.distr.distribution.size == 40000
def test_n_samples(self):
# Number of samples
assert self.distr.n_samples == 10000
def test_n_distr(self):
assert self.distr.shape == (4,)
def test_pdf_mean(self):
# Mean of each PDF
expected = np.mean(self.data, axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_mean, expected)
assert_quantity_allclose(self.distr.pdf_mean, [1, 2, 3, 4] * self.distr.unit, rtol=0.05)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_mean, Distribution)
assert isinstance(self.distr.pdf_mean, u.Quantity)
def test_pdf_std(self):
# Standard deviation of each PDF
expected = np.std(self.data, axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_std, expected)
assert_quantity_allclose(self.distr.pdf_std, [3, 2, 4, 5] * self.distr.unit, rtol=0.05)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_std, Distribution)
assert isinstance(self.distr.pdf_std, u.Quantity)
def test_pdf_var(self):
# Variance of each PDF
expected = np.var(self.data, axis=-1) * self.distr.unit**2
assert_quantity_allclose(self.distr.pdf_var, expected)
assert_quantity_allclose(self.distr.pdf_var, [9, 4, 16, 25] * self.distr.unit**2, rtol=0.1)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_var, Distribution)
assert isinstance(self.distr.pdf_var, u.Quantity)
def test_pdf_median(self):
# Median of each PDF
expected = np.median(self.data, axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_median, expected)
assert_quantity_allclose(self.distr.pdf_median, [1, 2, 3, 4] * self.distr.unit, rtol=0.1)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_median, Distribution)
assert isinstance(self.distr.pdf_median, u.Quantity)
@pytest.mark.skipif(not HAS_SCIPY, reason='no scipy')
def test_pdf_mad_smad(self):
# Median absolute deviation of each PDF
median = np.median(self.data, axis=-1, keepdims=True)
expected = np.median(np.abs(self.data - median), axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_mad, expected)
assert_quantity_allclose(self.distr.pdf_smad, self.distr.pdf_mad * SMAD_FACTOR, rtol=1e-5)
assert_quantity_allclose(self.distr.pdf_smad, [3, 2, 4, 5] * self.distr.unit, rtol=0.05)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_mad, Distribution)
assert isinstance(self.distr.pdf_mad, u.Quantity)
assert not isinstance(self.distr.pdf_smad, Distribution)
assert isinstance(self.distr.pdf_smad, u.Quantity)
def test_percentile(self):
expected = np.percentile(self.data, [10, 50, 90], axis=-1) * self.distr.unit
percs = self.distr.pdf_percentiles([10, 50, 90])
assert_quantity_allclose(percs, expected)
assert percs.shape == (3, 4)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(percs, Distribution)
assert isinstance(percs, u.Quantity)
def test_add_quantity(self):
distrplus = self.distr + [2000, 0, 0, 500] * u.pc
expected = (np.median(self.data, axis=-1) + np.array([2, 0, 0, 0.5])) * self.distr.unit
assert_quantity_allclose(distrplus.pdf_median, expected)
expected = np.var(self.data, axis=-1) * self.distr.unit**2
assert_quantity_allclose(distrplus.pdf_var, expected)
def test_add_distribution(self):
another_data = (np.random.randn(4, 10000)
* np.array([1000, .01, 80, 10])[:, np.newaxis]
+ np.array([2000, 0, 0, 500])[:, np.newaxis])
# another_data is in pc, but main distr is in kpc
another_distr = Distribution(another_data * u.pc)
combined_distr = self.distr + another_distr
expected = np.median(self.data + another_data/1000,
axis=-1) * self.distr.unit
assert_quantity_allclose(combined_distr.pdf_median, expected)
expected = np.var(self.data + another_data/1000, axis=-1) * self.distr.unit**2
assert_quantity_allclose(combined_distr.pdf_var, expected)
def test_quantity_init_T():
# Test that we can initialize directly from a Quantity
pq = np.random.poisson(np.array([1, 5, 30, 400]), (1000, 4)) * u.ct
Distribution(pq.T)
def test_quantity_init():
# Test that we can initialize directly from a Quantity
pq = np.random.poisson(np.array([1, 5, 30, 400])[:, np.newaxis],
(4, 1000)) * u.ct
Distribution(pq)
class TestDistributionStatistics():
def setup_class(self):
with NumpyRNGContext(12345):
self.data = np.random.normal(np.array([1, 2, 3, 4])[:, np.newaxis],
np.array([3, 2, 4, 5])[:, np.newaxis],
(4, 10000))
self.distr = Distribution(self.data * u.kpc)
def test_shape(self):
# Distribution shape
assert self.distr.shape == (4, )
assert self.distr.distribution.shape == (4, 10000)
def test_size(self):
# Total number of values
assert self.distr.size == 4
assert self.distr.distribution.size == 40000
def test_n_samples(self):
# Number of samples
assert self.distr.n_samples == 10000
def test_n_distr(self):
assert self.distr.shape == (4,)
def test_pdf_mean(self):
# Mean of each PDF
expected = np.mean(self.data, axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_mean, expected)
assert_quantity_allclose(self.distr.pdf_mean, [1, 2, 3, 4] * self.distr.unit, rtol=0.05)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_mean, Distribution)
assert isinstance(self.distr.pdf_mean, u.Quantity)
def test_pdf_std(self):
# Standard deviation of each PDF
expected = np.std(self.data, axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_std, expected)
assert_quantity_allclose(self.distr.pdf_std, [3, 2, 4, 5] * self.distr.unit, rtol=0.05)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_std, Distribution)
assert isinstance(self.distr.pdf_std, u.Quantity)
def test_pdf_var(self):
# Variance of each PDF
expected = np.var(self.data, axis=-1) * self.distr.unit**2
assert_quantity_allclose(self.distr.pdf_var, expected)
assert_quantity_allclose(self.distr.pdf_var, [9, 4, 16, 25] * self.distr.unit**2, rtol=0.1)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_var, Distribution)
assert isinstance(self.distr.pdf_var, u.Quantity)
def test_pdf_median(self):
# Median of each PDF
expected = np.median(self.data, axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_median, expected)
assert_quantity_allclose(self.distr.pdf_median, [1, 2, 3, 4] * self.distr.unit, rtol=0.1)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_median, Distribution)
assert isinstance(self.distr.pdf_median, u.Quantity)
@pytest.mark.skipif(not HAS_SCIPY, reason='no scipy')
def test_pdf_mad_smad(self):
# Median absolute deviation of each PDF
median = np.median(self.data, axis=-1, keepdims=True)
expected = np.median(np.abs(self.data - median), axis=-1) * self.distr.unit
assert_quantity_allclose(self.distr.pdf_mad, expected)
assert_quantity_allclose(self.distr.pdf_smad, self.distr.pdf_mad * SMAD_FACTOR, rtol=1e-5)
assert_quantity_allclose(self.distr.pdf_smad, [3, 2, 4, 5] * self.distr.unit, rtol=0.05)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(self.distr.pdf_mad, Distribution)
assert isinstance(self.distr.pdf_mad, u.Quantity)
assert not isinstance(self.distr.pdf_smad, Distribution)
assert isinstance(self.distr.pdf_smad, u.Quantity)
def test_percentile(self):
expected = np.percentile(self.data, [10, 50, 90], axis=-1) * self.distr.unit
percs = self.distr.pdf_percentiles([10, 50, 90])
assert_quantity_allclose(percs, expected)
assert percs.shape == (3, 4)
# make sure the right type comes out - should be a Quantity because it's
# now a summary statistic
assert not isinstance(percs, Distribution)
assert isinstance(percs, u.Quantity)
def test_add_quantity(self):
distrplus = self.distr + [2000, 0, 0, 500] * u.pc
expected = (np.median(self.data, axis=-1) + np.array([2, 0, 0, 0.5])) * self.distr.unit
assert_quantity_allclose(distrplus.pdf_median, expected)
expected = np.var(self.data, axis=-1) * self.distr.unit**2
assert_quantity_allclose(distrplus.pdf_var, expected)
def test_add_distribution(self):
another_data = (np.random.randn(4, 10000)
* np.array([1000, .01, 80, 10])[:, np.newaxis]
+ np.array([2000, 0, 0, 500])[:, np.newaxis])
# another_data is in pc, but main distr is in kpc
another_distr = Distribution(another_data * u.pc)
combined_distr = self.distr + another_distr
expected = np.median(self.data + another_data/1000,
axis=-1) * self.distr.unit
assert_quantity_allclose(combined_distr.pdf_median, expected)
expected = np.var(self.data + another_data/1000, axis=-1) * self.distr.unit**2
assert_quantity_allclose(combined_distr.pdf_var, expected)
def test_numpy_init_T():
rates = np.array([1, 5, 30, 400])
parr = np.random.poisson(rates, (1000, 4))
Distribution(parr.T)
def test_quantity_init_T(self):
# Test that we can initialize directly from a Quantity
pq = self.parr_t << u.ct
Distribution(pq.T)
def test_numpy_init():
# Test that we can initialize directly from a Numpy array
rates = np.array([1, 5, 30, 400])[:, np.newaxis]
parr = np.random.poisson(rates, (4, 1000))
Distribution(parr)
def test_numpy_init_T(self):
Distribution(self.parr_t.T)