def test_merge():
t = TDigest()
t2 = TDigest()
t3 = TDigest()
a = np.random.uniform(0, 1, N)
b = np.random.uniform(2, 3, N)
data = np.concatenate([a, b])
t2.update(a)
t3.update(b)
t2_centroids = t2.centroids()
t.merge(t2, t3)
assert t.min() == min(t2.min(), t3.min())
assert t.max() == max(t2.max(), t3.max())
assert t.size() == t2.size() + t3.size()
# Check no mutation of args
assert (t2.centroids() == t2_centroids).all()
# *Quantile
q = np.array([0.001, 0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 0.99, 0.999])
est = t.quantile(q)
q_est = quantiles_to_q(data, est)
np.testing.assert_allclose(q, q_est, atol=0.012, rtol=0)
# *CDF
x = q_to_x(data, q)
q_est = t.cdf(x)
np.testing.assert_allclose(q, q_est, atol=0.005)
with pytest.raises(TypeError):
t.merge(t2, 'not a tdigest')
def test_scale():
t = TDigest()
t.update(uniform)
for factor in [0.5, 2]:
t2 = t.scale(factor)
assert t is not t2
assert t.size() * factor == t2.size()
assert t.min() == t2.min()
assert t.max() == t2.max()
a = t.centroids()
b = t2.centroids()
np.testing.assert_array_equal(a['mean'], b['mean'])
np.testing.assert_allclose(a['weight'] * factor, b['weight'])
for val in [-0.5, 0, np.nan, np.inf]:
with pytest.raises(ValueError):
t.scale(val)
with pytest.raises(TypeError):
t.scale('foobar')
# Test scale compacts
eps = np.finfo('f8').eps
t = TDigest()
t.update([1, 2, 3, 4, 5],
[1, 1000, 1, 10000, 1])
t2 = t.scale(eps)
assert len(t2.centroids()) == 2
# Compacts to 0
t = TDigest()
t.update([1, 2, 3, 4, 5])
t2 = t.scale(eps)
assert len(t2.centroids()) == 0
def test_histogram():
t = TDigest()
data = np.random.normal(size=10000)
t.update(data)
hist, bins = t.histogram(100)
assert len(hist) == 100
assert len(bins) == 101
c = t.cdf(bins)
np.testing.assert_allclose((c[1:] - c[:-1]) * t.size(), hist)
min = t.min()
max = t.max()
eps = np.finfo('f8').eps
bins = np.array([min - 1, min - eps,
min, min + (max - min)/2, max,
max + eps, max + 1])
hist, bins2 = t.histogram(bins)
np.testing.assert_allclose(bins, bins2)
assert hist[0] == 0
assert hist[1] == 0
assert hist[-2] == 0
assert hist[-1] == 0
assert hist.sum() == t.size()
# range ignored when bins provided
hist2, bins2 = t.histogram(bins, range=(-5, -3))
np.testing.assert_allclose(hist, hist2)
np.testing.assert_allclose(bins, bins2)
def test_nonfinite():
t = TDigest()
data = gamma.copy()
data[::10] = np.nan
data[::7] = np.inf
t.update(data)
finite = data[np.isfinite(data)]
assert t.size() == len(finite)
assert t.min() == finite.min()
assert t.max() == finite.max()
t = TDigest()
t.add(np.nan)
t.add(np.inf)
t.add(-np.inf)
assert t.size() == 0
for w in [np.inf, -np.inf, np.nan]:
t = TDigest()
with pytest.raises(ValueError):
t.add(1, w)
w = np.array([1, 2, w, 3, 4])
t = TDigest()
with pytest.raises(ValueError):
t.update(np.ones(5), w)
def test_empty():
t = TDigest()
assert t.size() == 0
assert len(t.centroids()) == 0
assert np.isnan(t.min())
assert np.isnan(t.max())
assert np.isnan(t.quantile(0.5))
assert np.isnan(t.cdf(0.5))
def test_single():
t = TDigest()
t.add(10)
assert t.min() == 10
assert t.max() == 10
assert t.size() == 1
assert t.quantile(0) == 10
assert t.quantile(0.5) == 10
assert t.quantile(1) == 10
assert t.cdf(9) == 0
assert t.cdf(10) == 0.5
assert t.cdf(11) == 1
def test_distributions(data):
t = TDigest()
t.update(data)
assert t.size() == len(data)
assert t.min() == data.min()
assert t.max() == data.max()
check_valid_quantile_and_cdf(t)
# *Quantile
q = np.array([0.001, 0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 0.99, 0.999])
est = t.quantile(q)
q_est = quantiles_to_q(data, est)
np.testing.assert_allclose(q, q_est, atol=0.012, rtol=0)
# *CDF
x = q_to_x(data, q)
q_est = t.cdf(x)
np.testing.assert_allclose(q, q_est, atol=0.005)