def test_interpolator_bounds():
with pytest.raises(ValueError):
rq.Pipeline(
rq.LowPass(window=10,
portion=2,
subsample_rate=1,
quantile=0.5,
alpha=2.0))
with pytest.raises(ValueError):
rq.Pipeline(
rq.LowPass(window=10, portion=2, subsample_rate=1, quantile=2.5))
def test_median_array_input(window_size=71, length=1000):
pipe = rq.Pipeline(rq.LowPass(window=window_size,
portion=window_size // 2))
x = example_input(length)
y = pipe.feed(x)
z = pd.Series(x).rolling(window_size).median()
assert np.equal(y[window_size:], z.values[window_size:]).all(
) # exact equality, since no arithmetic is done on the numbers
def test_median_scalar_inputs(window_size=3,
length=100): # no interpolation yet
pipe = rq.Pipeline(rq.LowPass(window=window_size,
portion=window_size // 2))
v = example_input(length)
for i, x in enumerate(v):
y = pipe.feed(x)
if i >= window_size:
assert y == np.median(v[(i - window_size + 1):(i + 1)])
def test_basic_nans(window_size=5, length=20):
# make sure the pipeline effectively flushes its contents with NaNs
pipe = rq.Pipeline(rq.LowPass(window=window_size,
portion=window_size // 2))
x = example_input(length)
y = pipe.feed(x)
for i in range(window_size):
pipe.feed(np.nan)
z = pipe.feed(x)
assert np.equal(y, z).all()
def test_fancy_interpolation(
window_size=10,
n_trials=200
): # small windows may be more prone to boundary/edge-condition bugs
for trial in range(n_trials):
x = example_input(window_size)
quantile = np.random.uniform()
alpha, beta = np.random.uniform(size=2)
pipe = rq.Pipeline(
rq.LowPass(window=window_size,
quantile=quantile,
alpha=alpha,
beta=beta))
y = pipe.feed(x)
z = mquantiles(x, quantile, alphap=alpha, betap=beta)
assert z == y[-1]
# illustration of what my hypothetical API should look like
import numpy as np
import rolling_quantiles as rq
filter = rq.Pipeline( # stateful filter
rq.LowPass(window=100, portion=50, subsample_rate=2),
rq.HighPass(window=10, portion=3, subsample_rate=1))
# expose specialized pipelines like `rq.MedianFilter`
input = np.random.randn(1000)
output = filter.feed(
input
) # a single `ufunc` entry point that takes in arrays or scalars and spits out an appropriate amount of output
def measure_runtime(f):
start = time.perf_counter() # could also try time.monotonic()
res = f()
return time.perf_counter() - start, res
signal = np.cumsum(np.random.normal(size=100_000_000))
series = pd.Series(signal) # construct a priori for fairness
window_sizes = np.array([4, 10, 20, 30, 40, 50]) + 1 # odd
rq_times, sc_times, pd_times = [], [], []
for window_size in window_sizes:
pipe = rq.Pipeline(
rq.LowPass(window=window_size,
portion=window_size // 2,
subsample_rate=1))
rq_time, rq_res = measure_runtime(lambda: pipe.feed(signal))
sc_time, sc_res = measure_runtime(lambda: medfilt(signal, window_size))
pd_time, pd_res = measure_runtime(lambda: series.rolling(window_size).
quantile(0.5, interpolation="nearest"))
# rq_res and sc_res will differ slightly at the edges because medfilt pads both sides with zeros as if it were a convolution.
# I pad at the beginning only, since I employ an online algorithm.
offset = window_size // 2
discrepancy = rq_res[1000:2000] - sc_res[(1000 - offset):(2000 - offset)]
#print("maximum discrepancy between the two is", np.amax(np.abs(discrepancy)))
assert np.amax(np.abs(discrepancy)) < 1e-10
print("runtimes are", rq_time, "versus", sc_time, "versus", pd_time)
rq_times.append(rq_time)
sc_times.append(sc_time)
pd_times.append(pd_time)
def test_window_size():
with pytest.raises(ValueError):
rq.Pipeline(rq.LowPass())
def test_innocuous_interpolation(window_size=1001, length=10000):
pipe = rq.Pipeline(rq.LowPass(window=window_size, quantile=0.5))
x = example_input(length)
y = pipe.feed(x)
z = pd.Series(x).rolling(window_size).median()
assert np.equal(y[window_size:], z.values[window_size:]).all()
def test_typical_interpolation(window_size=40, quantile=0.2):
x = example_input(window_size) # one window only, due to scipy
pipe = rq.Pipeline(rq.LowPass(window=window_size, quantile=quantile))
y = pipe.feed(x)
z = mquantiles(x, quantile, alphap=1, betap=1)
assert z == y[-1]