def test_same_result_with_offset_index():
xx = np.arange(20, dtype=np.int64)
y = 3 * xx + 9
sol1 = l1_fit(xx, y)
xfit1 = sol1["x"]
xx = xx + 999
sol2 = l1_fit(xx, y)
xfit2 = sol2["x"]
assert (xfit1 == xfit2).all()
def test_gaps_work_on_line():
xx = np.arange(20, dtype=np.int64)
y = 3 * xx + 9
sol = l1_fit(xx, y)
xfit = sol["x"]
# this one has a big gap
xx = np.array(range(7) + range(15, 20))
y = 3 * xx + 9
sol = l1_fit(xx, y)
xfit2 = sol["x"]
for i, j in enumerate(xx):
diff = abs(xfit[j] - xfit2[i])
assert diff < 1e-6
def test_l1_fit_rand_with_permissive(
beta_d2=4.0, beta_d1=1.0, beta_seasonal=1.0, beta_step=2.5, period=12, noise=0, seed=3733, doplot=True, sea_amp=0.05
):
# print "seed=%s,noise=%s,beta_d2=%s,beta_d1=%s,beta_step=%s," \
# "beta_seasonal=%s" % (seed,noise,beta_d2,beta_d1,beta_step,beta_seasonal)
mock = make_l1tf_mock2(noise=noise, seed=seed, sea_amp=sea_amp)
y = mock["y_with_seasonal"]
xx = mock["x"]
step_permissives = [(30, 0.5)]
sol = l1_fit(
xx,
y,
beta_d2=beta_d2,
beta_d1=beta_d1,
beta_seasonal=beta_seasonal,
beta_step=beta_step,
period=period,
step_permissives=step_permissives,
)
if doplot:
plt.clf()
plt.plot(xx, y, linestyle="-", marker="o", markersize=4)
plt.plot(xx, sol["xbase"], label="base")
plt.plot(xx, sol["steps"], label="steps")
plt.plot(xx, sol["seas"], label="seasonal")
plt.plot(xx, sol["x"], label="full")
plt.legend(loc="upper left")
def test_l1_fit_monthly():
tol = 1e-12
xx = np.arange(4, dtype=np.int64)
y = 3 * xx + 9
sol1 = l1_fit(xx, y)
xfit1 = sol1["x"]
xx = np.array([date(2014, 1, 2), date(2014, 2, 5), date(2014, 3, 17), date(2014, 4, 1)])
sol2 = l1_fit_monthly(xx, y)
xfit2 = sol2["x"]
diff = abs(xfit1 - xfit2).max()
assert diff < tol
def test_l1_fit_linearity():
mock = make_l1tf_mock2(noise=0.05, seed=48457)
y = mock["y_with_seasonal"]
xx = mock["x"]
step_permissives = [(30, 0.5)]
sol = l1_fit(xx, y, period=6, step_permissives=step_permissives)
# some linear transformation should effect things only linearly
scale = 17.0
offset = -56.0
y = y * scale + offset
sol2 = l1_fit(xx, y, period=6, step_permissives=step_permissives)
x = sol["x"]
x2 = sol2["x"]
x_unscaled = (x2 - offset) / scale
tol = 1e-10
assert abs(x - x_unscaled).max() < tol
def test_l1_fit(beta_d2=4.0, beta_d1=1.0, beta_seasonal=1.0, beta_step=2.5, period=12, noise=0, seed=3733):
mock = make_l1tf_mock2(noise=noise, seed=seed)
y = mock["y_with_seasonal"]
xx = mock["x"]
plt.clf()
plt.plot(xx, y, linestyle="-", marker="o", markersize=4)
sol = l1_fit(
xx, y, beta_d2=beta_d2, beta_d1=beta_d1, beta_seasonal=beta_seasonal, beta_step=beta_step, period=period
)
plt.plot(xx, sol["xbase"], label="base")
plt.plot(xx, sol["steps"], label="steps")
plt.plot(xx, sol["seas"], label="seasonal")
plt.plot(xx, sol["x"], label="full")
plt.legend(loc="upper left")
def test_l1_fit_step():
# test that the place of the step and the specification of the step
# permissive do not have a off-by-one error
# note our steps occur from right to left
# the first one (right to left) that has jumped is the place
# where h is non-zero
n = 55
xx = np.arange(n)
y = xx * 0.0
y[xx <= 20] = 1000.0
# this means that y[20] includes the step-up and so 20 is where 'h' should be non-zero
sol = l1_fit(xx, y, beta_step=100.0, step_permissives=[(20, 0.1)])
print abs(sol["h"][19]) < 1e-10
print abs(sol["h"][21]) < 1e-10
print abs(sol["h"][20] - 1000.0) < 1e-10
def test_l1_fit_speed(beta_d2=4.0, beta_d1=1.0, beta_seasonal=1.0, beta_step=2.5, period=12, n=50, num=100):
mock = make_l1tf_mock2()
y = mock["y_with_seasonal"]
xx = mock["x"]
start = time.time()
for i in xrange(num):
sol = l1_fit(
xx[0:n],
y[0:n],
beta_d2=beta_d2,
beta_d1=beta_d1,
beta_seasonal=beta_seasonal,
beta_step=beta_step,
period=period,
)
fin = time.time()
runtime = fin - start
rate = num / runtime
print "num: %s, runtime: %s seconds, rate: %s per sec for %s points" % (num, runtime, rate, n)
return runtime
