def test_mass_data1(self):
t = np.loadtxt("./data/random_walk_data.csv")
q = t[:1000]
dp = utils.mass(q, t)
ans = np.loadtxt("./data/random_walk_data_distance_profile.csv")
assert len(dp) == len(t) - len(q) + 1, "mass_data1: result should have correct length"
assert np.allclose(dp, ans, atol=1e-5), "mass_data1: distance profile should be computed correctly"
def test_mass_sanity(self):
t = np.random.rand(1000)
m = np.random.randint(len(t) - 10)
k = np.random.randint(10, len(t) - m)
q = t[k:k+m]
dp = utils.mass(q, t)
assert dp[k] < 1e-5, "test_mass_sanity: distance of a series to itself should be zero"
def test_mass_random_data(self):
n = np.random.randint(100, 1000)
m = np.random.randint(10, n)
q = np.random.rand(m)
t = np.random.rand(n)
dp = utils.mass(q, t)
ans = helpers.naive_distance_profile(q, t)
assert np.allclose(dp, ans), "mass_random_data: distance profile should be computed correctly"
def _compute_matrix_profile(self):
"""
Compute the matrix profile using STAMP.
"""
try:
for n_iter, idx in enumerate(self._iterator):
D = utils.mass(self.ts2[idx: idx+self.window_size], self.ts1)
self._elementwise_min(D, idx)
except KeyboardInterrupt:
if self.verbose:
tqdm.write("Calculation interrupted at iteration {}. Approximate result returned.".format(n_iter))
def _compute_matrix_profile(self):
"""
Compute the matrix profile using PreSCRIMP.
"""
try:
mu_T, sigma_T = utils.rolling_avg_sd(self.ts1, self.window_size)
if self._same_ts:
mu_Q, sigma_Q = mu_T, sigma_T
else:
mu_Q, sigma_Q = utils.rolling_avg_sd(self.ts2, self.window_size)
for n_iter, idx in enumerate(self._iterator):
D = utils.mass(self.ts2[idx: idx+self.window_size], self.ts1)
self._elementwise_min(D, idx)
jdx = np.argmin(D) # the index of closest profile to the current idx
# compute diagonals until the next sampled point
q1 = self.ts2[idx:idx + self.sample_interval + self.window_size - 1]
q2 = self.ts1[jdx:jdx + self.sample_interval + self.window_size - 1]
lq = min(len(q1), len(q2))
q = q1[:lq] * q2[:lq]
q = utils.rolling_sum(q, self.window_size)
D = utils.calculate_distance_profile(q, self.window_size, mu_Q[idx:idx + len(q)], sigma_Q[idx:idx + len(q)],
mu_T[jdx:jdx + len(q)], sigma_T[jdx:jdx + len(q)])
self._index_profile[jdx: jdx + len(q)] = np.where(D < self._matrix_profile[jdx:jdx + len(q)],
np.arange(idx, idx + len(q)), self._index_profile[jdx:jdx + len(q)])
self._matrix_profile[jdx:jdx + len(q)] = np.minimum(D, self._matrix_profile[jdx:jdx + len(q)])
if self._same_ts:
self._index_profile[idx:idx + len(q)] = np.where(D < self._matrix_profile[idx:idx + len(q)],
np.arange(jdx, jdx + len(q)), self._index_profile[idx:idx + len(q)])
self._matrix_profile[idx:idx + len(q)] = np.minimum(D, self._matrix_profile[idx:idx + len(q)])
# compute diagonals until the previous sampled point
if idx != 0 and jdx != 0:
q1 = self.ts2[max(0, idx - self.sample_interval):(idx + self.window_size - 1)]
q2 = self.ts1[max(0, jdx - self.sample_interval):(jdx + self.window_size - 1)]
lq = min(len(q1), len(q2))
q = q1[-lq:] * q2[-lq:]
q = utils.rolling_sum(q, self.window_size)
D = utils.calculate_distance_profile(q, self.window_size, mu_Q[idx - len(q):idx],
sigma_Q[idx - len(q):idx],
mu_T[jdx - len(q):jdx], sigma_T[jdx - len(q):jdx])
self._index_profile[jdx - len(q): jdx] = np.where(D < self._matrix_profile[jdx - len(q):jdx],
np.arange(idx - len(q), idx),
self._index_profile[jdx - len(q):jdx])
self._matrix_profile[jdx - len(q):jdx] = np.minimum(D, self._matrix_profile[jdx - len(q):jdx])
if self._same_ts:
self._index_profile[idx - len(q):idx] = np.where(D < self._matrix_profile[idx - len(q):idx],
np.arange(jdx - len(q), jdx),
self._index_profile[idx - len(q):idx])
self._matrix_profile[idx - len(q):idx] = np.minimum(D, self._matrix_profile[idx - len(q):idx])
except KeyboardInterrupt:
if self.verbose:
tqdm.write("Calculation interrupted at iteration {}. Approximate result returned.".format(n_iter))
def update_ts2(self, pt):
"""
Update the time-series ts2 with a new data point at the end of the series. If doing self-join (ts1 == ts2),
both series will be updated.
:param float pt: The value of the new data point, to be attached to the end of ts2.
"""
if self._same_ts:
self.update_ts1(pt)
else:
self.ts2 = np.append(self.ts2, pt)
s = self.ts2[-self.window_size:]
idx = len(self.ts2) - self.window_size
D = utils.mass(s, self.ts1)
self._elementwise_min(D, idx)
def update_ts1(self, pt):
"""
Update the time-series ts1 with a new data point at the end of the series. If doing self-join (ts1 == ts2),
both series will be updated.
:param float pt: The value of the new data point, to be attached to the end of ts1.
"""
self.ts1 = np.append(self.ts1, pt)
if self._same_ts:
self.ts2 = np.copy(self.ts1)
s = self.ts1[-self.window_size:]
idx = len(self.ts1) - self.window_size
D = utils.mass(s, self.ts2)
if self._same_ts:
lower_ez_bound = max(0, idx - self.exclusion_zone)
upper_ez_bound = min(len(self.ts2), idx + self.exclusion_zone) + 1
D[lower_ez_bound:upper_ez_bound] = np.inf
self._index_profile[self._matrix_profile > D[:-1]] = idx
self._matrix_profile = np.minimum(self._matrix_profile, D[:-1])
min_idx = np.argmin(D)
self._index_profile = np.append(self._index_profile, min_idx)
self._matrix_profile = np.append(self._matrix_profile, D[min_idx])
def test_mass_query_too_long(self):
with pytest.raises(ValueError):
q = np.random.rand(1000)
t = np.random.rand(200)
dp = utils.mass(q, t)