def test_bug2():
with util_numpy.test_uses_numpy() as np:
s1 = np.array([0, 0, 1, 2, 1, 0, 1, 0, 0], dtype=np.double)
s2 = np.array([0.0, 1, 2, 0, 0, 0, 0, 0, 0])
d1a = dtw.distance_fast(s1, s2, window=2)
d1b = dtw.distance(s1, s2, window=2)
if directory:
fn = directory / "warpingpaths.png"
else:
file = tempfile.NamedTemporaryFile()
fn = Path(file.name + "_warpingpaths.png")
d2, paths = dtw.warping_paths(s1, s2, window=2)
best_path = dtw.best_path(paths)
if not dtwvis.test_without_visualization():
dtwvis.plot_warpingpaths(s1,
s2,
paths,
best_path,
filename=fn,
shownumbers=False)
print("Figure saved to", fn)
assert d1a == pytest.approx(d2)
assert d1b == pytest.approx(d2)
def quality_vs_trainsize(X_train, X_test, method, dist='dtw'):
method = clone(method)
def fit():
method.estimator.set_params(**{
'verbose': 0,
})
method.fit(X_train)
X_pred_codes = method.transform_codes(X_test)
X_pred = method.codes_to_signal(X_pred_codes)
rate = BaseBenchmark.compression_rate(X_test, X_pred_codes)
inv_rate = 1 / rate
return X_pred, rate, inv_rate
X_pred, rate, inv_rate = fit()
# Must truncate test timeseries to prediction timeseries
a1 = np.array(X_test)
a2 = np.array(X_pred)
assert a1.shape[0] >= a2.shape[0]
a1.resize(a2.shape)
# Compute distance
if dist == 'dtw':
d = dtw.distance_fast(a1, a2)
elif dist == 'rmsre':
d = np.sqrt(np.mean(np.square(np.divide(a1 - a2, a1))))
else:
raise ValueError(f'Unknown distance {dist}')
return X_train.shape[0], d, rate, inv_rate
def test_distance1_numpy():
with util_numpy.test_uses_numpy() as np:
s1 = np.array([0., 0, 1, 2, 1, 0, 1, 0, 0])
s2 = np.array([0., 1, 2, 0, 0, 0, 0, 0, 0])
d = dtw.distance_fast(s1, s2)
# print(f'd = {d}')
assert d == pytest.approx(math.sqrt(2))
def average_distance_to_templates(sequence, templates, warping_window):
len_seq = sequence.shape[0]
val_min = np.inf
label = 1
for i, t in enumerate(templates, start=1):
val = 0.0
for temp in t:
len_temp = temp.shape[0]
if warping_window is not None:
warping_window = int(
np.ceil(warping_window * max(len_seq, len_temp)))
# If the sequences have different lengths, the warping window cannot be smaller than the difference
# between the length of the sequences
warping_window = max(warping_window,
abs(len_seq - len_temp + 1))
d = (dtw.distance_fast(
sequence[:, 0], temp[:, 0],
window=warping_window)) / float(len_seq + len_temp)
val += d
val /= len(t)
if val < val_min:
val_min = val
label = i
return val_min, label
def test_psi_dtw_1d():
with util_numpy.test_uses_numpy() as np:
x = np.arange(0, 20, .5)
s1 = np.sin(x)
s2 = np.sin(x - 1)
random.seed(1)
for idx in range(len(s2)):
if random.random() < 0.05:
s2[idx] += (random.random() - 0.5) / 2
# print(f's1 = [' + ','.join(f'{vv:.2f}' for vv in s1) + ']')
# print(f's2 = [' + ','.join(f'{vv:.2f}' for vv in s2) + ']')
# print('distance_fast')
d1 = dtw.distance_fast(s1, s2, psi=2)
# print(f'{d1=}')
# print('warping_paths')
d2, paths = dtw.warping_paths(s1, s2, window=25, psi=2)
# print(f'{d2=}')
with np.printoptions(threshold=np.inf, linewidth=np.inf):
print(paths)
# print('warping_paths fast')
d3, paths = dtw.warping_paths_fast(s1, s2, window=25, psi=2)
# print(f'{d3=}')
# print(paths)
# print('best_path')
best_path = dtw.best_path(paths)
if not dtwvis.test_without_visualization():
if directory:
dtwvis.plot_warpingpaths(s1, s2, paths, best_path, filename=directory / "test_psi_dtw_1d.png")
np.testing.assert_almost_equal(d1, d2)
np.testing.assert_almost_equal(d1, d3)
def test_psi_dtw_1c():
with util_numpy.test_uses_numpy() as np:
x = np.arange(0, 20, .5)
s1 = np.sin(x)
s2 = np.sin(x - 1)
d = dtw.distance_fast(s1, s2, psi=2)
np.testing.assert_equal(d, 0.0)
def _reconizer_run(self, x, y):
"""
Roda o algotimo com a função de distância escolhida.
Esse método pode ser sobrescrito para utilizar implementações diferentes do DTW.
Como o AcceleratedDTW.
"""
return dtw.distance_fast(x, y, use_pruning=True)
def test_distance2_c():
with util_numpy.test_uses_numpy() as np:
dist_opts = {}
s1 = np.array([0.0, 0.0, 2.0, 1.0, 1.0, 0.0, 0.0])
s2 = np.array([0.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0])
d1 = dtw.distance(s1, s2, **dist_opts)
d2 = dtw.distance_fast(s1, s2, **dist_opts)
assert d1 == d2
assert d1 == pytest.approx(1.0)
def dist_matrix_py_fast(self,
x,
y,
rss=lambda x, y, z: np.sqrt(x**2 + y**2 + z**2)):
# Compute the distance matrix
# dm_count = 0
x_s = np.shape(x)
y_s = np.shape(y)
dm = np.zeros((x_s[0], y_s[0]))
# dm_size = x_s[0] * y_s[0]
# p = ProgressBar(dm_size)
for i in range(0, x_s[0]):
for j in range(0, y_s[0]):
# x_accelX = array.array('d', [round(n, 2) for n in x[i]["accelX"]])
# x_accelY = array.array('d', [round(n, 2) for n in x[i]["accelY"]])
# x_accelZ = array.array('d', [round(n, 2) for n in x[i]["accelZ"]])
# y_accelX = array.array('d', [round(n, 2) for n in y[j]["accelX"]])
# y_accelY = array.array('d', [round(n, 2) for n in y[j]["accelY"]])
# y_accelZ = array.array('d', [round(n, 2) for n in y[j]["accelZ"]])
x_accelX = array.array('d', x[i]["accelX"])
x_accelY = array.array('d', x[i]["accelY"])
x_accelZ = array.array('d', x[i]["accelZ"])
y_accelX = array.array('d', y[j]["accelX"])
y_accelY = array.array('d', y[j]["accelY"])
y_accelZ = array.array('d', y[j]["accelZ"])
# 3D modification
# dm[i, j] = rss(fastdtw(x_accelX, y_accelX, dist=euclidean),
# fastdtw(x_accelY, y_accelY, dist=euclidean),
# fastdtw(x_accelZ, y_accelZ, dist=euclidean))
dm[i, j] = rss(
dtw.distance_fast(x_accelX, y_accelX, use_pruning=True),
dtw.distance_fast(x_accelY, y_accelY, use_pruning=True),
dtw.distance_fast(x_accelZ, y_accelZ, use_pruning=True))
# dm[i, j] = self._dtw_distance(x[i, ::self.subsample_step],
# y[j, ::self.subsample_step])
# Update progress bar
# dm_count += 1
# p.animate(dm_count)
return dm
def test_distance2_bb():
with util_numpy.test_uses_numpy() as np:
dist_opts = {'max_step': 0.1}
s1 = np.array([0.0, 0.0, 2.0, 1.0, 1.0, 0.0, 0.0])
s2 = np.array([0.0, 1.0, 1.0, 1.0, 0.0, 0.0, 0.0])
d1 = dtw.distance(s1, s2, **dist_opts)
d2 = dtw.distance_fast(s1, s2, **dist_opts)
print(d1, d2)
assert d1 == d2
assert d1 == pytest.approx(np.inf)
def dta_dtw(signalA, signalB, **dtw_kwargs):
'''
The function bundles the path and distance of the dtaidistance package.
This is the underlying process to be applied in the HDTW process.
:param signalA: The first signal to apply DTW on
:param signalB: The second signal to apply DTW on
:param **dtw_kwargs: any key-word arguments to be propogated to the functions.
'''
return dtw.distance_fast(signalA, signalB,**dtw_kwargs), \
dtw.warping_path(signalA, signalB, **dtw_kwargs)
def test_distance1_a():
with util_numpy.test_uses_numpy() as np:
# dist_opts = {'max_dist': 0.201, 'max_step': 0.011, 'max_length_diff': 8, 'window': 3}
dist_opts = {'window': 3}
s1 = np.array(
[0., 0.01, 0., 0.01, 0., 0., 0., 0.01, 0.01, 0.02, 0., 0.])
s2 = np.array([0., 0.02, 0.02, 0., 0., 0.01, 0.01, 0., 0., 0., 0.])
d1 = dtw.distance(s1, s2, **dist_opts)
d2 = dtw.distance_fast(s1, s2, **dist_opts)
print("X")
assert d1 == d2
assert d1 == pytest.approx(0.02)
def test_distance1_b():
with util_numpy.test_uses_numpy() as np:
dist_opts = {}
s1 = np.array(
[0., 0.01, 0., 0.01, 0., 0., 0., 0.01, 0.01, 0.02, 0., 0.])
s2 = np.array([0., 0.02, 0.02, 0., 0., 0.01, 0.01, 0., 0., 0., 0.])
d1 = dtw.distance(s1, s2, **dist_opts)
d2 = dtw.distance_fast(s1, s2, **dist_opts)
d3, wps = dtw.warping_paths(s1, s2, **dist_opts)
print(np.power(wps, 2))
assert d1 == d2
assert d1 == d3
assert d1 == pytest.approx(0.02)
def quality_vs_width(X_train,
X_test,
method,
widths,
stride,
n_atoms,
dist='dtw'):
method = clone(method)
def fit(width):
method.set_params(**{
'width': width,
'stride': stride,
})
method.estimator.set_params(**{
'n_components': n_atoms,
'verbose': 0,
})
method.fit(X_train)
X_pred_codes = method.transform_codes(X_test)
X_pred = method.codes_to_signal(X_pred_codes)
rate = BaseBenchmark.compression_rate(X_test, X_pred_codes)
inv_rate = 1 / rate
return X_pred, rate, inv_rate
dists = []
rates = []
inv_rates = []
for width in tqdm(widths, leave=False):
X_pred, rate, inv_rate = fit(width)
# Must truncate test timeseries to prediction timeseries
a1 = np.array(X_test)
a2 = np.array(X_pred)
assert a1.shape[0] >= a2.shape[0]
a1.resize(a2.shape)
# Compute distance
if dist == 'dtw':
d = dtw.distance_fast(a1, a2)
elif dist == 'rmsre':
d = np.sqrt(np.mean(np.square(np.divide(a1 - a2, a1))))
else:
raise ValueError(f'Unknown distance {dist}')
dists.append(d)
rates.append(rate)
inv_rates.append(inv_rate)
return dists, rates, inv_rates
def helper_dtw_distance(sequence, templates, warping_window, index_tuple):
t = templates[index_tuple[0]][index_tuple[1]]
len_seq = sequence.shape[0]
len_temp = t.shape[0]
if warping_window is not None:
warping_window = int(np.ceil(warping_window * max(len_seq, len_temp)))
# If the sequences have different lengths, the warping window cannot be smaller than the difference
# between the length of the sequences
warping_window = max(warping_window, abs(len_seq - len_temp + 1))
d = (dtw.distance_fast(sequence[:, 0], t[:, 0],
window=warping_window)) / float(len_seq + len_temp)
return index_tuple[0], index_tuple[1], d
def distance(C, subim, S, m, rmin, cmin, factor):
from dtaidistance import dtw
"""
This function computes the spatial-temporal distance between
two pixels using the DTW distance.
Keyword arguments:
C : numpy.ndarray
ND-array containing cluster centres information
subim : numpy.ndarray
Cluster under analisis
S : float
Spacing
m : float
Compactness
rmin : float
Minimum row
cmin : float
Minimum column
factor : float
Corrective factor
Returns
-------
D: numpy.ndarray
ND-Array distance
"""
#Initialize submatrix
dc = numpy.zeros([subim.shape[1],subim.shape[2]])
ds = numpy.zeros([subim.shape[1],subim.shape[2]])
#get cluster centres
a2 = C[:subim.shape[0]] #Average time series
ic = (int(numpy.floor(C[subim.shape[0]])) - rmin) #X-coordinate
jc = (int(numpy.floor(C[subim.shape[0]+1])) - cmin) #Y-coordinate
# Critical Loop - need parallel implementation
for u in range(subim.shape[1]):
for v in range(subim.shape[2]):
a1 = subim[:,u,v] # Get pixel time series
dc[u,v] = dtw.distance_fast(a1.astype(float),a2.astype(float)) #Compute DTW distance
ds[u,v] = (((u-ic)**2 + (v-jc)**2)**0.5) #Calculate Spatial Distance
D = ( (dc)/m + (ds/S) )**0.5 #Calculate SPatial-temporal distance
return D
def find_closes_branch_ar(array_query, temp_exemplars):
closest_nodes = list()
bsf = 100000
for i in range(0, len(temp_exemplars)):
exemplars = np.asarray(temp_exemplars.__getitem__(i))
if DistanceMeasure.are_equal(exemplars, array_query):
return i
dist = dtw.distance_fast(array_query, exemplars)
if dist < bsf:
bsf = dist
closest_nodes.clear()
closest_nodes.append(i)
elif bsf == dist:
closest_nodes.append(i)
r = random.randint(0, len(closest_nodes) - 1)
return closest_nodes[r]
def test_distance4():
with util_numpy.test_uses_numpy(strict=False) as np:
try:
import pandas as pd
except ImportError:
# If no pandas, ignore test (not a required dependency)
return
s = [[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0.005, 0.01, 0.015, 0.02, 0.01, 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]]
p = np.array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.])
df = pd.DataFrame(data=s)
s = df.values
for i in range(s.shape[0]):
ss = s[i] # ss will not be C contiguous memory layout
d = dtw.distance_fast(ss, p)
def test_distance3_a():
with util_numpy.test_uses_numpy() as np:
dist_opts = {"penalty": 0.005, "max_step": 0.011, "window": 3}
s = np.array([
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.005, 0.01,
0.015, 0.02, 0.01, 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0.
])
p = np.array([
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.005, 0.01, 0.015,
0.02, 0.01, 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.
])
d1 = dtw.distance(s, p, **dist_opts)
d2 = dtw.distance_fast(s, p, **dist_opts)
assert d1 == pytest.approx(d2)
def find_closest_nodes(query, temp_exemplars: list):
array_query = np.asarray(query)
closest_nodes = list()
bsf = np.inf
if len(temp_exemplars) == 0:
return -1
for i in range(0, len(temp_exemplars)):
exemplars = np.asarray(temp_exemplars.__getitem__(i))
"""
if AppContext.AppContext.elastic_distance == "dtw":
# dist = dtw.distance_fast(array_query, exemplars, window=AppContext.AppContext.window_length,
# max_dist=bsf)
elif AppContext.AppContext.elastic_distance == "lcss":
esilon = LCSS.get_random_epsilon()
dist = LCSS.distance(array_query, exemplars, window_size=-np.inf, epsilon=esilon)
elif AppContext.AppContext.elastic_distance == "twe":
dist = TWE.distance(array_query, exemplars, TWE.get_random_nu(), TWE.get_random_lambda())
else:
"""
try:
dist = dtw.distance_fast(array_query, exemplars, window=2)
except RecursionError:
dist = DistanceMeasure._dtw_distance(array_query,
exemplars,
d=lambda x, y: abs(x - y))
if len(closest_nodes) == 0:
bsf = dist
closest_nodes.append(i)
else:
if dist < bsf:
bsf = dist
closest_nodes.clear()
closest_nodes.append(i)
elif bsf == dist:
closest_nodes.append(i)
if len(closest_nodes) == 0:
print("There are no closest Nodes")
r = 0
elif len(closest_nodes) == 1:
return closest_nodes[0]
else:
r = random.randint(0, len(closest_nodes) - 1)
return closest_nodes[r]
def bmu(self, x, windowSize, candidateWeights):
## function to find winning node
#: input observatiopn
# format input for use in this function --- dtw distance
# x = np.reshape(x[0], (1, 1, len(x[0])))
####################################
# calculate distances (in Euclidean and DTW it is the minimum). Iterate over all nodes to find distance
#x = np.reshape(x,(1,len(x[0])))
min_distance = float('inf')
dtw_Cals = 0
for i in candidateWeights:
#get candidate weight
#this needs to be a deep copy for dtw distance not to throw an error.
weights = copy.deepcopy(self.weights_Kohonen[i])
xCopy = copy.deepcopy(x)
#get dtw distance
distance = dtw.distance_fast(xCopy,
weights,
window=windowSize + 1,
use_pruning=True)
dtw_Cals += 1
#update min distance if new distance is lower
if distance < min_distance:
bmuIndex = i
min_distance = distance
#update min distance if new distance is equal to min distance
elif distance == min_distance and np.random.uniform(low=0,
high=1) < 0.3:
bmuIndex = i
min_distance = distance
return [bmuIndex, dtw_Cals]
def createUmatrix(self, windowSize):
normalizedWeights = self.weights_Kohonen
self.neighborhoodSize = 1
self.Umatrix = np.zeros((self.hiddenSize[0], self.hiddenSize[1]))
# Perform 2D convolution with input data and kernel to get sum of neighboring nodes
#for i in range(0,self.hiddenSize[0]):
#for j in range(0,self.hiddenSize[1]):
for idx in range(0, len(self.weights_Kohonen)):
nodeWeights = normalizedWeights[idx]
[i, j] = np.unravel_index(idx, self.hiddenSize)
#find all the neighbors for node at i,j
neighbors2d = self.find_neighbor_indices(i, j)
#remove self
neighbors2d.remove((i, j))
#get weights for node at i,j
neighbors = []
for neighbor in neighbors2d:
neighbors.append(
np.ravel_multi_index(neighbor, self.hiddenSize))
#temp = []
for neighbor in neighbors:
#for dtw
neighborWeights = normalizedWeights[neighbor]
distance = dtw.distance_fast(neighborWeights,
nodeWeights,
window=windowSize + 1,
use_pruning=True)
#distance2 = np.sqrt(sum((nodeWeights.flatten()-neighborWeights.flatten())**2))
#temp.append(distance)
self.Umatrix[i, j] += distance
#self.Umatrix[i, j] = min(temp)
return self.Umatrix
def distance_slic(C, subim, S, m, rmin, cmin):
#Initialize submatrix
dc = np.zeros([subim.shape[1], subim.shape[2]])
ds = np.zeros([subim.shape[1], subim.shape[2]])
#get cluster centres
a2 = C[:subim.shape[0]] #Average time series
ic = (int(np.floor(C[subim.shape[0]])) - rmin) #X-coordinate
jc = (int(np.floor(C[subim.shape[0] + 1])) - cmin) #Y-coordinate
# Critical Loop - need parallel implementation
for u in range(subim.shape[1]):
for v in range(subim.shape[2]):
a1 = subim[:, u, v] # Get pixel time series
dc[u,
v] = dtw.distance_fast(a1.astype(float),
a2.astype(float)) #Compute DTW distance
ds[u, v] = (
((u - ic)**2 + (v - jc)**2)**0.5) #Calculate Spatial Distance
D = (dc**2 + (m**2) *
(ds / S)**2)**0.5 #Calculate SPatial-temporal distance
return D
def d():
return dtw.distance_fast(s1, s2)
def test_distance6():
s1 = np.array([0, 0, 1, 2, 1, 0, 1, 0, 0], dtype=np.double)
s2 = np.array([0.0, 1, 2, 0, 0, 0, 0, 0, 0])
d = dtw.distance_fast(s1, s2, window=2)
print(d)
def d():
return dtw.distance_fast(s1, s2, use_pruning=True)
def test(amplitude, center, width, noise, target_norm,len_a,window,global_max, ii):
source_no_noise = amplitude*norm.pdf(range(0,400),center,width)
source_no_noise = np.random.normal(source_no_noise, scale=0)
source_max = max(abs(source_no_noise))
source_norm = amplitude*norm.pdf(range(0,400),center,width)
source_norm = np.random.normal(source_norm, scale=noise)
source_global_norm = source_norm/global_max
total_intensity = (abs(source_norm)).sum()
t,s = target_norm,source_norm
'''
Choose the normalisation method by commenting/uncommenting below
'''
target_norm,source_norm = target_norm/max(abs(target_norm)),source_norm/source_max
'''
target_norm = preprocessing.normalize([target_norm])
target_norm = target_norm.reshape((400,))
source_norm = preprocessing.normalize([source_norm])
source_norm = source_norm.reshape((400,))
'''
'''
'''
target,source = deriv(target_norm),deriv(source_global_norm)
target.insert(0,0)
target.insert(len(target),0)
source.insert(0,0)
source.insert(len(source),0)
target,source = np.asarray(target),np.asarray(source)
D = dtw.distance_fast(target,source,window,psi=0)
target_deriv,source_deriv = deriv(target_norm),deriv(source_norm)
target_deriv.insert(0,0)
target_deriv.insert(len(target_deriv),0)
source_deriv.insert(0,0)
source_deriv.insert(len(source_deriv),0)
target_deriv,source_deriv = np.asarray(target_deriv),np.asarray(source_deriv)
d, paths = dtw.warping_paths_fast(target_deriv, source_deriv, window, psi=0)
best_path = dtw.best_path(paths)
paths = np.array(best_path)
'''
euclidean = pow((target_norm[paths[:,0]] - source_norm[paths[:,1]]),2)
euclidean = np.sqrt(euclidean.sum())
'''
euclidean = abs(t[paths[:,0]] - s[paths[:,1]])
euclidean = euclidean.sum()
plt.figure(3)
plt.scatter(ii,euclidean)
plt.title('warping cost')
euclidean = np.exp(-euclidean/(total_intensity/D))
init_euclidean = abs(paths[:,0]-paths[:,1])
amplitude,center,width,noise = str(amplitude),str(center),str(width),str(noise)
plt.figure(1)
plt.plot(range(0,400),source_norm)
plt.figure(2)
plt.plot(range(0,400),source_deriv)
plt.plot(range(0,400),target_deriv,'k-')
plt.figure(4)
plt.scatter(ii,total_intensity)
plt.title('total intensity')
plt.figure(5)
plt.scatter(ii,D)
plt.title('D')
return euclidean, init_euclidean, target_norm
def dtw_distance(x, y):
return dtw.distance_fast(x, y, use_pruning=True)
img_name = "opencv_frame_{}.png".format(img_counter)
print("Please wait")
res = cv2.resize(gray, (100,100),interpolation = cv2.INTER_AREA)
hand = segment(res)
# check whether hand region is segmented
if hand is not None:
thresholded = hand
thresholded = cv2.erode(thresholded, None, iterations=2)
thresholded = cv2.dilate(thresholded, None, iterations=2)
cv2.imshow("Thesholded", thresholded)
dtw_feat = feat(thresholded)
final = []
for xtrain in x_train:
distance = dtw.distance_fast(dtw_feat,xtrain)
final.append(distance)
mini = final.index(min(final))
pred = y_train[mini]
print("Predicted:"+pred)
cv2.putText(clone,pred, (200,100), cv2.FONT_HERSHEY_SIMPLEX, 2, 127,3)
cv2.imwrite(img_name, clone)
print("{} written!".format(img_name))
img_counter += 1
cv2.imshow("Video Feed", clone)
def test_distance6():
with util_numpy.test_uses_numpy() as np:
s1 = np.array([0, 0, 1, 2, 1, 0, 1, 0, 0], dtype=np.double)
s2 = np.array([0.0, 1, 2, 0, 0, 0, 0, 0, 0])
d = dtw.distance_fast(s1, s2, window=2)