def dissCORT(x,y,q=1):
'''
Calculate the dissimilarity distance using the DTW and first and first order temporal correlation
x: A 2D N*M numpy array (data)
y: A vector M (a centroid)
q: tunning factor between DTW and first order temporal correlation 0<q<5
'''
p=len(y)
x1Index=np.ix_(range(len(x)),range(1,p))
x2Index=np.ix_(range(len(x)),range(p-1))
x=np.array(x)
y=np.array(y)
distance = [pydtw.dtw(i,y,pydtw.Settings(dist = 'euclid',
step = 'dp1',
window = 'palival_mod', param = 0.1,
compute_path = True)).get_dist() for i in x]
corrTempOrder=np.nansum((x[x1Index]-x[x2Index])*(y[1:]-y[:(p-1)]),axis=1) / (np.sqrt( np.nansum((x[x1Index]-x[x2Index])**2,axis=1) )*np.sqrt(np.nansum((y[1:]-y[:(p-1)])**2)))
if np.isnan(corrTempOrder).any():
for i in np.where(np.isnan(corrTempOrder))[0]:
corrTempOrder[i]=0
return((2/( 1+ np.exp(q*corrTempOrder)))*distance)
def warped_correlation(x, y):
# dist, path = fastdtw(z_scores_x, z_scores_y, dist=distance.euclidean)
path = pydtw.dtw(
x,
y,
pydtw.Settings(
step=
'p0sym', # Sakoe-Chiba symmetric step with slope constraint p = 0
window='palival', # type of the window
param=2.0, # window parameter
norm=False, # normalization
compute_path=True)).get_path()
x_path, y_path = zip(*path)
x_path = np.asarray(x_path)
y_path = np.asarray(y_path)
x_warped = x[x_path]
y_warped = y[y_path]
corr = np.corrcoef(x_warped, y_warped)[0, 1]
return corr
def dtw(sig1, sig2, window=5):
return pydtw.dtw(
sig1, sig2,
pydtw.Settings(step='p0sym',
window='palival',
param=window,
norm=False,
compute_path=False)).get_dist()
def fast_c_dtw_implementation(rec1, rec2, print_info=False):
s1 = rec1.np_values
s2 = rec2.np_values
if print_info:
print((s1))
print((rec1.values))
"""
Step class
Class containts different step patterns for dynamic time warping algorithm.
There are folowing step patterns available at the moment:
Usual step patterns:
'dp1'
'dp2'
'dp3'
Sakoe-Chiba classification:
'p0sym':
'p0asym':
'p05sym':
'p05asym':
'p1sym':
'p1asym':
'p2sym':
'p2asym':
You can see step pattern definition using print_step method
"""
d = pydtw.dtw(s1, s2, pydtw.Settings(dist='manhattan',
step='p2sym', # Sakoe-Chiba symmetric step with slope constraint p =
window='palival', # type of the window: scband, itakura, palival, itakura_mod
param=7.0, # window parameter
norm=True, # normalization
compute_path=print_info))
distance = d.get_dist()
if not np.isfinite(distance):
distance = 9999999.
if print_info:
print(distance)
d.plot_alignment()
#d.plot_mat_path()
#d.plot_seq_mat_path()
print(d.get_path())
print("next")
if print_info:
return float(distance), d.get_path()
return float(distance)
def warped_correlation(window_size, signals):
X = signals[0]
Y = signals[1]
cdtw = pydtw.dtw(
zscore(X),
zscore(Y),
pydtw.Settings(step = 'dp1',
window = 'scband',
param = window_size,
norm = False,
compute_path = True
)
)
dist = cdtw.get_dist()
path = len(cdtw.get_path())
warped_corellation = 1 - dist / (2 * path)
return warped_corellation
def fast_c_dtw_implementation(rec1, rec2):
s1 = rec1.np_values
s2 = rec2.np_values
d = pydtw.dtw(
s1,
s2,
pydtw.Settings(
step=
'p0sym', # Sakoe-Chiba symmetric step with slope constraint p = 0
window='palival', # type of the window
param=2.0, # window parameter
norm=True, # normalization
compute_path=False))
distance = d.get_dist()
if not np.isfinite(distance):
distance = 9999999.
return float(distance)
def dtw(src, tgt, stflag=None):
"""
DTW for the glossectomy patients
Args:
stflag (str): None or 'src' or 'tgt'
if set None, align freely
if set 'src', align according to the number of frames of src
if set 'tgt', align according to the number of frames of tgt
"""
# extract mcep only
src_mc = src[:, 0:cf.mcep_dim + 1]
tgt_mc = tgt[:, 0:cf.mcep_dim + 1]
# DTW calc only power component for distance cost
# when gllossectomy cases, it's difficult to matching by Mel-CD cost
src_pow = src_mc[:, 0]
tgt_pow = tgt_mc[:, 0]
"""
dist=距離計算関数
step=ステップ計算の形 ["dp1","dp2","dp3","p05sym","p05asym","p1sym","p1asym","p2sym","p2asym"]
window=制限窓の種類
param=制限窓の幅
試行錯誤の結果，"p05asym"が最も良く動作することが分かった
p05asym: Sakoe-Chiba classification p = 0.5, asymmetric step pattern
https://maxwell.ict.griffith.edu.au/spl/publications/papers/sigpro82_kkp_dtw.pdf
"""
d = pydtw.dtw(
src_pow, tgt_pow,
pydtw.Settings(dist='euclid',
step="p05asym",
window='nowindow',
compute_path=True))
twf = np.array(d.get_path()).T
if stflag == 'src':
sf, index = np.unique(twf[0], return_index=True)
twf = np.c_[sf, twf[1][index]].T
elif stflag == 'tgt':
tf, index = np.unique(twf[1], return_index=True)
twf = np.c_[twf[0][index], tf].T
# # DTW by melcd
# def distance_func(x, y): return melcd(x, y)
#
# _, path = fastdtw(src_mc, tgt_mc, dist=distance_func)
# twf = np.array(path).T
# alignment
src_aligned = src[twf[0]]
tgt_aligned = tgt[twf[1]]
# debug: check whether dtw works well
# plot matching path
# d.plot_alignment()
# exit()
global dtw_flag
if dtw_flag:
import matplotlib.pyplot as plt
plt.plot(src[:, 0])
plt.plot(tgt[:, 0])
plt.show()
plt.clf()
plt.plot(src_aligned[:, 0])
plt.plot(tgt_aligned[:, 0])
plt.show()
plt.clf()
dtw_flag = False
# import matplotlib.pyplot as plt
# plt.plot(src[:,0])
# plt.plot(tgt[:,0])
# plt.show()
# plt.clf()
#
# plt.plot(src_aligned[:,0])
# plt.plot(tgt_aligned[:,0])
# plt.show()
# plt.clf()
# exit()
return src_aligned, tgt_aligned
with open("bench_log.txt", "a") as text_file:
text_file.write("\n" + str(amp + 1) + "," + str(window + 1) +
",ucr," + str(ucrdist) + "," +
str(timet.microseconds) + ',' +
str(ucrloc + count) + ',' + str(ucrloc + count))
with open(
'paths/' + str(amp + 1) + "," + str(window + 1) +
'_loc_ucr.txt', "w") as text_file:
text_file.write(str(ucrloc))
#cydtw
timeb = datetime.now()
cdtw_master = pydtw.dtw(
x,
y,
pydtw.Settings(
step=
'p0sym', #Sakoe-Chiba symmetric step with slope constraint p = 0
window='palival', #type of the window
param=2.0, #window parameter
norm=False, #normalization
compute_path=True))
timet = datetime.now() - timeb
print("cdtw complete on amp " + str(amp + 1))
cdtwpath1, cdtwpath2 = zip(*cdtw_master.get_path())
cdtwpath1 = np.asarray(cdtwpath1)
cdtwpath2 = np.asarray(cdtwpath2)
with open("bench_log.txt", "a") as text_file:
text_file.write("\n" + str(amp + 1) + "," + str(window + 1) +
",cdtw," + str(cdtw_master.get_dist()) + "," +
str(timet.microseconds) + ',' +
str(cdtwpath2[0] + count) + ',' +
str(cdtwpath2[-1] + count))
else:
guyzz.append(np.genfromtxt('records/'+filename,delimiter=',',dtype=None))
n_waves=np.shape(guyzz[0])[1]
comb=list(itertools.combinations(range(np.shape(guyzz)[0]),2))
# ^^this is so f*****g brutal^^
matrix=np.zeros([np.shape(guyzz)[0],np.shape(guyzz)[0]])
for i in comb:
correlation_vector=0
firstguy=guyzz[i[0]]
secondguy=guyzz[i[1]]
minValue= np.min([np.shape(firstguy)[0],np.shape(secondguy)[0]])
firstguy2=firstguy[0:minValue][:]
secondguy2=secondguy[0:minValue][:] #trim the f**k out
for waves in range(n_waves):
d= pydtw.dtw(firstguy[:][waves], secondguy[:][waves],
pydtw.Settings(dist='manhattan',step='dp2',window='nowindow',
compute_path=True,norm=True))
correlation_vector+=d.get_dist()
correlation_vector/=n_waves
matrix[i[0]][i[1]]=correlation_vector
print(name + " : ")
print(matrix)
print('---------------')
