def compute_mat(data, sigma, t, verbose=True):
if verbose:
n = len(data)
m_0 = np.zeros((n,n))
for i in range(n):
print(str( (100. * i) / n ) + '%')
for j in range(n):
m_0[i,j] = np.exp( - ga.tga_dissimilarity(data[i], data[j], 2*sigma, t) )
else:
m_0 = np.array([[np.exp( - ga.tga_dissimilarity(d_1, d_2, 2*sigma, t) )
for d_1 in data]
for d_2 in data])
return m_0
def _GA(self, v1, v2):
sq1 = v1.reshape(1, len(v1)).astype('double')
sq2 = v2.reshape(1, len(v2)).astype('double')
myv = ga.tga_dissimilarity(sq1, sq2, self.sig, 0)
myv2 = np.exp(-myv)
sim = 1 - myv2
return sim
def TGAKernelXY(X, Y, sigma=0.1, triangular=10):
N = np.shape(X)[0]
M = np.shape(Y)[0]
KSeq = np.zeros((N, M))
for row1ind in range(N):
for row2ind in range(M):
val = ga.tga_dissimilarity(X[row1ind].T.copy(order='C'),
Y[row2ind].T.copy(order='C'), sigma,
triangular)
# KSeq[row1ind,row2ind]=val
KSeq[row1ind, row2ind] = np.exp(-val)
return KSeq
def test(d=10, T1=100, T2=120):
np.random.seed(1)
# generate two random time series of dimension d and duration T1, resp. T2
seq1 = np.random.rand(T1, d)
seq2 = np.random.rand(T2, d)
# define the sigma parameter
sigma = 0.5*(T1+T2)/2*np.sqrt((T1+T2)/2)
# compute the global alignment kernel value for different triangular parameters
diff_t = np.abs(T1-T2)
Ts = sorted(set([0, 1, diff_t/2, diff_t, diff_t+1, diff_t*3/2, diff_t*2]))
for triangular in Ts:
val = ga.tga_dissimilarity(seq1, seq2, sigma, triangular)
kval = np.exp(-val)
if 0 < triangular <= diff_t:
# for 0 < triangular <= diff_t, exp(-tga_d) == 0
assert kval == 0
print "T=%d \t exp(-tga_d)=%0.5f" % (triangular, kval)
def compare(artists):
seq1=artists[0]["event_sequence_with_0"]
seq2=artists[1]["event_sequence_with_0"]
T1=makeArray(seq1)
T2=makeArray(seq2)
# len_diff=max(len(seq1),len(seq2))-min(len(seq1),len(seq2))
# S1=makeArray(seq1 if len(seq1)>=len(seq2) else seq1+range(len_diff))
# S2=makeArray(seq2 if len(seq2)>=len(seq1) else seq2+range(len_diff))
# print S1
# print S2
# define the sigma parameter
sigma = 10*(len(seq1)+len(seq2))/2 * np.sqrt((len(seq1)+len(seq2))/2)
# compute the global alignment kernel value for different triangular parameters
val = ga.tga_dissimilarity(T1, T2, sigma, 0)
kval = np.exp(-val)
return kval
def test(d=10, T1=100, T2=120):
np.random.seed(1)
# generate two random time series of dimension d and duration T1, resp. T2
seq1 = np.random.rand(T1, d)
seq2 = np.random.rand(T2, d)
# define the sigma parameter
sigma = 0.5 * (T1 + T2) / 2 * np.sqrt((T1 + T2) / 2)
# compute the global alignment kernel value for different triangular parameters
diff_t = np.abs(T1 - T2)
Ts = sorted(
set([0, 1, diff_t / 2, diff_t, diff_t + 1, diff_t * 3 / 2,
diff_t * 2]))
for triangular in Ts:
val = ga.tga_dissimilarity(seq1, seq2, sigma, triangular)
kval = np.exp(-val)
if 0 < triangular <= diff_t:
# for 0 < triangular <= diff_t, exp(-tga_d) == 0
assert kval == 0
print "T=%d \t exp(-tga_d)=%0.5f" % (triangular, kval)
def compare(artists):
seq1 = artists[0]["event_sequence_with_0"]
seq2 = artists[1]["event_sequence_with_0"]
T1 = makeArray(seq1)
T2 = makeArray(seq2)
# len_diff=max(len(seq1),len(seq2))-min(len(seq1),len(seq2))
# S1=makeArray(seq1 if len(seq1)>=len(seq2) else seq1+range(len_diff))
# S2=makeArray(seq2 if len(seq2)>=len(seq1) else seq2+range(len_diff))
# print S1
# print S2
# define the sigma parameter
sigma = 10 * (len(seq1) + len(seq2)) / 2 * np.sqrt(
(len(seq1) + len(seq2)) / 2)
# compute the global alignment kernel value for different triangular parameters
val = ga.tga_dissimilarity(T1, T2, sigma, 0)
kval = np.exp(-val)
return kval