def extend_func_POIM(Q, P, M, L, r, mu, sigma, sm, maxdegree, small_k):
R = []
cR = []
for k in range(small_k):
cR.append(np.zeros((math.pow(4, k + 1) * L)))
R.append(np.zeros((math.pow(4, k + 1), L)))
for p in range(len(P)):
v = []
D = P[p] - small_k + 1
for m in range(M[p]):
vh = []
for msub in range(D):
vh.append(compute_w(L, r[p][m][:, msub:msub + small_k], sigma[p][m], mu[p][m] + msub, small_k))
v.append(vh)
cR[small_k - 1] += compute_scoring(mu[p], sigma[p], L, v, small_k, sm[p], D) #ein += statt = eingefuegt am 16.10
for i in range(small_k):
R[i] = utils.vec2matrix(cR[i], i + 1)
f = 0
for p in range(len(P)):#small_k):
for i in range(len(R[small_k - 1])):
for j in range(len(R[small_k - 1][0])):
f += math.pow(R[small_k - 1][i][j] - Q[small_k - 1][i][j], 2)
fval = 1. / 2 * f
return fval
def extend_gradient(Q, P, M, L, r, mu, sigma, sm, maxdegree, opt_var, small_k):
R = []
cR = []
for k in range(small_k):
cR.append(np.zeros((math.pow(4, k + 1) * L)))
R.append(np.zeros((math.pow(4, k + 1), L)))
Rgradmu = []
Rgradsig = []
Rgradpwm = []
Rgradsm = []
for p in range(len(P)):
D = P[p] - small_k + 1
K = P[p]
Rgradm = []
if opt_var[2] != 0:
Rgradsigm = []
Rgradpwmm = []
Rgradsmm = []
v = []
vgrad = []
vgradsig = []
vgradpwm = []
for m in range(M[p]):
vhh = []
vgradhh = []
vgradsighh = []
vgradpwmhh = []
for msub in range(D):
vh, vgradh, vgradsigh, vgradpwmh = compute_w(L, r[p][m][:, msub:msub + small_k], sigma[p][m], mu[p][m] + msub, small_k)
vhh.append(vh)
vgradhh.append(vgradh)
vgradsighh.append(vgradsigh)
vgradpwmhh.append(vgradpwmh)
v.append(vhh)
vgrad.append(vgradhh)
vgradsig.append(vgradsighh)
vgradpwm.append(vgradpwmhh)
R1, R2, R3, R4 = compute_scoring(P, M, L, r[p], mu[p], sigma[p], maxdegree, v, vgrad, vgradsig, vgradpwm, small_k, sm[p], opt_var, D)
for m in range(M[p]):
Rgradm.append(utils.vec2matrix(sm[p][m]*R2[m], small_k))
if opt_var[2] != 0:
Rgradsigm.append(utils.vec2matrix(sm[p][m]*R3[m], small_k))
if opt_var[1] != 0:
Rgradsmm.append(utils.vec2matrix(R1[m], small_k))
for n in range(4):
for u in range(P[p]):
Rgradpwmm.append(utils.vec2matrix(sm[p][m]*R4[m][n][u], small_k))
cR[small_k - 1] += sm[p][m]*R1[m]
Rgradmu.append(Rgradm)
if opt_var[2] != 0:
Rgradsig.append(Rgradsigm)
Rgradpwm.append(Rgradpwmm)
if opt_var[1] != 0:
Rgradsm.append(Rgradsmm)
for i in range(small_k):
R[i] = utils.vec2matrix(cR[i], i + 1)
gradmu = []
gradsig = []
gradpwm = []
gradsm = []
for d in range(len(P)):
gradmu.append([0.]* M[d])
gradsig.append([0.]* M[d])
gradpwm.append(np.zeros((M[d], 4, P[d])))
gradsm.append([0.]* M[d])
### here is the inner derivative multiplied with the outer derivative
for p in range(len(P)):
for i in range(len(R[small_k - 1])):
for j in range(len(R[small_k - 1][0])):
diff = R[small_k - 1][i][j] - Q[small_k - 1][i][j]
count = 0
for m in range(M[p]):
gradmu[p][m] += (diff * Rgradmu[p][m][i][j])
if opt_var[2] != 0:
gradsig[p][m] += (diff * Rgradsig[p][m][i][j])
if opt_var[1] != 0:
gradsm[p][m] += (diff * Rgradsm[p][m][i][j])
for n in range(4):
for u in range(P[p]):
gradpwm[p][m][n][u] += (diff * Rgradpwm[p][count][i][j])
count += 1
fobj2 = open("pruef_grad1", "wb")
pickle.dump([gradmu, gradsig, gradpwm, gradsm], fobj2)
fobj2.close()
return gradmu, gradsig, gradpwm, gradsm