def hill_climbing(quaternions, bindings, mediana, points, rot_points,
variancia, av):
print "\nHill climbing"
#vamos sempre trabalhar com os quaternioes antigos, e guardar nas estrutras novas!
new_set_quat = np.copy(quaternions)
new_rot_points = np.copy(rot_points)
medStart = np.copy(mediana)
# Rever o valor
to_much = 2
it = 0
accepted = 0
var = np.copy(variancia)
start_time = time.time()
while (it < number_it):
for n_q in range(0, bindings.shape[0]):
quat_mediana = bindings[n_q, 1, 0]
if (quat_mediana < mediana):
# arranjar randons
randoms = np.random.uniform(-1, 1, (number_of_randoms, 4))
# temos de mexer pouco nos quaternioes, por isso temos de somar valores baixos
randoms = np.divide(randoms, value_to_divide)
# alterar os valores pelo numeros random
# nesta posição bindings[n_q,0,0] está o numero do quaternião que está mais proximo
news_quats = np.sum((randoms, quaternions[int(bindings[n_q, 0,
0])]))
# mexemos nos quaterniões, temos de normalizalos
news_quats = qt.normalize_quat(news_quats)
new_points = qt.rotate_points(points, news_quats)
dist = qt.point_dists_mine(rot_points[n_q], new_points)
d = 0
for iz in range(len(dist)):
if (dist[iz] > quat_mediana) and (
dist[iz] < (mediana + to_much)) and (dist[iz] > d):
#guardar os novos pontos
new_rot_points[int(bindings[n_q, 0,
0])] = new_points[iz]
# guardar o novo quaterniao
new_set_quat[int(bindings[n_q, 0, 0])] = news_quats[iz]
d = dist[iz]
# não se guarda a nova distancia pois ela não será usada outra vez
# isto é, não se sabe como a mudança de um quat afetou o resto da distribuição
if stop_at_first == 1:
break
mins, bindings2 = qt.evaluate(new_rot_points)
var_old = np.copy(var)
varTemp = np.var(mins)
me = np.median(mins)
it = it + 1
if ((varTemp - var_old) < 0.):
rot_points = np.copy(new_rot_points)
quaternions = np.copy(new_set_quat)
var = np.copy(varTemp)
accepted = accepted + 1
bindings = np.copy(bindings2)
mediana = np.copy(me)
else:
var = np.copy(var_old)
mins = qt.evaluate_no_bindings(rot_points)
mediana2, av2, var2 = qt.draw_kde(
mins, 'distribution_new_' + str(quaternions.shape[0]) + '.png', 0.75)
print "Variancia antiga: ", variancia, " Variancia nova: ", var2, " diferença: ", var2 - variancia
print "Mediana antiga: ", medStart, " Nova mediana: ", mediana2
print "Media antiga: ", av, " Nova media: ", av2
#print "numero de aceitações: ", accepted
print "elapsed:", time.time() - start_time
return var2 - variancia, accepted, quaternions
number_it = 10
stop_at_first = 0
#print points
vdg = 0
acg = 0
start_time_global = time.time()
for times in range(0, 3):
start_time = time.time()
quats, rots = qt.spread_quaternions(points, number_of_quat,
quaternions_per_set)
print "elapsed:", time.time() - start_time
start_time = time.time()
mins, bindings = qt.evaluate(rots)
print "elapsed eval:", time.time() - start_time
mediana, av, var = qt.draw_kde(
mins, 'distribution_' + str(number_of_quat) + '.png')
print "Mediana: ", mediana
print "Media: ", av
print "Variancia: ", var
vd, a, quat_new = hill_climbing(quats, bindings, mediana, points, rots,
var, av)
if times == 0:
vdg = vd
acg = a
else:
vdg = (vd + vdg) / 2
acg = (a + acg) / 2
#points[ix,:] = dists
#print "number of quaternions: ",len(quats)
return points
if __name__ == '__main__':
ptsOnSphere = np.zeros((10, 3))
#ptsOnSphere = GetPointsEquiAngularlyDistancedOnSphere(500)
#ptsOnSphere = ptsOnSphere / ptsOnSphere.max(axis=0)
ptsOnSphere = spread_points(100, 100)
mins, bind = evaluate(ptsOnSphere)
#print bind
mediana, av, variancia = qt.draw_kde(
mins, 'distribution_' + str(ptsOnSphere.shape[0]) + '_new_dist.png',
np.max(ptsOnSphere) / ptsOnSphere.shape[0])
print mediana
print av
print variancia
if (True):
from numpy import *
import pylab as p
import mpl_toolkits.mplot3d.axes3d as p3
fig = p.figure()
ax = p3.Axes3D(fig)
x_s = []
y_s = []
z_s = []
def hill_climbing(quaternions, bindings, mediana, points, rot_points,
variancia, av):
print "\nHill climbing greedy"
#vamos sempre trabalhar com os quaternioes antigos, e guardar nas estrutras novas!
#new_set_quat = np.copy(quaternions)
new_rot_points = np.copy(rot_points)
mvar = 0
# Rever o valor
to_much = 2
it = 0
accepted = 0
var = np.copy(variancia)
start_time = time.time()
while (it < number_it):
for n_q in range(0, bindings.shape[0]):
quat_mediana = bindings[n_q, 1, 0]
if (quat_mediana < mediana):
# arranjar randons
randoms = np.random.uniform(-1, 1, (number_of_randoms, 4))
# temos de mexer pouco nos quaternioes, por isso temos de somar valores baixos
randoms = np.divide(randoms, value_to_divide)
# alterar os valores pelo numeros random
# nesta posição bindings[n_q,0,0] está o numero do quaternião que está mais proximo
news_quats = np.sum((randoms, quaternions[int(bindings[n_q, 0,
0])]))
# mexemos nos quaterniões, temos de normalizalos
news_quats = qt.normalize_quat(news_quats)
new_points = qt.rotate_points(points, news_quats)
dist = qt.point_dists_mine(rot_points[n_q], new_points)
point_new = 0
new_quat = 0
d = 0
for iz in range(len(dist)):
if (dist[iz] > quat_mediana) and (dist[iz] <
(mediana + to_much) and
(dist[iz] > d)):
#guardar os novos pontos
point_new = new_points[iz]
# guardar o novo quaterniao
new_quat = news_quats[iz]
d = dist[iz]
if stop_at_first == 1:
break
new_rot_points[int(bindings[n_q, 0, 0])] = point_new
# vamos verificar se este movimento teve influencia no sistema
# ou seja, calcular de novo as distancias, os bindings e mudar o quaternião usado
mins, bindings2 = qt.evaluate(new_rot_points)
var_old = np.copy(var)
varTemp = np.var(mins)
me = np.median(mins)
it = it + 1
if ((varTemp - var_old) < 0.):
#print 'new solution'
#print "Variancia antiga: ", var_old, " Variancia nova: ", varTemp, " diferença: ", varTemp-var_old
#não é preciso copiar tudo, só o que mudou
rot_points[int(bindings[n_q, 0, 0])] = point_new
quaternions[int(bindings[n_q, 0, 0])] = new_quat
var = np.copy(varTemp)
accepted = accepted + 1
bindings[n_q] = bindings2[n_q]
mediana = np.copy(me)
if times == 0:
mvar = varTemp - var_old
else:
mvar = ((varTemp - var_old) + mvar) / 2
else:
var = np.copy(var_old)
mins = qt.evaluate_no_bindings(rot_points)
mediana2, av2, var2 = qt.draw_kde(
mins,
'distribution_new_' + str(quaternions.shape[0]) + '_hill_greedy.png',
0.75)
print "Variancia antiga: ", variancia, " Variancia nova: ", var2, " diferença: ", var2 - variancia
#print "numero de aceitações: ", accepted
print "elapsed:", time.time() - start_time
print "media das mudanças: ", mvar
return var2 - variancia, accepted
def hill_climbing(quaternions, bindings, mediana, points, rot_points,
variancia, av):
print "\nHill climbing with memory"
#vamos sempre trabalhar com os quaternioes antigos, e guardar nas estrutras novas!
new_set_quat = np.copy(quaternions)
new_rot_points = np.copy(rot_points)
# REVER ESTE VALOR!
to_much = 2
it = 0
accepted = 0
var = np.copy(variancia)
start_time = time.time()
while (it < number_it):
already_moved = np.zeros(quaternions.shape[0])
for n_q in range(0, bindings.shape[0]):
# só queremos alterar alqueles que estão muito proximos
quat_mediana = bindings[n_q, 1, 0]
#print mediana
if ((quat_mediana < mediana) &
(already_moved[int(bindings[n_q, 0, 0])] == 0)):
# arranjar randons
randoms = np.random.uniform(-1, 1, (number_of_randoms, 4))
# temos de mexer pouco nos quaternioes, por isso temos de somar valores baixos
randoms = np.divide(randoms, value_to_divide)
# alterar os valores pelo numeros random
# nesta posição bindings[n_q,0,0] está o numero do quaternião que está mais proximo
news_quats = np.sum((randoms, quaternions[int(bindings[n_q, 0,
0])]))
# mexemos nos quaterniões, temos de normalizalos
news_quats = qt.normalize_quat(news_quats)
new_points = qt.rotate_points(points, news_quats)
dist = qt.point_dists_mine(rot_points[n_q], new_points)
d = 0
for iz in range(len(dist)):
if (dist[iz] > quat_mediana) and (
dist[iz] < (mediana + to_much)) and (dist[iz] > d):
#guardar os novos pontos
new_rot_points[int(bindings[n_q, 0,
0])] = new_points[iz]
#guardar o novo quaterniao
new_set_quat[int(bindings[n_q, 0, 0])] = news_quats[iz]
d = dist[iz]
if stop_at_first == 1:
break
already_moved[int(bindings[n_q, 0, 0])] = 1
mins, bindings2 = qt.evaluate(new_rot_points)
var_old = np.copy(var)
varTemp = np.var(mins)
me = np.median(mins)
it = it + 1
if ((varTemp - var_old) < 0.):
rot_points = np.copy(new_rot_points)
quaternions = np.copy(new_set_quat)
var = np.copy(varTemp)
accepted = accepted + 1
bindings = np.copy(bindings2)
mediana = np.copy(me)
else:
var = np.copy(var_old)
mins = qt.evaluate_no_bindings(rot_points)
mediana2, av2, var2 = qt.draw_kde(
mins,
'distribution_new_' + str(quaternions.shape[0]) + '_hill_memory.png',
0.75)
print "Variancia antiga: ", variancia, " Variancia nova: ", var2, " diferença: ", var2 - variancia
print "numero de aceitações: ", accepted
print "elapsed:", time.time() - start_time
return var2 - variancia, accepted
stop_at_first = 0
#print points
vdg = 0
acg = 0
start_time_global = time.time()
for times in range(0, 1):
start_time = time.time()
quats, rots = qt.spread_quaternions(points, number_of_quat,
quaternions_per_set)
print "elapsed:", time.time() - start_time
start_time = time.time()
mins, bindings = qt.evaluate(rots)
print "elapsed eval:", time.time() - start_time
mediana, av, var = qt.draw_kde(
mins,
'distribution_' + str(number_of_quat) + '_hill_all_to_median.png')
print "Mediana: ", mediana
print "Media: ", av
print "Variancia: ", var
vd, a, mediana2, av2, var2, quat_new = hill_climbing(
quats, bindings, mediana, points, rots, var, av)
if times == 0:
vdg = vd
acg = a
else:
vdg = (vd + vdg) / 2
acg = (a + acg) / 2