def transform(image, boxes, labels, dims=(300, 300), do_flip=True, mean=None):
if mean is None:
mean = [0.485, 0.456, 0.406]
new_image = image
new_boxes = boxes
new_labels = labels
new_image = photometric_distort(new_image)
new_image = to_tensor(new_image)
if rand_random() < 0.5:
new_image, new_boxes = expand(new_image, boxes, filler=mean)
new_image, new_boxes, new_labels = random_crop(new_image, new_boxes,
new_labels)
new_image = to_pil_image(new_image)
if do_flip and rand_random() < 0.5:
new_image, new_boxes = flip(new_image, new_boxes)
new_image, new_boxes = resize_and_maybe_transform_coordinates(new_image,
new_boxes,
dims=dims)
return new_image, new_boxes, new_labels
def randomFlipCoin(p):
""" Returns True with the *p* probability. If the *p* is 1.0,
the function will always return True, or if is 0.0, the
function will return always False.
Example:
>>> Util.randomFlipCoin(1.0)
True
:param p: probability, between 0.0 and 1.0
:rtype: True or False
"""
if p == 1.0: return True
if p == 0.0: return False
return True if rand_random() <= p else False
def randomFlipCoin(p):
""" Returns True with the *p* probability. If the *p* is 1.0,
the function will always return True, or if is 0.0, the
function will return always False.
Example:
>>> Util.randomFlipCoin(1.0)
True
:param p: probability, between 0.0 and 1.0
:rtype: True or False
"""
if p == 1.0: return True
if p == 0.0: return False
return rand_random() <= p
def photometric_distort(image):
new_image = image
distortions = [
adjust_brightness, adjust_contrast, adjust_saturation, adjust_hue
]
rand_shuffle(distortions)
for distortion in distortions:
if rand_random() < 0.5:
if distortion.__name__ is 'adjust_hue':
adjust_factor = rand_uniform(-18 / 255., 18 / 255.)
else:
adjust_factor = rand_uniform(0.5, 1.5)
new_image = distortion(new_image, adjust_factor)
return new_image
def G1DListCrossoverRealSBX(genome, **args):
""" Experimental SBX Implementation - Follows the implementation in NSGA-II (Deb, et.al)
Some implementation `reference <http://vision.ucsd.edu/~sagarwal/icannga.pdf>`_.
.. warning:: This crossover method is Data Type Dependent, which means that
must be used for 1D genome of real values.
"""
EPS = Consts.CDefG1DListSBXEPS
# Crossover distribution index
eta_c = Consts.CDefG1DListSBXEtac
gMom = args["mom"]
gDad = args["dad"]
# Get the variable bounds ('gDad' could have been used; but I love Mom:-))
lb = gMom.getParam("rangemin", Consts.CDefRangeMin)
ub = gMom.getParam("rangemax", Consts.CDefRangeMax)
sister = gMom.clone()
brother = gDad.clone()
sister.resetStats()
brother.resetStats()
for i in range(0,len(gMom)):
if math.fabs(gMom[i]-gDad[i]) > EPS:
if gMom[i] > gDad[i]:
#swap
temp = gMom[i]
gMom[i] = gDad[i]
gDad[i] = temp
#random number betwn. 0 & 1
u = rand_random()
beta = 1.0 + 2*(gMom[i] - lb)/(1.0*(gDad[i]-gMom[i]))
alpha = 2.0 - beta**(-(eta_c+1.0))
if u <= (1.0/alpha):
beta_q = (u*alpha)**(1.0/((eta_c + 1.0)*1.0))
else:
beta_q = (1.0/(2.0-u*alpha))**(1.0/(1.0*(eta_c + 1.0)))
brother[i] = 0.5*((gMom[i] + gDad[i]) - beta_q*(gDad[i]-gMom[i]))
beta = 1.0 + 2.0*(ub - gDad[i])/(1.0*(gDad[i]-gMom[i]))
alpha = 2.0 - beta**(-(eta_c+1.0))
if u <= (1.0/alpha):
beta_q = (u*alpha)**(1.0/((eta_c + 1)*1.0))
else:
beta_q = (1.0/(2.0-u*alpha))**(1.0/(1.0*(eta_c + 1.0)))
sister[i] = 0.5*((gMom[i] + gDad[i]) + beta_q*(gDad[i]-gMom[i]))
if brother[i] > ub: brother[i] = ub
if brother[i] < lb: brother[i] = lb
if sister[i] > ub: sister[i] = ub
if sister[i] < lb: sister[i] = lb
if rand_random() > 0.5:
# Swap
temp = sister[i]
sister[i] = brother[i]
brother[i] = temp
else:
sister[i] = gMom[i]
brother[i] = gDad[i]
return (sister, brother)
def G1DListCrossoverRealSBX(genome, **args):
""" Experimental SBX Implementation - Follows the implementation in NSGA-II (Deb, et.al)
Some implementation `reference <http://vision.ucsd.edu/~sagarwal/icannga.pdf>`_.
And another reference to the `Simulated Binary Crossover
<http://www.mitpressjournals.org/doi/abs/10.1162/106365601750190406>`_.
.. warning:: This crossover method is Data Type Dependent, which means that
must be used for 1D genome of real values.
"""
from . import Consts
EPS = Consts.CDefG1DListSBXEPS
# Crossover distribution index
eta_c = Consts.CDefG1DListSBXEtac
gMom = args["mom"]
gDad = args["dad"]
# Get the variable bounds ('gDad' could have been used; but I love Mom:-))
lb = gMom.getParam("rangemin", Consts.CDefRangeMin)
ub = gMom.getParam("rangemax", Consts.CDefRangeMax)
sister = gMom.clone()
brother = gDad.clone()
sister.resetStats()
brother.resetStats()
for i in range(0, len(gMom)):
if math.fabs(gMom[i] - gDad[i]) > EPS:
if gMom[i] > gDad[i]:
# swap
temp = gMom[i]
gMom[i] = gDad[i]
gDad[i] = temp
# random number betwn. 0 & 1
u = rand_random()
beta = 1.0 + 2 * (gMom[i] - lb) / (1.0 * (gDad[i] - gMom[i]))
alpha = 2.0 - beta**(-(eta_c + 1.0))
if u <= (1.0 / alpha):
beta_q = (u * alpha)**(1.0 / ((eta_c + 1.0) * 1.0))
else:
beta_q = (1.0 / (2.0 - u * alpha))**(1.0 / (1.0 *
(eta_c + 1.0)))
brother[i] = 0.5 * ((gMom[i] + gDad[i]) - beta_q *
(gDad[i] - gMom[i]))
beta = 1.0 + 2.0 * (ub - gDad[i]) / (1.0 * (gDad[i] - gMom[i]))
alpha = 2.0 - beta**(-(eta_c + 1.0))
if u <= (1.0 / alpha):
beta_q = (u * alpha)**(1.0 / ((eta_c + 1) * 1.0))
else:
beta_q = (1.0 / (2.0 - u * alpha))**(1.0 / (1.0 *
(eta_c + 1.0)))
sister[i] = 0.5 * ((gMom[i] + gDad[i]) + beta_q *
(gDad[i] - gMom[i]))
if brother[i] > ub:
brother[i] = ub
if brother[i] < lb:
brother[i] = lb
if sister[i] > ub:
sister[i] = ub
if sister[i] < lb:
sister[i] = lb
if rand_random() > 0.5:
# Swap
temp = sister[i]
sister[i] = brother[i]
brother[i] = temp
else:
sister[i] = gMom[i]
brother[i] = gDad[i]
return (sister, brother)
n_in=5
n_out=1
n_hid=1
n_hneu=[3]
n_neu=[n_in]+n_hneu+[n_out]
temp=0
sum=0
to_list=[]
momgenome=G1DConnections.G1DConnections()
dadgenome=G1DConnections.G1DConnections()
for i in xrange(1,len(n_neu)):
sum=sum+n_neu[i-1]+1
for toNode in xrange(sum,sum+n_neu[i]):
to_list=to_list+[toNode]
for fromNode in xrange(temp,sum-1):
momgenome.append((fromNode,toNode,rand_random()))
dadgenome.append((fromNode,toNode,rand_random()))
temp=sum
momgenome.genomeList=np.array(momgenome.genomeList, dtype=[('from','i'),('to','i'),('weight','f')])
dadgenome.genomeList=np.array(dadgenome.genomeList, dtype=[('from','i'),('to','i'),('weight','f')])
f = open('CrossoverResults.txt','w')
f.write( "###### Network setting ######"+'\n')
f.write( "Number of input: "+str(n_in)+'\n')
f.write( "Number of output: "+str(n_out)+'\n')
f.write( "Number of hidden layer: "+str(n_hid)+'\n')
f.write( "Number of neurons for each hidden layer: "+str(n_hneu)+'\n')
f.write( "###############################"+'\n')
f.write( "toNode list: "+str(to_list)+'\n')
f.write("###############################"+'\n')
f.write( " Parents:"+'\n')
f.write("###############################"+'\n')
def G1DListMutatorInteger_my(genome, **args):
#print genome[:-1]
if args["pmut"] <= 0.0: return 0
a=genome.My_get_listdic()
list_dic=copy.deepcopy(a[0])
list_margin=copy.deepcopy(a[1])
totalmoney=a[2]
list_zhonglei=a[3]
all_money=totalmoney
num = len(genome)
g_list=[0]*num
#print genome[0:],1
mutations = args["pmut"] * num
if rand_random()<mutations:
#''''IF':90000,'CU':30000,'RU':10000,'I':3500,'M':1500,'RB':1500,'SR':3000,'TA':2000,'Y':3500'''
first_lic=['CU','IF','RU']
second_lic=['I','M','RB','SR','TA','Y']
while(1):
if all_money<totalmoney*0.4 or all_money<10000:
break
len_cu=len(list_dic['CU'])
len_if=len(list_dic['IF'])
len_ru=len(list_dic['RU'])
if len_cu<=0 and len_if<=0 and len_ru<=0:
break
for i in first_lic:
len_dic=len(list_dic[i])
if len_dic<=0:
continue
if all_money>list_margin[i]:
if len_dic>1:
j=rand_randint(0,len_dic-1)
else:
j=0
money=list_margin[i]
all_lot=int(all_money/money)
lot=rand_randint(0,all_lot)
all_money=all_money-money*lot
n=list_dic[i][j]
g_list[n]=lot
del list_dic[i][j]
while(1):
if all_money<2000:
break
sum=0
for i in second_lic:
sum=sum+len(list_dic[i])
if sum<=0:
break
for i in second_lic:
len_dic=len(list_dic[i])
if len_dic<=0:
continue
if all_money>list_margin[i]:
if len_dic>1:
j=rand_randint(0,len_dic-1)
else:
j=0
money=list_margin[i]
all_lot=int(all_money/money)
lot=rand_randint(0,all_lot)
all_money=all_money-money*lot
n=list_dic[i][j]
g_list[n]=lot
del list_dic[i][j]
sum_l=0
for i in second_lic:
sum_l=sum_l+len(list_dic[i])
if sum==sum_l:
break
genome.genomeList=g_list
return int(mutations)