def sample(self, n_samples: int, space: Space, **kwargs):
kwargs = kwargs['kwargs']
if (self.cached_n_samples is not None
and self.cached_n_samples == n_samples
and self.cached_space is not None
and space == self.cached_space and self.cached_algo is not None
and self.cached_algo == kwargs['sample_algo']):
#lhs = lhsmdu.resample() # YC: this can fall into too clustered samples if there are many constraints
lhs = lhsmdu.sample(len(space), n_samples)
else:
self.cached_n_samples = n_samples
self.cached_space = space
self.cached_algo = kwargs['sample_algo']
if (kwargs['sample_algo'] == 'LHS-MDU'):
lhs = lhsmdu.sample(len(space), n_samples)
elif (kwargs['sample_algo'] == 'MCS'):
lhs = lhsmdu.createRandomStandardUniformMatrix(
len(space), n_samples)
else:
raise Excepetion(f"Unknown algorithm {kwargs['sample_algo']}")
lhs = np.array(
list(zip(*[np.array(lhs[k])[0] for k in range(len(lhs))])))
# print(lhs,'normalized',n_samples)
return lhs
def LHS_for_guesses(sample_count):
# we use latin hypercube sampling to obtain initial guesses for curve fitting
lhsmdu.setRandomSeed(None)
alpha_tau_sample = lhsmdu.sample(2,sample_count)
epsilon_mu_sample = lhsmdu.sample(2,sample_count)
alpha_tau_sample = alpha_tau_sample.tolist()
epsilon_mu_sample = epsilon_mu_sample.tolist()
# we then adjust the variables to the correct ranges
adjusted_sample = []
# for AT, we adjust to between 1 and 1/30
for var_dist in alpha_tau_sample:
var_dist = [(1 + var*(1/30-1)) for var in var_dist]
adjusted_sample.append(var_dist)
# for EM, we adjust to between 10 and 0.0001
# however, we actually use theta where EM = 10^theta, so theta is between 1 and -3
# prevents overweighting towards the top end of the spectrum
for var_dist in epsilon_mu_sample:
var_dist = [(1 + var*(-5-1)) for var in var_dist]
adjusted_sample.append(var_dist)
# then, for each pair of 4 variables, we run it through the fit and determine cost
# for every guess, we check to see if that's the lowest cost generated thus far
# if it is, we store it, and at the end, that's our result
lowest_cost = [10000000000000, [0,0,0,0]]
for i in range(sample_count):
test_guesses = []
for var_dist in adjusted_sample:
test_guesses.append(var_dist[i])
try:
# we have to rearrange test_guesses so that it goes alpha, epsilon, tau, mu
tau = test_guesses[1]
test_guesses[1] = test_guesses[2]
test_guesses[2] = tau
# we then adjust epsilon and mu to be 10^[their value], because it's currently theta
test_guesses[1] = 10**(test_guesses[1])
test_guesses[3] = 10**(test_guesses[3])
cost = SD_curve_fit(test_guesses).cost
print(f"{test_guesses} cost: {cost}")
if cost < lowest_cost[0]:
lowest_cost = [cost, test_guesses]
except OverflowError as e:
print("Overflow while running LHS for initial guesses: ", e)
except ValueError as e:
print("Residual error while running LHS for initial guesses: ", e)
print(f"LHS suggests that {lowest_cost[1]} is the best set of guesses")
return lowest_cost[1]
def get_rate_samples(self):
"""
Note: return type is matrix, access with rates[i, j] NOT rates[i][j]
"""
if self._algorithm == 'lhs':
rates = lhsmdu.sample(len(self._option_range_list),
self._sample_count)
return rates
if self._algorithm == 'mcs':
rates = lhsmdu.createRandomStandardUniformMatrix(
len(self._option_range_list), self._sample_count)
return rates
rates = lhsmdu.sample(len(self._option_range_list), self._sample_count)
return rates
def _initialise_population_latin_hypercube(self):
"""
Use a latin hypercube to seed the population of solutions.
:return: a list of Candidate objects (with no assigned fitnesses)
"""
k = lhsmdu.sample(len(self.solution_dimensions),
self.population_size) # returns numpy matrix
# scale the design for each dimension onto the correct range
for i, dim in enumerate(self.solution_dimensions):
dimension_range = dim.max_value - dim.min_value
k[i] = (k[i] * dimension_range
) + dim.min_value # operates on entire slice at once
# bind each dimension's putative value to valid intervals (granularity)
for i, dim in enumerate(self.solution_dimensions):
for c in range(self.population_size):
k[i, c] = dim.bind(k[i, c])
population = [] # a list of lists, each inner list being a candidate
# convert into format needed, a list of lists.
for c in range(self.population_size):
candidate_solution = [
k[d, c] for d in range(len(self.solution_dimensions))
]
candidate = Candidate(solution=candidate_solution)
population.append(candidate)
return population
def initialization_run(domain, num_DIM, num_init):
start_points_norm = lhsmdu.sample(num_DIM, num_init)
start_points_norm = start_points_norm.transpose()
start_points = np.zeros([num_init, num_DIM])
bounds = np.array(
[[domain[0].get('domain')[0], domain[0].get('domain')[1]],
[domain[1].get('domain')[0], domain[1].get('domain')[1]],
[domain[2].get('domain')[0], domain[2].get('domain')[1]],
[domain[3].get('domain')[0], domain[3].get('domain')[1]],
[domain[4].get('domain')[0], domain[4].get('domain')[1]],
[domain[5].get('domain')[0], domain[5].get('domain')[1]],
[domain[6].get('domain')[0], domain[6].get('domain')[1]],
[domain[7].get('domain')[0], domain[7].get('domain')[1]],
[domain[8].get('domain')[0], domain[8].get('domain')[1]],
[domain[9].get('domain')[0], domain[9].get('domain')[1]],
[domain[10].get('domain')[0], domain[10].get('domain')[1]],
[domain[11].get('domain')[0], domain[11].get('domain')[1]],
[domain[12].get('domain')[0], domain[12].get('domain')[1]]])
for i in range(num_init):
for j in range(num_DIM):
start_points[i, j] = bounds[
j, 0] + start_points_norm[i, j] * (bounds[j, 1] - bounds[j, 0])
return start_points
def lhs_mu_samples(self, seed=None):
seed = np.random.randint(99999) if seed is None else seed # set seed
samples = lhsmdu.sample(
self.d_enc, self.n_comp,
randomSeed=seed).A.T # get (n, d) array of samples
samples = samples * (self.mu_init_range[1] - self.mu_init_range[0]
) + self.mu_init_range[0] # scale up from [0,1]
return samples
def problem_2in1out( xlower = 0, xupper = 10, nsamples = 10, ylower = 0, yupper = 20 ):
'''n_dim_problem is a function to create a random, but well defined, problem
landscape created by generating random points in a 3d landscape and
fitting a gp to the points. By constraining the hyperparameters we are able
to produce a landscape with a controllable level of complexity.
example : my_problem_function = three_dim_problem( )
INPUTS: (keywords, not required)
xlower - Integer/Float - Defines the lower bound of all X's
xupper - Integer/Float - Defines the upper bound of all X's
nsamples - Integer - number of points along each dimension
ylower - Integer/Float - Defines the lower bound of Y
yupper - Integer/Float - Defines the upper bound of Y
OUTPUT:
fun - Function - A 3 dimensional function
example : yi = fun( [ xi1 , xi2 ] )
example : [yi,yj] = func( [ [ xi1 , xi2 ] , [ xj1 , xj2 ] )
'''
X1 = np.linspace( xlower , xupper , nsamples )
X2 = np.linspace( xlower , xupper , nsamples )
Y = np.ones( nsamples**2 )*np.random.uniform( ylower , yupper , nsamples**2 )
kernel = GPy.kern.RBF( input_dim = 2 , variance = 0.001 , lengthscale = (xupper-xlower)/5 )
X = np.array (list( it.product( X1 , X2 ) ) )
X = np.array(lhsmdu.sample(nsamples**2,2))*10
X = X.reshape( -1 , 2 )
Y = Y.reshape( -1 , 1 )
model = GPy.models.GPRegression( X, Y, noise_var = 1e-4 , kernel = kernel )
def fun(x):
assert( type(x) == list or type(x) == np.array , 'input must be a list or array')
assert(len(x) >= 2 , 'Input is in incorrect format')
value = None
try:
value = [ float( model.predict( np.array( [ [ i ] ] ).reshape( -1 , 2 ) )[ 0 ] ) for i in x ]
if value < ylower:
value = ylower
if value > yupper:
value = yupper
return( np.array(value) )
except:
pass
try:
value = float( model.predict( np.array( [ [ x ] ] ).reshape( -1 , 2 ) )[ 0 ] )
if value < ylower:
value = ylower
if value > yupper:
value = yupper
return( np.array(value) )
except:
print( '2in_1out problem ERROR - The x values do not match the required input size' )
return( None )
return( fun )
def random_lhs(min_bound, max_bound, dimensions, number_samples=1):
import lhsmdu
r_range = abs(max_bound - min_bound)
lolo = r_range * np.array(lhsmdu.sample(dimensions,
number_samples)) + min_bound
required_list = []
for ii in range(number_samples):
required_list.append(lolo[:, ii].reshape(-1))
return required_list
def genrandom(n_random):
### Use latin HyperCube sampling to choose intial points for nelder-mead search
random_points = np.asarray(
lhsmdu.sample(200 * n_random, len(
mp.domain.valid_ranges))) #Lhsmdu produce normalised numbers
# The next few lines makes the points fit to the search domain
mults = [vrange[1] - vrange[0] for vrange in mp.domain.valid_ranges]
rangemin = [vrange[0] for vrange in mp.domain.valid_ranges]
random_points = random_points * mults + rangemin
return (np.array(random_points).reshape(-1, len(mp.example)))
def __init__(self, nn, eps=0, train_eps=False, **kwargs):
self.aggr = 'laf'
seed = 92
atype = 'frac'
shared = True
embed_dim = '32'
if 'aggr' in kwargs.keys():
aggr = kwargs['aggr']
if 'seed' in kwargs.keys():
seed = kwargs['seed']
del kwargs['seed']
if 'style' in kwargs.keys():
atype = kwargs['style']
del kwargs['style']
if 'shared' in kwargs.keys():
shared = kwargs['shared']
del kwargs['shared']
if 'embed_dim' in kwargs.keys():
embed_dim = int(kwargs['embed_dim'])
del kwargs['embed_dim']
super(GINLafConv, self).__init__(aggr='add', **kwargs)
self.nn = nn
self.initial_eps = eps
if train_eps:
self.eps = torch.nn.Parameter(torch.Tensor([eps]))
else:
self.register_buffer('eps', torch.Tensor([eps]))
self.reset_parameters()
if shared:
params = torch.Tensor(lhsmdu.sample(13, 1, randomSeed=seed))
else:
params = torch.Tensor(lhsmdu.sample(13, 1, randomSeed=seed))
#params = torch.Tensor(lhsmdu.sample(13, 1, randomSeed=seed))
#par = torch.Tensor([[1, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1, 0, 0]] * out_channels)
#par = torch.Tensor([[1, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1, 0, 0]])
#params = par.t()
if atype == 'minus':
self.aggregation = ElementAggregationLayer(parameters=params)
elif atype == 'frac':
#self.aggregation = FractionalElementAggregationLayer(parameters=params)
self.aggregation = ScatterAggregationLayer(parameters=params)
def pertubate(point, beta, n=100):
### Use latin HyperCube sampling to choose intial points for nelder-mead search
print(beta)
print(n)
random_points = np.asarray(lhsmdu.sample(n, len(
mp.domain.valid_ranges))) #Lhsmdu produce normalised numbers
# The next few lines makes the points fit to the search domain
mins = conformpoint(np.array([p - beta for p in point]))
maxs = conformpoint(np.array([p + beta for p in point]))
mults = [maxs[i] - mins[i] for i in range(len(maxs))]
random_points = [
random_point * mults + mins for random_point in random_points
]
random_points = [list(conformpoint(point)) for point in random_points]
return (random_points)
def initial_evaluations(self, n_samples, x_dims, limits):
"""
Makes the initial evaluations of the parameter space and
evaluates the objective function at these locaitons
:param n_samples: number of initial samples to make
"""
np.random.seed(self.seed)
x = np.array(lhsmdu.sample(x_dims, n_samples, randomSeed=self.seed)).T\
* (limits[1]-limits[0])+limits[0]
# try:
y = np.array([self.objective_function(xi, *self.of_args).flatten() for xi in x])
# except :
# y = self.objective_function(x)
# update evaluation number
self.n_evaluations += n_samples
return x, y
def generate_local_data_lhs(y_sampler,
n_samples_per_dim,
step,
current_psi,
x_dim=1,
n_samples=2):
xs = y_sampler.x_dist.sample(
torch.Size([n_samples_per_dim * n_samples, x_dim])) # .to(device)
mus = torch.empty((len(xs), len(current_psi))) # .to(device)
mus = torch.tensor(
lhsmdu.sample(
len(current_psi), n_samples,
randomSeed=np.random.randint(1e5)).T).float() # .to(device)
mus = step * (mus * 2 - 1) + current_psi # .to(device)
mus = mus.repeat(1, n_samples_per_dim).reshape(-1, len(current_psi))
y_sampler.make_condition_sample({'mu': mus, 'X': xs})
data = y_sampler.condition_sample().detach() # .to(device)
return data.reshape(-1, 1), torch.cat([mus, xs], dim=1)
def init_weights(self, initialisation, nb_init_points):
print_sep("New Initialization")
nb_var = self.neuralnet.count_params()
if (initialisation == 0):
loc = 0
scale = 0
initial_weights = np.random.normal(loc=loc,
scale=scale,
size=(nb_var, nb_init_points))
if (initialisation == 1):
loc = 0
scale = 1
initial_weights = np.random.normal(loc=loc,
scale=scale,
size=(nb_var, nb_init_points))
if (initialisation == 2):
loc = (np.random.rand() - 0.5) * 10
scale = (np.random.rand()) * 10
initial_weights = np.random.normal(loc=loc,
scale=scale,
size=(nb_var, nb_init_points))
if (initialisation == 3):
loc = 0
scale = 10
initial_weights = np.array(
(lhsmdu.sample(nb_var, nb_init_points) - (0.5 + loc)) * scale)
if (initialisation in [0, 1, 2]):
print(
"Weights've been initialized on a normal distribution of mean=%.3f and scale=%.3f."
% (loc, scale))
else:
print(
"Weights've been initialized inside a LH of size=%.3f and center=%.3f."
% (scale, loc))
to_return = []
for i in initial_weights.T:
i = np.reshape(i, (-1, 1))
i = to_weights_shape(i, self.neuralnet.get_weights())
to_return += [i]
return to_return
def generate_LHS_draw(seed):
# Dictionary of stochastic settings
ST = {}
# hard code the number of stochastic settings here
# Number depends on stochastic cumulus scheme selected
ns = 20
nss = iter(range(ns))
LHS = sample(ns, nens, randomSeed=seed)
#### MICROPHYSICS #### (5)
# Some of these distrubtions could be beta like Hacker et al 2011
ST['nssl_ehw0'] = denormalise(LHS[next(nss), :], 0.4, 1.0)
ST['nssl_alphar'] = denormalise(LHS[next(nss), :], 0.0, 2.5)
ST['nssl_alphah'] = denormalise(LHS[next(nss), :], 0.0, 3.0)
ST['nssl_cccn'] = denormalise(LHS[next(nss), :], 0.3e9, 1.3e9)
# Hail collect eff. should be higher than graupel
# ehlw0 must be computed last because of constraint
ehlw0 = N.zeros([nens])
ehln = next(nss)
for xn, x in enumerate(ST['nssl_ehw0'][0, :].A1):
ehlw0[xn] = denormalise(LHS[ehln, xn], ST['nssl_ehw0'][0, xn], 1.0)
ST['nssl_ehlw0'] = N.matrix(ehlw0)
#### SHORTWAVE #### (0)
# No solar scattering any more
# swscat = denormalise(LHS[4,:],0.2,2.0)
##### CUMULUS ##### ( )
##### LAND SURFACE ##### (8)
# Radii - must be square root for uniform distribution round circle
for key in ('morphr_crs', 'morphr_btr', 'morphr_rad', 'morphr_tbot'):
ST[key] = N.sqrt(denormalise(LHS[next(nss), :], 0.0, 1.0))
# Angles - TODO: Must be square root!
for key in ('morphth_crs', 'morphth_btr', 'morphth_rad', 'morphth_tbot'):
ST[key] = denormalise(LHS[next(nss), :], 0.0, 2 * N.pi)
return ST
A dirty script to create our LHS input for a mass ML training dataset using CLFO data
'''
########## READ DATA
ignorecontinue = True
ignore = ['CO2','PXYL+MXYL','NO2','NOY','R','RO2','M','HO2','CH3O2','OH','TRICLETH','HONO','BUT2CHO', 'C3ME3CHO', 'C5H11CHO', 'CH2CL2', 'LIMONENE', 'MACR', 'MVK']
df = np.log10(pd.read_csv('../src/examples/clfo_describe.csv',index_col=0))
df = df[filter(lambda x: x not in ignore, df.columns)]
nruns = 1000#df.shape[1]**2
print(nruns)
# may be slow.
lhs = lhsmdu.sample(df.shape[1],nruns)
#lhs = lhsmdu.createRandomStandardUniformMatrix(df.shape[1],nruns)#monte carlo
np.save('lhs.npy',lhs)
print 'lhs ready'
#### END LHS
#### set ics BASE
'''
If using file use new ics, we are just creating a new one
'''
##ics = newics.newics(h5file ='clfo.h5',timestep=int(25*60*60))
ics = pd.DataFrame(
[
['ii', 'TIME', '0',str(24*60*60*5)],
def latin_hypercube(dic_factors, runs):
"""
Parameters:
dic_factors: The dictionary of factors to be included in the Latin Hypercube design.
runs: The number of runs to be used in the design.
Returns:
df: The dataframe containing the Latin Hypercube design.
Example:
>>> import design
>>> Factors = {'Height':[1.6,2],'Width':[0.2,0.4],'Depth':[0.2,0.3],'Temp':[10,20],'Pressure':[100,200]}
>>> design.latin_hypercube(Factors,50)
Height Width Depth Temp Pressure
0 1.814372 0.316126 0.203734 12.633408 150.994350
1 1.683852 0.327745 0.221157 10.833524 149.235694
2 1.952938 0.220208 0.212877 14.207334 177.737810
3 1.921001 0.306165 0.249451 13.747280 195.141219
4 1.709485 0.286836 0.214973 12.132761 144.060774
5 1.795442 0.339484 0.263747 16.494926 105.861897
6 1.849604 0.390856 0.229801 17.768834 157.379054
7 1.635933 0.295207 0.244843 15.561134 119.353027
8 1.800514 0.257358 0.232554 19.117071 114.431350
9 1.748656 0.311259 0.209185 19.573654 147.317771
10 1.610152 0.200320 0.269825 14.041168 192.787729
11 1.670380 0.283579 0.270421 11.422384 161.302466
12 1.914483 0.374190 0.273246 15.253950 110.213186
13 1.731642 0.363269 0.211263 15.011417 175.315691
14 1.864093 0.245809 0.235466 10.506234 123.998827
15 1.856580 0.314574 0.260263 11.787321 152.096424
16 1.651140 0.262106 0.289432 14.407869 121.954348
17 1.827840 0.278926 0.223818 12.824422 168.813816
18 1.780800 0.380327 0.252359 12.290440 171.741507
19 1.762333 0.224241 0.216475 18.386775 165.564771
20 1.949560 0.300988 0.285943 10.063231 155.134033
21 1.646881 0.248638 0.250362 16.701447 163.476898
22 1.974239 0.379487 0.279709 17.208315 181.757031
23 1.904317 0.216877 0.292985 18.829669 136.808281
24 1.899844 0.343903 0.230494 13.197326 198.654066
25 1.696839 0.329348 0.283741 18.193024 135.335187
26 1.689936 0.272728 0.218891 19.800988 131.615692
27 1.823893 0.299159 0.247030 10.790362 191.524570
28 1.841140 0.210635 0.286718 10.327824 167.595627
29 1.883991 0.385993 0.277186 18.773584 178.871167
30 1.932945 0.358221 0.294327 16.890948 125.635668
31 1.837620 0.370877 0.242782 17.103119 142.240418
32 1.740477 0.352914 0.265939 14.697769 129.088978
33 1.624078 0.347985 0.298516 13.933373 132.011517
34 1.786612 0.351899 0.225313 15.827930 188.649172
35 1.892142 0.206601 0.254650 14.805995 138.732923
36 1.656703 0.252798 0.205547 18.461586 184.792345
37 1.770805 0.270721 0.226262 11.940936 113.390934
38 1.672266 0.288289 0.275940 15.640371 186.777116
39 1.600629 0.240123 0.280908 17.934686 126.897387
40 1.995175 0.237031 0.240472 16.393982 116.475088
41 1.713062 0.265850 0.256147 17.418780 172.746504
42 1.964540 0.235473 0.266340 11.334520 196.454539
43 1.757516 0.366909 0.207040 13.488750 102.146392
44 1.942405 0.214971 0.290674 13.373628 109.206897
45 1.985601 0.229702 0.297658 12.435430 101.336426
46 1.617340 0.321384 0.200862 19.338525 159.238981
47 1.976837 0.393484 0.258497 16.167623 140.926988
48 1.877091 0.399951 0.239234 19.788923 182.759572
49 1.725652 0.332160 0.237414 11.136650 107.726667
"""
df = pd.DataFrame()
factor_names = []
count = 0
# Creates an array filled with a latin hypercube form 0 to 1
array = lhsmdu.sample(len(dic_factors), runs)
# This for loop converts the latin hypercube to have the levels entered into the dictionary of factors
for name in dic_factors:
factor_names.append(name)
low = min(dic_factors[name])
high = max(dic_factors[name])
# non_coded stored the array being mapped to fit the levels of the factors entered
non_coded = np.array(
list(map(lambda x: low + ((high - low) * x), array[count])))
# Converts non_coded (which is currently one column of the final dataframe) to a series
s_add = pd.Series(non_coded[0][0])
count += 1
# Adds the series to the dataframe
df = pd.concat([df, s_add], ignore_index=True, axis=1)
df = df.rename(columns=lambda y: factor_names[y])
return df
import lhsmdu lhsmdu.setRandomSeed(None) samples = lhsmdu.sample(1, 100) # initially, it is generated between 0 and 1 # we need to change it to a distribution between -100 and 100 adj = [] samples = samples.tolist() print(samples[0])
def runtime(numDimensions, numSamples):
''' Checks runtime using standard variables '''
start_time = time()
m = lhsmdu.sample(numDimensions, numSamples)
end_time = time()
print end_time - start_time
domain_flag = 'parsec' #Change this to change your domain
#######
if domain_flag == 'ffd':
sys.path.insert(0, cur_dir+'/domain/FFD_config')
if domain_flag == 'rastrigin':
sys.path.insert(0, cur_dir+'/domain/Rastrigin_config')
if domain_flag == 'simple2d':
sys.path.insert(0, cur_dir+'/domain/simple2d_config')
if domain_flag == 'parsec':
sys.path.insert(0, cur_dir+'/domain/Parsec_config')
if domain_flag == 'rosenbrock6d':
sys.path.insert(0, cur_dir+'/domain/Rosenbrock6d_config')
import domain_config as mp
import lhsmdu
import numpy as np
lhsx = lhsmdu.sample(10,1e6)
x = np.array(lhsx.T)
sum = 0
for xx in x:
sum = np.nansum(sum, fitness_fun(xx))
mean = sum/1e6
print('mean is ', mean)
# to convert lists to dictionary
parameters_dictionary = {}
for key in list_of_parameters:
for value in list_of_values:
parameters_dictionary[key] = value
list_of_values.remove(value)
break
# print(parameters_dictionary)
while k = 1:
try:
no_of_sample = int(input("enter number of samples needed")
k = 2
except:
print("Enter the value again")
k = 1
df = pd.DataFrame(index=list(range(no_of_sample)), columns=list_of_parameters)
z = len(df.columns)
for val in range(z):
k = lhsmdu.sample(1, 150)
k = (k*(list_of_maximum_values[val]-list_of_minimum_values[val]))+list_of_minimum_values[val]
k = k.transpose()
df_temp= pd.DataFrame(k)
df.iloc[:,val] = df_temp
df.to_excel("latin_hypercube_sampling.xlsx")
@author: WDN
Prior Data Generator
加入LHS采样方法，对空间中的数据进行采样，同时进行仿真，得到先验数据
"""
import WDNfeedback#反馈
import pandas as pd
import numpy as np
import lhsmdu
import math
pd.set_option('display.max_rows',None)
"利用LHS在空间中采样，这里用二维采样============================================="
k = lhsmdu.sample(6, 120) # Latin Hypercube Sampling with multi-dimensional uniformit
-"VBR其他流---------------------------------------------------------------------"
x1 = np.array(k[0])*100
x1=np.floor(x1)
vbrinterval=x1[0].tolist()
vbrinterval = [ math.ceil(x) for x in vbrinterval ]
vbrinterval
vbrinterval = [ x+1 for x in vbrinterval ]
vbrinterval
x1 = np.array(k[1])*64000
x1=np.floor(x1)
vbrsize=x1[0].tolist()
vbrsize = [ math.ceil(x) for x in vbrsize ]
vbrsize = [ x+1 for x in vbrsize ]
"Superapp视频流----------------------------------------------------------------"
x1 = np.array(k[2])*100
def main(N_parts, f_simulate, f_summaries, f_discrepancy, target_data,
params_mins, params_maxs, scale_param, desired_D, mesh, max_time):
#import numpy as np
import lhsmdu
from multiprocessing import Pool, cpu_count, get_context
import time
from functools import partial
import psutil
import time
from tqdm import tqdm
'''
%%%%%%%%%%%%%%%% VARIABLE INPUT %%%%%%%%%%%%%%%%%%%%%%%%%
'''
#Default options (can be changed here or adjusted on the fly by supplying options argument)
#resample_weighting = 'weighted';
# 'equal' - All kept particles are weighted equally for resampling steps
# 'weighted' - Particles are weighted according to their discrepancy values when resampling (taking
# worst kept particle as three sigma in a normal distribution, and then normalising
# weights so they sum to one)
#metric_weighting = 'variance';
# 'mahalanobis' - Mahalanobis distance is used for calculations of discrepancy
# 'variance' - Metrics are scaled according to their variance, but covariances are ignored
# 'none' - Euclidean distance is used
# (All options still include rescaling of angle metrics with respect to eccentricity)
keep_fraction = 0.5 #0.75;
# The proportion of particles to keep during each resample step (recommended ~0.5 as a starting value)
if N_parts < 100:
max_MCMC_steps = 50
else:
max_MCMC_steps = 300
# The maximum number of 'jiggle' steps applied to all particles in the attempt to find unique locations
verbose = 1
#Output information about particle uniqueness and discrepancy targets
'''
%%%%%%%%%%%%%%%% INSTANTIATE%%%%%%%%%%%%%%%%%%%%%%%%%
'''
process_count = cpu_count()
print('cpus: %d / %d' % (process_count, cpu_count()))
max_runtime = time.time() + max_time
print('beginning pool...')
start = time.time()
ESStarget = 0.8
burnidx = 50 #disabled
#5 #number of iterations to burn-in (use full R)
#generate and warm up pool
pool = get_context("spawn").Pool(processes=process_count,
maxtasksperchild=1)
time.sleep(1)
end = time.time()
print('time elapsed: ' + str(end - start))
print('warming up...')
start = time.time()
pool.map(warmup, range(process_count), chunksize=1)
time.sleep(1)
end = time.time()
print('time elapsed: ' + str(end - start))
print('parent precent memory used: ' + str(psutil.virtual_memory()[2]))
# Calculate the summaries for the target data
print('calculating target summaries...')
start = time.time()
target_summaries = np.asarray(f_summaries(target_data), dtype='f8')
end = time.time()
#print('calculated, time elapsed: '+str(end-start))
# Calculate the ordinal location of the last particle kept for convenience
worst_keep = int(np.rint(N_parts * keep_fraction) + 1)
# Read out the number of variables
N_theta = len(params_mins)
# Convert scaling parameters to log equivalents. The set of transformed
# parameters is termed 'theta' in this code
#print('scaling...')
scale_param = scale_param == 1
theta_mins = np.copy(params_mins).astype('f8')
theta_mins[scale_param] = np.log(theta_mins[scale_param])
theta_maxs = np.copy(params_maxs).astype('f8')
theta_maxs[scale_param] = np.log(theta_maxs[scale_param])
# Initialise particles using Latin Hypercube Sampling
#print('generating lhsmdu...')
part_lhs = lhsmdu.sample(N_parts, N_theta)
part_thetas = theta_mins + np.multiply(part_lhs, (theta_maxs - theta_mins))
# For simulation, convert scaling parameters back to true values
part_params = np.copy(part_thetas).astype('f8')
part_params[:, scale_param] = np.exp(part_params[:, scale_param])
# Load seed data (generated by a separate file)
#print('loading seed data...')
seedinfo = np.load("seedinfo.npy", allow_pickle=True)
Pt = np.asarray(seedinfo.item().get('permute_tables'), dtype='f8')
Ot = np.asarray(seedinfo.item().get('offset_tables'), dtype='f8')
#Count number of seeds created
N_seeds = np.shape(Pt)[0]
'''
%%%%%%%%%%%%%%%% PARPOOL 1: FIRST MCMC STEP %%%%%%%%%%%%%%%%%%%%%%%%%
'''
# Generate the seed information for each particle, then simulate the model
# using this seed information, storing both its output and the associated
# summary statistics
#Create an individual simulator object for this praticle, using the
# seed information
print('creating savespace...')
seed_nums = np.random.choice(a=range(N_seeds), size=N_parts)
part_outputs = np.zeros((N_parts, mesh['Nx'], mesh['Ny'])).astype('f8')
part_summaries = np.zeros((N_parts, 27)).astype('f8')
# Run the simulator and store summaries
time.sleep(5)
print('creating partial...')
#runSimLoop = partial(runSim, part_params=part_params, Pt=Pt, Ot=Ot, f_simulate=f_simulate, f_summaries=f_summaries, seed_nums=seed_nums)
iterable = []
for k in range(N_parts):
iterable.append((k, part_params[k, :], Pt[seed_nums[k], :, :],
Ot[seed_nums[k], :, :], f_simulate, f_summaries))
time.sleep(5)
#runSim_lambda = lambda k: runSim(k,part_params[k,:],Pt[seed_nums[k],:,:],Ot[seed_nums[k],:,:],f_simulate,f_summaries)
print('parent precent memory used: ' + str(psutil.virtual_memory()[2]))
print('beginning map...')
start_loop = time.time()
#results = pool.map(runSimLoop,range(N_parts), chunksize=1)
#results = []
#for i,k,part_out, part_sum in tqdm(enumerate(pool.imap_unordered(runSimLoop,range(N_parts),chunksize=int(N_parts/process_count)))):
#results.append(res)
chunksize = int(N_parts / process_count)
if chunksize < 1:
chunksize = 1
results = pool.starmap_async(runSim, iterable, chunksize=chunksize)
results_get = results.get()
for r in results_get:
k = r[0]
part_outputs[k, :, :] = r[1]
part_summaries[k, :] = r[2]
#pbar.update()
del iterable
print('time elapsed: ' + str(time.time() - start_loop))
print('parent precent memory used: ' + str(psutil.virtual_memory()[2]) +
'\n')
'''
%%%%%%%%%%%%%%%% WEIGHTING/RESAMPLING %%%%%%%%%%%%%%%%%%%%%%%%%
'''
# Calculate sample covariance matrix from the initial particles
C_S = np.cov(np.transpose(part_summaries))
# Use diagonals of this to extract only variances (no covariances)
C_S = np.diag(np.diag(C_S))
# Store the inverse, so that it only need be calculated once
invC_S = np.linalg.inv(C_S)
# Use discrepancy function with this inverse matrix included
f_discrep = partial(f_discrepancy,
target_metrics_old=target_summaries,
invC=invC_S)
# Calculate discrepancies for each particle based on this distance measure
part_Ds = np.asarray(f_discrepancy(part_summaries, target_summaries,
invC_S),
dtype='f8')
# Loop until stopping criteria is hit
looping = 1
testitcount = 1
while looping:
print('Iteration count: %g' % (testitcount, ))
if testitcount > burnidx:
burnin = 1
else:
burnin = 0
# First, sort all the particles according to their discrepancy
ordering = np.argsort(part_Ds)
part_Ds = part_Ds[ordering]
part_outputs = part_outputs[ordering, :, :]
part_thetas = part_thetas[ordering, :]
part_summaries = part_summaries[ordering, :]
# Now select the target discrepancy as the worst particle of the kept
# fraction
target_D = np.copy(part_Ds[worst_keep])
# Select which particles, from those that are kept, to resample the
# discarded particles onto.
# Calculate weights for each particle. These are found by
# assuming that the worst kept particle is three standard
# deviations away from D = 0. Due to normalisation of weights,
# which here takes place inside randsample, even if all
# particles are far from D = 0 it will not be an issue.
part_logweights = -9 * (part_Ds[:worst_keep]**
2) / part_Ds[worst_keep]**2
part_logweights = part_logweights + part_logweights.max()
#Normalise so largest weight is 1 (0 on log scale)
#part_weights = np.exp(part_logweights)
#part_weights = part_weights / np.sum(part_weights)
# Tempering and Normalizing
#via https://arxiv.org/pdf/1805.03317.pdf section 2.2
BetaSet = np.linspace(0, 1, num=100)
part_weights = np.exp(part_logweights)
weights = np.tile(part_weights, (len(BetaSet), 1))
weights = np.divide(
np.power(weights, BetaSet[:, None]),
np.sum(np.power(weights, BetaSet[:, None]), axis=1)[:, None])
ESS = (1 / np.sum(weights**2, axis=1)) / (N_parts * keep_fraction)
ESSind = np.argmin(np.abs(ESStarget - ESS))
alpha = 20 #cooling rate for beta*(iteration/alpha)
Beta = BetaSet[ESSind] * max(1, ((testitcount - 1) / alpha))
part_weights = part_weights**Beta / np.sum(part_weights**Beta)
print('Weight tempering exponent: %.3f' % (Beta))
# Now select particles from the kept particles, according to
# their weights, to decide which particles to copy ontio
selection = np.random.choice(a=range(worst_keep),
size=N_parts - worst_keep,
replace=False,
p=part_weights)
# Now perform the particle copy
part_thetas[worst_keep:, :] = part_thetas[selection, :]
part_outputs[worst_keep:, :, :] = part_outputs[selection, :, :]
part_summaries[worst_keep:, :] = part_summaries[selection, :]
part_Ds[worst_keep:] = part_Ds[selection]
accepted = np.zeros(np.shape(part_Ds))
# Now the particles are 'jiggled' so that we return to having a set of
# unique particles. This step can also be thought of as performing
# local exploration, and is sometimes called 'mutation'. MCMC steps are
# used to perform mutation, and an advantage of SMC approaches is that
# the positions of all particles can be used to decide on a sensible
# jumping distribution.
# Create a weighted covariance matrix to use in a multivariate
# normal jumping distribution. 2.38^2/N is the 'optimal' factor
# under some assumptions
Ctheta = (2.38**2 / N_theta) * np.cov(np.transpose(part_thetas))
'''
%%%%%%%%%%%%%%%% PARPOOL 2: LOOP MCMC STEP %%%%%%%%%%%%%%%%%%%%%%%%%
'''
#Create an individual simulator object for this praticle, using the
# Perform one MCMC step to evaluate how many are expected to be required
#MVN_loop_set = partial(MVN_move, N_moves=1, part_thetas=part_thetas, part_outputs=part_outputs, part_summaries=part_summaries, part_Ds=part_Ds, Ctheta=Ctheta, theta_mins=theta_mins, theta_maxs=theta_maxs, scale_param=scale_param, target_D=target_D, Pt=Pt, Ot=Ot, f_simulate=f_simulate, f_summaries=f_summaries, f_discrep=f_discrep, seed_nums=seed_nums)
iterable = []
for k in range(worst_keep, N_parts):
iterable.append(
(k, 1, np.array(part_thetas[k, :]).ravel(),
np.squeeze(part_outputs[k, :, :]),
np.squeeze(part_summaries[k, :]), np.squeeze(part_Ds[k]),
Ctheta, theta_mins, theta_maxs, scale_param, target_D,
np.squeeze(Pt[seed_nums[k], :, :]),
np.squeeze(Ot[seed_nums[k], :, :]), f_simulate, f_summaries,
f_discrep, burnin))
print('mvn1...')
start = time.time()
#for i,k, part_theta, part_output, part_summary, part_D, accept in tqdm(enumerate(pool.imap_unordered(MVN_loop_set,range(worst_keep,N_parts)))):
chunksize = int((N_parts - worst_keep) / process_count)
if chunksize < 1:
chunksize = 1
results = pool.starmap_async(MVN_move, iterable, chunksize=chunksize)
for r in results.get():
k = r[0]
part_thetas[k, :] = r[1]
part_outputs[k, :, :] = r[2]
part_summaries[k, :] = r[3]
part_Ds[k] = r[4]
accepted[k] = r[5]
del iterable
print('mvn1 duration: ' + str(time.time() - start))
# Calculate the acceptance rate, and ensure that the case where
# all or no particles are accepted does not result in a number
# of MCMC steps that cannot be calculated
est_accept_rate = np.mean(accepted)
est_accept_rate = np.where(est_accept_rate != 0, est_accept_rate,
1 * 10**(-6))
est_accept_rate = np.where(est_accept_rate != 1, est_accept_rate,
1 - 1 * 10**(-6))
# Calculate the expected number of MCMC steps required from the
# estimated acceptance rate. The number of steps cannot go
# above a user-specified value
R = np.ceil(np.log(0.05) / np.log(1 - est_accept_rate))
R = np.min([R, max_MCMC_steps])
print('MCMC steps: %d' % (R))
# Perform the remaining MCMC steps
#MVN_loop_set = partial(MVN_move, N_moves=int(R-1), part_thetas=part_thetas, part_outputs=part_outputs, part_summaries=part_summaries, part_Ds=part_Ds, Ctheta=Ctheta, theta_mins=theta_mins, theta_maxs=theta_maxs, scale_param=scale_param, target_D=target_D, Pt=Pt, Ot=Ot, f_simulate=f_simulate, f_summaries=f_summaries, f_discrep=f_discrep, seed_nums=seed_nums)
#print('Accept rate: %g'%(est_accept_rate))
#print('Running remaining MVN loops %d to %d '%(worst_keep,N_parts))
#print('Creating %d particles'%(int(R-1)))
iterable = []
for k in range(worst_keep, N_parts):
iterable.append(
(k, int(R - 1), np.array(part_thetas[k, :]).ravel(),
np.squeeze(part_outputs[k, :, :]),
np.squeeze(part_summaries[k, :]), np.squeeze(part_Ds[k]),
Ctheta, theta_mins, theta_maxs, scale_param, target_D,
np.squeeze(Pt[seed_nums[k], :, :]),
np.squeeze(Ot[seed_nums[k], :, :]), f_simulate, f_summaries,
f_discrep, burnin))
print('mvn2...')
start = time.time()
#for i,k, part_theta, part_output, part_summary, part_D, accept in tqdm(enumerate(pool.imap_unordered(MVN_loop_set,range(worst_keep,N_parts)))):
chunksize = int((N_parts - worst_keep) / process_count)
if chunksize < 1:
chunksize = 1
brokeout = 0
results = pool.starmap_async(MVN_move, iterable, chunksize=chunksize)
for r in results.get():
k = r[0]
part_thetas[k, :] = r[1]
part_outputs[k, :, :] = r[2]
part_summaries[k, :] = r[3]
part_Ds[k] = r[4]
accepted[k] = r[5]
brokeout += r[6]
del iterable
print('mvn2 duration: ' + str(time.time() - start))
brokepercent = 100 * brokeout / (N_parts - worst_keep)
print('Breakout = %.2f %%' % (brokepercent))
# Count the number of unique particles
#unique_thetas = np.unique(part_thetas, axis=0)
N_unique = np.shape(np.unique(part_thetas, axis=0))[0]
# Output information on discrepancy and uniqueness if flag is set
if verbose:
print(
"Current discrepancy target is %g, number of unique particles is %g"
% (target_D, N_unique))
# Check to see if the desired discrepancy target has been reached, and
# terminate the loop if so. The structure of the loop means that a
# mutation step will happen after the final resample, so we expect a
# full sample of N_parts particles all satisfying the desired
# discrepancy constraint
print("Discrepency differnce: %.2f\n" % (desired_D - target_D, ))
if target_D < desired_D:
looping = 0
#Also check to see if degeneracy is occurring - this happens when the
# MCMC steps fail to find new locations and thus remain copies of
# previous particles. This also terminates the loop when it becomes too
# severe, but with additional output of a warning
#print("Degeneracy differnce: %.2f\n" % (N_unique-N_parts))
if N_unique <= (N_parts / 2):
print(
"WARNING: SMC loop terminated due to particle degeneracy. Target not reached! \n"
)
looping = 0
'''
# Visualise Particles if flag is set
if visualise
% Call provided plotting function
f_visualise(part_thetas, part_outputs, part_summaries, part_Ds);
drawnow;
end
'''
'''
#Memory and runtime safety features
mem_used = psutil.virtual_memory()[2]
print('memory % used:', mem_used)
if mem_used > 90:
print('WARNING: High memory usage. Aborting!\n')
looping = 0
if time.time()>max_runtime:
#print('WARNING: Runtime exceeded. Aborting!')
looping = 0
testitcount+=1
'''
break
# Close the parallel pool now that it's use is finished
pool.close()
pool.join()
del pool
#Sort final unique output
ordering = np.argsort(part_Ds)
part_thetas = np.copy(np.unique(part_thetas[ordering, :], axis=0))
_, idx = np.unique(part_thetas, axis=0, return_index=True)
part_thetas = np.copy(part_thetas[idx, :])
part_Ds = np.copy(part_Ds[idx])
part_outputs = np.copy(np.asarray(part_outputs[idx, :, :], dtype='f8'))
part_summaries = np.copy(part_summaries[idx, :])
print("Best discrepancy: %.2f, Worst discrepancy: %.2f" %
(part_Ds[0], part_Ds[-1]))
print('SMCABC COMPLETE\n')
#print only unique output
return part_thetas, part_outputs, part_summaries, part_Ds
H_lims = np.asarray([0.025, 0.1]) # Disc aspect ratio
iH = 7
Md_lims = np.asarray([0.01, 0.1]) # Disc mass
iMd = 8
headers = 'Rp,Mpl,a,Ms1,Ms2,alpha,pindex,H,Mdisc' #column headers for csv later
lims = [
Rp_lims, Mpl_lims, a_lims, Ms1_lims, Ms2_lims, alpha_lims, pindex_lims,
H_lims, Md_lims
]
nparams = len(lims)
nsamp = 10
lhs = np.asarray(lhsmdu.sample(nparams,
nsamp)) # perform latin hypercube sampling
# scale random variables to be within desired limits
for i, lim in enumerate(lims):
lhs[i, :] = lhs[i, :] * (lim[1] - lim[0]) + lim[0]
lhs[i, :] = lhs[i, :]
# calculate required grid size and timestep vals
grid_rad = lhs[iRp, :] + 10 # 10 au greater than the planet location
torb = 2 * np.pi * grid_rad**(3. / 2.) # orbital period at grid outer radius
dt = torb / 100.
tvisc = 4. / 9. * (grid_rad**2 / lhs[ialpha, :]**2 / lhs[iH, :]**2) * np.sqrt(
grid_rad**3 / lhs[iMs1, :])
# calc sigma0 from disc mass and pindex
# NOTE: this assumes this disc mass extent is 100au NOT just the size of the grid
# with multi-dimensional uniformity. The following Python code calls two Latin Hypercube functions # where n data points are plotted on a square of side length L. # The grid and data set is color coded. Note: # The authors of the following code would use the Latin Hypercube function in sequence with # The Seer, to plot an unbiased amount of data points to train The Seer's neural network. # The origin of The Seer's test envrionment would be at 2.75 meters and # due to the radiation patterns and hardware implemented within The Seer. # The following code would account for this shift of origin from (0,0) to # (2.75,0), as well as a 0.5 meter padding space needed between The Seer's receiving antenna array # and the closest test data points. # Set number of sample points: n = 100 k = lhsmdu.sample(2,n) # Latin Hypercube Function with multi-dimensional uniformity # The project only utilized 2D but more dimensions could be included, check out # the lhsmdu project at pypi.org/project/lhsmdu/ k = np.array(k*5) # Set grind size or dimensions of of testing environment for first Latin Hypercube function # Uncomment to include a second set of LHS points #randSeed = 11 # Create random Latin Hypercube seed to have two unbiased data sets and samples # The random seed does not need to be LHS, other methods are available. #l = lhsmdu.sample(2,50, randomSeed=randSeed) #Second Latin Hypercube Function with (2D) uniformity #l = np.array(l*5.5) #Set grind size or dimensions of testing environment for second Latin Hypercube function fig = plt.figure() ax = fig.gca() ax.set_xticks(np.arange(0,7,.5)) # Set the number of ticks on the axis and the spacing between them ax.set_yticks(np.arange(0,7,.5)) # Set the number of ticks on the axis and the spacing between them
import numpy as np
import pandas as pd
from pythongp.test_functions import test_functions
import lhsmdu
lower = [100, 1000, 1000, 10, 10, 10, 10, 10]
higher = [10000, 10000, 10000, 1000, 1000, 1000, 1000, 1000]
lower = np.array(lower)
higher = np.array(higher)
d = 8
doe = np.array(np.transpose(lhsmdu.sample(d, d * 3)))
doe = doe * np.tile(higher - lower, (24, 1)) + np.tile(lower, (24, 1))
results = test_functions.g_10(doe)
doe = pd.DataFrame(doe, columns=['x{}'.format(i) for i in range(8)])
pd.concat((doe, results), axis=1).to_csv('/Users/sebastien/data/data_g10',
sep=';')
import lhsmdu
from mpl_toolkits.mplot3d import Axes3D
import matplotlib.pyplot as plt
import numpy
l = lhsmdu.sample(
3, 10) # Latin Hypercube Sampling of two variables, and 10 samples each.
k = lhsmdu.createRandomStandardUniformMatrix(3, 10) # Monte Carlo Sampling
pdir = numpy.matrix(
'55 55 115 55 115 55 115 80 70 60 100 84 63 108 73 94 85 63 98 112')
H = numpy.matrix(
'1.2 1.2 1.2 2.3 2.3 2.3 2.3 .7 1.2 1.7 1.4 1.1 2.1 .9 2.0 1.5 1.9 1.6 2.1 1.8'
)
F = numpy.matrix(
'.091 .125 .125 .167 .167 .091 .091 .091 .091 .140 .11 .131 .152 .116 .097 .149 .162 .144 .138 .169'
)
k[0] = 60 * k[0] + 55
l[0] = 60 * l[0] + 55
k[1] = 2.5 * k[1]
l[1] = 2.5 * l[1]
k[2] = .1 * k[2] + .085
l[2] = .1 * l[2] + .085
fig = plt.figure()
ax = fig.add_subplot(131, projection='3d')
ax.scatter([k[0]], [k[1]], [k[2]], color="b", label="MC")
ax.scatter([l[0]], [l[1]], [l[2]], color="r", label="LHS")
ax.scatter([pdir], [H], numpy.transpose(F), color="g", label="OURS")
ax.set_xlabel('Peak Dir')
ax.set_ylabel('Wave Height')
ax.set_zlabel('Frequency')
def GeneticAlgorithm(objFunc,
consFunc,
bndsArr,
halfpop=10,
maxit=100,
tol=1e-6,
mutationChance=0.005,
mutationScale=0.05,
verbose=0):
#objFunc is the objective function
#consFunc is the constraing function
#The constraint function must be of the form g(x) < 0, where x may be
#a vector
#bndsArr is the bounds array
#bndsArr has the form (# vars, 2)
#For example [[1, 2]
# [3, 4]]
#Means 2 variables. x1 is bounded by 1, 2, and x2 is bounded by 3, 4
#halfpop is half the population size. Half the population is used to
#ensure parity
#maxit is maximum iterations
#tol is tolerance per iteration. From one iteration to the next, if the change
#of the objective function is less than the tolerance, the algorithm
#will automatically terminate.
#mutationChance is the chance of a mutation occurring. It defaults to 0.5%
#mutationScale is the scale of the mutation, scale * (-1, 1)
#Establishes the total population
totPop = halfpop * 2
#Establishes the number of variables
numVars = len(bndsArr)
#Ensures the objective function is callable
if (not callable(objFunc)):
print("Error, objective function not specified correctly.")
return 0
#Ensures the constraint function is callable
elif (not callable(consFunc)):
print("Error, constraint function is not specified correctly.")
return 0
#Ensures the total population is an integer
elif (not isinstance(totPop, int)):
print("Error, population size specified as non-integer.")
return 0
#Ensures the maximum iterations is an integer
elif (not isinstance(maxit, int)):
print("Error, maximum iterations specified as non-integer.")
return 0
#Ensures the tolerance is an integer or decimal
elif ((not isinstance(tol, float)) and (not isinstance(tol, int))):
print("Error, tolerance not specified as a number.")
#Ensures the bounds array is the proper shape of (#vars, 2)
elif (np.shape(bndsArr) != (numVars, 2)):
print(
"Error, bounds array is not correct size. Should be of size (x , 2)"
)
return 0
#If all the above goes through, then it will execute the Genetic Algorithm
else:
print("Initialization success.")
print("Creating initial population with LHS...")
#Initializes an LHS array of size (#vars, total population)
LHSarr = lhsmdu.sample(numVars, totPop)
#Initializes an initial population of zeros of size
#(total population, #vars)
InitPop = np.zeros((totPop, numVars))
#Reassigns each population member with the LHS sample to between the
#lower bounds and upper bounds
for i in range(0, totPop): #i is population member index
for j in range(0, numVars): #j is variable number
InitPop[i, j] = LHSarr[j, i] * (bndsArr[j, 1] -
bndsArr[j, 0]) + bndsArr[j, 0]
#The results have the form (Population Member Index, Variable)
if (verbose == 1):
print("Initial population established.\n")
print("Testing initial population...")
currentEval = np.zeros(totPop)
for i in range(0, totPop):
currentEval[i] = ObjectiveFunction(InitPop[i, :])
minIndex = np.argmin(currentEval)
minVal = min(currentEval)
print("--------------------------------------------------------------")
print("Initial Population Evaluation")
print("The best objective function evaluation is ", minVal)
print("It ocurrs at population member ", minIndex)
print("--------------------------------------------------------------")
prevGenObj = minVal
for k in range(0, maxit):
if (verbose == 1):
print("Creating contest population.")
#Creates a copy of the initial population for the contest population
ContestPopulation = np.copy(InitPop)
#Creates an array that will be populated with indices for the contest
ContestIndexArray = np.zeros((totPop, 2), dtype=int)
#Creates an array of available contest indices
AvailableContestIndices = np.linspace(0,
totPop - 1,
totPop,
dtype=int)
#Assigns available contest indices randomly to the contest index array.
#Each population member is entered into the contest twice,
#so during the breeding phase the population size is maintained.
for i in range(0, halfpop):
for j in range(0, 2):
select = random.randrange(0, len(AvailableContestIndices))
ContestIndexArray[i, j] = AvailableContestIndices[select]
AvailableContestIndices = np.delete(
AvailableContestIndices, select)
#Reestablishes the available contest indices for the second pairing
AvailableContestIndices = np.linspace(0,
totPop - 1,
totPop,
dtype=int)
for i in range(0, halfpop):
for j in range(0, 2):
select = random.randrange(0, len(AvailableContestIndices))
ContestIndexArray[i + halfpop,
j] = AvailableContestIndices[select]
AvailableContestIndices = np.delete(
AvailableContestIndices, select)
EvaluatedConstraints = np.zeros(totPop)
EvaluatedObjectives = np.zeros(totPop)
for i in range(0, totPop):
EvaluatedConstraints[i] = consFunc(ContestPopulation[i, :])
EvaluatedObjectives[i] = objFunc(ContestPopulation[i, :])
ContestWinnerArr = np.zeros(totPop, dtype=int)
if (verbose == 1):
print(ContestIndexArray)
for i in range(0, totPop):
constraint1 = EvaluatedConstraints[ContestIndexArray[i, 0]]
constraint2 = EvaluatedConstraints[ContestIndexArray[i, 1]]
if ((constraint1 > 0) and (constraint2 > 0)):
if (constraint1 < constraint2):
ContestWinnerArr[i] = ContestIndexArray[i, 0]
else:
ContestWinnerArr[i] = ContestIndexArray[i, 1]
elif ((constraint1 < 0) and (constraint2 > 0)):
ContestWinnerArr[i] = ContestIndexArray[i, 0]
elif ((constraint1 > 0) and (constraint2 < 0)):
ContestWinnerArr[i] = ContestIndexArray[i, 1]
else:
evaluate1 = EvaluatedObjectives[ContestIndexArray[i, 0]]
evaluate2 = EvaluatedObjectives[ContestIndexArray[i, 1]]
if (evaluate1 < evaluate2):
ContestWinnerArr[i] = ContestIndexArray[i, 0]
else:
ContestWinnerArr[i] = ContestIndexArray[i, 1]
if (verbose == 1):
print("Contest finished.")
print("Winner indices: ", ContestWinnerArr)
print("Creating breeding pairs.")
AvailableBreeders = np.copy(ContestWinnerArr)
BreedingIndices = np.zeros((halfpop, 2), dtype=int)
for i in range(0, halfpop):
for j in range(0, 2):
select = random.randrange(0, len(AvailableBreeders))
BreedingIndices[i, j] = AvailableBreeders[select]
AvailableBreeders = np.delete(AvailableBreeders, select)
#Breed
for i in range(0, halfpop):
InitPop[i] = 0.5 * ContestPopulation[BreedingIndices[i,0]] + \
0.5 * ContestPopulation[BreedingIndices[i,1]]
InitPop[i + halfpop] = 2 * ContestPopulation[BreedingIndices[i,1]] - \
ContestPopulation[BreedingIndices[i,0]]
for i in range(0, totPop):
mutate = random.random()
if (mutate < mutationChance):
if (verbose == 1):
print("Mutation.")
mutationAmount = mutationScale * (2. * random.random() -
1.)
InitPop[i] += mutationAmount
#Evaluate change in best value
currentEval = np.zeros(totPop)
for i in range(0, totPop):
currentEval[i] = ObjectiveFunction(InitPop[i, :])
minIndex = np.argmin(currentEval)
minVal = min(currentEval)
print(
"--------------------------------------------------------------"
)
print("Iteration ", k + 1)
print("The best objective function evaluation is ", minVal)
print("It ocurrs at population member ", minIndex)
print(
"--------------------------------------------------------------"
)
currGenObj = minVal
if (abs(prevGenObj - currGenObj) <= tol):
print("Tolerance reached. Aborting algorithm.")
print("Best results found at ", InitPop[minIndex])
print("The value was: ", minVal)
return 0
else:
prevGenObj = currGenObj
print("Maximum iterations reached.")
return 0
def runtime(numDimensions, numSamples):
''' Checks runtime using standard variables '''
start_time = time()
m = lhsmdu.sample(numDimensions,numSamples)
end_time = time()
print end_time-start_time
for key in solvents:
s1, s2 = solvents[key][0], solvents[key][1]
distance = (x1 - s1)**2 + (x2 - s2)**2
dis_dict[distance] = key
dis_list[count] = distance
count += 1
min_dist = min(dis_list)
return dis_dict[min_dist]
# read in solvents and their polarity and mbo
sol = parse_data.solvent_parser('data_solvent.csv')
# number of initial samples
n_sample = 10
# sample uniformly from [0,1]^4
samples = lhsmdu.sample(4, n_sample)
# first two dimensions are {1,2,3}
samples[0] = np.floor(samples[0] * 3)
samples[1] = np.floor(samples[1] * 3)
# third dim is (0.6,1.4)
samples[2] = [0.6 for i in range(n_sample)] + \
samples[2] * (1.4-0.6)
# fourth dim is (5, 60)
samples[3] = [5. for i in range(n_sample)] + \
samples[3] * (60 - 5.)
# map samples to solution combo
solutions = [[0 for j in range(3)] for i in range(n_sample)]
for i in range(n_sample):
# halides
if samples[0, i] == 0:
solutions[i][0] = 'I'
if any(M_model) < np.min(range_M):
squared_sum += 100
# Step 5: Calculate cost-function
squared_sum = np.sum((yG_model - yG_vec))**2 + np.sum(
(yI_model - yI_vec)**2)
return squared_sum
## Hypercube set up
randSeed = 2 # random number of choice
lhsmdu.setRandomSeed(
randSeed) # Latin Hypercube Sampling with multi-dimensional uniformity
start = np.array(
lhsmdu.sample(18, 4)
) # Latin Hypercube Sampling with multi-dimensional uniformity (parameters, samples)
para, samples = start.shape
## intervals for the parameters
para_int = [0, 500]
minimum = (np.inf, None)
# Bounds for the model
bound_low = np.array(
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0.1, 0.1])
bound_upp = np.repeat(np.inf, para)
bounds = Bounds(bound_low, bound_upp)
if any(F_model < np.min(range_F)):
squared_sum += penalty
# Calculate cost-function
squared_sum += np.sum((yG_model - yG_vec)**2) + np.sum(
(yI_model - yI_vec)**2)
return squared_sum
## Hypercube set up
randSeed = 2 # random number of choice
lhsmdu.setRandomSeed(
randSeed) # Latin Hypercube Sampling with multi-dimensional uniformity
start = np.array(
lhsmdu.sample(8, 10)
) # Latin Hypercube Sampling with multi-dimensional uniformity (parameters, samples)
para, samples = start.shape
## intervals for the parameters
para_int = [0, 0.001, 0.01, 0.1, 1, 10, 100, 500]
# 0, 1, 2, 3, 4, 5, 6, 7
minimum = (np.inf, None)
# Bounds for the model
bound_low = np.repeat(0.0, para)
bound_upp = np.repeat(np.inf, para)
bounds = Bounds(bound_low, bound_upp)
fig = plt.figure()