def run(self):
'''
Run misoKG.
'''
if self.save_extra_files and self.overwrite:
os.system("rm %s %s" % (self.mu_fname, self.sig_fname))
# Error Handling
assert self.costs is not None, "Error - You must specify costs before running!"
assert len(self.costs) == len(
self.IS
), "Error - You must specify the same number of information sources as costs!"
if self.verbose:
print(
"\n-------------------------------------------------------------------------------------"
)
print("Beginning optimization.")
# print("\tParallel = %s" % str(self.parallel))
print(" Number of Information Sources = %d" % len(self.IS))
print(" Acquisition = "),
if self.acquisition == getNextSample_misokg:
print("misoKG with a cost list = %s" % str(self.costs))
elif self.acquisition == getNextSample_EI:
print("EI")
elif self.acquisition == getNextSample_kg:
print("KG")
else:
print("Custom!")
if self.hyperparameter_objective == MLE:
obj_name = "MLE"
elif self.hyperparameter_objective == MAP:
obj_name = "MAP"
else:
obj_name = "Custom"
if self.loglike == gaussian_loglike:
loglike_name = "Gaussian"
elif self.loglike == bonilla_loglike:
loglike_name = "Bonilla"
else:
loglike_name = "Custom"
print(
"The Hyperparameter Objective is %s with %d starting samples."
% (obj_name, self.n_start))
print("The loglikelihood method is %s." % loglike_name)
if self.dynamic_pc:
print(
"Will use a dynamic pearson correlation coefficient for rho."
)
print("Will optimize the following parameters:")
print(" " + ', '.join(self.theta.hp_names))
print(
"-------------------------------------------------------------------------------------"
)
# Start - TIMER
self.t0 = time.time()
# Step 1 - Ensure we have our historical training set. If not, then
# generate one.
if self.fname_historical is None:
self.fname_historical = "historical.dat"
if self.numerical:
self.sample_numerical()
else:
self.sample()
else:
self.historical = pickle.load(open(self.fname_historical, 'r'))
if self.numerical:
if len(self.historical[0]) != len(self.domain) + 1:
raise Exception(
"The historical data seems to be incorrect for misoKG. Maybe the IS associated with each point was not included?"
)
if len(self.historical[0]) not in [10, 17]:
raise Exception(
"The historical data seems to be incorrect for misoKG. Maybe the IS associated with each point was not included?"
)
# Step 2 - Generate a full list of our sample space if it has not been given
if self.mixed_solvents and not self.numerical:
raise Exception(
"Mixed Solvents have not been implemented properly.")
else:
if self.combinations is None and not self.numerical:
self.combinations = self.get_combos_pure_solvent()
if self.all_X is None and not self.numerical:
self.all_X = pal_strings.alphaToNum(
self.combinations, solvents, mixed_halides=self.mixed_halides)
self.all_solvent_properties = np.array(self.all_X)[:, -3:-1]
self.all_Y = np.array([0 for i in range(len(self.all_X))])
# Step 2.5 - Store our X and Y points
if not self.numerical:
self.assign_samples()
# Store a list of samples that have been sampled at all information sources
self.indices_overlap = range(len(self.sampled_X))
# Step 3 - Get our hyperparameters. As we don't have initial ones,
# don't use_theta for this instance.
self.updateHPs(use_theta=False)
# Step 3.5 - Initialize indices_overlap variables
self.indices_overlap_len = len(self.indices_overlap)
self.indices_overlap_changed = False
# Step 4 - Update the posterior based on the historical data.
self.updatePosterior()
if not self.numerical:
# Save combinations and default save actions
if self.combos_fname is not None:
fptr = open(self.combos_fname, 'w')
for i, c in enumerate(self.combinations):
fptr.write("%d\t%s\n" % (i, c))
fptr.close()
self.save()
# Step 5 - Begin the main loop
start, stop = len(self.sampled_X), len(self.combinations)
else:
start, stop = len(self.sampled_X), len(self.all_X)
best_found_in = start
best_value = max(
np.array(self.sampled_objectives)[self._get_info_source_map(
self.sampled_X)[0]])
best_index = self.sampled_indices[self.sampled_objectives.index(
best_value)]
best_name = self.combinations[best_index]
# Initialize our costs based on the sampled so far
self.total_cost = sum([self.costs[int(x[0])] for x in self.sampled_X])
best_prediction = max(
np.array(self.mu)[self._get_info_source_map(self.all_X)[0]])
best_prediction = list(
np.array(self.mu)[self._get_info_source_map(
self.all_X)[0]]).index(best_prediction)
best_prediction = self._get_info_source_map(
self.all_X)[0][best_prediction]
recommendation = self.combinations[best_prediction]
if self.save_extra_files and self.sample_fname is not None:
fptr = open(self.sample_fname, 'a')
for v in zip(self.sampled_names, self.sampled_objectives):
fptr.write("%s\t%.4f\n" % v)
# Begin the main loop
fully_sampled = False
recommendation_kill_flag = False
iteration_kill_flag = False
cost_kill_flag = False
for index in range(start, stop):
if self.iteration_kill_switch is not None and index >= self.iteration_kill_switch:
iteration_kill_flag = True
break
# If we have sampled all the IS0, and noise doesn't exist, then we gracefully exit
if not self.noise and all([
i_IS0 in self.sampled_indices
for i_IS0 in self._get_info_source_map(self.all_X)[0]
]):
fully_sampled = True
break
# Step 6 - acquisition Function. Decide on next point(s) to sample.
next_point = self.acquisition(
self.mu,
#self.theta.rho_matrix(self.all_X) * self.K,
self.K,
max(self.sampled_objectives),
len(self.combinations),
self.costs,
self.all_X,
self.sampled_indices,
save=self.acquisition_fname)
if next_point in self.sampled_indices:
print(
"\nFAILURE!!!! SAMPLED # %s - Index = %d# POINT TWICE!\n" %
(self.combinations[next_point], next_point))
print("K Diagonal = %s" %
' '.join(["%f" % v for v in np.diag(self.K)]))
print("K[%d] = %s" % (next_point, ' '.join(
["%f" % v for v in self.K[next_point]])))
print("Sampled Points = %s" % str(self.sampled_indices))
raise Exception(
"Error - acquisition function grabbed an already sampled point!"
)
if self.verbose:
r = -1.23
if "[0, 1]" in self.theta.rho:
r = self.theta.rho["[0, 1]"]
suffix = "(iter %d) %s = %.4f, sampling %s. Recommendation = %s, Current Cost = %.2f, Rho = %.3f" % (
best_found_in, best_name, best_value,
self.combinations[next_point], recommendation,
self.total_cost, r)
ppb(index, stop, prefix='Running', suffix=suffix, pad=True)
if self.logger_fname is not None:
fptr = open(self.logger_fname, 'a')
fptr.write(suffix + "\n")
fptr.close()
if self.recommendation_kill_switch is not None and recommendation == self.recommendation_kill_switch:
recommendation_kill_flag = True
break
# Step 7 - Sample point(s)
self.sampled_indices.append(next_point)
self.sampled_names.append(self.combinations[next_point])
if not self.numerical:
self.sampled_X.append(
pal_strings.alphaToNum(
self.sampled_names[-1],
solvents,
mixed_halides=self.mixed_halides)[0])
h, c, _, s, info_lvl = pal_strings.parseName(
self.sampled_names[-1])
self.sampled_objectives.append(self.IS[info_lvl](h, c[0], s))
else:
x = self.all_X[next_point]
self.sampled_X.append(x)
info_lvl = int(x[0])
self.sampled_objectives.append(self.IS[info_lvl](*x[1:]))
if self.save_extra_files and self.sample_fname is not None:
fptr = open(self.sample_fname, 'a')
fptr.write(
"%s\t%.4f\n" %
(self.sampled_names[-1], self.sampled_objectives[-1]))
# Ensure we get an array of all sampled indices that have been sampled
# at ALL information source levels
chk = self.sampled_X[-1][1:]
if self.numerical:
found = [
i for i, v in enumerate(self.sampled_X)
if all(chk == v[1:])
]
else:
found = [
i for i, v in enumerate(self.sampled_X) if chk == v[1:]
]
# Assume we have 4 IS. If we find 4 of chk, then we now have fully sampled chk.
if len(found) == len(self.IS):
for f in found:
if f not in self.indices_overlap:
self.indices_overlap.append(f)
self.indices_overlap_changed = self.indices_overlap_len != len(
self.indices_overlap)
self.indices_overlap_len = len(self.indices_overlap)
# Step 7.5 - Maybe re-opt the hyperparameters
# Note, we do this in a two step approach. First, we optimize all HPs based on
# only data points that exist at all levels of theory. Then we optimize
# only at the highest level of theory sampled so far (IS0).
if index != start and (self.reopt is not None
and index % self.reopt == 0) or (
self.ramp_opt is not None
and index < self.ramp_opt):
self.updateHPs()
# Step 8a - Update the posterior completely if we are reoptimizing the HPs
self.updatePosterior()
else:
# Step 8b - Update the posterior with only the newest sampled point
self.updatePosterior(
(self.sampled_indices[-1], self.sampled_objectives[-1]))
self.save()
# Count the cost of this iteration
self.total_cost += self.costs[info_lvl]
if self.cost_kill_switch is not None and self.total_cost > self.cost_kill_switch:
cost_kill_flag = True
break
# Get our recommendation from max(mu) for only IS0
best_prediction = max(
np.array(self.mu)[self._get_info_source_map(self.all_X)[0]])
best_prediction = list(
np.array(self.mu)[self._get_info_source_map(
self.all_X)[0]]).index(best_prediction)
best_prediction = self._get_info_source_map(
self.all_X)[0][best_prediction]
recommendation = self.combinations[best_prediction]
# Get the best sampled so far
potential_best = max(
np.array(self.sampled_objectives)[self._get_info_source_map(
self.sampled_X)[0]])
if potential_best > best_value:
best_found_in = index
best_value = potential_best
best_index = self.sampled_indices[
self.sampled_objectives.index(best_value)]
best_name = self.combinations[best_index]
# END TIMER
self.t1 = time.time()
if self.verbose:
print("-----------------------")
print("PAL Optimizer has completed in %.2f s" %
(self.t1 - self.t0))
if fully_sampled:
print("Optimizer quit early as IS0 was fully sampled")
if recommendation_kill_flag:
print("Optimizer quit early due to recommendation of %s." %
self.recommendation_kill_switch)
if iteration_kill_flag:
print("Optimizer quit early due to exceeding %d iterations." %
self.iteration_kill_switch)
if cost_kill_flag:
print("Optimizer quit early due to exceeding %.4f cost." %
self.cost_kill_switch)
print("-----------------------")
print("Best combination: %s" % best_name)
print(" Objective: %.4f" % best_value)
print(" Maximized: %d" % best_found_in)
print("-----------------------")
print self.theta
print(
"-------------------------------------------------------------------------------------\n"
)
def sample(self,
specify=None,
debug=False,
MAX_LOOP=10,
allow_reduced=False):
'''
This function will run, in parallel, N_samples of the objective functions for
historical data generation. Note, these are run for EVERY information source.
'''
if debug:
print("Collecting LHS samples...")
if specify is None:
counter, samples = 0, []
while len(
samples) != self.historical_nsample and counter < MAX_LOOP:
# Grab a latin hypercube sample
samples = doe_lhs.lhs(
int(self.mixed_halides) * 2 + 2, self.historical_nsample)
# Round the LHS and figure out the samples
solvent_ranges = [
i * 1.0 / len(self.S)
for i in range(1,
len(self.solvents) + 1)
]
solv = lambda v: self.S[[v <= s
for s in solvent_ranges].index(True)]
trio = lambda v: [
int(v > (chk - 1.0 / 3.0) and v <= chk)
for chk in [1. / 3., 2. / 3., 1.0]
]
# Grab our samples
if self.mixed_halides:
halides = [sorted([s[0], s[1], s[2]]) for s in samples]
halides = [[trio(h) for h in hh] for hh in halides]
samples = [
h[0] + h[1] + h[2] + trio(s[3]) + [
self.solvents[solv(s[-1])]["density"],
self.solvents[solv(s[-1])]["dielectric"],
self.solvents[solv(s[-1])]["index"]
] for h, s in zip(halides, samples)
]
else:
samples = [
trio(s[0]) + trio(s[1]) + [
self.solvents[solv(s[-1])]["density"],
self.solvents[solv(s[-1])]["dielectric"],
self.solvents[solv(s[-1])]["index"]
] for s in samples
]
# Ensure no duplicates
samples = sorted(samples, key=lambda x: x[-1])
samples = [tuple(s) for s in samples]
samples = [list(s) for s in set(samples)]
counter += 1
else:
if isinstance(specify, int):
specify = [specify]
self.historical_nsample = len(specify)
samples = [self.combinations[i] for i in specify]
samples = pal_strings.alphaToNum(samples,
solvents,
mixed_halides=self.mixed_halides,
name_has_IS=True)
samples = [s[1:] for s in samples
] # Remove the IS label from the descriptor
if allow_reduced:
print(
"Warning - Will sample from subspace due to duplicates (%d instead of %d)."
% (len(samples), self.historical_nsample))
self.historical_nsample = len(samples)
elif specify is None:
assert counter < MAX_LOOP, "Error - Unable to sample from space without duplicates!"
if debug:
print("Will sample %s" % str(samples))
# Now, run these simulations to get the sample points
jobs = []
for i, sample in enumerate(samples):
if debug:
print "Running %s..." % sample
s = pal_strings.parseNum(sample,
self.solvents,
mixed_halides=self.mixed_halides,
num_has_IS=False)
hat, cat, _, solv, _ = pal_strings.parseName(s, name_has_IS=False)
cat = cat[0]
if not self.mixed_halides:
hat = hat[0]
if debug:
print("\tAdding %s to sample runs..." % s)
for j, obj in enumerate(self.IS):
jobs.append([[j] + copy.deepcopy(sample), obj(hat, cat, solv)])
# Now, get results from each simulation
samples = []
for sample, j in jobs:
if not isinstance(j, float):
j.wait()
samples.append(sample + [j.get_result()])
# In special situations, when we are reading from a list for example, we don't need to worry
# about a job object, and can just assign the value directly.
else:
samples.append(sample + [j])
s = pal_strings.parseNum(samples[-1][:-1],
self.solvents,
mixed_halides=self.mixed_halides,
num_has_IS=True)
if debug:
print("\t%s was found as %lg" % (s, samples[-1][-1]))
# Save the sampled data
fptr = open(self.fname_historical, "w")
pickle.dump(samples, fptr)
fptr.close()
self.historical = samples
if debug:
print("Done Collecting Samples\n")
sim.combinations = [k[1] + "Pb" + k[0] + "_" + k[2] + "_" + str(IS) for k in [key.split() for key in IS0.keys()] for IS in range(len(sim.IS))]
combos_no_IS = [k[1] + "Pb" + k[0] + "_" + k[2] for k in [key.split() for key in IS0.keys()]]
# Because we do this, we should also generate our own historical sample
sim.historical_nsample = len(combos_no_IS)
choices = combos_no_IS
tmp_data = pal_strings.alphaToNum(
choices,
solvents,
mixed_halides=True,
name_has_IS=False)
data = []
for IS in range(len(sim.IS)):
for i, d in enumerate(tmp_data):
h, c, _, s, _ = pal_strings.parseName(pal_strings.parseNum(d, solvents, mixed_halides=True, num_has_IS=False), name_has_IS=False)
c = c[0]
data.append([IS] + d + [sim.IS[IS](h, c, s)])
IS0 = np.array([x[-1] for x in data if x[0] == 0])
IS1 = np.array([x[-1] * 1.8 for x in data if x[0] == 1])
IS0, IS1 = zip(*sorted(zip(IS0, IS1)))
# IS0 = IS0 / np.linalg.norm(IS0)
# IS1 = IS1 / np.linalg.norm(IS1)
# print IS0
# print IS1
import matplotlib.pyplot as plt
def run_misokg(run_index):
# Store data for debugging
IS0 = pickle.load(open("enthalpy_N1_R3_Ukcal-mol", 'r'))
IS1 = pickle.load(open("enthalpy_N1_R2_Ukcal-mol", 'r'))
# Generate the main object
sim = Optimizer()
# Assign simulation properties
#sim.hyperparameter_objective = MAP
sim.hyperparameter_objective = MLE
###################################################################################################
# File names
sim.fname_out = "enthalpy_misokg.dat"
sim.fname_historical = None
# Information sources, in order from expensive to cheap
sim.IS = [
lambda h, c, s: -1.0 * IS0[' '.join([''.join(h), c, s])],
lambda h, c, s: -1.0 * IS1[' '.join([''.join(h), c, s])]
]
sim.costs = [1.0, 0.1]
sim.logger_fname = "data_dumps/%d_misokg.log" % run_index
if os.path.exists(sim.logger_fname):
os.system("rm %s" % sim.logger_fname)
os.system("touch %s" % sim.logger_fname)
sim.obj_vs_cost_fname = "data_dumps/%d_misokg.dat" % run_index
sim.mu_fname = "data_dumps/%d_mu_misokg.dat" % run_index
sim.sig_fname = "data_dumps/%d_sig_misokg.dat" % run_index
sim.combos_fname = "data_dumps/%d_combos_misokg.dat" % run_index
sim.hp_fname = "data_dumps/%d_hp_misokg.dat" % run_index
sim.acquisition_fname = "data_dumps/%d_acq_misokg.dat" % run_index
sim.save_extra_files = True
########################################
# Override the possible combinations with the reduced list of IS0
# Because we do this, we should also generate our own historical sample
combos_no_IS = [
k[1] + "Pb" + k[0] + "_" + k[2]
for k in [key.split() for key in IS0.keys()]
]
sim.historical_nsample = 10
choices = np.random.choice(combos_no_IS,
sim.historical_nsample,
replace=False)
tmp_data = pal_strings.alphaToNum(choices,
solvents,
mixed_halides=True,
name_has_IS=False)
data = []
for IS in range(len(sim.IS)):
for i, d in enumerate(tmp_data):
h, c, _, s, _ = pal_strings.parseName(pal_strings.parseNum(
d, solvents, mixed_halides=True, num_has_IS=False),
name_has_IS=False)
c = c[0]
data.append([IS] + d + [sim.IS[IS](h, c, s)])
sim.fname_historical = "data_dumps/%d.history" % run_index
pickle.dump(data, open(sim.fname_historical, 'w'))
simple_data = [d for d in data if d[0] == 0]
pickle.dump(simple_data,
open("data_dumps/%d_reduced.history" % run_index, 'w'))
########################################
sim.n_start = 10 # The number of starting MLE samples
sim.reopt = 20
sim.ramp_opt = None
sim.parallel = False
# Possible compositions by default
sim.A = ["Cs", "MA", "FA"]
sim.B = ["Pb"]
sim.X = ["Cl", "Br", "I"]
sim.solvents = copy.deepcopy(solvents)
sim.S = list(set([v["name"] for k, v in sim.solvents.items()]))
sim.mixed_halides = True
sim.mixed_solvents = False
# Parameters for debugging and overwritting
sim.debug = False
sim.verbose = True
sim.overwrite = True # If True, warning, else Error
sim.acquisition = getNextSample_misokg
# Functional forms of our mean and covariance
# MEAN: 4 * mu_alpha + mu_zeta
# COV: sig_alpha * |X><X| + sig_beta * I_N + sig_zeta + MaternKernel(S, weights, sig_m)
SCALE = [2.0, 4.0][int(sim.mixed_halides)]
# _1, _2, _3 used as dummy entries
def mean(X, Y, theta):
mu = np.array([SCALE * theta.mu_alpha + theta.mu_zeta for _ in Y])
return mu
sim.mean = mean
def cov_old(X, Y, theta):
A = theta.sig_alpha * np.dot(
np.array(X)[:, 1:-3],
np.array(X)[:, 1:-3].T)
B = theta.sig_beta * np.diag(np.ones(len(X)))
C = theta.sig_zeta
D = mk52(np.array(X)[:, -3:-1], [theta.l1, theta.l2], theta.sig_m)
return theta.rho_matrix(X) * (A + B + C + D)
def cov(X0, Y, theta):
A = theta.sig_alpha * np.dot(
np.array(X0)[:, :-3],
np.array(X0)[:, :-3].T)
B = theta.sig_beta * np.diag(np.ones(len(X0)))
C = theta.sig_zeta
D = mk52(np.array(X0)[:, -3:-1], [theta.l1, theta.l2], theta.sig_m)
Kx = A + B + C + D
Ks = np.array([
np.array(
[theta.rho[str(sorted([i, j]))] for j in range(theta.n_IS)])
for i in range(theta.n_IS)
])
if theta.normalize_Ks:
Ks = Ks / np.linalg.norm(Ks)
e = np.diag(np.array([theta.e1, theta.e2]))
Ks = e.dot(Ks.dot(e))
return np.kron(Ks, Kx)
sim.cov = cov
sim.theta.bounds = {}
sim.theta.mu_alpha, sim.theta.bounds['mu_alpha'] = None, (
1E-3, lambda _, Y: max(Y))
sim.theta.sig_alpha, sim.theta.bounds['sig_alpha'] = None, (
1E-2, lambda _, Y: 10.0 * np.var(Y))
sim.theta.sig_beta, sim.theta.bounds['sig_beta'] = None, (
1E-2, lambda _, Y: 10.0 * np.var(Y))
sim.theta.mu_zeta, sim.theta.bounds['mu_zeta'] = None, (
1E-3, lambda _, Y: max(Y))
sim.theta.sig_zeta, sim.theta.bounds['sig_zeta'] = None, (
1E-2, lambda _, Y: 10.0 * np.var(Y))
sim.theta.sig_m, sim.theta.bounds['sig_m'] = None, (1E-2,
lambda _, Y: np.var(Y))
sim.theta.l1, sim.theta.bounds['l1'] = None, (1E-1, 1)
sim.theta.l2, sim.theta.bounds['l2'] = None, (1E-1, 1)
sim.theta.e1, sim.theta.bounds['e1'] = None, (1E-1, 1.0)
sim.theta.e2, sim.theta.bounds['e2'] = None, (1E-1, 1.0)
# # NOTE! This is a reserved keyword in misoKG. We will generate a list of the same length
# # of the information sources, and use this for scaling our IS.
sim.theta.rho = {"[0, 0]": 1.0, "[0, 1]": 0.96, "[1, 1]": 1.0}
sim.theta.bounds['rho [0, 0]'] = (0.1, 1.0)
sim.theta.bounds['rho [0, 1]'] = (0.1, 1.0)
sim.theta.bounds['rho [1, 1]'] = (0.1, 1.0)
sim.theta.set_hp_names()
sim.primary_rho_opt = False
sim.update_hp_only_with_IS0 = False
sim.update_hp_only_with_overlapped = False
sim.theta.normalize_L = False
sim.theta.normalize_Ks = False
# This was a test feature that actually over-wrote rho to be PSD
# sim.force_rho_psd = True
sim.recommendation_kill_switch = "FAPbBrBrCl_THTO_0"
###################################################################################################
# Start simulation
sim.run()
def run_misokg(run_index):
# Store data for debugging
IS0 = pickle.load(open("enthalpy_N1_R3_Ukcal-mol", 'r'))
IS1 = pickle.load(open("enthalpy_N1_R2_Ukcal-mol", 'r'))
# Generate the main object
sim = Optimizer()
# Assign simulation properties
#sim.hyperparameter_objective = MAP
sim.hyperparameter_objective = MLE
###################################################################################################
# File names
sim.fname_out = "enthalpy_misokg.dat"
sim.fname_historical = None
# Information sources, in order from expensive to cheap
sim.IS = [
lambda h, c, s: -1.0 * IS0[' '.join([''.join(h), c, s])],
lambda h, c, s: -1.0 * IS1[' '.join([''.join(h), c, s])]
]
sim.costs = [
1.0,
0.1,
]
sim.logger_fname = "data_dumps/%d_misokg.log" % run_index
if os.path.exists(sim.logger_fname):
os.system("rm %s" % sim.logger_fname)
os.system("touch %s" % sim.logger_fname)
sim.obj_vs_cost_fname = "data_dumps/%d_misokg.dat" % run_index
sim.mu_fname = "data_dumps/%d_mu_misokg.dat" % run_index
sim.sig_fname = "data_dumps/%d_sig_misokg.dat" % run_index
sim.combos_fname = "data_dumps/%d_combos_misokg.dat" % run_index
sim.hp_fname = "data_dumps/%d_hp_misokg.dat" % run_index
sim.acquisition_fname = "data_dumps/%d_acq_misokg.dat" % run_index
sim.save_extra_files = True
########################################
# Override the possible combinations with the reduced list of IS0
# Because we do this, we should also generate our own historical sample
combos_no_IS = [
k[1] + "Pb" + k[0] + "_" + k[2]
for k in [key.split() for key in IS0.keys()]
]
#sim.historical_nsample = 240
sim.historical_nsample = 10
choices = np.random.choice(combos_no_IS,
sim.historical_nsample,
replace=False)
tmp_data = pal_strings.alphaToNum(choices,
solvents,
mixed_halides=True,
name_has_IS=False)
data = []
for IS in range(len(sim.IS)):
for i, d in enumerate(tmp_data):
h, c, _, s, _ = pal_strings.parseName(pal_strings.parseNum(
d, solvents, mixed_halides=True, num_has_IS=False),
name_has_IS=False)
c = c[0]
data.append([IS] + d + [sim.IS[IS](h, c, s)])
sim.fname_historical = "data_dumps/%d.history" % run_index
pickle.dump(data, open(sim.fname_historical, 'w'))
simple_data = [d for d in data if d[0] == 0]
pickle.dump(simple_data,
open("data_dumps/%d_reduced.history" % run_index, 'w'))
########################################
sim.n_start = 10 # The number of starting MLE samples
sim.reopt = 10
sim.ramp_opt = None
sim.parallel = False
# Possible compositions by default
sim.A = ["Cs", "MA", "FA"]
sim.B = ["Pb"]
sim.X = ["Cl", "Br", "I"]
sim.solvents = copy.deepcopy(solvents)
sim.S = list(set([v["name"] for k, v in sim.solvents.items()]))
sim.mixed_halides = True
sim.mixed_solvents = False
# Parameters for debugging and overwritting
sim.debug = False
sim.verbose = True
sim.overwrite = True # If True, warning, else Error
sim.acquisition = getNextSample_misokg
# Functional forms of our mean and covariance
# MEAN: 4 * mu_alpha + mu_zeta
# COV: sig_alpha * |X><X| + sig_beta * I_N + sig_zeta + MaternKernel(S, weights, sig_m)
SCALE = [2.0, 4.0][int(sim.mixed_halides)]
# _1, _2, _3 used as dummy entries
def mean(X, Y, theta):
mu = np.array([SCALE * theta.mu_alpha + theta.mu_zeta for _ in Y])
return mu
sim.mean = mean
def cov_old(X, Y, theta):
A = theta.sig_alpha * np.dot(
np.array(X)[:, 1:-3],
np.array(X)[:, 1:-3].T)
B = theta.sig_beta * np.diag(np.ones(len(X)))
C = theta.sig_zeta
D = mk52(np.array(X)[:, -3:-1], [theta.l1, theta.l2], theta.sig_m)
return theta.rho_matrix(X) * (A + B + C + D)
def cov_old2(X, Y, theta):
A = theta.sig_alpha * np.dot(
np.array(X)[:, 1:-3],
np.array(X)[:, 1:-3].T)
B = theta.sig_beta * np.diag(np.ones(len(X)))
C = theta.sig_zeta
D = mk52(np.array(X)[:, -3:-1], [theta.l1, theta.l2], theta.sig_m)
return theta.rho_matrix(X, use_psd=True) * (A + B + C + D)
def cov_new(X, Y, theta):
# Get a list of all unique X, removing initial IS identifier
X0 = []
for x in X:
if not any(
[all([a == b for a, b in zip(x[1:], xchk)]) for xchk in X0]):
X0.append(x[1:])
A = theta.sig_alpha * np.dot(
np.array(X0)[:, :-3],
np.array(X0)[:, :-3].T)
B = theta.sig_beta * np.diag(np.ones(len(X0)))
C = theta.sig_zeta
D = mk52(np.array(X0)[:, -3:-1], [theta.l1, theta.l2], theta.sig_m)
Kx = A + B + C + D
L = np.array([
np.array([
theta.rho[str(sorted([i, j]))] if i >= j else 0.0
for j in range(theta.n_IS)
]) for i in range(theta.n_IS)
])
# Normalize L to stop over-scaling values small
L = L / np.linalg.norm(L)
# Force it to be positive semi-definite
Ks = L.dot(L.T)
return np.kron(Ks, Kx)
#K = np.kron(Ks, Kx)
# Now, we get the sub-covariance matrix for the specified sampled X and Y
indices = []
for l in range(theta.n_IS):
for i, x in enumerate(X0):
test = [l] + list(x)
if any(
[all([a == b for a, b in zip(test, xchk)]) for xchk in X]):
indices.append(l * len(X0) + i)
K_local = K[np.ix_(indices, indices)]
return K_local
sim.cov = cov_new
sim.theta.bounds = {}
sim.theta.mu_alpha, sim.theta.bounds['mu_alpha'] = None, (
1E-3, lambda _, Y: max(Y))
sim.theta.sig_alpha, sim.theta.bounds['sig_alpha'] = None, (
1E-2, lambda _, Y: 10.0 * np.var(Y))
sim.theta.sig_beta, sim.theta.bounds['sig_beta'] = None, (
1E-2, lambda _, Y: 10.0 * np.var(Y))
sim.theta.mu_zeta, sim.theta.bounds['mu_zeta'] = None, (
1E-3, lambda _, Y: max(Y))
sim.theta.sig_zeta, sim.theta.bounds['sig_zeta'] = None, (
1E-2, lambda _, Y: 10.0 * np.var(Y))
sim.theta.sig_m, sim.theta.bounds['sig_m'] = None, (1E-2,
lambda _, Y: np.var(Y))
sim.theta.l1, sim.theta.bounds['l1'] = None, (1E-1, 1)
sim.theta.l2, sim.theta.bounds['l2'] = None, (1E-1, 1)
# # NOTE! This is a reserved keyword in misoKG. We will generate a list of the same length
# # of the information sources, and use this for scaling our IS.
# sim.theta.rho = {"[0, 0]": 1, "[0, 1]": None, "[1, 1]": 1}
# sim.theta.bounds['rho [0, 1]'] = (-1.0, 1.0)
# sim.theta.bounds['rho [0, 0]'] = (1, 1)
# sim.theta.bounds['rho [1, 1]'] = (1, 1)
sim.theta.rho = {"[0, 0]": None, "[0, 1]": None, "[1, 1]": None}
sim.theta.bounds['rho [0, 0]'] = (0.1, 1.0)
sim.theta.bounds['rho [0, 1]'] = (0.1, 1.0)
sim.theta.bounds['rho [1, 1]'] = (0.1, 1.0)
sim.theta.set_hp_names()
sim.primary_rho_opt = False
#sim.update_hp_only_with_IS0 = True
sim.update_hp_only_with_overlapped = True
###################################################################################################
# Start simulation
sim.run()