def run_pso(lum, drug_conc, time, weights):
# create global versions of lum, drug_conc, and time variables that can be used in cost function
global _lum, _drug_conc, _time, _weights
_lum = np.array(lum)
_drug_conc = np.array(drug_conc)
_time = np.array(time)
_weights = np.array(weights)
##########
# _dc, idx = np.unique(_drug_conc, return_index=True)
# _lum_avg = np.array([np.mean(_lum[idx[i]:idx[i+1]], axis=0) for i in range(len(idx)-1)] +
# [np.mean(_lum[idx[-1]:], axis=0)])
# colors = cmap(np.linspace(0, 1, len(_dc)))
# plt.figure()
# u_idx = 0
# for i in range(len(_lum)):
# if i <= idx[-1] and i == idx[u_idx+1]:
# u_idx += 1
# plt.plot(_time[i], _lum[i], 'o', color=colors[u_idx])
# plt.plot(_time[idx[u_idx]], _lum_avg[u_idx], '--', lw=2, color=colors[u_idx])
# plt.show()
##########
# estimate slope m and y-intercept nC0 from control luminescence data
global _m, _nC0
_m = 0
_nC0 = 0
# find indices of control experiments
indices = np.where(_drug_conc == 0)[0]
for idx in indices:
fit = linregress(_time[idx], _lum[idx])
_m += fit.slope
_nC0 += fit.intercept
_m /= len(indices)
_nC0 /= len(indices)
# run PSO to estimate the other three parameters: kdiv-kdeath, kdiv*-kdeath*, koff/kon
pso = PSO(save_sampled=True, verbose=True, shrink_steps=False)
pso.set_start_position([2, 2, 2])
# allows particles to move +/- 2 orders of magnitude
pso.set_bounds(2)
# sets maximum speed that a particle can travel
pso.set_speed(-0.1, 0.1)
pso.run(
num_particles=100,
num_iterations=1000,
stop_threshold=0,
num_processors=1,
max_iter_no_improv=1000,
cost_function=cost
)
return np.array([_m, _nC0] + list(pso.best.pos)) # [m, nC0, kdiv-kdeath, kdiv*-kdeath*, koff/kon]
def test_population_creation():
known_sol = [8.6998, 6.7665]
pso = PSO(start=[0, 0], verbose=False, shrink_steps=False)
pso.set_bounds(lower=[-100, -100], upper=[100, 100])
pso.set_speed(-10, 10)
pso.run(num_iterations=100, num_particles=10, cost_function=h1, )
pso.return_ranked_populations()
error = np.sum((pso.best.pos - known_sol) ** 2)
print('True value: [8.6998, 6.7665]. Found:{0}. Error^2 = {1}'.format(
pso.best.pos, error))
assert (error < 0.1)
def run_pso(run, iterations, bd):
pso = PSO(save_sampled=False, verbose=True, shrink_steps=False)
#pso.set_cost_function(costfunction)
pso.set_start_position(starting_position)
pso.set_bounds(bd)
pso.set_speed(-.1, .1)
pso.run(num_particles=20, num_iterations=iterations, stop_threshold=1e-5,
cost_function=costfunction, max_iter_no_improv=50,
num_processors=20, save_samples=True)
param_sets = convert_to_flat_array(pso, model)
#print(param_sets)
param_sets.to_csv('run'+run+'.csv')
def test_population_creation():
pso = PSO(cost_function=h1, start=[10, 0], verbose=False)
pso.set_bounds(lower=[-100, -100], upper=[100, 100])
pso.run(num_iterations=100, num_particles=10)
pso.return_ranked_populations()
error = np.sum((pso.best - [8.6998, 6.7665]) ** 2)
print('True value: [8.6998, 6.7665]. Found:{0}. Error^2 = {1}'.format(pso.best, error))
assert (error < 0.1)
def __call__(self, pso_kwargs=None,
cost_type='norm_logpdf', custom_cost=None):
"""Call the SwarmIt instance to construct to instance of the NestedSampling object.
Args:
pso_kwargs (dict): Dictionary of any additional optional keyword
arguments to pass to the PSO object constructor.
Defaults to dict().
cost_type (str): Define the type of cost estimator
to use. Options are 'norm_logpdf'=>Compute the cost using
the normal distribution estimator, 'mse'=>Compute the
cost using the negative mean squared error estimator,
'sse'=>Compute the cost using the negative sum of
squared errors estimator. Defaults to 'norm_logpdf'.
Returns:
type: Description of returned object.
"""
if pso_kwargs is None:
pso_kwargs = dict()
# self.ns_version = ns_version
self._pso_kwargs = pso_kwargs
#population_size = pso_population_size
if cost_type == 'mse':
cost = self.mse_cost
elif cost_type == 'sse':
cost = self.sse_cost
elif (cost_type == 'custom') and (custom_cost is not None):
self.set_custom_cost(custom_cost)
cost = self.custom_cost
else:
cost = self.norm_logpdf_cost
# Construct the PSO
if 'save_sampled' not in pso_kwargs.keys():
pso_kwargs['save_sampled'] = False
if 'verbose' not in pso_kwargs.keys():
pso_kwargs['verbose'] = False
pso = PSO(**pso_kwargs)
pso.set_start_position(self._starting_position)
pso.set_cost_function(cost)
pso.set_bounds(lower=self._lower, upper=self._upper)
pso.set_speed(-.25, .25)
return pso
def run_example():
# create PSO object
pso = PSO(save_sampled=False, verbose=True, shrink_steps=False)
pso.set_start_position(starting_position)
# allows particles to move +/- 2 orders of magnitude
pso.set_bounds(2)
# sets maximum speed that a particle can travel
pso.set_speed(-.1, .1)
pso.run(num_particles=24,
num_iterations=100,
stop_threshold=1e-5,
num_processors=18,
max_iter_no_improv=20,
cost_function=likelihood)
display(pso.best.pos, save_name='best_fit')
np.savetxt("pso_fit_for_model.csv", pso.best.pos)
create_gif_of_model_training(pso)
def run_pso(run, iterations, bd, outdir='', suffix=''):
pso = PSO(save_sampled=True, verbose=True, shrink_steps=False)
#pso.set_cost_function(costfunction)
pso.set_start_position(starting_params)
pso.set_bounds(bd)
pso.set_speed(-0.1, 0.1)
pso.run(num_particles=20,
num_iterations=iterations,
stop_threshold=1e-5,
cost_function=costfunction,
max_iter_no_improv=500,
num_processors=20,
save_samples=True)
param_sets = convert_to_flat_array(pso, model)
#print(param_sets)
outfile = os.path.join(outdir, 'run%d%s.csv' % (run, suffix))
param_sets.to_csv(outfile, index_label='iter')
def test_himmelblau():
""" test to see if PSO can find simple minimum
"""
minimums = [[3.0, 2.0],
[-2.805118, 3.131312],
[-3.779310, -3.283186],
[3.584428, -1.848126]]
pso = PSO(cost_function=himmelblau, start=[10, 0], verbose=False)
pso.set_bounds(lower=[-100, -100], upper=[100, 100])
pso.run(num_iterations=100, num_particles=10)
good_min = False
for i in minimums:
if np.sum((pso.best - i) ** 2) < .1:
good_min = True
error = np.sum((pso.best - i) ** 2)
found_min = i
if good_min:
print('Found minimum')
print('True value: {0}. Found:{1}. Error^2 = {2}'.format(found_min, pso.best, error))
def run_example():
# Here we initial the class
# We must proivde the cost function and a starting value
optimizer = PSO(cost_function=obj_function, start=start_position, verbose=True)
# We also must set bounds. This can be a single scalar or an array of len(start_position)
optimizer.set_bounds(parameter_range=3)
optimizer.set_speed(speed_min=-.5, speed_max=.5)
optimizer.run(num_particles=25, num_iterations=100)
if plot:
display(start_position, optimizer.best)
print("Original values {0}".format(log10_original_values ** 10))
print("Starting values {0}".format(start_position ** 10))
print("Best PSO values {0}".format(optimizer.best ** 10))
fig = plt.figure()
fig.add_subplot(221)
plt.scatter(log10_original_values[0], log10_original_values[1], marker='>', color='b', label='ideal')
plt.scatter(start_position[0], start_position[1], marker='^', color='r', label='start')
plt.scatter(optimizer.history[:, 0], optimizer.history[:, 1], c=optimizer.values, cmap=plt.cm.coolwarm)
fig.add_subplot(223)
plt.scatter(log10_original_values[0], log10_original_values[2], marker='>', color='b', label='ideal')
plt.scatter(start_position[0], start_position[2], marker='^', color='r', label='start')
plt.scatter(optimizer.history[:, 0], optimizer.history[:, 2], c=optimizer.values, cmap=plt.cm.coolwarm)
fig.add_subplot(222)
plt.scatter(log10_original_values[1], log10_original_values[2], marker='>', color='b', label='ideal')
plt.scatter(start_position[1], start_position[2], marker='^', color='r', label='start')
plt.scatter(optimizer.history[:, 1], optimizer.history[:, 2], c=optimizer.values, cmap=plt.cm.coolwarm)
fig.add_subplot(224)
plt.legend(loc=0)
plt.colorbar()
plt.tight_layout()
plt.savefig('population.png')
def run_example():
# Runs the cost function to calculate error between model and data
print("Error at start = {}".format(likelihood(starting_position)[0]))
# Displays the model with defaul positions
display(starting_position, save_name='starting_position')
# create PSO object
pso = PSO(save_sampled=False, verbose=True, num_proc=4)
pso.set_cost_function(likelihood)
pso.set_start_position(starting_position)
# allows particles to move +/- 2 orders of magnitude
pso.set_bounds(2)
# sets maximum speed that a particle can travel
pso.set_speed(-.25, .25)
pso.run(num_particles=25, num_iterations=50, stop_threshold=1e-5)
display(pso.best, save_name='best_fit')
np.savetxt("pso_fit_for_model.csv", pso.best)
def test_mismatched_bounds():
pso = PSO(start=[10, 0], cost_function=himmelblau, verbose=False)
pso.set_bounds(lower=[-100, 0, -100], upper=[100, 100])
pso.run(num_iterations=100, num_particles=10)
def run_example_multiple():
best_pars = np.zeros((100, len(model.parameters)))
counter = 0
for i in range(100):
pso = PSO(save_sampled=False, verbose=False, num_proc=4)
pso.set_cost_function(likelihood)
nominal_random = xnominal + np.random.uniform(-1, 1, len(xnominal))
pso.set_start_position(nominal_random)
pso.set_bounds(2.5)
pso.set_speed(-.25, .25)
pso.run(25, 100)
if pso.best.fitness.values[0] < 0.066:
Y = np.copy(pso.best)
param_values[rates_of_interest_mask] = 10**Y
best_pars[counter] = param_values
counter += 1
print(i, counter)
# display(pso.best)
np.save('jnk3_noASK1_ncalibrated_pars_1h', best_pars)
def run_example():
pso = PSO(save_sampled=False, verbose=True, num_proc=4)
pso.set_cost_function(likelihood)
pso.set_start_position(xnominal)
pso.set_bounds(lower=lower, upper=upper)
pso.set_speed(-.25, .25)
pso.run(25, 200)
display(pso.best)
np.save('calibrated_pars_pso1', pso.best)
def run_example():
# Runs the cost function to calculate error between model and data
print("Error at start = {}".format(likelihood(starting_position)[0]))
# Displays the model with defaul positions
display(starting_position, save_name='starting_position')
# create PSO object
pso = PSO(save_sampled=False, verbose=True, num_proc=4)
pso.set_cost_function(likelihood)
pso.set_start_position(starting_position)
# allows particles to move +/- 2 orders of magnitude
pso.set_bounds(2)
# sets maximum speed that a particle can travel
pso.set_speed(-.25, .25)
pso.run(num_particles=25, num_iterations=50, stop_threshold=1e-5)
display(pso.best, save_name='best_fit')
np.savetxt("pso_fit_for_model.csv", pso.best)
def run_example():
pso = PSO(save_sampled=False, verbose=True, num_proc=4)
pso.set_cost_function(likelihood)
pso.set_start_position(xnominal)
pso.set_bounds(2)
pso.set_speed(-.25, .25)
pso.run(25, 100)
display(pso.best)
def run_pso(run, iterations, bd):
pso = PSO(save_sampled=False,
verbose=True,
shrink_steps=False,
num_proc=14)
pso.set_cost_function(costfunction)
pso.set_start_position(starting_position)
pso.set_bounds(bd)
pso.set_speed(-.1, .1)
pso.run(num_particles=200, num_iterations=iterations, stop_threshold=1e-5)
#print('best pos: ', pso.best.pos)
print('history ', pso.history)
print('run ', run)
print('fit ', pso.best.fitness)
print('all fitness ', pso.values)
np.savetxt("H841_params_" + run + ".txt", 10**pso.history, delimiter=",")
np.savetxt("H841_fit_" + run + ".txt", pso.values, delimiter=",")
def run_example():
# Here we initial the class
# We must proivde the cost function and a starting value
optimizer = PSO(cost_function=obj_function,
start=start_position,
verbose=True)
# We also must set bounds. This can be a single scalar or an array of len(start_position)
optimizer.set_bounds(parameter_range=3)
optimizer.set_speed(speed_min=-.5, speed_max=.5)
optimizer.run(num_particles=25, num_iterations=100)
if plot:
display(start_position, optimizer.best)
print("Original values {0}".format(log10_original_values**10))
print("Starting values {0}".format(start_position**10))
print("Best PSO values {0}".format(optimizer.best**10))
fig = plt.figure()
fig.add_subplot(221)
plt.scatter(log10_original_values[0],
log10_original_values[1],
marker='>',
color='b',
label='ideal')
plt.scatter(start_position[0],
start_position[1],
marker='^',
color='r',
label='start')
plt.scatter(optimizer.history[:, 0],
optimizer.history[:, 1],
c=optimizer.values,
cmap=plt.cm.coolwarm)
fig.add_subplot(223)
plt.scatter(log10_original_values[0],
log10_original_values[2],
marker='>',
color='b',
label='ideal')
plt.scatter(start_position[0],
start_position[2],
marker='^',
color='r',
label='start')
plt.scatter(optimizer.history[:, 0],
optimizer.history[:, 2],
c=optimizer.values,
cmap=plt.cm.coolwarm)
fig.add_subplot(222)
plt.scatter(log10_original_values[1],
log10_original_values[2],
marker='>',
color='b',
label='ideal')
plt.scatter(start_position[1],
start_position[2],
marker='^',
color='r',
label='start')
plt.scatter(optimizer.history[:, 1],
optimizer.history[:, 2],
c=optimizer.values,
cmap=plt.cm.coolwarm)
fig.add_subplot(224)
plt.legend(loc=0)
plt.colorbar()
plt.tight_layout()
plt.savefig('population.png')
log_original_values = np.log10(param_values[rate_mask])
if '__main__' == __name__:
# We will use a best guess starting position for the model,
start_position = log_original_values + \
np.random.uniform(-1, 1,
size=len(log_original_values))
display(start_position, "Before optimization")
plt.tight_layout()
plt.savefig("fit_before_pso.png", bbox_inches='tight')
logger.info("Saving pre-fit figure as fit_before_pso.png")
# Here we initial the class
# We must proivde the cost function and a starting value
optimizer = PSO(start=start_position, verbose=True, shrink_steps=False)
# We also must set bounds of the parameter space, and the speed PSO will
# travel (max speed in either direction)
optimizer.set_bounds(parameter_range=4)
optimizer.set_speed(speed_min=-.05, speed_max=.05)
# Now we run the pso algorithm
optimizer.run(num_particles=50,
num_iterations=500,
num_processors=12,
cost_function=obj_function,
max_iter_no_improv=25)
best_params = optimizer.best.pos
display(best_params, "After optimization")
def test_no_bounds():
pso = PSO(start=[10, 0], cost_function=himmelblau, verbose=False)
pso.run(num_iterations=100, num_particles=10)
def obj_function(params):
params_tmp = np.copy(params)
param_values[rate_mask] = 10 ** params_tmp # don't need to change
result = solver1.run(param_values=param_values)
ysim_array1 = result.observables['MLKLa_obs'][:]
ysim_norm1 = normalize(ysim_array1)
e1 = np.sum((ydata_norm - ysim_norm1) ** 2)
return e1,
def run_example():
print('run_example')
best_pars = np.zeros((1000, len(model.parameters)))
counter = 0
# Here we initial the class
# We must proivde the cost function and a starting value
for i in range(1000):
optimizer = PSO(cost_function=obj_function,start = log10_original_values, verbose=True)
# We also must set bounds. This can be a single scalar or an array of len(start_position)
optimizer.set_bounds(parameter_range=2)
optimizer.set_speed(speed_min=-.25, speed_max=.25)
optimizer.run(num_particles=75, num_iterations=25)
best_pars[i] = optimizer.best
print(optimizer.best)
# print(i, counter)
np.save('optimizer_best_75_50_100TNF',best_pars)
if '__main__' == __name__:
run_example()
def test_missing_cost_function():
pso = PSO(start=[10, 0], verbose=False)
pso.set_bounds(lower=[-100, -100], upper=[100, 100])
pso.run(num_iterations=100, num_particles=10)
# USER-Set: must appropriately update cost function!
def cost(position):
Y = np.copy(position)
param_values[calibrate_mask] = 10**Y
sim = solver.run(param_values=param_values).all
logp_data = np.sum(like_data.logpdf(sim['observable']))
if np.isnan(logp_data):
logp_data = np.inf
return -logp_data,
# Setup the particle swarm optimization run
# Set the number of particles in the swarm.
num_particles = 25
# Set the number of iterations for PSO run.
num_iterations = 50
# Construct the optimizer
pso = PSO(save_sampled=False, verbose=True, num_procs=1)
pso.set_cost_function(cost)
starting_position = swarm_param.centers()
pso.set_start_position(starting_position)
pso.set_bounds(lower=swarm_param.lower(), upper=swarm_param.upper())
# sets maximum speed that a particle can travel
pso.set_speed(-.25, .25)
# run it
pso.run(num_particles, num_iterations, stop_threshold=1e-5)
print("Best parameters: ", pso.best)
