def runGibbs(prior, posterior_object_for_sample, ndim, niters, niters_burn, outdir, snapshot=False, datafname=None, diagnose=False):
print('-------- BEGINNING BURN-IN CHAIN --------')
print('burn in iterations: ', niters_burn)
prior = prior
ndat = posterior_object_for_sample.ndat
# set up empty D column-stacked latent variable array
D = []
for i in range(ndat):
D.append(np.zeros((1,3)))
# empty chain for parameter vectors
chains = []
loglike = 0
param = prior # start parameter vector as the prior vector
n_accept_cosmo = 0.0
n_accept_B = 0.0
chain_tab = pd.DataFrame(columns=['alpha', 'beta', 'rx', 'rc', 'sigma_res', 'cstar', 'xstar', 'mstar', 'omegam', 'w', 'h'])
entry_list = []
# create positive semi-definite covariance matrix for proposal distributions
from sklearn import datasets
#cov_proposal_cosmo = datasets.make_spd_matrix(n_dim = 2, random_state=None) * 0.1 # EDIT: just sample omegam, omegade
cov_proposal_cosmo = np.array([[5, -0.02],
[-0.02, 1]]) * (1./80) * ((2.38**2)/2) # estimated from previous inference
# now compute proposal sigmas for SALT-2 params
#cov_proposal_B = datasets.make_spd_matrix(n_dim = 2, random_state=None) * 0.1
cov_proposal_B = np.array([[0.02, -0.023],
[-0.023, 0.2]]) * (2.38**2 / 2) # estimated from inference
# We run our initial chains to get an estimate of the covariance between cosmo_params and cov between B_params
for iter in range(niters_burn):
param1,n_accept_cosmo,loglike,log_corr = step_one(posterior_object_for_sample, loglike, param, ndim, cov_proposal_cosmo, n_accept_cosmo)
param2,n_accept_B,loglike,log_corr = step_two(posterior_object_for_sample, loglike, param1, ndim, cov_proposal_B, n_accept_B)
D,param3 = step_three(posterior_object_for_sample, D, param2, ndim) # update D
param4 = step_four(posterior_object_for_sample, D, param3, ndim)
param5 = step_five(posterior_object_for_sample, D, param4, ndim)
param6 = step_six(posterior_object_for_sample, D, param5, ndim)
entry_list.append(pd.Series(param6, index=chain_tab.columns)) # move on in sample space
param = param6
#print(param)
#print('proposal estimation acceptance fraction = ', n_accept_cosmo / niters)
# add to chains dataframe
chain_tab = chain_tab.append(entry_list, ignore_index=True)
# compute proposal sigmas for C and B:
omegam = chain_tab['omegam'].values
w = chain_tab['w'].values
h = chain_tab['h'].values
alpha = chain_tab['alpha'].values
beta = chain_tab['beta'].values
cosmo_proposal_cov = np.cov(np.matrix([omegam, w])) # EDIT: just sample omegam, omegade
B_proposal_cov = np.cov(np.matrix([alpha, beta]))
print('cosmo proposal covariance: ', cosmo_proposal_cov)
#print('B proposal cov: ', B_proposal_cov)
print('-------- STARTING ACTUAL SAMPLER CHAIN -------- ')
#print('gibbs sampling for {} iterations'.format(niters))
# set up empty D column-stacked latent variable array
D = []
for i in range(ndat):
D.append(np.zeros((1,3)))
# empty chain for parameter vectors
chains = []
loglike = 0
D_chain = [] # for making diagnostic plot for subset of SN1a
# make prior cube
cube = gibbs_library.makePriorCube(ndim)
param = gibbs_library.vanillaPrior(cube) # start with prior again
# reset acceptance fraction
n_accept_cosmo = 0.0
n_accept_B = 0.0
true_param = [0.13, 2.56,
1.0, 0.1, 0.1, 0., 0., -19.3, 0.3, 0.7, 0.72]
param[2:8] = true_param[2:8]
print('og param', param)
# open write file in case the PBS job gets killed part way:
fname = outdir + 'post_chains_in_prog.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for iter in range(niters):
#t0 = time.time()
param,n_accept_cosmo,loglike,log_correction = step_one(posterior_object_for_sample, loglike, param, ndim, cosmo_proposal_cov, n_accept_cosmo)
#t1 = time.time()
#print('step 1 comp time: ', t1-t0)
# t0 = time.time()
param,n_accept_B,loglike,log_correction = step_two(posterior_object_for_sample, loglike, param, ndim, cov_proposal=B_proposal_cov, n_accept_B=n_accept_B)
#t1 = time.time()
# print('step 2 comp time: ', t1-t0)
# t0 = time.time()
#param0 = param
#D,param0 = step_three(posterior_object_for_sample, D, param0, ndim) # update D
# t1 = time.time()
# print('step 3 comp time: ', t1-t0)
#param = step_four(posterior_object_for_sample, D, param, ndim)
#param = step_five(posterior_object_for_sample, D, param, ndim)
#param = step_six(posterior_object_for_sample, D, param, ndim)
#print(param)
loglike = [loglike]
param_loglike = param + loglike + [log_correction]
chains.append(np.asarray(param_loglike)) # move on in sample space
if iter % 10 == 0:
print('current parameter vector \nand likelihood for iter {}: '.format(iter), param_loglike)
if diagnose == True:
# latent vals from chain
c = D[0::3]
x1 = D[1::3]
mB = D[2::3]
# take all SN1a
#c = c[::10]
#x1 = x1[::10]
#mB = mB[::10]
D_save = [] # thinned latent vector to be saved
for i in range(len(c)):
D_save.append(c[i])
D_save.append(x1[i])
D_save.append(mB[i])
D_chain.append(D_save) # add to trace chain
wr.writerow(param_loglike) # write the parameter step as we go.
if snapshot == True:
if iter == niters-1:
latent_plots.plot_attributes(D, param, iter)
print('sampling complete: cosmo param acceptance fraction = ', n_accept_cosmo/niters)
print('sampling complete: SALT-2 param acceptance fraction = ', n_accept_B/niters)
fname = outdir + 'post_chains.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in chains:
wr.writerow(i)
if diagnose == True:
fname = outdir + 'D_latent.csv'
# save chain output for diagnostic purposes
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in D_chain:
wr.writerow(i)
return n_accept_cosmo / niters, n_accept_B / niters
def runGibbs(prior,
posterior_object_for_sample,
ndim,
niters,
outdir,
snapshot=False,
datafname=None,
diagnose=False):
ndat = posterior_object_for_sample.ndat
# set up empty D column-stacked latent variable array
D = []
for i in range(ndat):
D.append(np.zeros((1, 3)))
# empty chain for parameter vectors
chains = []
param = prior # start parameter vector as the prior vector
# evaluate log-likelihood at the original parameter vector
loglike = posterior_object_for_sample.log_like_selection(param)
# define acceptance fractions
n_accept_cosmo = 0.0 # step 1
n_accept_B = 0.0 # step 2
n_accept_dstar = 0.0 # step 3
n_accept_sigmares = 0.0 # step 4
n_accept_rx = 0.0 # step 5
n_accept_rc = 0.0 # step 6
# define jump proposals
cov_proposal_cosmo = np.array(
[[0.00862212, 0.01333633], [0.01333633, 0.02501767]]) * (
(2.38**2) / 2) * 10 # EDIT: just sample omegam, omegade
cov_proposal_B = np.array([[7.51338227e-05, -7.81496651e-07],
[-7.81496651e-07, 7.63401552e-03]
]) * (2.38**2 / 2) * 10
cov_proposal_p = np.array([[
2.26136092e-05, 4.55322597e-06, -8.29879371e-06
], [
4.55322597e-06, 2.29976669e-03, 3.91206305e-05
], [-8.29879371e-06, 3.91206305e-05, 4.14424789e-04]]) * (2.38**2 / 3) * 10
cov_proposal_sr = 0.00013154 * (2.38**2) * 10
cov_proposal_rx = 0.00119617 * (2.38**2) * 10
cov_proposal_rc = 1.27621525e-05 * (2.38**2) * 10
print('-------- STARTING ACTUAL SAMPLER CHAIN -------- ')
print('gibbs sampling for {} iterations'.format(niters))
# empty chain for parameter vectors
chains = []
D_chain = [] # for making diagnostic plot for subset of SN1a
# start with prior again
true_param = [0.13, 2.56, 1.0, 0.1, 0.1, 0., 0., -19.3, 0.3, 0.7, 0.72]
# open write file in case the PBS job gets killed part way:
fname = outdir + 'post_chains_in_prog.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for iter in range(niters):
param, n_accept_cosmo, loglike = step_one(
posterior_object_for_sample, loglike, param, ndim,
cov_proposal_cosmo, n_accept_cosmo)
param, n_accept_B, loglike = step_two(posterior_object_for_sample,
loglike,
param,
ndim,
cov_proposal=cov_proposal_B,
n_accept_B=n_accept_B)
D, param, n_accept_dstar, loglike = step_three(
posterior_object_for_sample, loglike, D, param, ndim,
cov_proposal_p, n_accept_dstar) # update D
param, n_accept_sigmares, loglike = step_four(
posterior_object_for_sample, loglike, D, param, ndim,
cov_proposal_sr, n_accept_sigmares)
param, n_accept_rx, loglike = step_five(
posterior_object_for_sample, loglike, D, param, ndim,
cov_proposal_rx, n_accept_rx)
param, n_accept_rc, loglike = step_six(posterior_object_for_sample,
loglike, D, param, ndim,
cov_proposal_rc,
n_accept_rc)
param_loglike = param + [loglike]
chains.append(np.asarray(param_loglike)) # move on in sample space
if iter % 1 == 0:
print(
'current parameter vector \nand likelihood for iter {}: '.
format(iter), param_loglike)
if diagnose == True:
# latent values from chain
c = D[0::3]
x1 = D[1::3]
mB = D[2::3]
D_save = [] # thinned latent vector to be saved
for i in range(len(c)):
D_save.append(c[i])
D_save.append(x1[i])
D_save.append(mB[i])
D_chain.append(D_save) # add to trace chain
wr.writerow(param_loglike) # write the parameter step as we go.
if snapshot == True:
if iter == niters - 1:
latent_plots.plot_attributes(D, param, iter)
acc_numbs = [
n_accept_cosmo, n_accept_B, n_accept_dstar, n_accept_sigmares,
n_accept_rx, n_accept_rc
]
print('acc numbs: ', acc_numbs)
acceptance_fracs = []
acceptance_fracs = [i / niters for i in acc_numbs]
print('sampling complete')
for step in range(len(acc_numbs)):
print('acceptance fraction for step {}: '.format(step + 1),
acceptance_fracs[step])
fname = outdir + 'post_chains.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in chains:
wr.writerow(i)
if diagnose == True:
fname = outdir + 'D_latent.csv'
# save chain output for diagnostic purposes
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in D_chain:
wr.writerow(i)
return acceptance_fracs
def runGibbs(prior, posterior_object_for_sample, ndim, niters, niters_burn, outdir, snapshot=False, datafname=None, diagnose=False):
print('-------- BEGINNING BURN-IN CHAIN --------')
print('burn in iterations: ', niters_burn)
param = prior
ndat = posterior_object_for_sample.ndat
# set up empty D column-stacked latent variable array
D = []
for i in range(ndat):
D.append(np.zeros((1,3)))
# empty chain for parameter vectors
chains = []
# compute loglike to initialize sampling
loglike = posterior_object_for_sample.log_likelihood(param)
param = prior # start parameter vector as the prior vector
n_accept_cosmo = 0.0
n_accept_B = 0.0
chain_tab = pd.DataFrame(columns=['alpha', 'beta', 'rx', 'rc', 'sigma_res', 'cstar', 'xstar', 'mstar', 'omegam', 'w', 'h'])
entry_list = []
# Estimated from MultiNest runs
cov_proposal_cosmo = np.array([[0.00862212, 0.01333633],
[0.01333633, 0.02501767]]) * (2.38**2 / 2) # EDIT: just sample omegam, omegade
cov_proposal_B = np.array([[ 7.51338227e-05, -7.81496651e-07],
[-7.81496651e-07, 7.63401552e-03]]) * (2.38**2 / 2)
# We run our initial chains to get an estimate of the covariance between cosmo_params and cov between B_params
for iter in range(niters_burn):
param,n_accept_cosmo,loglike = step_one(posterior_object_for_sample, D, loglike, param, ndim, cov_proposal_cosmo, n_accept_cosmo)
param,n_accept_B,loglike = step_two(posterior_object_for_sample, D, loglike, param, ndim, cov_proposal_B, n_accept_B)
D,param = step_three(posterior_object_for_sample, D, param, ndim) # update D
param = step_four(posterior_object_for_sample, D, param, ndim)
param = step_five(posterior_object_for_sample, D, param, ndim)
param = step_six(posterior_object_for_sample, D, param, ndim)
entry_list.append(pd.Series(param, index=chain_tab.columns)) # move on in sample space
#print(param)
#print('proposal estimation acceptance fraction = ', n_accept_cosmo / niters)
# add to chains dataframe
chain_tab = chain_tab.append(entry_list, ignore_index=True)
# compute proposal sigmas for C and B:
omegam = chain_tab['omegam'].values
w = chain_tab['w'].values
h = chain_tab['h'].values
alpha = chain_tab['alpha'].values
beta = chain_tab['beta'].values
cosmo_proposal_cov = np.cov(np.matrix([omegam, w])) # EDIT: just sample omegam, omegade
B_proposal_cov = np.cov(np.matrix([alpha, beta]))
print('cosmo proposal covariance: ', cosmo_proposal_cov)
#print('B proposal cov: ', B_proposal_cov)
print('-------- STARTING ACTUAL SAMPLER CHAIN -------- ')
#print('gibbs sampling for {} iterations'.format(niters))
# set up empty D column-stacked latent variable array
D = []
for i in range(ndat):
D.append(np.zeros((1,3)))
# empty chain for parameter vectors
chains = []
# reset log-likelihood
loglike = posterior_object_for_sample.log_likelihood(prior)
D_chain = [] # for making diagnostic plot for subset of SN1a
# reset acceptance fraction
n_accept_cosmo = 0.0
n_accept_B = 0.0
# open write file in case the PBS job gets killed part way:
fname = outdir + 'post_chains_in_prog.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for iter in range(niters):
param,n_accept_cosmo,loglike = step_one(posterior_object_for_sample, D, loglike, param, ndim, cosmo_proposal_cov, n_accept_cosmo)
param,n_accept_B,loglike = step_two(posterior_object_for_sample, D, loglike, param, ndim, cov_proposal=B_proposal_cov, n_accept_B=n_accept_B)
D,param = step_three(posterior_object_for_sample, D, param, ndim) # update D
param = step_four(posterior_object_for_sample, D, param, ndim)
param = step_five(posterior_object_for_sample, D, param, ndim)
param = step_six(posterior_object_for_sample, D, param, ndim)
param_loglike = param + [loglike]
chains.append(np.asarray(param_loglike)) # move on in sample space
if iter % 100 == 0:
print('current parameter vector \nand likelihood for iter {}: '.format(iter), param_loglike)
if diagnose == True:
# latent vals from chain
c = D[0::3]
x1 = D[1::3]
mB = D[2::3]
D_save = [] # thinned latent vector to be saved
for i in range(len(c)):
D_save.append(c[i])
D_save.append(x1[i])
D_save.append(mB[i])
D_chain.append(D_save) # add to trace chain
wr.writerow(param_loglike) # write the parameter step as we go.
if snapshot == True:
if iter == niters-1:
latent_plots.plot_attributes(D, param, iter)
print('sampling complete: cosmo param acceptance fraction = ', n_accept_cosmo/niters)
print('sampling complete: SALT-2 param acceptance fraction = ', n_accept_B/niters)
fname = outdir + 'post_chains.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in chains:
wr.writerow(i)
fname = outdir + 'D_latent.csv'
# save chain output for diagnostic purposes
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in D_chain:
wr.writerow(i)
return n_accept_cosmo / niters, n_accept_B / niters
def runGibbs(prior, posterior_object_for_sample, ndim, niters, niters_burn, outdir, snapshot=False, datafname=None, diagnose=False):
print('-------- BEGINNING BURN-IN CHAIN --------')
print('burn in iterations: ', niters_burn)
ndat = posterior_object_for_sample.ndat
# set up empty D column-stacked latent variable array
D = []
for i in range(ndat):
D.append(np.zeros((1,3)))
# empty chain for parameter vectors
chains = []
param = prior # start parameter vector as the prior vector
# evaluate log-likelihood at the original parameter vector
loglike = posterior_object_for_sample.log_like_selection(param)
# define acceptance fractions
n_accept_cosmo = 0.0 # step 1
n_accept_B = 0.0 # step 2
n_accept_dstar = 0.0 # step 3
n_accept_sigmares = 0.0 # step 4
n_accept_rx = 0.0 # step 5
n_accept_rc = 0.0 # step 6
chain_tab = pd.DataFrame(columns=['alpha', 'beta', 'rx', 'rc', 'sigma_res', 'cstar', 'xstar', 'mstar', 'omegam', 'w', 'h'])
entry_list = []
# create positive semi-definite covariance matrix for proposal distributions
#from sklearn import datasets
#cov_proposal_cosmo = np.diag([np.random.rand()*5, np.random.rand()*1]) * (1./80) * ((2.38**2)/2) # EDIT: just sample omegam, omegade
#cov_proposal_cosmo = np.array([[5, -0.02],
# [-0.02, 1]]) * (1./80) * ((2.38**2)/2) # estimated from previous inference
# now compute proposal sigmas for SALT-2 params
#cov_proposal_B = np.diag([np.random.rand()*0.02, np.random.rand()*0.2]) * (2.38**2 / 2)
#cov_proposal_B = np.array([[0.02, -0.023],
# [-0.023, 0.2]]) * (2.38**2 / 2) # estimated from inference
# Estimated from MultiNest runs
cov_proposal_cosmo = np.array([[0.00862212, 0.01333633],
[0.01333633, 0.02501767]]) * ((2.38**2)/2) # EDIT: just sample omegam, omegade
cov_proposal_B = np.array([[ 7.51338227e-05, -7.81496651e-07],
[-7.81496651e-07, 7.63401552e-03]]) * (2.38**2 / 2)
# We run our initial chains to get an estimate of the covariance between cosmo_params and cov between B_params
for iter in range(niters_burn):
param1, n_accept_cosmo, loglike = step_one(posterior_object_for_sample, loglike, param, ndim, cov_proposal_cosmo, n_accept_cosmo)
param2, n_accept_B,loglike = step_two(posterior_object_for_sample, loglike, param1, ndim, cov_proposal_B, n_accept_B)
D,param3, n_accept_dstar, loglike = step_three(posterior_object_for_sample, loglike, D, param2, ndim, n_accept_dstar) # update D
param4, n_accept_sigmares, loglike = step_four(posterior_object_for_sample, loglike, D, param3, ndim, n_accept_sigmares)
param5, n_accept_rx, loglike = step_five(posterior_object_for_sample, loglike, D, param4, ndim, n_accept_rx)
param6, n_accept_rc, loglike = step_six(posterior_object_for_sample, loglike, D, param5, ndim, n_accept_rc)
entry_list.append(pd.Series(param6, index=chain_tab.columns)) # move on in sample space
param = param6
# add to chains dataframe
chain_tab = chain_tab.append(entry_list, ignore_index=True)
# compute proposal sigmas for C and B:
omegam = chain_tab['omegam'].values
w = chain_tab['w'].values
h = chain_tab['h'].values
alpha = chain_tab['alpha'].values
beta = chain_tab['beta'].values
#cosmo_proposal_cov = ((2.38)**2 / 2) * np.cov((chain_tab['omegam'], chain_tab['w'])) # EDIT: just sample omegam, omegade
#B_proposal_cov = ((2.38)**2 / 2) * np.cov((chain_tab['alpha'], chain_tab['beta']))
# estimated from multinest
cosmo_proposal_cov = np.array([[0.00862212, 0.01333633],
[0.01333633, 0.02501767]]) * ((2.38**2)/2) # EDIT: just sample omegam, omegade
B_proposal_cov = np.array([[ 7.51338227e-05, -7.81496651e-07],
[-7.81496651e-07, 7.63401552e-03]]) * (2.38**2 / 2)
print('cosmo proposal covariance: ', cosmo_proposal_cov)
print('-------- STARTING ACTUAL SAMPLER CHAIN -------- ')
print('gibbs sampling for {} iterations'.format(niters))
# empty chain for parameter vectors
chains = []
D_chain = [] # for making diagnostic plot for subset of SN1a
# start with prior again
param = prior
# re-calculate log-like
loglike = posterior_object_for_sample.log_like_selection(param)
# reset acceptance fraction
n_accept_cosmo = 0.0 # step 1
n_accept_B = 0.0 # step 2
n_accept_dstar = 0.0 # step 3
n_accept_sigmares = 0.0 # step 4
n_accept_rx = 0.0 # step 5
n_accept_rc = 0.0 # step 6
true_param = [0.13, 2.56,
1.0, 0.1, 0.1, 0., 0., -19.3, 0.3, 0.7, 0.72]
# open write file in case the PBS job gets killed part way:
fname = outdir + 'post_chains_in_prog.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for iter in range(niters):
param,n_accept_cosmo,loglike = step_one(posterior_object_for_sample, loglike, param, ndim, cosmo_proposal_cov, n_accept_cosmo)
param,n_accept_B,loglike = step_two(posterior_object_for_sample, loglike, param, ndim, cov_proposal=B_proposal_cov, n_accept_B=n_accept_B)
D,param,n_accept_dstar,loglike = step_three(posterior_object_for_sample, loglike, D, param, ndim, n_accept_dstar) # update D
param,n_accept_sigmares,loglike = step_four(posterior_object_for_sample, loglike, D, param, ndim, n_accept_sigmares)
param,n_accept_rx,loglike = step_five(posterior_object_for_sample, loglike, D, param, ndim, n_accept_rx)
param,n_accept_rc,loglike = step_six(posterior_object_for_sample, loglike, D, param, ndim, n_accept_rc)
param_loglike = param + [loglike]
chains.append(np.asarray(param_loglike)) # move on in sample space
if iter % 1 == 0:
print('current parameter vector \nand likelihood for iter {}: '.format(iter), param_loglike)
if diagnose == True:
# latent values from chain
c = D[0::3]
x1 = D[1::3]
mB = D[2::3]
D_save = [] # thinned latent vector to be saved
for i in range(len(c)):
D_save.append(c[i])
D_save.append(x1[i])
D_save.append(mB[i])
D_chain.append(D_save) # add to trace chain
wr.writerow(param_loglike) # write the parameter step as we go.
if snapshot == True:
if iter == niters-1:
latent_plots.plot_attributes(D, param, iter)
acc_numbs = [n_accept_cosmo, n_accept_B, n_accept_dstar, n_accept_sigmares, n_accept_rx, n_accept_rc]
print('acc numbs: ', acc_numbs)
acceptance_fracs = []
acceptance_fracs = [i / niters for i in acc_numbs]
print('sampling complete')
for step in range(len(acc_numbs)):
print('acceptance fraction for step {}: '.format(step + 1), acceptance_fracs[step])
fname = outdir + 'post_chains.csv'
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in chains:
wr.writerow(i)
if diagnose == True:
fname = outdir + 'D_latent.csv'
# save chain output for diagnostic purposes
with open(fname, 'w') as myfile:
wr = csv.writer(myfile, quoting=csv.QUOTE_ALL)
for i in D_chain:
wr.writerow(i)
return acceptance_fracs