def make_first_proposal(log_target,
# MCMC parameters
k=10, N_burn_in=10**3, N_MCMC=2*10**4, N_adapt=500, scale=.1, upper=3., lower=-3.,
# long_patch parameter
K_g=15, critical_r=2.,
N_thin=10):
'''Run a black box algorithm to obtain a gaussian mixture suitable
to importance sample the target. Also return samples obtained
by multiple (``k``) Markov-chains and the `GaussianInference` object.
``log_target`` is the log of the target function. For additional
parameters refer to the functions ``create_mcmc_samples``,
``make_long_patch_gaussian_mixture`` and the code in this function.
Dimension is read as ``log_target.dim``
'''
mcmc_data = create_mcmc_samples(log_target, k, N_burn_in, N_MCMC, N_adapt, scale, upper, lower)
mcmcmix = make_long_patch_gaussian_mixture(mcmc_data, K_g, critical_r)
vb = pypmc.mix_adapt.variational.GaussianInference(np.vstack(mcmc_data)[::N_thin], initial_guess=mcmcmix)
vb.run(1000, prune=.5*len(vb.data)/len(mcmcmix), rel_tol=1e-10, abs_tol=1e-5, verbose=True)
return vb.make_mixture(), mcmc_data, vb
from load_save import save_first_proposal
# load parameters
from params_pure_variational import params
locals().update(params)
# load the target function
from target import LogTarget; log_target = LogTarget(dim)
# load pypmc
import pypmc
from pypmc.tools.convergence import perp, ess
from pypmc.density.mixture import create_gaussian_mixture
# create data from markov-chains
values = create_mcmc_samples(log_target, k, N_burn_in, N_MCMC, N_adapt)
# ***********************************************************************
# ****************** nothing above should be changed ! ******************
# ***********************************************************************
# developer notes:
# save all parameter you want to save as:
# params.update( (('param_1', value_1), (param_2, value_2)) )
# params will be written into the database and into the proposal as
# attribute `creation_parameters`
# example: save the number of samples used for the perp and ess estimate; save the command line call
N_perp_ess = 10**5
params.update( [('N_perp_ess', N_perp_ess), ('sys.argv', sys.argv)] )
