def forward_sample_from_prior(
inf_seed,
n_samples,
M_c,
T,
probe_columns=(0, ),
ROW_CRP_ALPHA_GRID=(),
COLUMN_CRP_ALPHA_GRID=(),
S_GRID=(),
MU_GRID=(),
do_multiprocessing=True,
N_GRID=default_n_grid,
):
helper = functools.partial(
_forward_sample_from_prior,
M_c=M_c,
T=T,
probe_columns=probe_columns,
ROW_CRP_ALPHA_GRID=ROW_CRP_ALPHA_GRID,
COLUMN_CRP_ALPHA_GRID=COLUMN_CRP_ALPHA_GRID,
S_GRID=S_GRID,
MU_GRID=MU_GRID,
N_GRID=N_GRID,
)
cpu_count = 1 if not do_multiprocessing else multiprocessing.cpu_count()
with gu.MapperContext(do_multiprocessing) as mapper:
seeds = numpy.random.randint(32676, size=cpu_count)
n_samples_list = gu.divide_N_fairly(n_samples, cpu_count)
forward_sample_data_list = mapper(helper, zip(seeds, n_samples_list))
forward_sample_data = condense_diagnostics_data_list(
forward_sample_data_list)
return forward_sample_data
def run_posterior_chains(
M_c,
T,
num_chains,
num_iters,
probe_columns,
row_crp_alpha_grid,
column_crp_alpha_grid,
s_grid,
mu_grid,
N_GRID=default_n_grid,
CT_KERNEL=0,
):
# run geweke: transition-erase loop
helper = functools.partial(
run_posterior_chain,
M_c=M_c,
T=T,
num_iters=num_iters,
probe_columns=probe_columns,
ROW_CRP_ALPHA_GRID=row_crp_alpha_grid,
COLUMN_CRP_ALPHA_GRID=column_crp_alpha_grid,
S_GRID=s_grid,
MU_GRID=mu_grid,
N_GRID=N_GRID,
CT_KERNEL=CT_KERNEL,
# this breaks with multiprocessing
plot_rand_idx=(num_chains == 1),
)
seeds = list(range(num_chains))
do_multiprocessing = num_chains != 1
with gu.MapperContext(do_multiprocessing) as mapper:
diagnostics_data_list = mapper(helper, seeds)
pass
return diagnostics_data_list
def get_log_density_series(values, bins):
bin_widths = numpy.diff(bins)
#
with gu.MapperContext() as mapper:
counts = mapper(get_count, [(el, bins) for el in values])
pass
counts = numpy.vstack(counts).cumsum(axis=0)
#
ratios = counts / numpy.arange(1., len(counts) + 1.)[:, numpy.newaxis]
densities = ratios / bin_widths[numpy.newaxis, :]
log_densities = numpy.log(densities)
return log_densities
def get_fixed_gibbs_kl_series(forward, not_forward):
forward, not_forward = make_same_length(forward, not_forward)
forward, not_forward = map(numpy.array, (forward, not_forward))
grid = numpy.array(sorted(set(forward).union(not_forward)))
kls = numpy.repeat(numpy.nan, len(forward))
try:
bins = numpy.append(grid, grid[-1] + numpy.diff(grid)[-1])
#
log_true_series = get_log_density_series(forward, bins)
log_inferred_series = get_log_density_series(not_forward, bins)
arg_tuples = [(grid, x, y)
for x, y in zip(log_true_series, log_inferred_series)]
with gu.MapperContext() as mapper:
kls = mapper(_get_kl_tuple, arg_tuples)
pass
except Exception as e:
# this definitley happens if len(grid) == 1; as in column crp alpha for
# single column model
pass
return kls