def test_proposal_COB_complete():
n, n_obs = (5, 500)
tree_prior = dists.TreePriors.Hubs(n)
ct_prior = dists.CountPriors.TruncatedNB(1, .5)
cob_freq = 10
prior = dists.Priors.BasisCount(n, GraphAndBasis, ct_prior, tree_prior)
temp = Graph(5)
temp.SetComplete()
g = GraphAndBasis(5,
temp._dol,
tree=tree_prior.Sample(),
basis=np.ones((n - 1) * (n - 2) // 2, dtype=bool))
data = generate_data(n, n_obs, g, seed=123)
lik = dists.Likelihoods.Delta(g.copy())
prop = dists.Proposals.BasisWalk(n,
GraphAndBasis,
tree_prior,
skip=cob_freq)
np.random.seed(123)
sampler = MCMC_Sampler(prior, prop, lik, data)
sampler.run(5000, g)
assert (tree_changes_only_every_k_iter(sampler, cob_freq))
assert (if_both_empty_or_complete_cob_always_happen(sampler, cob_freq))
assert (only_propose_one_basis_difference(sampler, cob_freq))
assert (lookup_contains_all_params(sampler))
assert (sampler_types_are_correct(n, sampler))
def test_MCMC_edge_1():
n, n_obs = (5, 1000)
basis = edge_basis(n)
tree_prior = None
ct_prior = dists.CountPriors.Uniform()
np.random.seed(123)
for i in range(3):
g = GraphAndBasis(n)
g.SetRandom()
sampler = run_MCMC(g,
n,
n_obs,
tree_prior,
ct_prior,
basis=basis,
seed=123,
fixed_init=True)
print(sampler.res['LIK_'][-1])
median_graph = str_list_to_median_graph(n,
sampler.res['SAMPLES'][500:],
.5)
assert (np.sum(median_graph != g.GetAdjM()) < 3)
assert (only_propose_one_edge_difference(sampler))
assert (lookup_contains_all_params(sampler))
def test_MCMC_edge_0():
n, n_obs = (5, 1000)
basis = edge_basis(n)
tree_prior = None
ct_prior = dists.CountPriors.Uniform()
# Empty
g = GraphAndBasis(n)
sampler = run_MCMC(g,
n,
n_obs,
tree_prior,
ct_prior,
basis=basis,
seed=123)
print(sampler.res['LIK_'][-1])
median_graph = str_list_to_median_graph(n, sampler.res['SAMPLES'][500:])
assert ((median_graph == g.GetAdjM()).all())
assert (only_propose_one_edge_difference(sampler))
assert (lookup_contains_all_params(sampler))
# Circle
g = GraphAndBasis(n, circle_graph(n))
sampler = run_MCMC(g,
n,
n_obs,
tree_prior,
ct_prior,
basis=basis,
seed=123)
print(sampler.res['LIK_'][-1])
median_graph = str_list_to_median_graph(n, sampler.res['SAMPLES'][500:],
.75)
assert ((median_graph == g.GetAdjM()).all())
assert (only_propose_one_edge_difference(sampler))
assert (lookup_contains_all_params(sampler))
def test_MCMC_cb_0():
n, n_obs = (5, 1000)
tree_prior = dists.TreePriors.Uniform(n)
ct_prior = dists.CountPriors.TruncatedNB(n, .75)
cob_freq = 2
# Empty
g = GraphAndBasis(n)
sampler = run_MCMC(g,
n,
n_obs,
tree_prior,
ct_prior,
skip=cob_freq,
seed=123)
print(sampler.res['LIK_'][-1])
median_graph = str_list_to_median_graph(n, sampler.res['SAMPLES'][500:])
assert ((median_graph == g.GetAdjM()).all())
assert (only_propose_one_basis_difference(sampler, cob_freq))
assert (lookup_contains_all_params(sampler))
# Circle
g = GraphAndBasis(n, circle_graph(n))
sampler = run_MCMC(g,
n,
n_obs,
tree_prior,
ct_prior,
skip=cob_freq,
seed=123)
print(sampler.res['LIK_'][-1])
median_graph = str_list_to_median_graph(n, sampler.res['SAMPLES'][500:])
assert ((median_graph == g.GetAdjM()).all())
assert (only_propose_one_basis_difference(sampler, cob_freq))
assert (lookup_contains_all_params(sampler))
from tests.test_utils.generate_data import generate_data_and_save
from utils.generate_basis import circle_graph
import numpy as np
import networkx as nx
# Number of nodes and observations
n, n_obs = 20, 2
# Graphs to test on
graphs = []
# Empty Graph
graphs.append(Param(n))
# Circle Graph
g = Param(n, circle_graph(n))
graphs.append(g.copy())
# Random Graph Cycle Space
np.random.seed(123)
from dists.TreePriors import Hubs as TreePrior
from dists.Priors import BasisInclusion as Prior
tree_prior = TreePrior(n)
prior = Prior(n, Param, alpha=.5, tree_prior=tree_prior)
g = Param(n)
n_edges = n * (n - 1) // 2
while g.EdgeCount() <= int(.1 * n_edges) or g.EdgeCount() > int(.4 * n_edges):
g = prior.Sample()
graphs.append(g.copy())
g = Param(n)
from dists.Params import GraphAndBasis
from dists.CountPriors import TruncatedNB
from dists.Priors import BasisCount, BasisInclusion
from dists.TreePriors import Uniform
from utils.generate_basis import circle_graph
from tests.test_utils.generate_data import generate_data_and_save
import numpy as np
import networkx as nx
import pickle
# Number of nodes and observations
n, n_obs = 100, 1000
# Empty Graph
empty = GraphAndBasis(n)
# Circle Graph
circle = GraphAndBasis(n, circle_graph(n))
# Random Graphs from the Cycle space
print("Generating Graphs from the Cycle space...")
ct_prior = TruncatedNB(1, .5)
prior = BasisCount(n, GraphAndBasis, ct_prior, Uniform(n))
np.random.seed(123)
random0 = prior.Sample()
random1 = prior.Sample()
# Random Graphs from the Cycle space complement
print("Generating Graphs from the Cycle space complement...")
random2 = GraphAndBasis(
def get_log_posteriors(config):
n = config['n']
n_obs = config['n_obs']
name = config['true_graph']
with open(
f"data/graph_{config['true_graph']}_{config['n']}_{config['n_obs']}.pkl",
'rb') as handle:
g = pickle.load(handle)
g = Param(n, g._dol)
if n <= 5:
g_list = Graph.get_all_graphs(n)
delta, D = 3, np.eye(n)
rv = nbinom(1, .5)
idx = g.GetID()
data = np.loadtxt(f"data/{name}_{n}_{n_obs}.dat", delimiter=',')
U = data.transpose() @ data
assert (U.shape == D.shape)
liks = list(
map(
lambda x: laplace_approx(x._dol, delta + n_obs, D + U) -
laplace_approx(x._dol, delta, D), g_list))
diffs = list(
map(
lambda x: len(
set(g.GetEdgeL()).symmetric_difference(x.GetEdgeL())),
g_list))
priors = list(map(lambda x: np.log(rv.pmf(x)), diffs))
posteriors = np.array(liks) + np.array(priors)
return np.array(g._basis_active,
dtype=int), [(np.array(list(g_list[i].GetID()),
dtype=int), posteriors[i])
for i in range(len(g_list))]
else:
with open(config_to_path(config), 'rb') as handle:
sampler = pickle.load(handle)
if config['cob_freq'] is not None:
warnings.warn(
"Bases change throughout the simulation, all graphs is represented w.r.t. the last basis"
)
tree = sampler.last_params._tree
g = Param(len(g), g._dol, tree)
gba = np.array(g._basis_active, dtype=int)
log_posts = []
ids = []
for i in range(sampler.iter):
if sampler.res['PARAMS'][i] not in ids:
g_ba = np.array(list(
sampler.res['PARAMS_PROPS'][i]['BASIS_ID']),
dtype=int)
log_posts.append(
(g_ba, sampler.res['LIK_'][i] + sampler.res['PRIOR'][i]))
ids.append(sampler.res['PARAMS'][i])
return gba, log_posts