def _model(x, asn, sampler=_hybrid, alpha=0.1, annealing=None):
"""Create clustering model."""
return bmcc.BayesianMixture(
data=x,
sampler=sampler,
component_model=bmcc.NormalWishart(df=ds.dim,
scale=np.identity(ds.dim) *
np.sqrt(ds.dim)),
mixture_model=bmcc.DPM(alpha=alpha, use_eb=False),
annealing=annealing,
assignments=asn,
thinning=10)
def run_sample(args):
"""Run MCMC sampling"""
# Unpack
path, method_name = args
dst = path.replace(
BASE_DIR, os.path.join(RESULT_DIR, method_name))
# Ignore if test already run (file present)
if os.path.exists(dst):
return
# Load dataset
dataset = bmcc.GaussianMixture(path, load=True)
# Fetch method
method = METHODS[method_name]
cm = bmcc.NormalWishart(
df=dataset.d,
scale=np.identity(dataset.d) if SCALE_MATRIX else None)
mm = method["mixture"](dataset.k)
# Create model
model = bmcc.BayesianMixture(
data=dataset.data,
sampler=method["sampler"],
component_model=cm,
mixture_model=mm,
assignments=np.zeros(dataset.n).astype(np.uint16),
thinning=5)
# Run iterations (break on exceeding limit)
try:
for i in range(5000):
model.iter()
if np.max(model.assignments) > CLUSTERS_LIMIT:
break
except Exception as e:
print("Exception in {} / {}:".format(method_name, path))
print(e)
np.savez(dst, hist=model.hist)
def hybrid(*args, **kwargs):
for _ in range(5):
bmcc.gibbs(*args, **kwargs)
bmcc.split_merge(*args, **kwargs)
mm = bmcc.MFM(gamma=1, prior=lambda k: poisson.logpmf(k, 3))
# mm = bmcc.DPM(alpha=1, use_eb=False)
cm = bmcc.NormalWishart(df=3)
# Create mixture model
model = bmcc.BayesianMixture(data=dataset.data,
sampler=hybrid,
component_model=cm,
mixture_model=mm,
assignments=np.zeros(POINTS).astype(np.uint16),
thinning=THINNING)
# Run Iterations
start = time.time()
for i in tqdm(range(ITERATIONS)):
model.iter()
print("gibbs_iterate: {:.2f}s [{:.2f} ms/iteration]".format(
time.time() - start, (time.time() - start) * 1000 / ITERATIONS))
# Select Least Squares clustering
start = time.time()
res = model.select_lstsq(burn_in=BURN_IN)
res.evaluate(dataset.assignments,
oracle=dataset.oracle,
ITERATIONS = 2000
K = 3
Q = np.identity(K) * 0.2 + np.ones((K, K)) * 0.1
ds = bmcc.StochasticBlockModel(n=N, k=K, r=1, a=0.8, b=1, shuffle=False, Q=Q)
print(ds.Q)
# plt.matshow(ds.data)
# plt.show()
start = time.time()
model = bmcc.BayesianMixture(
data=ds.data,
sampler=bmcc.gibbs,
component_model=bmcc.SBM(a=1, b=1),
mixture_model=bmcc.MFM(gamma=1, prior=lambda k: poisson.logpmf(k, K)),
# mixture_model=bmcc.DPM(alpha=1),
assignments=np.zeros(N).astype(np.uint16),
thinning=1)
for _ in tqdm(range(ITERATIONS)):
model.iter()
print(time.time() - start)
res = model.select_lstsq(burn_in=1500)
fig, axs = plt.subplots(2, 2)
axs[0][0].matshow(ds.data)
axs[0][1].matshow(res.matrix)
axs[1][0].plot(res.num_clusters)