def gen_ggm_trajectory(dataframe, n_samples, D=None, delta=1.0, cache={}, alpha=0.5, beta=0.5, **args):
p = dataframe.shape[1]
if D is None:
D = np.identity(p)
sd = seqdist.GGMJTPosterior()
sd.init_model(np.asmatrix(dataframe), D, delta, cache)
return mh(alpha, beta, n_samples, sd)
def sample_trajectories_ggm_parallel(dataframe, n_samples, randomize=[1000], D=None, delta=1.0,
reps=1, output_directory=".", **args):
p = dataframe.shape[1]
if D is None:
D = np.identity(p)
queue = multiprocessing.Queue()
processes = []
rets = []
for _ in range(reps):
for r in randomize:
for T in n_samples:
sd = seqdist.GGMJTPosterior()
sd.init_model(np.asmatrix(dataframe), D, delta, {})
print("Starting: " + str((T, r, str(sd), True)))
proc = Process(target=trajectory_to_queue,
args=(T, r,
sd, queue, True))
proc.start()
processes.append(proc)
time.sleep(2)
for _ in processes:
ret = queue.get() # will block
rets.append(ret)
for p in processes:
p.join()
return rets
def sample_trajectory_ggm(dataframe, n_samples, randomize=1000, D=None, delta=1.0, cache={}, **args):
p = dataframe.shape[1]
if D is None:
D = np.identity(p)
sd = seqdist.GGMJTPosterior()
sd.init_model(np.asmatrix(dataframe), D, delta, cache)
return sample_trajectory(n_samples, randomize, sd)
def smc_ggm_graphs(N, alpha, beta, radius, X, D, delta):
cache = {}
seq_dist = seqdist.GGMJTPosterior()
seq_dist.init_model(X, D, delta, cache)
(trees, log_w) = approximate(N, alpha, beta, radius, seq_dist)
log_w_rescaled = np.array(log_w.T)[seq_dist.p - 1] - \
max(np.array(log_w.T)[seq_dist.p - 1])
norm_w = np.exp(log_w_rescaled) / sum(np.exp(log_w_rescaled))
graphs = [jtlib.graph(tree) for tree in trees]
return (graphs, norm_w)
def from_json(self, mcmc_json):
graphs = [json_graph.node_link_graph(js_graph)
for js_graph in mcmc_json["trajectory"]]
self.set_trajectory(graphs)
self.set_time(mcmc_json["run_time"])
self.optional = mcmc_json["optional"]
self.sampling_method = mcmc_json["sampling_method"]
if mcmc_json["model"]["name"] == "ggm_jt_post":
self.seqdist = sd.GGMJTPosterior()
elif mcmc_json["model"]["name"] == "loglin_jt_post":
self.seqdist = sd.LogLinearJTPosterior()
self.seqdist.init_model_from_json(mcmc_json["model"])
def sample_trajectories_ggm_to_file(dataframe, n_samples, randomize=[1000], D=None, delta=1.0,
reps=1, output_directory=".", **args):
p = dataframe.shape[1]
if D is None:
D = np.identity(p)
graph_trajectories = []
for _ in range(reps):
for r in randomize:
for T in n_samples:
sd = seqdist.GGMJTPosterior()
sd.init_model(np.asmatrix(dataframe), D, delta, {})
graph_trajectory = trajectory_to_file(T, r, sd, dir=output_directory)
graph_trajectories.append(graph_trajectory)
return graph_trajectories
def sample_trajectories_ggm_parallel(dataframe,
n_particles,
n_samples,
D=None,
delta=1.0,
alphas=[0.5],
betas=[0.5],
radii=[None],
reset_cache=True,
reps=1,
**args):
p = dataframe.shape[1]
if D is None:
D = np.identity(p)
if radii == [None]:
radii = [p]
queue = multiprocessing.Queue()
processes = []
rets = []
for _ in range(reps):
for N in n_particles:
for T in n_samples:
for rad in radii:
for alpha in alphas:
for beta in betas:
sd = seqdist.GGMJTPosterior()
sd.init_model(np.asmatrix(dataframe), D, delta, {})
print("Starting: " +
str((N, T, alpha, beta, rad, str(sd),
reset_cache, True)))
proc = Process(target=trajectory_to_queue,
args=(N, T, alpha, beta, rad, sd,
queue, reset_cache, True))
processes.append(proc)
proc.start()
time.sleep(2)
for _ in processes:
ret = queue.get() # will block
rets.append(ret)
for p in processes:
p.join()
return rets
def sample_trajectory_ggm(dataframe,
n_particles,
n_samples,
D=None,
delta=1.0,
alpha=0.5,
beta=0.5,
radius=None,
reset_cache=True,
**args):
""" Particle Gibbs for approximating distributions over
Gaussian graphical models.
Args:
n_particles (int): Number of particles in SMC in each Gibbs iteration
n_samples (int): Number of Gibbs iterations (samples)
alpha (float): sparsity parameter for the Christmas tree algorithm
beta (float): sparsity parameter for the Christmas tree algorithm
radius (float): defines the radius within which ned nodes are selected
dataframe (np.matrix): row matrix of data
D (np.matrix): matrix parameter for the hyper inverse wishart prior
delta (float): degrees of freedom for the hyper inverse wishart prior
cache (dict): cache for clique likelihoods
Returns:
Trajectory: Markov chain of the underlying graphs of the junction trees sampled by pgibbs.
"""
p = dataframe.shape[1]
if D is None:
D = np.identity(p)
if radius is None:
radius = p
sd = seqdist.GGMJTPosterior()
sd.init_model(np.asmatrix(dataframe), D, delta, {})
return sample_trajectory(n_particles,
alpha,
beta,
radius,
n_samples,
sd,
reset_cache=reset_cache)
def sample_trajectories_ggm_to_file(dataframe,
n_particles,
n_samples,
D=None,
delta=1.0,
alphas=[0.5],
betas=[0.5],
radii=[None],
reset_cache=True,
reps=1,
output_directory=".",
output_filename="trajectory.json",
**args):
p = dataframe.shape[1]
if D is None:
D = np.identity(p)
if radii == [None]:
radii = [p]
graph_trajectories = []
for _ in range(reps):
for N in n_particles:
for T in n_samples:
for rad in radii:
for alpha in alphas:
for beta in betas:
sd = seqdist.GGMJTPosterior()
sd.init_model(np.asmatrix(dataframe), D, delta, {})
graph_trajectory = trajectory_to_file(
N,
T,
alpha,
beta,
rad,
sd,
reset_cache=reset_cache,
output_filename=output_filename,
dir=output_directory)
graph_trajectories.append(graph_trajectory)
return graph_trajectories
pool = Pool(processes=len(self.classes))
async_results = [None for _ in self.same_graph_groups]
for g, group in enumerate(self.same_graph_groups):
x_centered = np.array([])
# concatenate the data, centered by mean in each class
for c in group:
c_inds = (np.array(self.y).ravel() == c)
xc = self.x[np.ix_(c_inds, range(self.p))]
xc_centered = xc - np.mean(xc, axis=0)
if len(x_centered) == 0:
x_centered = xc_centered
else:
x_centered = np.concatenate((x_centered, xc_centered),
axis=0)
seq_dist = seqjtdist.GGMJTPosterior()
cache = {}
seq_dist.init_model(x_centered, self.hyper_tau[g],
self.hyper_alpha[g], cache)
if async is True:
async_results[g] = pool.apply_async(
trilearn.pgibbs.sample_trajectory,
(n_particles, cta_alpha, cta_beta, self.smc_radius,
n_pgibbs_samples, seq_dist))
else:
self.ggm_trajs[g] = trilearn.pgibbs.sample_trajectory(
n_particles, cta_alpha, cta_beta, self.smc_radius,
n_pgibbs_samples, seq_dist)
if async is True:
for g in range(len(self.same_graph_groups)):
self.ggm_trajs[g] = async_results[g].get()