true_std_devs = (20, 5, 0.25)
noisy_data = gen_data_from_multi_src(eval_model, true_std_devs, plot=True)
num_samples = 50000
burnin = int(num_samples / 2)
num_params = 5
init_inputs = np.array([[1] * num_params, [2] * num_params])
cov = np.eye(num_params)
src_num_pts = (NUM_DATA_PTS, NUM_DATA_PTS, NUM_DATA_PTS)
src_std_devs = (None, None, None) # estimate all 3 source std devs
log_like_args = [src_num_pts, src_std_devs]
log_like_func = MultiSourceNormal
priors = [ImproperUniform(0., 6.), ImproperUniform(0., 6.)] + \
[ImproperUniform(0, None)] * 3 # priors for all 3 source std devs
vector_mcmc = VectorMCMC(eval_model, noisy_data, priors, log_like_args,
log_like_func)
chain = vector_mcmc.metropolis(init_inputs,
num_samples,
cov,
adapt_interval=200,
adapt_delay=5000,
progress_bar=True)
plot_mcmc_chain(chain,
param_labels=['a', 'b', 'std1', 'std2', 'std3'],
burnin=burnin,
include_kde=True)
from smcpy.mcmc.vector_mcmc import VectorMCMC
from smcpy.utils.plotter import plot_mcmc_chain
from run_example_known_std import *
if __name__ == '__main__':
np.random.seed(200)
noisy_data = generate_data(eval_model, plot=False)
num_samples = 10000
burnin = 5000
std_dev = None # estimate it; requires mod to init_inputs, cov, and priors
init_inputs = np.array([[1, 1, 1], [2, 2, 2]])
cov = np.eye(3)
priors = [uniform(0., 6.), uniform(0., 6.), uniform(0, 10)]
vector_mcmc = VectorMCMC(eval_model, noisy_data, priors, std_dev)
chain = vector_mcmc.metropolis(init_inputs,
num_samples,
cov,
adapt_interval=200,
adapt_delay=100,
progress_bar=True)
plot_mcmc_chain(chain,
param_labels=['a', 'b', 'std'],
burnin=burnin,
include_kde=True)
from smcpy.mcmc.vector_mcmc import VectorMCMC
from smcpy.mcmc.vector_mcmc_kernel import VectorMCMCKernel
from smcpy.utils.plotter import plot_pairwise
from run_example_known_std import *
if __name__ == '__main__':
np.random.seed(200)
noisy_data = generate_data(eval_model, plot=False)
std_dev = None # estimate it; requires prior for std_dev and param_order
priors = [uniform(0., 6.), uniform(0., 6.), uniform(0, 10)]
vector_mcmc = VectorMCMC(eval_model, noisy_data, priors, std_dev)
mcmc_kernel = VectorMCMCKernel(vector_mcmc, param_order=('a', 'b', 'std'))
smc = SMCSampler(mcmc_kernel)
phi_sequence = np.linspace(0, 1, 20)
step_list, mll_list = smc.sample(num_particles=500,
num_mcmc_samples=5,
phi_sequence=phi_sequence,
ess_threshold=1.0,
progress_bar=True)
print('marginal log likelihood = {}'.format(mll_list[-1]))
print('parameter means = {}'.format(step_list[-1].compute_mean()))
plot_pairwise(step_list[-1].params, step_list[-1].weights,
['a', 'b', 'std'])
def vector_mcmc(stub_model, data, priors):
return VectorMCMC(stub_model, data, priors)
if __name__ == "__main__":
np.random.seed(2)
true_mean = 1
true_std = 2
n_data_pts = 100
n_samples = 6000
burnin = 0
param_names = ['mean', 'std']
model = lambda x: np.tile(x[:, 0], (1, n_data_pts))
data = np.random.normal(1, 2, n_data_pts)
priors = [ImproperUniform(), ImproperUniform(0, None)]
x0 = np.array([[-3, 7]])
cov = np.eye(2) * 0.001
vmcmc = VectorMCMC(model, data, priors)
chain = vmcmc.metropolis(x0,
n_samples,
cov,
adapt_interval=100,
adapt_delay=1000,
progress_bar=True)
burnin, z = compute_geweke(chain[0], window_pct=10, step_pct=1)
plot_mcmc_chain(chain, param_names)
plot_geweke(burnin, z, param_names)