def test_shape_add_remove_sample(self):
# test if add_remove_part traverses the sample space correctly
# look at how much probability mass objects with a certain number of parts get
maxn = 3
s = ShapeTestHypothesis(parts=[CuboidPrimitive()], maxn=3)
sample_count = 40000
p = DeterministicMixtureProposal(
moves={'add_remove': shape_add_remove_part},
params={'MAX_PART_COUNT': maxn})
sampler = MHSampler(initial_h=s,
data=None,
proposal=p,
burn_in=5000,
sample_count=sample_count,
best_sample_count=1,
thinning_period=1,
report_period=4000)
run = sampler.sample()
# count how many times we sampled each part_count
k = np.array([len(sample.parts) for sample in run.samples.samples])
# since the prior p(h) = 1 / part_count and part_count is 1,2 or 3, we expect 6/11 of all samples to have a
# single part, 3/11 to have 2 parts, and 2/11 to have 3 parts.
self.assertAlmostEqual(np.mean(k == 1), 6.0 / 11.0, places=1)
self.assertAlmostEqual(np.mean(k == 2), 3.0 / 11.0, places=1)
self.assertAlmostEqual(np.mean(k == 3), 2.0 / 11.0, places=1)
def test_bdaooss_add_remove_sample(self):
# test if add_remove samples correctly
# sample from the prior using add_remove and see if we get each tree with its expected probability.
# here we constrain depth to 1; so, there are 4 possible trees: P, P->P, P->[P, P], P->[P, P, P].
# p({P}) = 1/4, p({P->P}) = 1/16, p({P->{P, P}) = 1/64, p({P->{P,P,P}) = 1/256
# therefore, we expect to see P 64/85 of the time, P->P 16/85 of the time, P->[P, P] 4/85 of the time, and
# P->[P, P, P] 1/85 of the time.
h = BDAoOSSShapeTestHypothesis(shape=self.h1.shape)
p = DeterministicMixtureProposal(moves={"add_remove": bdaooss_add_remove_part}, params={"MAX_DEPTH": 2})
sampler = MHSampler(
initial_h=h,
proposal=p,
data=None,
burn_in=0,
sample_count=50000,
best_sample_count=1,
thinning_period=1,
report_period=5000,
)
run = sampler.sample()
k = np.array([len(s.shape.spatial_model.spatial_states) for s in run.samples.samples])
self.assertAlmostEqual(np.mean(k == 1), 64 / 85.0, places=1)
self.assertAlmostEqual(np.mean(k == 2), 16 / 85.0, places=1)
self.assertAlmostEqual(np.mean(k == 3), 4 / 85.0, places=1)
self.assertAlmostEqual(np.mean(k == 4), 1 / 85.0, places=1)
def test_bdaooss_add_remove_sample(self):
# test if add_remove samples correctly
# sample from the prior using add_remove and see if we get each tree with its expected probability.
# here we constrain depth to 1; so, there are 4 possible trees: P, P->P, P->[P, P], P->[P, P, P].
# p({P}) = 1/4, p({P->P}) = 1/16, p({P->{P, P}) = 1/64, p({P->{P,P,P}) = 1/256
# therefore, we expect to see P 64/85 of the time, P->P 16/85 of the time, P->[P, P] 4/85 of the time, and
# P->[P, P, P] 1/85 of the time.
h = BDAoOSSShapeTestHypothesis(shape=self.h1.shape)
p = DeterministicMixtureProposal(
moves={'add_remove': bdaooss_add_remove_part},
params={'MAX_DEPTH': 2})
sampler = MHSampler(initial_h=h,
proposal=p,
data=None,
burn_in=0,
sample_count=50000,
best_sample_count=1,
thinning_period=1,
report_period=5000)
run = sampler.sample()
k = np.array([
len(s.shape.spatial_model.spatial_states)
for s in run.samples.samples
])
self.assertAlmostEqual(np.mean(k == 1), 64 / 85.0, places=1)
self.assertAlmostEqual(np.mean(k == 2), 16 / 85.0, places=1)
self.assertAlmostEqual(np.mean(k == 3), 4 / 85.0, places=1)
self.assertAlmostEqual(np.mean(k == 4), 1 / 85.0, places=1)
def test_voxel_sample_prior(self):
# test if sampling from the prior produces a set of samples with expected frequency statistics.
# for the prior defined in VoxelBasedShapeMaxD, we would expect to see an equal number of samples for each
# number of partial voxels. Here we constrain depth to 2, so a tree can have at least 0 and at most 9 partial
# voxels. Therefore, we would expect to see each number of partial voxels 0.1 of the time.
h = VoxelBasedShapeTestHypothesis(voxel=self.v1, max_depth=2)
moves = {
'flip_empty_vs_partial': voxel_based_shape_flip_empty_vs_partial,
'flip_full_vs_partial': voxel_based_shape_flip_full_vs_partial,
'flip_full_vs_empty': voxel_based_shape_flip_full_vs_empty
}
proposal = RandomMixtureProposal(moves=moves, params={'MAX_DEPTH': 2})
sampler = MHSampler(initial_h=h,
data=None,
proposal=proposal,
burn_in=0,
thinning_period=1,
sample_count=50000,
best_sample_count=1,
report_period=5000)
run = sampler.sample()
partial_voxel_counts = np.array([
s.voxel.count_voxels_by_status(PARTIAL_VOXEL)
for s in run.samples.samples
])
for i in range(10):
self.assertAlmostEqual(np.mean(partial_voxel_counts == i),
1.0 / 10.0,
places=1)
def test_paperclip_add_remove_joint_sample(self):
# test if add_remove_joint traverses the sample space correctly
# look at how much probability mass objects with a certain number of parts get
h = PaperClipTestHypothesis(joint_positions=[
np.array((-0.4, 0.0, 0.0)),
np.array((0.4, 0.0, 0.0))
],
mid_segment_id=0)
h.min_joints = 2
h.max_joints = 4
sample_count = 40000
p = DeterministicMixtureProposal(
moves={'add_remove': paperclip_shape_add_remove_joint},
params={'MAX_NEW_SEGMENT_LENGTH': 0.6})
sampler = MHSampler(initial_h=h,
data=None,
proposal=p,
burn_in=5000,
sample_count=sample_count,
best_sample_count=1,
thinning_period=1,
report_period=4000)
run = sampler.sample()
# count how many times we sampled each joint
k = np.array([sample.joint_count for sample in run.samples.samples])
# since the prior is uniform and we min_joints=5, max_joints=6,
# we expect to see an equal number of samples with 5, 6, or 7 joints.
self.assertAlmostEqual(np.mean(k == 2), 1.0 / 3.0, places=1)
self.assertAlmostEqual(np.mean(k == 3), 1.0 / 3.0, places=1)
self.assertAlmostEqual(np.mean(k == 4), 1.0 / 3.0, places=1)
def test_shape_add_remove_sample(self):
# test if add_remove_part traverses the sample space correctly
# look at how much probability mass objects with a certain number of parts get
maxn = 3
s = ShapeTestHypothesis(parts=[CuboidPrimitive()], maxn=3)
sample_count = 40000
p = DeterministicMixtureProposal(moves={'add_remove': shape_add_remove_part}, params={'MAX_PART_COUNT': maxn})
sampler = MHSampler(initial_h=s, data=None, proposal=p, burn_in=5000, sample_count=sample_count,
best_sample_count=1, thinning_period=1, report_period=4000)
run = sampler.sample()
# count how many times we sampled each part_count
k = np.array([len(sample.parts) for sample in run.samples.samples])
# since the prior p(h) = 1 / part_count and part_count is 1,2 or 3, we expect 6/11 of all samples to have a
# single part, 3/11 to have 2 parts, and 2/11 to have 3 parts.
self.assertAlmostEqual(np.mean(k == 1), 6.0 / 11.0, places=1)
self.assertAlmostEqual(np.mean(k == 2), 3.0 / 11.0, places=1)
self.assertAlmostEqual(np.mean(k == 3), 2.0 / 11.0, places=1)
def test_voxel_sample_prior(self):
# test if sampling from the prior produces a set of samples with expected frequency statistics.
# for the prior defined in VoxelBasedShapeMaxD, we would expect to see an equal number of samples for each
# number of partial voxels. Here we constrain depth to 2, so a tree can have at least 0 and at most 9 partial
# voxels. Therefore, we would expect to see each number of partial voxels 0.1 of the time.
h = VoxelBasedShapeTestHypothesis(voxel=self.v1, max_depth=2)
moves = {'flip_empty_vs_partial': voxel_based_shape_flip_empty_vs_partial,
'flip_full_vs_partial': voxel_based_shape_flip_full_vs_partial,
'flip_full_vs_empty': voxel_based_shape_flip_full_vs_empty}
proposal = RandomMixtureProposal(moves=moves, params={'MAX_DEPTH': 2})
sampler = MHSampler(initial_h=h, data=None, proposal=proposal, burn_in=0, thinning_period=1,
sample_count=50000, best_sample_count=1, report_period=5000)
run = sampler.sample()
partial_voxel_counts = np.array([s.voxel.count_voxels_by_status(PARTIAL_VOXEL) for s in run.samples.samples])
for i in range(10):
self.assertAlmostEqual(np.mean(partial_voxel_counts == i), 1.0 / 10.0, places=1)
def test_paperclip_add_remove_joint_sample(self):
# test if add_remove_joint traverses the sample space correctly
# look at how much probability mass objects with a certain number of parts get
h = PaperClipTestHypothesis(joint_positions=[np.array((-0.4, 0.0, 0.0)), np.array((0.4, 0.0, 0.0))],
mid_segment_id=0)
h.min_joints = 2
h.max_joints = 4
sample_count = 40000
p = DeterministicMixtureProposal(moves={'add_remove': paperclip_shape_add_remove_joint},
params={'MAX_NEW_SEGMENT_LENGTH': 0.6})
sampler = MHSampler(initial_h=h, data=None, proposal=p, burn_in=5000, sample_count=sample_count,
best_sample_count=1, thinning_period=1, report_period=4000)
run = sampler.sample()
# count how many times we sampled each joint
k = np.array([sample.joint_count for sample in run.samples.samples])
# since the prior is uniform and we min_joints=5, max_joints=6,
# we expect to see an equal number of samples with 5, 6, or 7 joints.
self.assertAlmostEqual(np.mean(k == 2), 1.0 / 3.0, places=1)
self.assertAlmostEqual(np.mean(k == 3), 1.0 / 3.0, places=1)
self.assertAlmostEqual(np.mean(k == 4), 1.0 / 3.0, places=1)
def run_chain(name,
sampler,
initial_h,
data,
kernel,
burn_in,
thinning_period,
sample_count,
best_sample_count,
report_period,
results_folder,
temperatures=None):
"""Run an MCMC chain and save results.
This function is used by run_experiment scripts to run chains and save results.
Parameters:
name (str): name of the chain. Used as the folder name to save sample images
sampler (str): Sampler to use. 'mh' for Metropolis-Hastings, 'pt' for Parallel Tempering
initial_h (I3DHypothesis): Initial hypothesis
data (numpy.ndarray): Observed data
kernel (mcmclib.Proposal): Transition kernel of the chain
burn_in (int): Number of burn in iterations
thinning_period (int): Keep every ith sample
sample_count (int): Number of samples to take
best_sample_count (int): Size of the best samples list
report_period (int): Report the status of the chain every report_period iterations
results_folder (str): Folder to save the results
temperatures (list): Temperatures of each chain for Parallel Tempering sampler
Returns:
dict: results
"""
if sampler == 'mh':
from mcmclib.mh_sampler import MHSampler
sampler = MHSampler(initial_h=initial_h,
data=data,
proposal=kernel,
burn_in=burn_in,
sample_count=sample_count,
best_sample_count=best_sample_count,
thinning_period=thinning_period,
report_period=report_period)
elif sampler == 'pt':
if temperatures is None:
raise ValueError(
'ParallelTempering sampler requires temperatures parameter.')
chain_count = len(temperatures)
from mcmclib.parallel_tempering_sampler import ParallelTemperingSampler
sampler = ParallelTemperingSampler(
initial_hs=[initial_h] * chain_count,
data=data,
proposals=[kernel] * chain_count,
temperatures=temperatures,
burn_in=burn_in,
sample_count=sample_count,
best_sample_count=best_sample_count,
thinning_period=int(thinning_period / chain_count),
report_period=int(report_period / chain_count))
else:
raise ValueError('Unknown sampler. Possible choices are mh and pt.')
start = time.time()
run = sampler.sample()
end = time.time()
# generate a random run id
run_id = np.random.randint(1000000)
run_file = "{0:s}/{1:s}_{2:s}_{3:06d}.pkl".format(
results_folder, name,
time.strftime("%Y%m%d_%H%M%S", time.localtime(start)), run_id)
run.save(run_file)
# save images of samples to disk
fwm2 = vfm.VisionForwardModel(render_size=(300, 300))
try:
os.mkdir("{0:s}/{1:s}".format(results_folder, name))
except OSError as e:
warnings.warn(e.message)
for i, sample in enumerate(run.samples.samples):
fwm2.save_render(
"{0:s}/{1:s}/s{2:d}.png".format(results_folder, name, i), sample)
for i, sample in enumerate(run.best_samples.samples):
fwm2.save_render(
"{0:s}/{1:s}/b{2:d}.png".format(results_folder, name, i), sample)
sample_lls = [
sample.log_likelihood(data) for sample in run.samples.samples
]
best_lls = [
sample.log_likelihood(data) for sample in run.best_samples.samples
]
mse_best = -2 * initial_h.params['LL_VARIANCE'] * np.max(best_lls)
mse_mean = -2 * initial_h.params['LL_VARIANCE'] * np.mean(best_lls)
mse_sample = -2 * initial_h.params['LL_VARIANCE'] * np.mean(sample_lls)
# form the results dictionary
results = {
'run_id': run_id,
'run_file': run_file,
'mean_acceptance_rate': run.run_log.IsAccepted.mean(),
'start_time': start,
'end_time': end,
'duration': (end - start) / 60.0,
'best_posterior': np.max(run.best_samples.log_probs),
'best_ll': np.max(best_lls),
'mse': mse_best,
'mean_best_posterior': np.mean(run.best_samples.log_probs),
'mean_best_ll': np.mean(best_lls),
'mse_mean': mse_mean,
'mean_sample_posterior': np.mean(run.samples.log_probs),
'mean_sample_ll': np.mean(sample_lls),
'mse_sample': mse_sample
}
# add acceptance rate by move to results
acc_rate_by_move = run.acceptance_rate_by_move()
acc_rates = dict(
zip(acc_rate_by_move.MoveType, acc_rate_by_move.AcceptanceRate))
results.update(acc_rates)
return results
def run_chain(name, sampler, initial_h, data, kernel, burn_in, thinning_period, sample_count, best_sample_count,
report_period, results_folder, temperatures=None):
"""Run an MCMC chain and save results.
This function is used by run_experiment scripts to run chains and save results.
Parameters:
name (str): name of the chain. Used as the folder name to save sample images
sampler (str): Sampler to use. 'mh' for Metropolis-Hastings, 'pt' for Parallel Tempering
initial_h (I3DHypothesis): Initial hypothesis
data (numpy.ndarray): Observed data
kernel (mcmclib.Proposal): Transition kernel of the chain
burn_in (int): Number of burn in iterations
thinning_period (int): Keep every ith sample
sample_count (int): Number of samples to take
best_sample_count (int): Size of the best samples list
report_period (int): Report the status of the chain every report_period iterations
results_folder (str): Folder to save the results
temperatures (list): Temperatures of each chain for Parallel Tempering sampler
Returns:
dict: results
"""
if sampler == 'mh':
from mcmclib.mh_sampler import MHSampler
sampler = MHSampler(initial_h=initial_h, data=data, proposal=kernel, burn_in=burn_in, sample_count=sample_count,
best_sample_count=best_sample_count, thinning_period=thinning_period,
report_period=report_period)
elif sampler == 'pt':
if temperatures is None:
raise ValueError('ParallelTempering sampler requires temperatures parameter.')
chain_count = len(temperatures)
from mcmclib.parallel_tempering_sampler import ParallelTemperingSampler
sampler = ParallelTemperingSampler(initial_hs=[initial_h]*chain_count, data=data, proposals=[kernel]*chain_count,
temperatures=temperatures, burn_in=burn_in, sample_count=sample_count,
best_sample_count=best_sample_count,
thinning_period=int(thinning_period / chain_count),
report_period=int(report_period / chain_count))
else:
raise ValueError('Unknown sampler. Possible choices are mh and pt.')
start = time.time()
run = sampler.sample()
end = time.time()
# generate a random run id
run_id = np.random.randint(1000000)
run_file = "{0:s}/{1:s}_{2:s}_{3:06d}.pkl".format(results_folder, name,
time.strftime("%Y%m%d_%H%M%S", time.localtime(start)),
run_id)
run.save(run_file)
# save images of samples to disk
fwm2 = vfm.VisionForwardModel(render_size=(300, 300))
try:
os.mkdir("{0:s}/{1:s}".format(results_folder, name))
except OSError as e:
warnings.warn(e.message)
for i, sample in enumerate(run.samples.samples):
fwm2.save_render("{0:s}/{1:s}/s{2:d}.png".format(results_folder, name, i), sample)
for i, sample in enumerate(run.best_samples.samples):
fwm2.save_render("{0:s}/{1:s}/b{2:d}.png".format(results_folder, name, i), sample)
sample_lls = [sample.log_likelihood(data) for sample in run.samples.samples]
best_lls = [sample.log_likelihood(data) for sample in run.best_samples.samples]
mse_best = -2 * initial_h.params['LL_VARIANCE'] * np.max(best_lls)
mse_mean = -2 * initial_h.params['LL_VARIANCE'] * np.mean(best_lls)
mse_sample = -2 * initial_h.params['LL_VARIANCE'] * np.mean(sample_lls)
# form the results dictionary
results = {'run_id': run_id, 'run_file': run_file, 'mean_acceptance_rate': run.run_log.IsAccepted.mean(),
'start_time': start, 'end_time': end, 'duration': (end - start) / 60.0,
'best_posterior': np.max(run.best_samples.log_probs), 'best_ll': np.max(best_lls), 'mse': mse_best,
'mean_best_posterior': np.mean(run.best_samples.log_probs),
'mean_best_ll': np.mean(best_lls), 'mse_mean': mse_mean,
'mean_sample_posterior': np.mean(run.samples.log_probs),
'mean_sample_ll': np.mean(sample_lls), 'mse_sample': mse_sample}
# add acceptance rate by move to results
acc_rate_by_move = run.acceptance_rate_by_move()
acc_rates = dict(zip(acc_rate_by_move.MoveType, acc_rate_by_move.AcceptanceRate))
results.update(acc_rates)
return results