def test_noise2(self):
observation_noise = 0.01
odometry_noise = 0.02
sim = new_simulation(point_dim=3,
landmark_dim=3,
seed=20,
observation_noise=observation_noise,
odometry_noise=odometry_noise,
noise_matrix='diag')
fg = factorgraph.GaussianFactorGraph()
for f in sim.factors():
fg.add_factor(f)
optimizer = optim.Occam(fg)
optimizer.optimize()
optimizer.update()
m_hat = np.concatenate([m.position for m in fg.landmarks])
p_hat = np.concatenate([p.position for p in fg.points])
m, p = np.ravel(sim.landmark_positions), np.ravel(sim.point_positions)
self.assertTrue(
np.linalg.norm(optimizer.res_d)**2 < len(optimizer.res_d) * 4 *
observation_noise**2)
self.assertTrue(
np.linalg.norm(optimizer.res_t)**2 < len(optimizer.res_t) * 4 *
odometry_noise**2)
self.assertTrue(np.allclose(p, p_hat, atol=0.1))
def test_no_noise2(self):
sim = new_simulation(point_dim=3, landmark_dim=3, seed=42)
fg = factorgraph.GaussianFactorGraph()
for f in sim.factors():
fg.add_factor(f)
landmarks = fg.landmarks
self.assertTrue(sim.num_unique_landmarks < len(landmarks))
optimizer = optim.Occam(fg)
optimizer.optimize()
optimizer.update()
unique_landmarks = fg.landmarks
self.assertEqual(sim.num_unique_landmarks, len(unique_landmarks))
sim_groups, _ = sim.equivalences()
fg_groups = set(
frozenset(g) for g in fg.correspondence_map.set_map().values())
diff = sim_groups.symmetric_difference(fg_groups)
self.assertTrue(len(diff) == 0)
m_hat = np.concatenate([m.position for m in fg.landmarks])
p_hat = np.concatenate([p.position for p in fg.points])
unique_order = [
sim.unique_index_map[landmarks.index(k)] for k in unique_landmarks
]
m, p = np.ravel(
sim.unique_landmark_positions[unique_order, :]), np.ravel(
sim.point_positions)
self.assertTrue(np.allclose(p, p_hat))
self.assertTrue(np.allclose(m, m_hat))
def test_no_measurements(self):
fg = factorgraph.GaussianFactorGraph()
optimizer = optim.Occam(fg)
optimizer.optimize()
optimizer.update()
def occam_trial(seed=12, max_observations=4):
observation_noise = 0.01
odometry_noise = 0.02
sim = new_simulation(point_dim=3,
landmark_dim=1,
num_points=NUM_POINTS,
num_landmarks=NUM_LANDMARKS,
seed=seed,
observation_noise=observation_noise,
odometry_noise=odometry_noise,
noise_matrix='identity')
fg = factorgraph.GaussianFactorGraph()
for f in sim.factors(max_observations=max_observations):
fg.add_factor(f)
start = time.time()
optimizer = optim.Occam(fg)
optimizer.optimize()
optimizer.update(merge=False)
end = time.time()
time_elapsed = end - start
m_hat = np.concatenate([m.position for m in fg.landmarks])
p_hat = np.concatenate([p.position for p in fg.points])
m, p = np.ravel(
sim.landmark_positions[observed_landmark_index(fg, sim), :]), np.ravel(
sim.point_positions)
me = np.mean(np.concatenate([(m - m_hat), (p - p_hat)]))
return me, time_elapsed
def test_sequential_noise(self):
num_points = 1000
observation_noise = 0.02
odometry_noise = 0.01
sim = new_simulation(point_dim=3,
landmark_dim=3,
num_points=1000,
seed=67,
observation_noise=observation_noise,
odometry_noise=odometry_noise,
noise_matrix='diag')
num_partitions = 100
# partition_indptr = np.sort(
# np.concatenate([[0], np.random.choice(sim.num_points, num_partitions - 1), [sim.num_points]]))
partition_indptr = np.linspace(0,
num_points,
num_partitions + 1,
dtype=np.int)
fpw = int(np.max(partition_indptr[1:] - partition_indptr[:-1]) + 5)
fg = factorgraph.GaussianFactorGraph(free_point_window=fpw)
landmarks = sim.landmark_variables
optimizer = optim.Occam(fg)
for i in range(num_partitions):
for f in sim.factors(
(partition_indptr[i], partition_indptr[i + 1])):
fg.add_factor(f)
optimizer.optimize()
optimizer.update()
unique_landmarks = fg.landmarks
unique_order = [
sim.unique_index_map[landmarks.index(k)]
for k in unique_landmarks
]
sim_groups, _ = sim.equivalences((0, partition_indptr[i + 1]))
fg_groups = set(
frozenset(g) for g in fg.correspondence_map.set_map().values())
diff = sim_groups.symmetric_difference(fg_groups)
self.assertTrue(len(diff) == 0)
p_hat = np.concatenate([p.position for p in fg.points])
m_hat = np.concatenate([m.position for m in fg.landmarks])
m = np.ravel(sim.unique_landmark_positions[unique_order, :])
p = np.ravel(sim.point_positions[:partition_indptr[i + 1], :])
self.assertTrue(np.allclose(p, p_hat, atol=0.1))
self.assertTrue(np.allclose(m, m_hat, atol=0.1))
def test_odom_only(self):
import factor
sim = new_simulation(point_dim=3,
landmark_dim=1,
seed=1776,
observation_noise=0.0,
odometry_noise=0.0)
fg = factorgraph.GaussianFactorGraph()
for f in sim.factors():
if isinstance(f, factor.ObservationFactor):
continue
fg.add_factor(f)
optimizer = optim.Occam(fg)
optimizer.optimize()
optimizer.update()
p_hat = np.concatenate([p.position for p in fg.points])
p = np.ravel(sim.point_positions)
self.assertTrue(np.allclose(p, p_hat, atol=1e-3))