loaded = np.load(DATA_FILE)
raw_training_vectors = loaded['input_geoms']
raw_target_vectors = loaded['intersection']
training_vectors = []
target_vectors = []
# skip non-intersecting geometries
for train, target in zip(raw_training_vectors, raw_target_vectors):
if not target[0, 0] == 0: # a zero coordinate designates an empty geometry
training_vectors.append(train)
target_vectors.append(target)
means = localized_mean(training_vectors)
training_vectors = localized_normal(training_vectors, means, 1e4)
target_vectors = localized_normal(target_vectors, means, 1e4)
(data_points, max_points, INPUT_VECTOR_LEN) = training_vectors.shape
# Expand the target vectors to gaussian mixture model size compatible with the prediction format
component_1 = target_vectors[:, :, 0:5] # The first gaussian mixture model component for each point
component_1 = np.append(component_1, np.zeros((data_points, max_points, 1)), axis=2) # Add pi feature
target_vectors = np.append(
np.reshape(
np.repeat(component_1, GAUSSIAN_MIXTURE_COMPONENTS, axis=1),
(data_points, max_points, 6 * GAUSSIAN_MIXTURE_COMPONENTS)),
target_vectors[:, :, 5:], axis=2)
output_size = GAUSSIAN_MIXTURE_COMPONENTS * 6 + ONE_HOT_LEN
Loss = GaussianMixtureLoss(GAUSSIAN_MIXTURE_COMPONENTS, max_points)
osm_vectors = []
surface_area_vectors = []
combined_geom_vectors = []
# skip non-intersecting geometries
for brt, osm, target, combined in zip(raw_brt_vectors, raw_osm_vectors, raw_surface_area_vectors,
raw_combined_geom_vectors):
if not target[0] == 0: # a zero coordinate designates an empty geometry
brt_vectors.append(brt)
osm_vectors.append(osm)
surface_area_vectors.append(target)
combined_geom_vectors.append(combined)
# data whitening
means = localized_mean(combined_geom_vectors)
brt_vectors = localized_normal(brt_vectors, means, 1e4)
osm_vectors = localized_normal(osm_vectors, means, 1e4)
surface_area_vectors = np.array(surface_area_vectors)
# shape determination
data_points, brt_max_points, brt_seq_len = brt_vectors.shape
_, osm_max_points, osm_seq_len = osm_vectors.shape
brt_inputs = Input(shape=(brt_max_points, brt_seq_len))
brt_model = LSTM(brt_max_points * 2, activation='relu')(brt_inputs)
osm_inputs = Input(shape=(osm_max_points, osm_seq_len))
osm_model = LSTM(osm_max_points * 2, activation='relu')(osm_inputs)
concat = concatenate([brt_model, osm_model])
def test_dummy_geom(self):
dummy_means = localized_mean(dummy_geom)
dummy_localized = localized_normal(dummy_geom, dummy_means)
self.assertEqual(dummy_localized[0, 0, FULL_STOP_INDEX], 1)
0.,
0.,
]]])
dummy_geom = np.array(
[[[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]]])
means = localized_mean(input_geom)
scaled = localized_normal(input_geom, means)
class TestGeomScaler(unittest.TestCase):
def test_scaling(self):
self.assertAlmostEqual(scaled[0, 0, 0], -0.4763)
self.assertAlmostEqual(scaled[0, 0, 1], 0.3646)
self.assertAlmostEqual(scaled[0, 9, 0], -0.4263)
self.assertAlmostEqual(scaled[0, 9, 1], 0.3136)
def test_localized_mean(self):
self.assertAlmostEqual(means[0, 0, 0], 6.7232059)
self.assertAlmostEqual(means[0, 0, 1], 52.1477995)
corrected = input_geom
corrected[:, 0:10, 0:2] -= means
corrected[:, 0:10, 0:2] *= 1e4
raw_intersection_vectors = loaded['intersection']
brt_vectors = []
osm_vectors = []
intersection_vectors = []
# skip non-intersecting geometries
for brt, osm, target in zip(raw_brt_vectors, raw_osm_vectors, raw_intersection_vectors):
if not target[0, 0] == 0: # a zero coordinate designates an empty geometry
brt_vectors.append(brt)
osm_vectors.append(osm)
intersection_vectors.append(target)
# data whitening
means = localized_mean(intersection_vectors)
brt_vectors = localized_normal(brt_vectors, means, 1e4)
osm_vectors = localized_normal(osm_vectors, means, 1e4)
intersection_vectors = localized_normal(intersection_vectors, means, 1e4)
# shape determination
(data_points, brt_max_points, BRT_INPUT_VECTOR_LEN) = brt_vectors.shape
(_, osm_max_points, OSM_INPUT_VECTOR_LEN) = osm_vectors.shape
target_max_points = intersection_vectors.shape[1]
output_seq_length = (GAUSSIAN_MIXTURE_COMPONENTS * 6) + ONE_HOT_LEN
output_size_2d = target_max_points * output_seq_length
Loss = GaussianMixtureLoss(num_components=GAUSSIAN_MIXTURE_COMPONENTS, num_points=target_max_points)
brt_inputs = Input(shape=(brt_max_points, BRT_INPUT_VECTOR_LEN))
brt_model = LSTM(brt_max_points * 2, activation='relu')(brt_inputs)
loaded = np.load(DATA_FILE)
raw_training_vectors = loaded['input_geoms']
raw_target_vectors = loaded['intersection']
input_vectors = []
target_vectors = []
# skip non-intersecting geometries
for train, target in zip(raw_training_vectors, raw_target_vectors):
if not target[0, 0] == 0: # a zero coordinate designates an empty geometry
input_vectors.append(train)
target_vectors.append(target)
print('Preprocessing vectors...')
means = localized_mean(input_vectors)
input_vectors = localized_normal(input_vectors, means, 1e4)
input_vectors = np.array(
[GeoVectorizer.interpolate(vector, DENSIFIED) for vector in input_vectors])
target_vectors = localized_normal(target_vectors, means, 1e4)
target_vectors = np.array(
[GeoVectorizer.interpolate(vector, 50) for vector in target_vectors])
(data_points, max_input_points, INPUT_VECTOR_LEN) = input_vectors.shape
(_, max_target_points, _) = target_vectors.shape
output_size = GAUSSIAN_MIXTURE_COMPONENTS * 6 + GEOM_TYPE_LEN + RENDER_LEN
Loss = GaussianMixtureLoss(GAUSSIAN_MIXTURE_COMPONENTS, max_target_points)
inputs = Input(shape=(max_input_points, INPUT_VECTOR_LEN))
model = Dense(INPUT_VECTOR_LEN, activation='relu')(inputs)