def single_segment_as_gif(segment,
directory="/tmp/online_matcher/visuals/animated/", frames=60, black_and_white=False):
import os
if not os.path.exists(directory):
os.makedirs(directory)
import voxelize
rotations = voxelize.create_rotations([voxelize.recenter_segment(segment)], n_angles=frames)
segments_as_gif(rotations, filename='segment',
directory=directory, black_and_white=black_and_white)
def single_segment_as_gif(segment,
directory="/tmp/online_matcher/visuals/animated/", frames=60, black_and_white=False):
import os
if not os.path.exists(directory):
os.makedirs(directory)
import voxelize
rotations = voxelize.create_rotations([voxelize.recenter_segment(segment)], n_angles=frames)
segments_as_gif(rotations, filename='segment',
directory=directory, black_and_white=black_and_white)
def single_segment_rotations_reconstruction_as_gif(segment, vae, confidence=0.3,
directory="/tmp/online_matcher/visuals/animated/",
frames=120, black_and_white=False):
import os
if not os.path.exists(directory):
os.makedirs(directory)
import voxelize
rotations = voxelize.create_rotations([segment], n_angles=frames)
import autoencoder.model
VOXEL_SIDE = vae.MP.INPUT_SHAPE[0]
rotations_vox, features_voxel_scale = voxelize.voxelize(rotations, VOXEL_SIDE)
reconstruction_vox = vae.batch_encode_decode([np.reshape(sample, vae.MP.INPUT_SHAPE) for sample in rotations_vox], batch_size=120)
reconstruction_vox = [np.reshape(vox, [VOXEL_SIDE, VOXEL_SIDE, VOXEL_SIDE]) for vox in reconstruction_vox]
from voxelize import unvoxelize
reconstruction = [unvoxelize(vox > confidence) for vox in reconstruction_vox]
reconstruction = [voxelize.recenter_segment(segment*scale) for (segment, scale) in zip(reconstruction, features_voxel_scale)]
segments_as_gif(reconstruction, rotate_YP=(2*np.pi/frames), filename='reconstruction_rot',
directory=directory, black_and_white=black_and_white)
def single_segment_rotations_reconstruction_as_gif(segment, vae, confidence=0.3,
directory="/tmp/online_matcher/visuals/animated/",
frames=120, black_and_white=False):
import os
if not os.path.exists(directory):
os.makedirs(directory)
import voxelize
rotations = voxelize.create_rotations([segment], n_angles=frames)
import autoencoder.model
VOXEL_SIDE = vae.MP.INPUT_SHAPE[0]
rotations_vox, features_voxel_scale = voxelize.voxelize(rotations, VOXEL_SIDE)
reconstruction_vox = vae.batch_encode_decode([np.reshape(sample, vae.MP.INPUT_SHAPE) for sample in rotations_vox], batch_size=120)
reconstruction_vox = [np.reshape(vox, [VOXEL_SIDE, VOXEL_SIDE, VOXEL_SIDE]) for vox in reconstruction_vox]
from voxelize import unvoxelize
reconstruction = [unvoxelize(vox > confidence) for vox in reconstruction_vox]
reconstruction = [voxelize.recenter_segment(segment*scale) for (segment, scale) in zip(reconstruction, features_voxel_scale)]
segments_as_gif(reconstruction, rotate_YP=(2*np.pi/frames), filename='reconstruction_rot',
directory=directory, black_and_white=black_and_white)
features_nn, confusion_nn = vae.batch_encode(
[np.reshape(sample, MP.INPUT_SHAPE) for sample in segments_vox])
fnames_nn = [
'autoencoder_feature' + str(i + 1)
for i, _ in enumerate(features_nn[0])
]
F = features_nn
C = classes
# In[ ]:
if not RUN_AS_PY_SCRIPT:
print("Rotating segments")
from voxelize import create_rotations
rotated_segments, rotated_classes = create_rotations(segments,
N_ROTATION_ANGLES,
classes=classes)
if False: # walls_vs_cars
print("Removing unknowns")
rotated_segments = [
segment
for segment, class_ in zip(rotated_segments, rotated_classes)
if class_ != "unknown"
]
rotated_classes = [
class_ for class_ in rotated_classes if class_ != "unknown"
]
print("Voxelizing rotations")
from voxelize import voxelize
rotated_segments_vox, rotated_segments_scale = voxelize(
rotated_segments, VOXEL_SIDE)
# ## Create Voxelized Segment Dataset - With Rotated Copies
# In[ ]:
## Split into training and val data
split_at = min(VAL_EXAMPLES, int(0.2 * len(ids)))
val = segments[:split_at]
train = segments[split_at:]
# In[ ]:
if not TRAIN_TWINS:
print("Rotating segments")
from voxelize import create_rotations
train = create_rotations(train, N_ROTATION_ANGLES, ROTATION_OFFSET)
val = create_rotations(val, 12, ROTATION_OFFSET)
print("Voxelizing training data")
from voxelize import voxelize
train_vox, _ = voxelize(train,VOXEL_SIDE)
val_vox, _ = voxelize(val ,VOXEL_SIDE)
train_twins_vox = None
val_twins_vox = None
if train_vox[0].shape != MP.INPUT_SHAPE:
print("Reshaping")
train_vox=[np.reshape(vox, MP.INPUT_SHAPE) for vox in train_vox]
val_vox=[np.reshape(vox, MP.INPUT_SHAPE) for vox in val_vox]
del train # Save some memory
# In[ ]:
if not RUN_AS_PY_SCRIPT:
print("Computing Features for Segments")
features_nn, confusion_nn = vae.batch_encode([np.reshape(sample, MP.INPUT_SHAPE) for sample in segments_vox])
fnames_nn = ['autoencoder_feature'+str(i+1) for i, _ in enumerate(features_nn[0])]
F = features_nn
C = classes
# In[ ]:
if not RUN_AS_PY_SCRIPT:
print("Rotating segments")
from voxelize import create_rotations
rotated_segments, rotated_classes = create_rotations(segments, N_ROTATION_ANGLES, classes=classes)
if False: # walls_vs_cars
print("Removing unknowns")
rotated_segments = [segment for segment, class_ in zip(rotated_segments, rotated_classes) if class_ != "unknown"]
rotated_classes = [class_ for class_ in rotated_classes if class_ != "unknown"]
print("Voxelizing rotations")
from voxelize import voxelize
rotated_segments_vox, rotated_segments_scale = voxelize(rotated_segments, VOXEL_SIDE)
print("Computing Features for rotations")
rotated_features, _ = vae.batch_encode([np.reshape(sample, MP.INPUT_SHAPE) for sample in rotated_segments_vox])
F = rotated_features
C = rotated_classes
# ## T-SNE