def get_random_batch(self, batch_size):
data_ids_a = np.zeros(batch_size, 'int32')
data_ids_b = np.zeros(batch_size, 'int32')
viewpoint_ids_a = torch.zeros(batch_size)
viewpoint_ids_b = torch.zeros(batch_size)
for i in range(batch_size):
class_id = np.random.choice(self.class_ids)
object_id = np.random.randint(0, self.num_data[class_id])
viewpoint_id_a = np.random.randint(0, 24)
viewpoint_id_b = np.random.randint(0, 24)
data_id_a = (object_id + self.pos[class_id]) * 24 + viewpoint_id_a
data_id_b = (object_id + self.pos[class_id]) * 24 + viewpoint_id_b
data_ids_a[i] = data_id_a
data_ids_b[i] = data_id_b
viewpoint_ids_a[i] = viewpoint_id_a
viewpoint_ids_b[i] = viewpoint_id_b
images_a = torch.from_numpy(self.images[data_ids_a].astype('float32') /
255.)
images_b = torch.from_numpy(self.images[data_ids_b].astype('float32') /
255.)
distances = torch.ones(batch_size).float() * self.distance
elevations_a = torch.ones(batch_size).float() * self.elevation
elevations_b = torch.ones(batch_size).float() * self.elevation
viewpoints_a = srf.get_points_from_angles(distances, elevations_a,
-viewpoint_ids_a * 15)
viewpoints_b = srf.get_points_from_angles(distances, elevations_b,
-viewpoint_ids_b * 15)
return images_a, images_b, viewpoints_a, viewpoints_b
def get_random_batch(self, batch_size):
data_ids_a = np.zeros(batch_size, 'int32')
data_ids_b = np.zeros(batch_size, 'int32')
viewpoint_ids_a = torch.zeros(batch_size)
viewpoint_ids_b = torch.zeros(batch_size)
batch_a = []
batch_b = []
for i in range(batch_size):
class_idx = np.random.randint(0, len(self.class_ids))
class_id = self.class_ids[class_idx]
object_id = np.random.randint(0, self.num_data[class_id])
viewpoint_id_a = np.random.randint(0, 24)
viewpoint_id_b = np.random.randint(0, 24)
batch_a.append(self.images[class_idx][object_id, viewpoint_id_a])
batch_b.append(self.images[class_idx][object_id, viewpoint_id_b])
viewpoint_ids_a[i] = viewpoint_id_a
viewpoint_ids_b[i] = viewpoint_id_b
images_a = torch.from_numpy(np.stack(batch_a).astype('float32') / 255.)
images_b = torch.from_numpy(np.stack(batch_b).astype('float32') / 255.)
distances = torch.ones(batch_size).float() * self.distance
elevations_a = torch.ones(batch_size).float() * self.elevation
elevations_b = torch.ones(batch_size).float() * self.elevation
viewpoints_a = srf.get_points_from_angles(distances, elevations_a,
-viewpoint_ids_a * 15)
viewpoints_b = srf.get_points_from_angles(distances, elevations_b,
-viewpoint_ids_b * 15)
return images_a, images_b, viewpoints_a, viewpoints_b
def get_all_batches_for_evaluation(self, batch_size, class_id):
data_ids = np.arange(self.num_data[class_id]) + self.pos[class_id]
viewpoint_ids = np.tile(np.arange(24), data_ids.size)
data_ids = np.repeat(data_ids, 24) * 24 + viewpoint_ids
distances = torch.ones(data_ids.size).float() * self.distance
elevations = torch.ones(data_ids.size).float() * self.elevation
viewpoints_all = srf.get_points_from_angles(
distances, elevations,
-torch.from_numpy(viewpoint_ids).float() * 15)
for i in range((data_ids.size - 1) // batch_size + 1):
images = torch.from_numpy(
self.images_rgba[data_ids[i * batch_size:(i + 1) *
batch_size]].astype('float32') /
255.)
images_depth = torch.from_numpy(
self.images_depth[data_ids[i * batch_size:(i + 1) *
batch_size]].astype('float32'))
images_normal = torch.from_numpy(
self.images_normal[data_ids[i * batch_size:(i + 1) *
batch_size]].astype('float32'))
voxels = torch.from_numpy(
self.voxels[data_ids[i * batch_size:(i + 1) * batch_size] //
24].astype('float32'))
yield images, images_depth, images_normal, voxels
def get_specific_view_batch(self, batch_size,view_id):
data_ids_a = np.zeros(batch_size, 'int32')
viewpoint_ids_a = torch.zeros(batch_size)
for i in range(batch_size):
class_id = np.random.choice(self.class_ids)
object_id = np.random.randint(0, self.num_data[class_id])
viewpoint_id_a = view_id
data_id_a = (object_id + self.pos[class_id]) * 24 + viewpoint_id_a
data_ids_a[i] = data_id_a
viewpoint_ids_a[i] = viewpoint_id_a
images_a = torch.from_numpy(self.images[data_ids_a].astype('float32') / 255.)
distances = torch.ones(batch_size).float() * self.distance
elevations_a = torch.ones(batch_size).float() * self.elevation
viewpoints_a = nr.get_points_from_angles(distances, elevations_a, -viewpoint_ids_a * 15 + 90)
views_a = torch.cat([distances.unsqueeze(0),elevations_a.unsqueeze(0),-viewpoint_ids_a.unsqueeze(0) * 15 + 90])
return images_a, viewpoints_a,views_a
def get_random_single_batch(self, batch_size):
data_ids_a = np.zeros(batch_size, 'int32')
viewpoint_ids_a = torch.zeros(batch_size)
for i in range(batch_size):
class_id = np.random.choice(self.class_ids)
object_id = np.random.randint(0, self.num_data[class_id])
if self.view_range == None:
viewpoint_id_a = np.random.randint(0,24)
else:
viewpoint_id_a = np.random.choice(self.view_range)
viewpoint_id_a = np.random.randint(0, 24)
data_id_a = (object_id + self.pos[class_id]) * 24 + viewpoint_id_a
data_ids_a[i] = data_id_a
viewpoint_ids_a[i] = viewpoint_id_a
images_a = torch.from_numpy(self.images[data_ids_a].astype('float32') / 255.)
distances = torch.ones(batch_size).float() * self.distance
elevations_a = torch.ones(batch_size).float() * self.elevation
noise = (torch.rand(batch_size)-0.5) * 2 * self.view_noise
viewpoints_a = nr.get_points_from_angles(distances, elevations_a, -viewpoint_ids_a * 15 + 90 + noise)
views_a = torch.cat([distances.unsqueeze(0),elevations_a.unsqueeze(0),-viewpoint_ids_a.unsqueeze(0) * 15 + 90 + noise])
return images_a, viewpoints_a, views_a
def set_eyes_from_angles(self, distances, elevations, azimuths):
if self.camera_mode not in ['look', 'look_at']:
raise ValueError('Projection does not need to set eyes')
self.transformer._eye = srf.get_points_from_angles(
distances, elevations, azimuths)
def set_eyes_from_angles(self, distances, elevations, azimuths):
self._eye = srf.get_points_from_angles(distances, elevations, azimuths)
