def cluster_grasps(grasps, qualities):
cluster_indexes = np.asarray(
utils.farthest_points(grasps, num_clusters,
utils.distance_by_translation_grasp))
output_grasps = []
output_qualities = []
for i in range(num_clusters):
indexes = np.where(cluster_indexes == i)[0]
if ratio_of_grasps_to_be_used < 1:
num_grasps_to_choose = max(
1,
int(ratio_of_grasps_to_be_used * float(len(indexes))))
if len(indexes) == 0:
raise NoPositiveGraspsException
indexes = np.random.choice(indexes,
size=num_grasps_to_choose,
replace=False)
output_grasps.append(grasps[indexes, :, :])
output_qualities.append(qualities[indexes])
output_grasps = np.asarray(output_grasps)
output_qualities = np.asarray(output_qualities)
return output_grasps, output_qualities
def apply_dropout(self, pc):
if self.opt.occlusion_nclusters == 0 or self.opt.occlusion_dropout_rate == 0.:
return np.copy(pc)
labels = utils.farthest_points(pc, self.opt.occlusion_nclusters,
utils.distance_by_translation_point)
removed_labels = np.unique(labels)
removed_labels = removed_labels[np.random.rand(removed_labels.shape[0])
< self.opt.occlusion_dropout_rate]
if removed_labels.shape[0] == 0:
return np.copy(pc)
mask = np.ones(labels.shape, labels.dtype)
for l in removed_labels:
mask = np.logical_and(mask, labels != l)
return pc[mask]