def test_scale_vertices(self):
# Run and check results visually.
# - Simple mesh.
mesh = pymesh.load_mesh(
os.path.join(current_dir, '../common_data/simple_mesh.ply'))
mesh = augmentation(mesh=mesh,
vertices_scale_mean=1.0,
vertices_scale_var=0.1)
if (not os.path.exists(os.path.join(current_dir, '../output_data/'))):
os.mkdir(os.path.join(current_dir, '../output_data'))
pymesh.save_mesh(mesh=mesh,
filename=os.path.join(
current_dir,
'../output_data/simple_mesh_scale_vertices.ply'),
ascii=True)
# - Modified non-flat version of simple mesh.
mesh = pymesh.load_mesh(
os.path.join(current_dir,
'../common_data/simple_mesh_nonflat.ply'))
mesh = augmentation(mesh=mesh,
vertices_scale_mean=1.0,
vertices_scale_var=0.1)
pymesh.save_mesh(
mesh=mesh,
filename=os.path.join(
current_dir,
'../output_data/simple_mesh_nonflat_scale_vertices.ply'),
ascii=True)
def process(self):
# Verify that no previous data is being inadvertently erased.
processed_data_folder = osp.join(self.processed_dir, self.__split)
if (osp.exists(processed_data_folder)):
if (os.listdir(processed_data_folder)):
raise IOError(
"The folder containing the processed data, "
f"'{processed_data_folder}', is not empty. Most likely the "
"root folder you have set for the current dataset "
"coincides with that of a previously-generated dataset, "
"and the current dataset has parameters not fully "
"compatible with those used to generate the data already "
"in the folder. Please double-check the dataset parameters "
"or delete the content of the folder/specify a different "
"root folder of the dataset.")
# Each category is assigned an index, to be used as target in
# classification.
category_indices = [
self.valid_categories.index(category)
for category in self.__categories
]
for category, category_index in zip(self.__categories,
category_indices):
process_subfolder = osp.join(processed_data_folder, category)
if (not osp.exists(process_subfolder)):
os.makedirs(process_subfolder)
paths = glob.glob(
osp.join(self.raw_dir, f'coseg_{category}', self.__split,
'*.ply'))
for path in paths:
# Mesh name without extension.
mesh_name = path.rpartition('/')[2].split('.')[0]
# Load mesh.
mesh = pymesh.load_mesh(path)
# Load per-face class label.
face_label_file = osp.join(self.raw_dir, f'coseg_{category}',
'seg', f"{mesh_name}.eseg")
with open(face_label_file, 'r') as f:
y = np.loadtxt(f, dtype='long')
# Preprocess mesh.
mesh = preprocess_mesh(
input_mesh=mesh,
prevent_nonmanifold_edges=self.__prevent_nonmanifold_edges)
# Perform data augmentation and post-augmentation.
for augmentation_idx in range(self.__num_augmentations):
augmented_mesh = augmentation(
mesh=mesh,
vertices_scale_mean=self.__vertices_scale_mean,
vertices_scale_var=self.__vertices_scale_var,
edges_flip_fraction=self.__edges_flip_fraction)
postaugmented_mesh = post_augmentation(
mesh=augmented_mesh,
slide_vertices_fraction=self.__slide_vertices_fraction)
# Convert the mesh to dual-primal graphs.
graph_creator = GraphCreator(
mesh=postaugmented_mesh,
single_dual_nodes=self.__single_dual_nodes,
undirected_dual_edges=self.__undirected_dual_edges,
primal_features_from_dual_features=self.
__primal_features_from_dual_features,
prevent_nonmanifold_edges=self.
__prevent_nonmanifold_edges)
primal_graph, dual_graph = graph_creator.create_graphs()
(primal_edge_to_dual_node_idx
) = graph_creator.primal_edge_to_dual_node_idx
# Add the ground-truth per-face class label to the primal
# graph.
assert (len(y) == primal_graph.num_nodes)
primal_graph.y = torch.from_numpy(y).long()
# Save the graphs and the dictionary.
torch.save(
primal_graph,
osp.join(
process_subfolder,
f"{mesh_name}_aug_{augmentation_idx}_primal.pt"))
torch.save(
dual_graph,
osp.join(
process_subfolder,
f"{mesh_name}_aug_{augmentation_idx}_dual.pt"))
petdni_filename = osp.join(
process_subfolder,
f"{mesh_name}_aug_{augmentation_idx}_petdni.pkl")
pymesh.save_mesh(
osp.join(process_subfolder,
f"{mesh_name}_aug_{augmentation_idx}.obj"),
postaugmented_mesh)
try:
with open(petdni_filename, 'wb') as f:
pkl.dump(primal_edge_to_dual_node_idx, f)
except IOError:
raise IOError("Error while writing file "
f"'{petdni_filename}'. Exiting.")
if (self.pre_filter is not None):
raise NotImplementedError
if (self.pre_transform is not None):
raise NotImplementedError
# Save the input parameters of the dataset, so that when using it
# without repreprocessing the data, one can make sure that the input
# parameters match those with which the preprocessed data saved to disk
# was generated.
dataset_parameters_filename = osp.join(
self.processed_dir, f'processed_data_params_{self.__split}.pkl')
try:
with open(dataset_parameters_filename, 'wb') as f:
pkl.dump(self.input_parameters, f)
except IOError:
raise IOError("Error while writing file dataset parameter file "
f"'{dataset_parameters_filename}'. Exiting.")