def test_wrong_method():
with pytest.raises(ValueError):
mcubes.smooth(
np.zeros((10, 10), dtype=np.bool_),
method='wrong',
max_iters=500,
rel_tol=1e-4
)
def test_sphere():
# Create sphere with radius 25 centered at (50, 50, 50)
x, y, z = np.mgrid[:100, :100, :100]
levelset = np.sqrt((x - 50)**2 + (y - 50)**2 + (z - 50)**2) - 25
# vertices, triangles = mcubes.marching_cubes(levelset, 0)
# mcubes.export_obj(vertices, triangles, 'sphere1.obj')
binary_levelset = levelset > 0
smoothed_levelset = mcubes.smooth(
binary_levelset,
method='constrained',
max_iters=500,
rel_tol=1e-4
)
vertices, _ = mcubes.marching_cubes(smoothed_levelset, 0.0)
# Check all vertices have same distance to (50, 50, 50)
dist = np.sqrt(np.sum((vertices - [50, 50, 50])**2, axis=1))
assert dist.min() > 24.5 and dist.max() < 25.5
assert np.all(np.abs(smoothed_levelset - levelset) < 1)
assert np.all((smoothed_levelset > 0) == binary_levelset)
def test_wrong_ndim():
binary_levelset = np.random.uniform(size=(10)) < 0.5
with pytest.raises(ValueError):
mcubes.smooth(
binary_levelset,
method='constrained',
max_iters=500,
rel_tol=1e-4
)
binary_levelset = np.random.uniform(size=(10, 10, 10, 10)) < 0.5
with pytest.raises(ValueError):
mcubes.smooth(
binary_levelset,
method='constrained',
max_iters=500,
rel_tol=1e-4
)
def test_circle():
x, y = np.mgrid[:100, :100]
levelset = np.sqrt((x - 50)**2 + (y - 50)**2) - 25
binary_levelset = levelset > 0
smoothed_levelset = mcubes.smooth(
binary_levelset,
max_iters=500,
rel_tol=1e-4
)
assert np.all(np.abs(smoothed_levelset - levelset) < 1)
assert np.all((smoothed_levelset > 0) == binary_levelset)
def smooth(volume: np.ndarray, mode: str) -> Tuple[np.ndarray, float]:
isovalue = 0.0
if mode == 'auto':
volume = mcubes.smooth(volume)
elif mode == 'constrained':
volume = mcubes.smooth_constrained(volume, max_iters=500)
elif mode == 'gaussian':
volume = mcubes.smooth_gaussian(volume)
elif mode == 'no':
isovalue = 0.5
else:
raise ValueError("unknown mode '{}'".format(mode))
return volume, isovalue
def test_gaussian_smoothing():
# Create sphere with radius 25 centered at (50, 50, 50)
x, y, z = np.mgrid[:100, :100, :100]
levelset = np.sqrt((x - 50)**2 + (y - 50)**2 + (z - 50)**2) - 25
binary_levelset = levelset > 0
smoothed_levelset = mcubes.smooth(
binary_levelset,
method='gaussian',
sigma=3
)
vertices, _ = mcubes.marching_cubes(smoothed_levelset, 0.0)
# Check all vertices have same distance to (50, 50, 50)
dist = np.sqrt(np.sum((vertices - [50, 50, 50])**2, axis=1))
assert dist.min() > 24 and dist.max() < 25
def marchingCubes(cls,
voxel_model: VoxelModel,
smooth: bool = False,
color: Tuple[float, float, float,
float] = (0.8, 0.8, 0.8, 1)):
"""
Generate a mesh object from a VoxelModel object using a marching cubes algorithm.
This meshing approach is best suited to high resolution models where some smoothing is acceptable.
Args:
voxel_model: VoxelModel object to be converted to a mesh
smooth: Enable smoothing
color: Mesh color in the format (r, g, b, a)
Returns:
None
"""
voxel_model_fit = voxel_model.fitWorkspace().getOccupied()
voxels = voxel_model_fit.voxels.astype(np.uint16)
x, y, z = voxels.shape
model_offsets = voxel_model_fit.coords
voxels_padded = np.zeros((x + 2, y + 2, z + 2))
voxels_padded[1:-1, 1:-1, 1:-1] = voxels
if smooth:
voxels_padded = mcubes.smooth(voxels_padded)
levelset = 0
else:
levelset = 0.5
verts, tris = mcubes.marching_cubes(voxels_padded, levelset)
# Shift model to align with origin
verts = np.subtract(verts, 0.5)
verts[:, 0] = np.add(verts[:, 0], model_offsets[0])
verts[:, 1] = np.add(verts[:, 1], model_offsets[1])
verts[:, 2] = np.add(verts[:, 2], model_offsets[2])
verts_colors = generateColors(len(verts), color)
return cls(voxels_padded, verts, verts_colors, tris,
voxel_model.resolution)
def compute_isosurfaces(
bands: Dict[Spin, np.ndarray],
kpoint_dim: Tuple[int, int, int],
fermi_level: float,
reciprocal_space: ReciprocalCell,
decimate_factor: Optional[float] = None,
decimate_method: str = "quadric",
smooth: bool = False,
) -> Dict[Spin, List[Tuple[np.ndarray, np.ndarray]]]:
"""
Compute the isosurfaces at a particular energy level.
Args:
bands: The band energies, given as a dictionary of ``{spin: energies}``, where
energies has the shape (nbands, nkpoints).
kpoint_dim: The k-point mesh dimensions.
fermi_level: The energy at which to calculate the Fermi surface.
reciprocal_space: The reciprocal space representation.
decimate_factor: If method is "quadric", factor is the scaling factor by which
to reduce the number of faces. I.e., final # faces = initial # faces *
factor. If method is "cluster", factor is the voxel size in which to
cluster points. Default is None (no decimation).
decimate_method: Algorithm to use for decimation. Options are "quadric" or
"cluster".
smooth: If True, will smooth resulting isosurface. Requires PyMCubes. Smoothing
algorithm will use constrained smoothing algorithm to preserve fine details
if input dimension is lower than (500, 500, 500), otherwise will apply a
Gaussian filter.
Returns:
A dictionary containing a list of isosurfaces as ``(vertices, faces)`` for
each spin channel.
"""
rlat = reciprocal_space.reciprocal_lattice
spacing = 1 / (np.array(kpoint_dim) - 1)
if smooth and mcubes is None:
smooth = False
warnings.warn("Smoothing disabled, install PyMCubes to enable smoothing.")
if decimate_factor is not None and open3d is None:
decimate_factor = None
warnings.warn("Decimation disabled, install open3d to enable decimation.")
isosurfaces = {}
for spin, ebands in bands.items():
ebands -= fermi_level
spin_isosurface = []
for band in ebands:
# check if band crosses fermi level
if np.nanmax(band) > 0 > np.nanmin(band):
band_data = band.reshape(kpoint_dim)
if smooth:
#MDF_COMMENT trying smooth
# smoothing algorithm requires input to have surface at 0.5
#MDF_COMMENT actually the smooth algorith takes the whole range of
#MDF_COMMENT values, you give the isosurface later, and the band_data is already
#MDF_COMMENT translated to the fermi level.
#MDF_COMMENT in the example at
#MDF_COMMENT https://github.com/pmneila/PyMCubes/tree/b6ce8cc7a4e86c9b4b1bedea7073f573b6c7b739
# 0.5 is the level of the sphere they want to show, but you can set
# whatever offset you want. you can also not set an offset and then ask
# for any other isolevel to the marching_cubes routine.
smoothed_band_data = mcubes.smooth(band_data) #-band_data.max() +0.5 )#MDF_COMMENT - 0.5)
# and outputs embedding array with values 0 and 1
verts, faces = mcubes.marching_cubes(smoothed_band_data, 0)
# have to manually set spacing with PyMCubes
verts = verts * spacing
# comes out as np.uint64, but trimesh doesn't like this
faces = faces.astype(np.int32)
else:
verts, faces, _, _ = marching_cubes(band_data, 0, spacing=spacing)
if decimate_factor:
verts, faces = decimate_mesh(
verts, faces, decimate_factor, method=decimate_method
)
if isinstance(reciprocal_space, WignerSeitzCell):
verts = np.dot(verts - 0.5, rlat) * 3
verts, faces = _trim_surface(reciprocal_space, verts, faces)
else:
# convert coords to cartesian
verts = np.dot(verts - 0.5, rlat)
spin_isosurface.append((verts, faces))
isosurfaces[spin] = spin_isosurface
return isosurfaces
def _calculate_band_isosurfaces(
spin: Spin,
band_idx: int,
energies: np.ndarray,
kpoint_dim: Tuple[int, int, int],
spacing: np.ndarray,
reference: np.ndarray,
reciprocal_space: ReciprocalCell,
decimate_factor: Optional[float],
decimate_method: str,
smooth: bool,
calculate_dimensionality: bool,
property_interpolator: Optional[LinearInterpolator],
):
"""Helper function to calculate the connected isosurfaces for a band."""
from skimage.measure import marching_cubes
from ifermi.analysis import (
connected_subsurfaces,
equivalent_surfaces,
isosurface_dimensionality,
)
rlat = reciprocal_space.reciprocal_lattice
if smooth and mcubes is None:
smooth = False
warnings.warn(
"Smoothing disabled, install PyMCubes to enable smoothing.")
if decimate_factor is not None and open3d is None:
decimate_factor = None
warnings.warn(
"Decimation disabled, install open3d to enable decimation.")
band_data = energies.reshape(kpoint_dim)
if smooth:
smoothed_band_data = mcubes.smooth(band_data)
verts, faces = mcubes.marching_cubes(smoothed_band_data, 0)
# have to manually set spacing with PyMCubes
verts *= spacing
# comes out as np.uint64, but trimesh doesn't like this
faces = faces.astype(np.int32)
else:
verts, faces, _, _ = marching_cubes(band_data, 0, spacing=spacing)
if decimate_factor:
verts, faces = decimate_mesh(verts,
faces,
decimate_factor,
method=decimate_method)
verts += reference
# break the isosurface into connected subsurfaces
subsurfaces = connected_subsurfaces(verts, faces)
if calculate_dimensionality:
# calculate dimensionality of periodically equivalent surfaces.
dimensionalities = {}
mapping = equivalent_surfaces([s[0] for s in subsurfaces])
for idx in mapping:
dimensionalities[idx] = isosurface_dimensionality(
*subsurfaces[idx])
else:
dimensionalities = None
mapping = np.zeros(len(subsurfaces))
isosurfaces = []
dimensionality = None
orientation = None
properties = None
for (subverts, subfaces), idx in zip(subsurfaces, mapping):
# convert vertices to cartesian coordinates
subverts = np.dot(subverts, rlat)
subverts, subfaces = trim_surface(reciprocal_space, subverts, subfaces)
if len(subverts) == 0:
# skip surfaces that do not enter the reciprocal space boundaries
continue
if calculate_dimensionality:
dimensionality, orientation = dimensionalities[idx]
if property_interpolator is not None:
properties = face_properties(property_interpolator, subverts,
subfaces, band_idx, spin, rlat)
isosurface = Isosurface(
vertices=subverts,
faces=subfaces,
band_idx=band_idx,
properties=properties,
dimensionality=dimensionality,
orientation=orientation,
)
isosurfaces.append(isosurface)
return isosurfaces
def npy2point(folder='ct_train', to_save='v', number_points=300, dim=3, crop_size=224, tocrop=False):
"""
convert .npy to point cloud
:param folder: the folder name of the data set
:param to_save: choose which oen to save. 'v' represents vertices, 'p' represents plots, '' represents all.
:param number_points: number of points for each point cloud
:param dim: the dimension of the point clouds
:param crop_size: the size of the cropped masks / gt
:param tocrop: whether to crop the mask / gt
:return:
"""
assert to_save=='' or to_save=='v' or to_save=='p'
import mcubes
crop_from = 128 - crop_size//2
crop_to = 128 + crop_size//2
vertices_fold = os.path.join('../../input/PnpAda_release_data/', folder, 'vertices/')
plots_fold = os.path.join('../../input/PnpAda_release_data/', folder, 'plots/')
if not os.path.exists(vertices_fold):
os.mkdir(vertices_fold)
if not os.path.exists(plots_fold):
os.mkdir(plots_fold)
folder_path = os.path.join('../../input/PnpAda_release_data/', folder, "mask/")
for path in tqdm(glob.glob(folder_path + '*.npy')):
filename = os.path.splitext(os.path.basename(path))[0]
vertices_path = os.path.join(vertices_fold, filename + '.npy')
plot_path = os.path.join(plots_fold, filename + '.npy')
if not os.path.exists(vertices_path):
mask = np.load(path)
if args.toplot:
from matplotlib import pyplot as plt
temp = mask[...,0][crop_from:crop_to, crop_from:crop_to] if tocrop else mask[...,0]
plt.imshow(temp)
plt.show()
mask = np.where(mask > 0, 1, 0)
mask = np.moveaxis(mask, -1, 0)
if tocrop:
mask = crop_volume(mask, crop_size=crop_size)
mask = np.concatenate([mask, mask, mask], axis=0)
point_cloud = np.zeros((crop_size, crop_size)) if tocrop else np.zeros((256, 256))
vertices_array = np.zeros((number_points, dim))
if mask.sum() > 50:
vol = mcubes.smooth(mask)
vertices, triangles = mcubes.marching_cubes(vol, 0)
try:
vertices = graipher(vertices, number_points, dim=dim)
except:
print(filename)
exit()
vertices_array = np.array(vertices, dtype=np.int)
if args.toplot:
fig = plt.figure()
from mpl_toolkits.mplot3d import Axes3D
ax = fig.add_subplot(111, projection='3d')
ax.scatter(vertices[:, 0], vertices[:, 1], vertices[:, 2], s=10)
plt.show()
point_cloud[vertices_array[:,1], vertices_array[:,2]] = 1
if args.toplot:
plt.imshow(point_cloud, cmap='gray')
plt.show()
if to_save=='v' or to_save=='':
np.save(vertices_path, vertices_array)
if to_save=='p' or to_save=='':
np.save(plot_path, point_cloud)
# mcubes.export_mesh(vertices, triangles, "heart_single_slice.dae", "MyHeart_s")
print("finish")
def array2mesh(array, thresh=0., dim=3, coords=None, bbox=np.array([[-1,-1,-1],[1,1,1]]), return_coords=False, \
if_decimate=False, decimate_face=4096, cart_coord=True, gaussian_sigma=None):
"""from 1-D array to 3D mesh
Args:
array (np.ndarray): 1-D array
thresh (float, optional): threshold. Defaults to 0..
dim (int, optional): 2 or 3, curve or mesh. Defaults to 3.
coords (np.ndarray, optional): array's coordinates (num_points, x_dim). Defaults to None.
bbox (np.ndarray, optional): bounding box of coords. Defaults to np.array([[-1,-1],[1,1]]).
return_coords (bool, optional): whether return the coords. Defaults to False.
decimate_face (int, optional): whether to simplify the mesh. Defaults to 4096.
cart_coord (bool, optional): cartesian coordinate in array form, x->i, y->j,... and all varibles increases monotonically. Defaults to True.
gaussian_sigma (float, optional): sigma value for gaussian filter (set None if there is no filter)
Returns:
tuple: `verts`, `faces`, `coords` or `verts`, `faces` according to `return_coords`
"""
grid = nputil.array2NDCube(array, N=dim)
if gaussian_sigma is not None:
#from scipy.ndimage import gaussian_filter
#grid = gaussian_filter(grid.astype(float), sigma=gaussian_sigma)
grid = mcubes.smooth(grid)
if dim == 3:
verts, faces = mcubes.marching_cubes(grid, thresh)
if cart_coord == False:
verts = verts[:, [1, 0, 2]]
verts = verts / (grid.shape[0] - 1) # rearrange order and rescale
elif dim == 2:
contours = find_contours(grid, thresh)
vcount, points, edges = 0, [], []
for contour in contours:
ind = np.arange(len(contour))
points.append(contour)
edges.append(np.c_[vcount + ind,
vcount + (ind + 1) % len(contour)])
vcount += len(contour)
if len(contours) == 0:
return None, None
verts = np.concatenate(points, axis=0)[:, [1, 0]] / (grid.shape[0] - 1)
#verts = verts[:,[1,0]]
faces = np.concatenate(edges, axis=0)
#levelset_samples = igl.sample_edges(points, edges, 10)
if coords is not None:
bbmin, bbmax = nputil.arrayBBox(coords)
else:
bbmin, bbmax = bbox
coords = nputil.makeGrid(bb_min=bbmin, bb_max=bbmax, shape=grid.shape)
verts = verts * (bbmax - bbmin) + bbmin
verts, faces = verts, faces.astype(int)
if if_decimate == True:
if dim != 3:
print("Warning! decimation only works for 3D")
elif faces.shape[0] > decimate_face: # Only decimate when appropriate
reached, verts, faces, _, _ = igl.decimate(verts, faces,
decimate_face)
faces = faces.astype(int)
if return_coords == True:
return verts, faces, coords
else:
return verts, faces
