from cmmlib import cmm
from cmmlib.inout import load_mesh
from cmmlib.vis.weights import show_weights
K1 = 8 + 6 + 12
K2 = 4
verts, tris = load_mesh('meshes/bumpy_cube6.obj')
Phi_cpr1 = cmm.compressed_manifold_modes(verts, tris, K1, mu=20., scaled=False)
Phi_cpr2 = cmm.compressed_manifold_modes(verts, tris, K2, mu=20., scaled=False)
Phi_vari1 = cmm.varimax_modes(verts, tris, K1)
Phi_vari2 = cmm.varimax_modes(verts, tris, K2)
# establish some consistent ordering of varimax modes just for visualization
i = (Phi_cpr2[:, 0]**2).argmax() # select one vertex
# permute according to activation strength of that vertex
Phi_vari2 = Phi_vari2[:, (Phi_vari2[i]**2).argsort()[::-1]]
Phi_vari1 = Phi_vari1[:, (Phi_vari1[i]**2).argsort()[::-1]]
Phi_cpr2 = Phi_cpr2[:, (Phi_cpr2[i]**2).argsort()[::-1]]
Phi_cpr1 = Phi_cpr1[:, (Phi_cpr1[i]**2).argsort()[::-1]]
show_weights(verts,
tris, (Phi_cpr1, Phi_cpr2, Phi_vari1, Phi_vari2),
('CMM, K=%d' % K1, 'CMM, K=%d' % K2, 'Variax, K=%d' % K1,
'Varimax, K=%d' % K2),
contours=7,
show_labels=True)
from cmmlib import cmm
from cmmlib.inout import load_mesh
from cmmlib.vis.weights import show_weights
K = 8 + 6 + 12
mu = 20.0
filename = 'meshes/bumpy_cube6.obj'
verts, tris = load_mesh(filename)
Phi_cpr = cmm.compressed_manifold_modes(verts, tris, K, mu=mu)
Phi_dense = cmm.manifold_harmonics(verts, tris, K)
Phi_vari = cmm.varimax_modes(verts, tris, K)
show_weights(verts, tris, (Phi_cpr, Phi_dense, Phi_vari),
('CMM', 'MH', 'Varimax'), show_labels=True)
from cmmlib import cmm
from cmmlib.inout import load_mesh
from cmmlib.vis.weights import show_weights
K1 = 8 + 6 + 12
K2 = 4
verts, tris = load_mesh('meshes/bumpy_cube6.obj')
Phi_cpr1 = cmm.compressed_manifold_modes(verts, tris, K1, mu=20., scaled=False)
Phi_cpr2 = cmm.compressed_manifold_modes(verts, tris, K2, mu=20., scaled=False)
Phi_vari1 = cmm.varimax_modes(verts, tris, K1)
Phi_vari2 = cmm.varimax_modes(verts, tris, K2)
# establish some consistent ordering of varimax modes just for visualization
i = (Phi_cpr2[:, 0]**2).argmax() # select one vertex
# permute according to activation strength of that vertex
Phi_vari2 = Phi_vari2[:, (Phi_vari2[i]**2).argsort()[::-1]]
Phi_vari1 = Phi_vari1[:, (Phi_vari1[i]**2).argsort()[::-1]]
Phi_cpr2 = Phi_cpr2[:, (Phi_cpr2[i]**2).argsort()[::-1]]
Phi_cpr1 = Phi_cpr1[:, (Phi_cpr1[i]**2).argsort()[::-1]]
show_weights(
verts, tris,
(Phi_cpr1, Phi_cpr2, Phi_vari1, Phi_vari2),
('CMM, K=%d' % K1, 'CMM, K=%d' % K2, 'Variax, K=%d' % K1, 'Varimax, K=%d' % K2),
contours=7, show_labels=True)
elif init == 1:
Phi_init = 1 * (color2[:, 0] > 200).astype(np.float).reshape(len(verts), K)
else:
Phi_init = np.random.uniform(-1, 1, (len(verts), K))
Phis = [Phi_init]
def callback(H, mu, Phi, E, S, **kwargs):
global i
i += 1
if i in at_iters:
Phis.append(Phi)
i = 0
cmm.compressed_manifold_modes(
verts, tris, K, mu=20, init=Phi_init, scaled=True,
callback=callback, tol_abs=0, tol_rel=0, maxiter=max(at_iters)+1,
check_interval=1)
all_Phis.append(Phis)
Phi_at_iters = np.array([np.array(Phis) for Phis in all_Phis])
show_weights(
verts, tris,
Phi_at_iters.reshape((num_inits*len(at_iters), len(verts), K)),
["Iteration %d (%d)" % (it, exp) if it > 0 else "Initialization (%d)" % exp
for exp, it in product(range(num_inits), at_iters)],
show_labels=True, label_offset_axis=1, offset_spacing2=1.3,
label_offset=0.8, num_columns=len(at_iters),
)
def main(input_filename, K, mu, output_dir=None, visualize=False, scaled=False,
maxiter=None, ply=False, off=False):
if (off or ply) and not output_dir:
print ("please specify an output directory")
return 1
if output_dir and not path.exists(output_dir):
print("%s does not exist" % output_dir)
return 2
verts, tris = load_mesh(input_filename, normalize=True)
print ("%d vertices, %d faces" % (len(verts), len(tris)))
Phi_cpr, D = cmm.compressed_manifold_modes(
verts, tris, K, mu=mu, scaled=scaled,
maxiter=maxiter, verbose=100, return_D=True)
if D is None:
D_diag = np.ones(len(verts))
D = sparse.eye(len(verts))
else:
D_diag = D.data
if visualize:
from cmmlib.vis.weights import show_weights
show_weights(verts, tris, Phi_cpr)
if output_dir:
# save in simple text format
np.savetxt(path.join(output_dir, 'phi.txt'), Phi_cpr, fmt='%f')
np.savetxt(path.join(output_dir, 'D_diag.txt'), D_diag, fmt='%f')
# save in matlab format
savemat(path.join(output_dir, 'phi.mat'),
dict(verts=verts, tris=tris+1, phi=Phi_cpr, D=D))
# save HDF5 format if possible
try:
import h5py
except ImportError:
print ("Cannot save as HDF5, please install the h5py module")
else:
with h5py.File(path.join(output_dir, 'phi.h5'), 'w') as f:
f['verts'] = verts
f['tris'] = tris
f['phi'] = Phi_cpr
f['d_diag'] = D_diag
# save NPY format
np.save(path.join(output_dir, 'phi.npy'), Phi_cpr)
np.save(path.join(output_dir, 'D_diag.npy'), Phi_cpr)
if off or ply:
# map phi scalars to colors
from mayavi.core.lut_manager import LUTManager
from cmmlib.vis.weights import _centered
lut = LUTManager(lut_mode='RdBu').lut.table.to_array()[:, :3]
colors = [
lut[(_centered(Phi_cpr[:, k]) * (lut.shape[0]-1)).astype(int)]
for k in range(K)]
# save in a single scene as a collage
w = int(np.ceil(np.sqrt(K))) if K > 6 else K
spacing = 1.2 * verts.ptp(axis=0)
all_verts = [verts + spacing * (1.5, 0, 0)]
all_tris = [tris]
all_color = [np.zeros(verts.shape, np.int) + 127]
for k in range(K):
all_verts.append(verts + spacing * (-(k % w), 0, int(k / w)))
all_tris.append(tris + len(verts) * (k+1))
all_color.append(colors[k])
if off:
save_coff(path.join(output_dir, 'input.off'),
verts.astype(np.float32), tris)
for k in range(K):
save_coff(path.join(output_dir, 'cmh_%03d.off' % k),
verts.astype(np.float32), tris, colors[k])
save_coff(path.join(output_dir, 'all.off'),
np.vstack(all_verts), np.vstack(all_tris),
np.vstack(all_color))
if ply:
from tvtk.api import tvtk
pd = tvtk.PolyData(
points=np.vstack(all_verts).astype(np.float32),
polys=np.vstack(all_tris).astype(np.uint32))
pd.point_data.scalars = np.vstack(all_color).astype(np.uint8)
pd.point_data.scalars.name = 'colors'
ply = tvtk.PLYWriter(
file_name=path.join(output_dir, 'all.ply'),
input=pd, color=(1, 1, 1))
ply.array_name = 'colors'
ply.write()
def main():
# prepare data
verts, tris = load_mesh('meshes/bumpy_cube6.obj')
K = 8 + 6 + 12
mu = 20.0
maxiter = 500
num_experiments = 1
# prepare data structures
ResultData = namedtuple('ResultData', 'Phi info logger')
data_per_algo = defaultdict(list)
# parameters passed to algorithms - to be fair,
# turn off automatic penalty adjustment in our method
# set check_interval = 1 to force callback to be called in every iteration
algorithms = [
('osher', cmm.solve_compressed_osher,
dict()),
('ours', cmm.solve_compressed_splitorth,
dict(auto_adjust_penalty=False, check_interval=1)),
]
# run the algorithm with same input mesh and
# different random initializations
print "will now run the method with %d different initializations" % num_experiments
print "the primal residual with our method should consistently be lower compared to Oshers et al."
for random_seed in range(num_experiments):
np.random.seed(random_seed)
Phi_init = np.random.random((len(verts), K))
print "--"
print "running experiment %d" % random_seed
for name, method, options in algorithms:
Phi, info, logger = test_convergence(
verts, tris, K, mu,
Phi_init, method, maxiter, **options)
print name, info['r_primal']
data_per_algo[name].append(ResultData(Phi, info, logger))
# visualize results
import matplotlib
if len(sys.argv) > 1:
matplotlib.use('pgf') # TODO PGF
else:
matplotlib.use('WxAgg')
import pylab as pl
pl.rc('text', usetex=True)
pl.rc('font', size=6)
pl.figure(figsize=(0.9, 1.1))
plot_fun = pl.semilogy
plot_fun(data_per_algo['osher'][0].logger.r_primals, c='r',
label=r'osher')
plot_fun(data_per_algo['ours'][0].logger.r_primals, c='g',
label=r'ours')
pl.legend(loc='center right').draggable()
pl.locator_params(nbins=4, axis='x')
pl.xlabel('iteration')
pl.ylabel(r'$ \left\Vert r \right\Vert_2$')
pl.gca().yaxis.set_major_locator(matplotlib.ticker.LogLocator(numticks=4))
if len(sys.argv) > 2:
pl.savefig(sys.argv[1]) #, bbox_inches='tight')
else:
pl.ion()
pl.show()
# let us first visualize the modes where things go wrong
# which are most probably those where the primal residual is very high
Q = data_per_algo['osher'][0].info['Q']
P = data_per_algo['osher'][0].info['P']
worst_first = ((Q - P)**2).sum(axis=0).argsort()[::-1]
show_weights(
verts, tris,
(data_per_algo['osher'][0].Phi[:, worst_first],
data_per_algo['ours'][0].Phi[:, worst_first]),
('osher', 'ours')
)
def main(input_filename, K, mu, output_dir=None, visualize=False, scaled=False,
maxiter=None, ply=False, off=False):
if (off or ply) and not output_dir:
print "please specify an output directory"
return 1
if output_dir and not path.exists(output_dir):
print "%s does not exist" % output_dir
return 2
verts, tris = load_mesh(input_filename, normalize=True)
print "%d vertices, %d faces" % (len(verts), len(tris))
Phi_cpr, D = cmm.compressed_manifold_modes(
verts, tris, K, mu=mu, scaled=scaled,
maxiter=maxiter, verbose=100, return_D=True)
if D is None:
D_diag = np.ones(len(verts))
D = sparse.eye(len(verts))
else:
D_diag = D.data
if visualize:
from cmmlib.vis.weights import show_weights
show_weights(verts, tris, Phi_cpr)
if output_dir:
# save in simple text format
np.savetxt(path.join(output_dir, 'phi.txt'), Phi_cpr, fmt='%f')
np.savetxt(path.join(output_dir, 'D_diag.txt'), D_diag, fmt='%f')
# save in matlab format
savemat(path.join(output_dir, 'phi.mat'),
dict(verts=verts, tris=tris+1, phi=Phi_cpr, D=D))
# save HDF5 format if possible
try:
import h5py
except ImportError:
print "Cannot save as HDF5, please install the h5py module"
else:
with h5py.File(path.join(output_dir, 'phi.h5'), 'w') as f:
f['verts'] = verts
f['tris'] = tris
f['phi'] = Phi_cpr
f['d_diag'] = D_diag
# save NPY format
np.save(path.join(output_dir, 'phi.npy'), Phi_cpr)
np.save(path.join(output_dir, 'D_diag.npy'), Phi_cpr)
if off or ply:
# map phi scalars to colors
from mayavi.core.lut_manager import LUTManager
from cmmlib.vis.weights import _centered
lut = LUTManager(lut_mode='RdBu').lut.table.to_array()[:, :3]
colors = [
lut[(_centered(Phi_cpr[:, k]) * (lut.shape[0]-1)).astype(int)]
for k in xrange(K)]
# save in a single scene as a collage
w = int(np.ceil(np.sqrt(K))) if K > 6 else K
spacing = 1.2 * verts.ptp(axis=0)
all_verts = [verts + spacing * (1.5, 0, 0)]
all_tris = [tris]
all_color = [np.zeros(verts.shape, np.int) + 127]
for k in xrange(K):
all_verts.append(verts + spacing * (-(k % w), 0, int(k / w)))
all_tris.append(tris + len(verts) * (k+1))
all_color.append(colors[k])
if off:
save_coff(path.join(output_dir, 'input.off'),
verts.astype(np.float32), tris)
for k in xrange(K):
save_coff(path.join(output_dir, 'cmh_%03d.off' % k),
verts.astype(np.float32), tris, colors[k])
save_coff(path.join(output_dir, 'all.off'),
np.vstack(all_verts), np.vstack(all_tris),
np.vstack(all_color))
if ply:
from tvtk.api import tvtk
pd = tvtk.PolyData(
points=np.vstack(all_verts).astype(np.float32),
polys=np.vstack(all_tris).astype(np.uint32))
pd.point_data.scalars = np.vstack(all_color).astype(np.uint8)
pd.point_data.scalars.name = 'colors'
ply = tvtk.PLYWriter(
file_name=path.join(output_dir, 'all.ply'),
input=pd, color=(1, 1, 1))
ply.array_name = 'colors'
ply.write()
from cmmlib import cmm
from cmmlib.inout import load_mesh
from cmmlib.vis.weights import show_weights
K = 8 + 6 + 12
mu = 20.0
filename = 'meshes/bumpy_cube6.obj'
verts, tris = load_mesh(filename)
Phi_cpr = cmm.compressed_manifold_modes(verts, tris, K, mu=mu)
Phi_dense = cmm.manifold_harmonics(verts, tris, K)
Phi_vari = cmm.varimax_modes(verts, tris, K)
show_weights(verts,
tris, (Phi_cpr, Phi_dense, Phi_vari), ('CMM', 'MH', 'Varimax'),
show_labels=True)
import numpy as np
from cmmlib import cmm
from cmmlib.align import optimal_permutation
from cmmlib.inout import load_mesh
from cmmlib.vis.weights import show_weights
verts, tris = load_mesh('meshes/hand_4054.obj')
K = 6
mus = [1/100., 1., 10.]
all_Phi = [cmm.manifold_harmonics(verts, tris, K)] \
+ [cmm.compressed_manifold_modes(verts, tris, K, mu, scaled=False)
for mu in mus]
# permute modes so we can visualize them side-by-side
all_Phi_aligned = \
[np.dot(optimal_permutation(Phi.T, all_Phi[-1].T, allow_reflection=True), Phi.T).T
for Phi in all_Phi[:-1]] + \
[all_Phi[-1]]
show_weights(verts, tris, all_Phi_aligned, names=['dense'] + map(str, mus))
tris,
K,
mu=20,
init=Phi_init,
scaled=True,
callback=callback,
tol_abs=0,
tol_rel=0,
maxiter=max(at_iters) + 1,
check_interval=1)
all_Phis.append(Phis)
Phi_at_iters = np.array([np.array(Phis) for Phis in all_Phis])
show_weights(
verts,
tris,
Phi_at_iters.reshape((num_inits * len(at_iters), len(verts), K)),
[
"Iteration %d (%d)" %
(it, exp) if it > 0 else "Initialization (%d)" % exp
for exp, it in product(range(num_inits), at_iters)
],
show_labels=True,
label_offset_axis=1,
offset_spacing2=1.3,
label_offset=0.8,
num_columns=len(at_iters),
)
