def optimization(ratio):
lr0 = 0.0005
lr = lr0
folder_name = os.getcwd() + '/progress-b-1-%f/' % ratio
if not os.path.isdir(folder_name):
os.mkdir(folder_name)
filename = os.getcwd() + '/../exp_bunny/setup/bunny_transient.mat'
setup = scipy.io.loadmat(filename)
gt_transient = np.array(setup['gt_transient'], dtype=np.double, order='C')
gt_mesh = MESH()
gt_mesh.v = igl.eigen.MatrixXd(np.array(setup['gt_v'], dtype=np.double))
gt_mesh.f = igl.eigen.MatrixXd(np.array(setup['gt_f'],
dtype=np.double)).castint()
opt = OPT(20000)
opt.max_distance_bin = gt_transient.shape[1]
opt.smooth_weight = 0.001
mesh = MESH()
mesh_init_location = os.getcwd() + '/init/cnlos_bunny_threshold.obj'
igl.readOBJ(mesh_init_location, mesh.v, mesh.f)
mesh.v = np.array(mesh.v, dtype=np.float32, order='C')
mesh.f = np.array(mesh.f, dtype=np.int32, order='C')
opt.resolution = 64
opt.smooth_ratio = ratio
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
weight = rendering.create_weighting_function(gt_transient, opt.gamma)
global_counter = 0
old_v = np.array(mesh.v)
old_f = np.array(mesh.f)
global_counter, convergence_flag, l2_record = optimize_parameters.optimize_shape(
mesh, gt_transient, weight, opt, 15, lr, gt_mesh, global_counter,
folder_name)
grad_v = np.array(mesh.v)
rendering.el_topo_gradient(mesh, old_v)
rendering.el_topo_remeshing(mesh, .5 / opt.resolution)
eltopo_v = np.array(mesh.v)
eltopo_f = np.array(mesh.f)
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
isotropic_v = np.array(mesh.v)
isotropic_f = np.array(mesh.f)
filename = folder_name + 'data.mat'
scipy.io.savemat(filename,
mdict={
'old_v': old_v,
'old_f': old_f,
'grad_v': grad_v,
'eltopo_v': eltopo_v,
'eltopo_f': eltopo_f,
'isotropic_v': isotropic_v,
'isotropic_f': isotropic_f
})
def optimization(ratio):
lr0 = 0.0001
lr = lr0
folder_name = os.getcwd() + '/progress-b-1-%f/' % ratio
if not os.path.isdir(folder_name):
os.mkdir(folder_name)
filename = os.getcwd() + '/setup/bunny_transient_64_0.mat'
setup = scipy.io.loadmat(filename)
gt_transient = np.array(setup['gt_transient'], dtype=np.double, order='C')
gt_mesh = MESH()
gt_mesh.v = igl.eigen.MatrixXd(np.array(setup['gt_v'], dtype=np.double))
gt_mesh.f = igl.eigen.MatrixXd(np.array(setup['gt_f'],
dtype=np.double)).castint()
opt = OPT(20000)
opt.max_distance_bin = gt_transient.shape[1]
opt.smooth_weight = 0.001
mesh = MESH()
mesh_init_location = os.getcwd() + '/init/cnlos_bunny_threshold_64.obj'
igl.readOBJ(mesh_init_location, mesh.v, mesh.f)
mesh.v = np.array(mesh.v, dtype=np.float32, order='C')
mesh.f = np.array(mesh.f, dtype=np.int32, order='C')
opt.resolution = 64
opt.smooth_ratio = ratio
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
weight = rendering.create_weighting_function(gt_transient, opt.gamma)
optimization_alpha = torch.Tensor([mesh.alpha])
optimization_alpha.requires_grad_()
optimizer_alpha = optim.Adam([optimization_alpha], lr=opt.alpha_lr)
dummy_loss2 = optimization_alpha**2
dummy_loss2.backward()
global_counter = 0
for t in range(3):
if mesh.f.shape[0] > 250000:
break
global_counter, l2_record = optimize_parameters.optimize_alpha(
mesh, gt_transient, weight, optimization_alpha, optimizer_alpha,
opt, 50, global_counter)
if t == 0:
l2_0 = l2_record
lr = (l2_record / l2_0) * lr0 * ((0.99)**(t / 50))
print('new lr %f' % lr)
old_v = np.array(mesh.v)
global_counter, convergence_flag = optimize_parameters.optimize_shape(
mesh, gt_transient, weight, opt, 15, lr, gt_mesh, global_counter)
if convergence_flag:
if opt.testing_flag == 1:
opt.testing_flag = 0
else:
opt.testing_flag = 1
rendering.el_topo_gradient(mesh, old_v)
rendering.el_topo_remeshing(mesh, .5 / opt.resolution)
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
rendering.compute_mesh_affinity(mesh)
rendering.removeTriangle(mesh, opt)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
def optimization(ratio):
lr0 = 0.0001
lr = lr0
folder_name = os.getcwd() + '/progress-b-1-%f/' % ratio
if not os.path.isdir(folder_name):
os.mkdir(folder_name)
filename = os.getcwd() + '/transient.mat'
setup = scipy.io.loadmat(filename)
gt_transient = np.array(setup['transient'], dtype=np.double, order='C')
opt = OPT(20000)
opt.max_distance_bin = gt_transient.shape[1]
opt.smooth_weight = 0.001
opt.lighting = np.array(setup['lighting'], dtype=np.float32, order='C')
opt.sensor = np.array(setup['lighting'], dtype=np.float32, order='C')
opt.gt_mesh = False
gt_mesh = MESH()
mesh = MESH()
mesh_init_location = os.getcwd() + '/cnlos_s_threshold.obj'
igl.readOBJ(mesh_init_location, mesh.v, mesh.f)
mesh.v = np.array(mesh.v, dtype=np.float32, order='C')
mesh.f = np.array(mesh.f, dtype=np.int32, order='C')
opt.resolution = 64
opt.smooth_ratio = ratio
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
weight = rendering.create_weighting_function(gt_transient, opt.gamma)
mesh.albedo = optimize_parameters.initial_fitting_albedo(
mesh, gt_transient, weight, opt)
global_counter = 0
l2 = np.empty(400)
albedo = np.empty(400)
resolution_cnt = 0
for t in range(400):
if mesh.f.shape[0] > 250000:
break
if t < 30 or resolution_cnt < 5:
global_counter, l2_record = optimize_parameters.optimize_albedo(
mesh, gt_transient, weight, opt, 50, global_counter,
folder_name)
if t == 0:
l2_0 = l2_record
opt.albedo_lr = (l2_reccord / l2_0) * opt_albedo_lr0 * (
(0.99)**(t / 5))
lr = (l2_record / l2_0) * lr0 * ((0.99)**(t / 5))
print('new lr %f' % lr)
old_v = np.array(mesh.v)
global_counter, convergence_flag, l2_record = optimize_parameters.optimize_shape(
mesh, gt_transient, weight, opt, 15, lr, gt_mesh, global_counter,
folder_name)
resolution_cnt += 1
if resolution_cnt == 50:
resolution_cnt = 0
opt.resolution *= 1.5
opt.sample_num *= 1.5
if convergence_flag:
if opt.testing_flag == 1:
opt.testing_flag = 0
opt.smooth_ratio = ratio / 10 + t / 100
print('shading')
else:
opt.testing_flag = 1
opt.smooth_ratio = ratio + t / 10
rendering.el_topo_gradient(mesh, old_v)
rendering.el_topo_remeshing(mesh, .5 / opt.resolution)
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
rendering.compute_mesh_affinity(mesh)
rendering.removeTriangle(mesh, opt)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
filename = folder_name + '%05d.mat' % (t)
scipy.io.savemat(filename,
mdict={
'v': mesh.v,
'f': mesh.f,
'albedo': mesh.albedo
})
l2[t] = l2_record
albedo[t] = mesh.albedo
filename = folder_name + 'progress.mat'
scipy.io.savemat(filename, mdict={'albedo': albedo, 'l2': l2})
def optimization(ratio):
lr0 = 0.0001
lr = lr0
folder_name = os.getcwd() + '/progress-b-3-%f/' % ratio
if not os.path.isdir(folder_name):
os.mkdir(folder_name)
test_num = 1
filename = os.getcwd() + '/setup/bunny_transient_64_0_%d.mat' % test_num
setup = scipy.io.loadmat(filename)
gt_transient = np.array(setup['gt_transient'], dtype=np.double, order='C')
gt_mesh = MESH()
gt_mesh.v = igl.eigen.MatrixXd(np.array(setup['gt_v'], dtype=np.double))
gt_mesh.f = igl.eigen.MatrixXd(np.array(setup['gt_f'],
dtype=np.double)).castint()
opt = OPT(20000)
opt.max_distance_bin = gt_transient.shape[1]
opt.smooth_weight = 0.001
mesh = MESH()
mesh_init_location = os.getcwd(
) + '/init/cnlos_bunny_threshold_64_0_%d.obj' % test_num
igl.readOBJ(mesh_init_location, mesh.v, mesh.f)
mesh.v = np.array(mesh.v, dtype=np.float32, order='C')
mesh.f = np.array(mesh.f, dtype=np.int32, order='C')
opt.resolution = 64
opt.smooth_ratio = ratio
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
weight = rendering.create_weighting_function(gt_transient, opt.gamma)
global_counter = 0
l2 = np.empty(500)
alpha = np.empty(500)
for t in range(500):
if mesh.f.shape[0] > 250000:
break
global_counter, l2_record = optimize_parameters.optimize_alpha(
mesh, gt_transient, weight, opt, 50, global_counter)
if t == 0:
l2_0 = l2_record
lr = (l2_record / l2_0) * lr0 * ((0.99)**(t / 2))
print('new lr %f' % lr)
old_v = np.array(mesh.v)
global_counter, convergence_flag, l2_record = optimize_parameters.optimize_shape(
mesh, gt_transient, weight, opt, 15, lr, gt_mesh, global_counter)
if convergence_flag:
if opt.testing_flag == 1:
opt.testing_flag = 0
opt.smooth_ratio = ratio / 10 + t / 100
print('shading')
else:
opt.testing_flag = 1
opt.smooth_ratio = ratio + t / 10
rendering.el_topo_gradient(mesh, old_v)
rendering.el_topo_remeshing(mesh, .5 / opt.resolution)
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
rendering.compute_mesh_affinity(mesh)
rendering.removeTriangle(mesh, opt)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
filename = folder_name + '%05d.mat' % (t)
scipy.io.savemat(filename,
mdict={
'v': mesh.v,
'f': mesh.f,
'alpha': mesh.alpha
})
l2[t] = l2_record
alpha[t] = mesh.alpha
filename = folder_name + 'progress.mat'
scipy.io.savemat(filename, mdict={'alpha': alpha, 'l2': l2})
def optimization(ratio):
lr0 = 0.0001 / 3
lr = lr0
folder_name = os.getcwd() + '/progress-b-1-%f/' % ratio
if not os.path.isdir(folder_name):
os.mkdir(folder_name)
resolution = 256
batch = 16
opt = OPT(20000, resolution)
a = np.arange(opt.lighting.shape[0])
b = np.split(a, batch)
gt_transient = np.empty(
(opt.resolution * opt.resolution, opt.max_distance_bin),
dtype=np.double,
order='C')
for i, x in zip(range(len(b)), b):
filename = os.getcwd() + '/setup/bunny_transient_%d_%d.mat' % (
resolution, i)
setup = scipy.io.loadmat(filename)
gt_transient[x, :] = np.array(setup['gt_transient'],
dtype=np.double,
order='C')
gt_mesh = MESH()
gt_mesh.v = igl.eigen.MatrixXd(np.array(setup['gt_v'], dtype=np.double))
gt_mesh.f = igl.eigen.MatrixXd(np.array(setup['gt_f'],
dtype=np.double)).castint()
opt.smooth_weight = 0.001
mesh = MESH()
mesh_init_location = os.getcwd(
) + '/init/cnlos_bunny_threshold_%d.obj' % resolution
igl.readOBJ(mesh_init_location, mesh.v, mesh.f)
mesh.v = np.array(mesh.v, dtype=np.float32, order='C')
mesh.f = np.array(mesh.f, dtype=np.int32, order='C')
opt.smooth_ratio = ratio
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
optimization_v = torch.from_numpy(mesh.v[mesh.v_edge == 0, :])
optimization_v.requires_grad_()
optimization_v_edge = torch.from_numpy(mesh.v[mesh.v_edge == 1, :])
optimization_v_edge.requires_grad_()
weight = rendering.create_weighting_function(gt_transient, opt.gamma)
optimizer = Adam_Modified([{
'params': optimization_v
}, {
'params': optimization_v_edge,
'lr': lr * opt.edge_lr_ratio
}],
lr=lr)
dummy_loss = torch.sum(optimization_v) + torch.sum(optimization_v_edge)
dummy_loss.backward()
scale_flag = False
remesh_flag = False
weight_flag = True
run_count = 0
l2_record = np.empty(opt.T)
v2_record = np.empty(opt.T)
l2_original_record = np.empty(opt.T)
for t in range(opt.T):
if remesh_flag:
print('remesh')
if mesh.f.shape[0] > 250000:
l2_record = l2_record[range(t)]
l2_original_record = l2_original_record[range(t)]
v2_record = v2_record[range(t)]
break
eltopo_tic = time.time()
rendering.el_topo_gradient(mesh, old_v)
mesh.v = np.array(mesh.v, dtype=np.double, order='C')
rendering.el_topo_remeshing(mesh, .5 / opt.resolution)
rendering.isotropic_remeshing(mesh, .5 / opt.resolution)
rendering.compute_mesh_affinity(mesh)
rendering.removeTriangle(mesh, opt)
#rendering.keep_largest_connected_component(mesh)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
optimization_v = torch.from_numpy(mesh.v[mesh.v_edge == 0, :])
optimization_v.requires_grad_()
optimization_v_edge = torch.from_numpy(mesh.v[mesh.v_edge == 1, :])
optimization_v_edge.requires_grad_()
weight = rendering.create_weighting_function(
gt_transient, opt.gamma)
optimizer = Adam_Modified([{
'params': optimization_v
}, {
'params': optimization_v_edge,
'lr': lr * opt.edge_lr_ratio
}],
lr=lr)
dummy_loss = torch.sum(optimization_v) + torch.sum(
optimization_v_edge)
dummy_loss.backward()
remesh_flag = False
run_count = 0
print('t = %d' % t)
tic = time.time()
transient = np.zeros(gt_transient.shape, dtype=np.double, order='C')
grad = np.zeros(mesh.v.shape, dtype=np.double, order='C')
tmp_opt = copy.deepcopy(opt)
for i, x in zip(range(len(b)), b):
tmp_opt.lighting = opt.lighting[x, :]
tmp_opt.lighting_normal = opt.lighting_normal[x, :]
transient[x, :], grad_tmp, length = rendering.inverseRendering(
mesh, gt_transient[x, :], weight[x, :], tmp_opt)
grad += grad_tmp
grad /= len(b)
smoothing_val, smoothing_grad = rendering.renderStreamedNormalSmoothing(
mesh)
l2, original_l2 = rendering.evaluate_loss_with_normal_smoothness(
gt_transient, weight, transient, smoothing_val, mesh, opt)
if weight_flag:
opt.smooth_weight = original_l2 / smoothing_val / opt.smooth_ratio
weight_flag = False
print('new smooth weight %f' % opt.smooth_weight)
if t > 0:
lr = (original_l2 / l2_original_record[0]) * lr0 * (
(0.99)**(t / 15))
print('new lr %f' % lr)
grad += opt.smooth_weight * smoothing_grad
v2 = rendering.compute_v2(mesh.v, gt_mesh)
print(
'%05d update time: %8.8f L2 loss: %8.8f old_l2 loss: %8.8f v2: %8.8f'
% (t, time.time() - tic, l2, original_l2, v2))
l2_record[t] = l2
l2_original_record[t] = original_l2
v2_record[t] = v2
filename = folder_name + '%05d.mat' % (t)
scipy.io.savemat(filename,
mdict={
'v': mesh.v,
'f': mesh.f,
'transient': transient,
'l2': l2,
'l2_original': original_l2,
'grad': grad,
'smoothing_grad': smoothing_grad,
'sample_num': opt.sample_num
})
run_count += 1
if run_count > 2:
if ((l2_original_record[t - 1] - original_l2) /
l2_original_record[t - 1]) < opt.loss_epsilon or (
(l2_record[t - 1] - l2) /
l2_record[t - 1]) < opt.loss_epsilon:
if opt.testing_flag == 1:
opt.testing_flag = 0
opt.smooth_ratio = 10 + t / 100
print('shading based')
else:
opt.testing_flag = 1
opt.resolution *= 1.5
opt.sample_num *= 1.5
opt.loss_epsilon /= 2
opt.smooth_ratio = ratio + t / 10
print('remesh %d' % opt.resolution)
remesh_flag = True
weight_flag = True
#opt.gamma
#weight = rendering.create_weighting_function(gt_transient, opt.gamma)
#opt.smooth_ratio *= 1.5
continue
optimization_v.grad.data = torch.from_numpy(
grad[mesh.v_edge == 0, :]).float()
optimization_v_edge.grad.data = torch.from_numpy(
grad[mesh.v_edge == 1, :]).float()
optimizer.step()
mesh.v[mesh.v_edge == 0, :] = np.array(optimization_v.data.numpy(),
dtype=np.float32,
order='C')
mesh.v[mesh.v_edge == 1, :] = np.array(
optimization_v_edge.data.numpy(), dtype=np.float32, order='C')
if run_count == 15:
remesh_flag = True
filename = folder_name + 'loss_val.mat'
scipy.io.savemat(filename,
mdict={
'l2': l2_record,
'l2_original_record': l2_original_record,
'v2_record': v2_record,
'weight': weight
})
exp_folder = os.getcwd() + '/weight_test2/'
if not os.path.isdir(exp_folder):
os.mkdir(exp_folder)
#gamma_list = [-2,-0.5,0, 1, 2.5, 5, 7.5, 10]
gamma_list = [-2, -0.5, 0, 5, 10]
for gamma in gamma_list:
print('%f'%(gamma))
folder_name = exp_folder + 'progress_%f/'%(gamma)
if not os.path.isdir(folder_name):
os.mkdir(folder_name)
#else:
# continue
weight = rendering.create_weighting_function(gt_transient, gamma, 'i')
mesh = MESH()
z = .38
v = np.array([[-.25, -.25, z], [.25, -.25, z], [.25, .25, z], [-.25, .25, z], [0, -.25, z], [.25, 0, z], [0, .25, z], [-.25,0,z], [0,0,z]])
f = np.array([[0, 8,4], [0,7,8], [4,8,1],[1,8,5],[8,2,5],[8,6,2],[7,3,8],[3,6,8]])
mesh.v = np.array(v, dtype=np.float32, order = 'C')
mesh.f = np.array(f, dtype=np.int32, order = 'C')
#mesh_init_location = os.getcwd() + '/cnlos_bunny_threshold.obj'
#igl.readOBJ(mesh_init_location, mesh.v, mesh.f)
#mesh.v = np.array(mesh.v, dtype=np.float32, order = 'C')
#mesh.f = np.array(mesh.f, dtype=np.int32, order = 'C')
mesh_optimization = MESH()
mesh_optimization.v = torch.from_numpy(mesh.v)
mesh_optimization.v.requires_grad_()
def optimization(ratio):
lr0 = 0.0001
lr = lr0
folder_name = os.getcwd() + '/progress-b-1-%f/'% ratio
if not os.path.isdir(folder_name):
os.mkdir(folder_name)
filename = os.getcwd() + '/transient.mat'
setup = scipy.io.loadmat(filename)
gt_transient = np.array(setup['rect_data'], dtype=np.double, order = 'C')
gt_transient[:,:,range(600}] = 0
gt_transient = np.reshape((gt_transient, opt.resolution*opt.resolution, 2048))
opt = OPT(20000)
mesh = MESH()
mesh_init_location = os.getcwd() + '/cnlos_s_threshold.obj'
igl.readOBJ(mesh_init_location, mesh.v, mesh.f)
mesh.v = np.array(mesh.v, dtype=np.float32, order = 'C')
mesh.f = np.array(mesh.f, dtype=np.int32, order = 'C')
rendering.isotropic_remeshing(mesh, .5/opt.resolution)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
optimization_v = torch.from_numpy(mesh.v[mesh.v_edge==0,:])
optimization_v.requires_grad_()
optimization_v_edge = torch.from_numpy(mesh.v[mesh.v_edge==1,:])
optimization_v_edge.requires_grad_()
weight = rendering.create_weighting_function(gt_transient, opt.gamma)
optimizer = Adam_Modified([{'params':optimization_v}, {'params': optimization_v_edge, 'lr': lr*opt.edge_lr_ratio}], lr = lr)
dummy_loss = torch.sum(optimization_v) + torch.sum(optimization_v_edge)
dummy_loss.backward()
scale_flag = False
remesh_flag = False
weight_flag = True
run_count = 0
l2_record = np.empty(opt.T)
v2_record = np.empty(opt.T)
l2_original_record = np.empty(opt.T)
for t in range(opt.T):
if remesh_flag:
print('remesh')
if mesh.f.shape[0] > 250000:
l2_record = l2_record[range(t)]
l2_original_record = l2_original_record[range(t)]
v2_record = v2_record[range(t)]
break
eltopo_tic = time.time()
rendering.el_topo_gradient(mesh, old_v)
mesh.v = np.array(mesh.v, dtype=np.double, order='C')
rendering.el_topo_remeshing(mesh, .5/opt.resolution)
rendering.isotropic_remeshing(mesh, .5/opt.resolution)
rendering.removeTriangle(mesh,opt)
#rendering.keep_largest_connected_component(mesh)
old_v = np.array(mesh.v)
rendering.compute_mesh_affinity(mesh)
rendering.border_indicator(mesh)
optimization_v = torch.from_numpy(mesh.v[mesh.v_edge==0,:])
optimization_v.requires_grad_()
optimization_v_edge = torch.from_numpy(mesh.v[mesh.v_edge==1,:])
optimization_v_edge.requires_grad_()
weight = rendering.create_weighting_function(gt_transient, opt.gamma)
optimizer = Adam_Modified([{'params':optimization_v}, {'params': optimization_v_edge, 'lr': lr*opt.edge_lr_ratio}], lr = lr)
dummy_loss = torch.sum(optimization_v) + torch.sum(optimization_v_edge)
dummy_loss.backward()
remesh_flag = False
run_count = 0
print('t = %d'% t)
tic = time.time()
transient, grad, length = rendering.inverseRendering(mesh, gt_transient, weight, opt)
smoothing_val, smoothing_grad = rendering.renderStreamedNormalSmoothing(mesh)
l2, original_l2 = rendering.evaluate_loss_with_normal_smoothness(gt_transient, weight, transient, smoothing_val, mesh, opt)
if weight_flag:
opt.smooth_weight = original_l2/smoothing_val/opt.smooth_ratio
weight_flag = False
print('new smooth weight %f' %opt.smooth_weight)
if t > 0:
lr = (original_l2/l2_original_record[0]) * lr0 * ((0.99)**(t/15))
print('new lr %f' % lr)
grad += opt.smooth_weight * smoothing_grad
print('%05d update time: %8.8f L2 loss: %8.8f old_l2 loss: %8.8f'% (t, time.time() - tic, l2, original_l2))
l2_record[t] = l2
l2_original_record[t] = original_l2
v2_record[t] = v2
filename = folder_name + '%05d.mat'%(t)
scipy.io.savemat(filename, mdict={'v':mesh.v, 'f':mesh.f, 'transient':transient, 'l2':l2, 'l2_original':original_l2, 'grad': grad, 'smoothing_grad':smoothing_grad, 'sample_num': opt.sample_num})
run_count += 1
if run_count > 2:
if ((l2_original_record[t-1] - original_l2)/l2_original_record[t-1])< opt.loss_epsilon or ((l2_record[t-1] - l2)/l2_record[t-1])< opt.loss_epsilon:
if opt.testing_flag == 1:
opt.testing_flag = 0
opt.smooth_ratio = 10 + t/100
print('shading based')
else:
opt.testing_flag = 1
opt.resolution *= 1.5
opt.sample_num *= 1.5
opt.loss_epsilon /= 2
opt.smooth_ratio = ratio + t/10
print('remesh %d'%opt.resolution)
remesh_flag = True
weight_flag = True
continue
optimization_v.grad.data = torch.from_numpy(grad[mesh.v_edge==0,:]).float()
optimization_v_edge.grad.data = torch.from_numpy(grad[mesh.v_edge==1,:]).float()
optimizer.step()
mesh.v[mesh.v_edge==0,:] = np.array(optimization_v.data.numpy(), dtype=np.float32, order='C')
mesh.v[mesh.v_edge==1,:] = np.array(optimization_v_edge.data.numpy(), dtype=np.float32, order='C')
if run_count == 15:
remesh_flag = True
filename = folder_name + 'loss_val.mat'
scipy.io.savemat(filename, mdict={'l2':l2_record, 'l2_original_record': l2_original_record, 'weight':weight})