def generate_figure(z, num_samples, empty=False, disable={}, verbose=True):
degree, num_control_points, dim, is_closed = (z['degree'],
z['num_control_points'],
z['dim'], z['is_closed'])
if verbose:
print(' degree:', degree)
print(' num_control_points:', num_control_points)
print(' dim:', dim)
print(' is_closed:', is_closed)
c = UniformBSpline(degree, num_control_points, dim, is_closed=is_closed)
Y, w, u, X = [np.array(z[k]) for k in 'YwuX']
if verbose:
print(' num_data_points:', Y.shape[0])
kw = {}
if Y.shape[1] == 3:
kw['projection'] = '3d'
f = plt.figure()
if empty:
ax = f.add_axes((0, 0, 1, 1), **kw)
ax.set_xticks([])
ax.set_yticks([])
if Y.shape[1] == 3:
ax.set_zticks([])
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
for spine in ax.spines.values():
spine.set_visible(False)
else:
ax = f.add_subplot(111, **kw)
ax.set_aspect('equal')
def plot(X, *args, **kwargs):
ax.plot(*(tuple(X.T) + args), **kwargs)
if 'Y' not in disable:
plot(Y, '.', c=C['r'])
if 'u' not in disable:
for m, y in zip(c.M(u, X), Y):
plot(np.r_['0,2', m, y], '-', c=C['o'])
if 'X' not in disable:
plot(X, 'o--', ms=6.0, c='k', mec='k')
if 'M' not in disable:
plot(c.M(c.uniform_parameterisation(num_samples), X),
'-',
c=C['b'],
lw=3.0)
if not empty:
e = z.get('e')
if e is not None:
ax.set_title('Energy: {:.7e}'.format(e))
return f
def generate_figure(z, num_samples, empty=False, disable={}, verbose=True):
degree, num_control_points, dim, is_closed = (
z['degree'], z['num_control_points'], z['dim'], z['is_closed'])
if verbose:
print(' degree:', degree)
print(' num_control_points:', num_control_points)
print(' dim:', dim)
print(' is_closed:', is_closed)
c = UniformBSpline(degree, num_control_points, dim, is_closed=is_closed)
Y, w, u, X = [np.array(z[k]) for k in 'YwuX']
if verbose:
print(' num_data_points:', Y.shape[0])
kw = {}
if Y.shape[1] == 3:
kw['projection'] = '3d'
f = plt.figure()
if empty:
ax = f.add_axes((0, 0, 1, 1), **kw)
ax.set_xticks([])
ax.set_yticks([])
if Y.shape[1] == 3:
ax.set_zticks([])
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
for spine in ax.spines.values():
spine.set_visible(False)
else:
ax = f.add_subplot(111, **kw)
ax.set_aspect('equal')
def plot(X, *args, **kwargs):
ax.plot(*(tuple(X.T) + args), **kwargs)
if 'Y' not in disable:
plot(Y, '.', c=C['r'])
if 'u' not in disable:
for m, y in zip(c.M(u, X), Y):
plot(np.r_['0,2', m, y], '-', c=C['o'])
if 'X' not in disable:
plot(X, 'o--', ms=6.0, c='k', mec='k')
if 'M' not in disable:
plot(c.M(c.uniform_parameterisation(num_samples), X), '-',
c=C['b'], lw=3.0)
if not empty:
e = z.get('e')
if e is not None:
ax.set_title('Energy: {:.7e}'.format(e))
return f
def main():
parser = argparse.ArgumentParser()
parser.add_argument('input_path')
parser.add_argument('output_path')
parser.add_argument(
'solver_type',
nargs='?',
default='dn',
choices=UniformBSplineLeastSquaresOptimiser.SOLVER_TYPES)
parser.add_argument('--output-all', default=False, action='store_true')
parser.add_argument('--max-num-iterations', type=int, default=100)
parser.add_argument('--min-radius', type=float, default=1e-9)
parser.add_argument('--max-radius', type=float, default=1e12)
parser.add_argument('--initial-radius', type=float, default=1e4)
args = parser.parse_args()
print('Input:', args.input_path)
with open(args.input_path, 'r') as fp:
z = json.load(fp)
degree, num_control_points, dim, is_closed = (z['degree'],
z['num_control_points'],
z['dim'], z['is_closed'])
print(' degree:', degree)
print(' num_control_points:', num_control_points)
print(' dim:', dim)
print(' is_closed:', is_closed)
c = UniformBSpline(degree, num_control_points, dim, is_closed=is_closed)
Y, w, u, X = [np.array(z[k]) for k in 'YwuX']
lambda_ = z['lambda_']
print(' num_data_points:', Y.shape[0])
print(' lambda_:', lambda_)
print('UniformBSplineLeastSquaresOptimiser:')
print(' solver_type:', args.solver_type)
print(' max_num_iterations:', args.max_num_iterations)
print(' min_radius: {:g}'.format(args.min_radius))
print(' max_radius: {:g}'.format(args.max_radius))
print(' initial_radius: {:g}'.format(args.initial_radius))
print('UniformBSplineLeastSquaresOptimiser Output:')
(u1, X1, has_converged, states,
num_iterations, time_taken) = UniformBSplineLeastSquaresOptimiser(
c,
args.solver_type).minimise(Y,
w,
lambda_,
u,
X,
max_num_iterations=args.max_num_iterations,
min_radius=args.min_radius,
max_radius=args.max_radius,
initial_radius=args.initial_radius,
return_all=True)
print(' has_converged:', has_converged)
print(' num_iterations:', num_iterations)
print(' num_successful_iterations:', len(states) - 1)
print(' initial_energy: {:.3e}'.format(states[0][2]))
print(' final_energy: {:.3e}'.format(states[-1][2]))
print(' time_taken: {:.3e}s'.format(time_taken))
print(' per_iteration: {:.3e}s'.format(time_taken / num_iterations))
print('Output:', args.output_path)
if args.output_all:
if not os.path.exists(args.output_path):
os.makedirs(args.output_path)
for i, (u, X, e, radius) in enumerate(states):
z['u'], z['X'] = u.tolist(), X.tolist()
z['e'], z['radius'] = e, radius
output_path = os.path.join(args.output_path, '{}.json'.format(i))
print(' ', output_path)
with open(output_path, 'w') as fp:
fp.write(json.dumps(z, indent=4))
else:
z['u'], z['X'] = u1.tolist(), X1.tolist()
z['e'], z['radius'] = states[-1][2:]
with open(args.output_path, 'w') as fp:
fp.write(json.dumps(z, indent=4))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('num_data_points', type=int)
parser.add_argument('w', type=float_tuple)
parser.add_argument('lambda_', type=float)
parser.add_argument('degree', type=int)
parser.add_argument('num_control_points', type=int)
parser.add_argument('output_path')
parser.add_argument('--alpha', type=float, default=1.0 / (2.0 * np.pi))
parser.add_argument('--dim', type=int, choices={2, 3}, default=2)
parser.add_argument('--frequency', type=float, default=1.0)
parser.add_argument('--num-init-points', type=int, default=16)
parser.add_argument('--sigma', type=float, default=0.05)
parser.add_argument('--seed', type=int)
args = parser.parse_args()
print('Parameters:')
print(' alpha:', args.alpha)
print(' frequency:', args.frequency)
x = np.linspace(0.0, 2.0 * np.pi, args.num_data_points)
y = np.exp(-args.alpha * x) * np.sin(args.frequency * x)
if args.dim == 2:
Y = np.c_[x, y]
else:
Y = np.c_[x, y, np.linspace(0.0, 1.0, args.num_data_points)]
# Initialise `X` so that the uniform B-spline linearly interpolates between
# the first and last noise-free data points.
t = np.linspace(0.0, 1.0, args.num_control_points)[:, np.newaxis]
X = Y[0] * (1 - t) + Y[-1] * t
c = UniformBSpline(args.degree, args.num_control_points, args.dim)
m0, m1 = c.M(c.uniform_parameterisation(2), X)
x01 = 0.5 * (X[0] + X[-1])
X = (np.linalg.norm(Y[0] - Y[-1]) / np.linalg.norm(m1 - m0)) * (X -
x01) + x01
# Add isotropic zero-mean Gaussian noise to the data.
if args.seed is not None:
print(' seed:', args.seed)
np.random.seed(args.seed)
print(' sigma:', args.sigma)
Y += args.sigma * np.random.randn(Y.size).reshape(Y.shape)
# Set `w`.
if np.any(np.asarray(args.w) < 0):
raise ValueError('w <= 0.0 (= {})'.format(args.w))
if len(args.w) == 1:
w = np.empty((args.num_data_points, args.dim), dtype=float)
w.fill(args.w[0])
elif len(args.w) == args.dim:
w = np.tile(args.w, (args.num_data_points, 1))
else:
raise ValueError('len(w) is invalid (= {})'.format(len(args.w)))
# Check `lambda_`.
if args.lambda_ <= 0.0:
raise ValueError('lambda_ <= 0.0 (= {})'.format(args.lambda_))
# Initialise `u`.
u0 = c.uniform_parameterisation(args.num_init_points)
D = scipy.spatial.distance.cdist(Y, c.M(u0, X))
u = u0[D.argmin(axis=1)]
# Save.
to_list = lambda _: _.tolist()
z = dict(degree=args.degree,
num_control_points=args.num_control_points,
dim=args.dim,
is_closed=False,
Y=to_list(Y),
w=to_list(w),
lambda_=args.lambda_,
u=to_list(u),
X=to_list(X))
print('Output:', args.output_path)
with open(args.output_path, 'w') as fp:
fp.write(json.dumps(z, indent=4))
def fit_bspline(x, y, dim = 2, degree=2, num_control_points = 20,
is_closed = False, num_init_points=1000):
''' Fits and returns a bspline curve to the given x and y points
Parameters
----------
x : list
data x-coordinates
y : list
data y-coordinates
dim : int
the dimensionality of the dataset (default: 2)
degree : int
the degree of the b-spline polynomial (default: 2)
num_control_points : int
the number of b-spline control points (default: 20)
is_closed : boolean
should the b-spline be closed? (default: false)
num_init_points : int
number of initial points to use in the b-spline parameterization
when starting the regression. (default: 1000)
Returns
-------
c: a UniformBSpline object containing the optimized b-spline
'''
# TODO: extract dimensionality from the x,y dataset itself
num_data_points = len(x)
c = UniformBSpline(degree, num_control_points, dim, is_closed=is_closed)
Y = np.c_[x, y] # Data num_points by dimension
# Now we need weights for all of the data points
w = np.empty((num_data_points, dim), dtype=float)
# Currently, every point is equally important
w.fill(1) # Uniform weight to the different points
# Initialize `X` so that the uniform B-spline linearly interpolates between
# the first and last noise-free data points.
t = np.linspace(0.0, 1.0, num_control_points)[:, np.newaxis]
X = Y[0] * (1 - t) + Y[-1] * t
# NOTE: Not entirely sure if the next three lines are necessary or not
m0, m1 = c.M(c.uniform_parameterisation(2), X)
x01 = 0.5 * (X[0] + X[-1])
X = (np.linalg.norm(Y[0] - Y[-1]) / np.linalg.norm(m1 - m0)) * (X - x01) + x01
# Regularization weight on the control point distance
# This specifies a penalty on having the b-spline control points close
# together, and in some sense prevents over-fitting. Change this is the
# curve doesn't capture the curve variation well or smoothly enough
lambda_ = 0.5
# These parameters affect the regression solver.
# Presently, they are disabled below, but you can think about enabling them
# if that would be useful for your use case.
max_num_iterations = 1000
min_radius = 0
max_radius = 400
initial_radius = 100
# Initialize U
u0 = c.uniform_parameterisation(num_init_points)
D = cdist(Y, c.M(u0, X))
u = u0[D.argmin(axis=1)]
# Run the solver
(u, X, has_converged, states, num_iterations,
time_taken) = UniformBSplineLeastSquaresOptimiser(c,'lm').minimise(
Y, w, lambda_, u, X,
#max_num_iterations = max_num_iterations,
#min_radius = min_radius,
#max_radius = max_radius,
#initial_radius = initial_radius,
return_all=True)
return c,u0,X
def fit_bspline(x,
y,
dim=2,
degree=2,
num_control_points=20,
is_closed=False,
num_init_points=1000):
''' Fits and returns a bspline curve to the given x and y points
Parameters
----------
x : list
data x-coordinates
y : list
data y-coordinates
dim : int
the dimensionality of the dataset (default: 2)
degree : int
the degree of the b-spline polynomial (default: 2)
num_control_points : int
the number of b-spline control points (default: 20)
is_closed : boolean
should the b-spline be closed? (default: false)
num_init_points : int
number of initial points to use in the b-spline parameterization
when starting the regression. (default: 1000)
Returns
-------
c: a UniformBSpline object containing the optimized b-spline
'''
# TODO: extract dimensionality from the x,y dataset itself
num_data_points = len(x)
c = UniformBSpline(degree, num_control_points, dim, is_closed=is_closed)
Y = np.c_[x, y] # Data num_points by dimension
# Now we need weights for all of the data points
w = np.empty((num_data_points, dim), dtype=float)
# Currently, every point is equally important
w.fill(1) # Uniform weight to the different points
# Initialize `X` so that the uniform B-spline linearly interpolates between
# the first and last noise-free data points.
t = np.linspace(0.0, 1.0, num_control_points)[:, np.newaxis]
X = Y[0] * (1 - t) + Y[-1] * t
# NOTE: Not entirely sure if the next three lines are necessary or not
m0, m1 = c.M(c.uniform_parameterisation(2), X)
x01 = 0.5 * (X[0] + X[-1])
X = (np.linalg.norm(Y[0] - Y[-1]) / np.linalg.norm(m1 - m0)) * (X -
x01) + x01
# Regularization weight on the control point distance
# This specifies a penalty on having the b-spline control points close
# together, and in some sense prevents over-fitting. Change this if the
# curve doesn't capture the curve variation well or smoothly enough
lambda_ = 0.5
# These parameters affect the regression solver.
# Presently, they are disabled below, but you can think about enabling them
# if that would be useful for your use case.
# max_num_iterations = 1000
# min_radius = 0
# max_radius = 400
# initial_radius = 100
# Initialize U
u0 = c.uniform_parameterisation(num_init_points)
D = cdist(Y, c.M(u0, X))
u = u0[D.argmin(axis=1)]
# Run the solver
(
u, X, has_converged, states, num_iterations, time_taken
) = UniformBSplineLeastSquaresOptimiser(c, 'lm').minimise(
Y,
w,
lambda_,
u,
X,
#max_num_iterations = max_num_iterations,
#min_radius = min_radius,
#max_radius = max_radius,
#initial_radius = initial_radius,
return_all=True)
return c, u0, X
def main():
parser = argparse.ArgumentParser()
parser.add_argument('num_data_points', type=int)
parser.add_argument('w', type=float_tuple)
parser.add_argument('lambda_', type=float)
parser.add_argument('degree', type=int)
parser.add_argument('num_control_points', type=int)
parser.add_argument('output_path')
parser.add_argument('--alpha', type=float, default=1.0 / (2.0 * np.pi))
parser.add_argument('--dim', type=int, choices={2, 3}, default=2)
parser.add_argument('--frequency', type=float, default=1.0)
parser.add_argument('--num-init-points', type=int, default=16)
parser.add_argument('--sigma', type=float, default=0.05)
parser.add_argument('--seed', type=int)
args = parser.parse_args()
print('Parameters:')
print(' alpha:', args.alpha)
print(' frequency:', args.frequency)
x = np.linspace(0.0, 2.0 * np.pi, args.num_data_points)
y = np.exp(-args.alpha * x) * np.sin(args.frequency * x)
if args.dim == 2:
Y = np.c_[x, y]
else:
Y = np.c_[x, y, np.linspace(0.0, 1.0, args.num_data_points)]
# Initialise `X` so that the uniform B-spline linearly interpolates between
# the first and last noise-free data points.
t = np.linspace(0.0, 1.0, args.num_control_points)[:, np.newaxis]
X = Y[0] * (1 - t) + Y[-1] * t
c = UniformBSpline(args.degree, args.num_control_points, args.dim)
m0, m1 = c.M(c.uniform_parameterisation(2), X)
x01 = 0.5 * (X[0] + X[-1])
X = (np.linalg.norm(Y[0] - Y[-1]) / np.linalg.norm(m1 - m0)) * (X - x01) + x01
# Add isotropic zero-mean Gaussian noise to the data.
if args.seed is not None:
print(' seed:', args.seed)
np.random.seed(args.seed)
print(' sigma:', args.sigma)
Y += args.sigma * np.random.randn(Y.size).reshape(Y.shape)
# Set `w`.
if np.any(np.asarray(args.w) < 0):
raise ValueError('w <= 0.0 (= {})'.format(args.w))
if len(args.w) == 1:
w = np.empty((args.num_data_points, args.dim), dtype=float)
w.fill(args.w[0])
elif len(args.w) == args.dim:
w = np.tile(args.w, (args.num_data_points, 1))
else:
raise ValueError('len(w) is invalid (= {})'.format(len(args.w)))
# Check `lambda_`.
if args.lambda_ <= 0.0:
raise ValueError('lambda_ <= 0.0 (= {})'.format(args.lambda_))
# Initialise `u`.
u0 = c.uniform_parameterisation(args.num_init_points)
D = scipy.spatial.distance.cdist(Y, c.M(u0, X))
u = u0[D.argmin(axis=1)]
# Save.
to_list = lambda _: _.tolist()
z = dict(degree=args.degree,
num_control_points=args.num_control_points,
dim=args.dim,
is_closed=False,
Y=to_list(Y),
w=to_list(w),
lambda_=args.lambda_,
u=to_list(u),
X=to_list(X))
print('Output:', args.output_path)
with open(args.output_path, 'w') as fp:
fp.write(json.dumps(z, indent=4))
