Skip to content

Instantly share code, notes, and snippets.

@jgillis
Last active April 2, 2024 10:11
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save jgillis/e5eb8dd1450417cb308ae440ee69dfb4 to your computer and use it in GitHub Desktop.
Save jgillis/e5eb8dd1450417cb308ae440ee69dfb4 to your computer and use it in GitHub Desktop.
bspline fitting
from casadi import *
# Generate 3D data to fit
x = np.linspace(0,2,10)
y = np.linspace(2,4,10)
z = np.linspace(3,5,10)
[X,Y,Z] = np.meshgrid(x,y,z)
xyz_flat = np.vstack((X.ravel(),Y.ravel(),Z.ravel())).T
D = (X+Y)*Z
# Define type of BSpline fit
degree = [1,2,3]
n_dims = len(degree)
knots_precursor = [[0,0.5,1,1.5,2], [2,3,4], [3,4,5]]
# Now with multiplicities
knots = [([p[0]]*d)+p+([p[-1]]*d) for p,d in zip(knots_precursor,degree)]
# The most efficient way to fit a large amount of data to a bspline
# First, figure out the sparse numeric mapping between coefficients and outputs
J = MX.bspline_dual(xyz_flat.ravel(),knots,degree)
# Construct and solve a fitting problem to figure out the coefficients
opti = Opti()
coeff = opti.variable(J.size2())
opti.minimize(sumsqr(J @ coeff - D.ravel()))
# Note: if you want to incorporate spline derivatives into the object/constraints,
# there is manual work needed
opti.solver("ipopt")
sol = opti.solve()
coeff_fitted = sol.value(coeff)
# Numeric fitting is finished
# Now construct a BSpline CasADi Function
x = MX.sym("x",n_dims)
y = bspline(x,DM(coeff_fitted),knots,degree,1)
f = Function('spline',[x],[y])
# Test
print(f(vertcat(0,2,3)),(0+2)*3)
print(f(vertcat(0.17,2.8,3.5)),(0.17+2.8)*3.5)
from casadi import *
# Generate 3D data to fit
x = np.linspace(0,2,10)
y = np.linspace(2,4,10)
z = np.linspace(3,5,10)
[X,Y,Z] = np.meshgrid(x,y,z)
xyz_flat = np.vstack((X.ravel(),Y.ravel(),Z.ravel())).T
D = (X+Y)*Z
# Define type of BSpline fit
degree = [1,2,3]
n_dims = len(degree)
knots_precursor = [[0,0.5,1,1.5,2], [2,3,4], [3,4,5]]
# Now with multiplicities
knots = [([p[0]]*d)+p+([p[-1]]*d) for p,d in zip(knots_precursor,degree)]
# Construct a differentiable BSpline CasADi Function
x = MX.sym("x",n_dims)
C = MX.sym("C",100)
y = bspline(x,C,knots,degree,1,{"inline":True})
f = Function("spline",[x,C],[y])
# Fit in opti
opti = Opti()
coeff = opti.variable(100)
opti.minimize(sumsqr(f(xyz_flat.T,coeff).T-D.ravel()))
# Note: if you want to incorporate spline derivatives into the object/constraints,
# you can simply use CasADi AD on f
opti.solver("ipopt")
sol = opti.solve()
coeff_fitted = sol.value(coeff)
# Numeric fitting is finished
# Now construct a BSpline CasADi Function
x = MX.sym("x",n_dims)
y = bspline(x,DM(coeff_fitted),knots,degree,1)
f = Function('spline',[x],[y])
# Test
print(f(vertcat(0,2,3)),(0+2)*3)
print(f(vertcat(0.17,2.8,3.5)),(0.17+2.8)*3.5)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment