tf-bspline

A TensorFlow implementation of a b-spline interpolator. This is done so that TensorFlow can follow the operations (i.e. for use in calculating derivatives). This utilizes the cardinal b-spline interpolation from tensorflow-graphics, however, only select files are taken from that package because it takes a very long time to import. Be aware that updates to tensorflow-graphics will need to be manually incorporated in tf-bspline.

This package will not exceed the speed of scipy's interpolation operations and is in fact much slower.

BSpline()

BSpline(start_position, end_position, num_internal_knots, degree)

Constructor.

Creates a Tensorflow-based b-spline object for use with x,y coordinate data.

start_position-> start x-value of the bspline.

end_position-> end x-value of the bspline.

num_internal_knots-> minimum number of knots that will be placed in b-spline range (more will actually be used in total, including those outside of specified range). Actual knots used may be obtained from get_knots().

degree-> degree of the b-spline. Only up to degree 4 is supported.

set_knot_values(knot_y_values)

Manually set the knot y-values.

knot_y_values-> new y-values. Must include exactly the same number of values as number of knots given by get_knots().

get_knots()

Returns-> current knot positions and values.

fit_points(x, y)

Fits b-spline to given x, y data by adjusting internal knot y-values

x-> x values

y-> y values

_call_(positions=None, raster=None)

Returns the calculated b-spline points. MUST SPECIFY ONLY POSITIONS OR RASTER, NOT BOTH

positions-> specific x-values to return values for.

raster-> division length to divide full b-spline range by.

Returns-> b-spline points with shape=(2, num_points). i.e. ([x], [y])

Example

Here is an example test script with output.

from tfbspline import BSpline
import numpy as np
import matplotlib.pyplot as plt

# Create junk data calculated with y = x^2 + noise
start = -5
end = 2
data_x = np.linspace(start, end, 40)
noise = np.random.normal(0, 0.05, 40)
data_y = np.square(np.linspace(-1, 1, 40)) + noise

# create b-spline
spline = BSpline(start, end, num_internal_knots=3, degree=3)
# store original knots to plot later
original_knots = spline.get_knots()

# fit spline to data
spline.fit_points(data_x, data_y)

# get fitted knots
fitted_knots = spline.get_knots()

# plot fitted spline
spline_points = spline(raster=0.001)
plt.plot(spline_points[:, 0], spline_points[:, 1], 'g')

# plot knots
plt.plot(fitted_knots[0], fitted_knots[1], 'r.')

# plot original spline
spline.set_knot_values(original_knots[1])
spline_points = spline(raster=0.001)
plt.plot(spline_points[:, 0], spline_points[:, 1], 'y')

# plot data
plt.plot(data_x, data_y, 'b.')

# add legend
plt.legend(['fitted b-spline', 'knots', 'unfitted b-spline', 'data'], loc='best')
plt.show()

psomers3 / tf-bspline