NumPy Broadcasting: Calculating sum of squared differences between two arrays - python

I have the following code. It is taking forever in Python. There must be a way to translate this calculation into a broadcast...
def euclidean_square(a,b):
squares = np.zeros((a.shape[0],b.shape[0]))
for i in range(squares.shape[0]):
for j in range(squares.shape[1]):
diff = a[i,:] - b[j,:]
sqr = diff**2.0
squares[i,j] = np.sum(sqr)
return squares

You can use np.einsum after calculating the differences in a broadcasted way, like so -
ab = a[:,None,:] - b
out = np.einsum('ijk,ijk->ij',ab,ab)
Or use scipy's cdist with its optional metric argument set as 'sqeuclidean' to give us the squared euclidean distances as needed for our problem, like so -
from scipy.spatial.distance import cdist
out = cdist(a,b,'sqeuclidean')

I collected the different methods proposed here, and in two other questions, and measured the speed of the different methods:
import numpy as np
import scipy.spatial
import sklearn.metrics
def dist_direct(x, y):
d = np.expand_dims(x, -2) - y
return np.sum(np.square(d), axis=-1)
def dist_einsum(x, y):
d = np.expand_dims(x, -2) - y
return np.einsum('ijk,ijk->ij', d, d)
def dist_scipy(x, y):
return scipy.spatial.distance.cdist(x, y, "sqeuclidean")
def dist_sklearn(x, y):
return sklearn.metrics.pairwise.pairwise_distances(x, y, "sqeuclidean")
def dist_layers(x, y):
res = np.zeros((x.shape[0], y.shape[0]))
for i in range(x.shape[1]):
res += np.subtract.outer(x[:, i], y[:, i])**2
return res
# inspired by the excellent https://github.com/droyed/eucl_dist
def dist_ext1(x, y):
nx, p = x.shape
x_ext = np.empty((nx, 3*p))
x_ext[:, :p] = 1
x_ext[:, p:2*p] = x
x_ext[:, 2*p:] = np.square(x)
ny = y.shape[0]
y_ext = np.empty((3*p, ny))
y_ext[:p] = np.square(y).T
y_ext[p:2*p] = -2*y.T
y_ext[2*p:] = 1
return x_ext.dot(y_ext)
# https://stackoverflow.com/a/47877630/648741
def dist_ext2(x, y):
return np.einsum('ij,ij->i', x, x)[:,None] + np.einsum('ij,ij->i', y, y) - 2 * x.dot(y.T)
I use timeit to compare the speed of the different methods. For the comparison, I use vectors of length 10, with 100 vectors in the first group, and 1000 vectors in the second group.
import timeit
p = 10
x = np.random.standard_normal((100, p))
y = np.random.standard_normal((1000, p))
for method in dir():
if not method.startswith("dist_"):
continue
t = timeit.timeit(f"{method}(x, y)", number=1000, globals=globals())
print(f"{method:12} {t:5.2f}ms")
On my laptop, the results are as follows:
dist_direct 5.07ms
dist_einsum 3.43ms
dist_ext1 0.20ms <-- fastest
dist_ext2 0.35ms
dist_layers 2.82ms
dist_scipy 0.60ms
dist_sklearn 0.67ms
While the two methods dist_ext1 and dist_ext2, both based on the idea of writing (x-y)**2 as x**2 - 2*x*y + y**2, are very fast, there is a downside: When the distance between x and y is very small, due to cancellation error the numerical result can sometimes be (very slightly) negative.

Another solution besides using cdist is the following
difference_squared = np.zeros((a.shape[0], b.shape[0]))
for dimension_iterator in range(a.shape[1]):
difference_squared = difference_squared + np.subtract.outer(a[:, dimension_iterator], b[:, dimension_iterator])**2.

Related

How do I speed up my RBF kernel function in my k-means algorithm

I am tring to implement the RBF Kernel Function for my kernel k-means alg. Here is my formula.
And then I implement it with Numpy, but there's a two-layer for loop, and I'm thinking about how to turn it into a matrix operation. Because if I could do matrix operations, it would be a lot faster to process my 784-dimensional data. Or maybe my implemention is not correct? Can someone help me?
import numpy as np
def get_gamma(X, Y):
gamma = 0
for x in X:
for y in Y:
tmp = x - y
gamma += tmp**2
gamma = gamma / (length**2)
return gamma
def kernel(X, Y, gamma):
up = np.sum(np.power(X-Y, 2))
res = np.exp(-*up/gamma)
return res
def kernel_distance(X, Y):
gamma = get_gamma(X, Y)
a = kernel(X, X, gamma)
b = kernel(Y, Y, gamma)
c = kernel(X, Y, gamma)
return np.sqrt(a+b-2*c)
That's odd if I run your code it gives me a number for k. But shouldn't it be an array? Also shouldn't X and Y be 2d since those are basically a list of your points? Anyways if I take my own X and Y
from scipy.spatial.distance import cdist
import numpy as np
n = 10
X = np.random.random((n,3))
Y = np.random.random((n,3))
I can solve your problem like this
norms_sq = cdist(X,Y,'sqeuclidean')
two_sigma_sq = 1/n**2*np.sum(norms_sq)
k = np.exp(-norms_sq/two_sigma_sq)

How to do n-D distance and nearest neighbor calculations on numpy arrays

This question is intended to be a canonical duplicate target
Given two arrays X and Y of shapes (i, n) and (j, n), representing lists of n-dimensional coordinates,
def test_data(n, i, j, r = 100):
X = np.random.rand(i, n) * r - r / 2
Y = np.random.rand(j, n) * r - r / 2
return X, Y
X, Y = test_data(3, 1000, 1000)
what are the fastest ways to find:
The distance D with shape (i,j) between every point in X and every point in Y
The indices k_i and distance k_d of the k nearest neighbors against all points in X for every point in Y
The indices r_i, r_j and distance r_d of every point in X within distance r of every point j in Y
Given the following sets of restrictions:
Only using numpy
Using any python package
Including the special case:
Y is X
In all cases distance primarily means Euclidean distance, but feel free to highlight methods that allow other distance calculations.
#1. All Distances
only using numpy
The naive method is:
D = np.sqrt(np.sum((X[:, None, :] - Y[None, :, :])**2, axis = -1))
However this takes up a lot of memory creating an (i, j, n)-shaped intermediate matrix, and is very slow
However, thanks to a trick from #Divakar (eucl_dist package, wiki), we can use a bit of algebra and np.einsum to decompose as such: (X - Y)**2 = X**2 - 2*X*Y + Y**2
D = np.sqrt( # (X - Y) ** 2
np.einsum('ij, ij ->i', X, X)[:, None] + # = X ** 2 \
np.einsum('ij, ij ->i', Y, Y) - # + Y ** 2 \
2 * X.dot(Y.T)) # - 2 * X * Y
Y is X
Similar to above:
XX = np.einsum('ij, ij ->i', X, X)
D = np.sqrt(XX[:, None] + XX - 2 * X.dot(X.T))
Beware that floating-point imprecision can make the diagonal terms deviate very slightly from zero with this method. If you need to make sure they are zero, you'll need to explicitly set it:
np.einsum('ii->i', D)[:] = 0
Any Package
scipy.spatial.distance.cdist is the most intuitive builtin function for this, and far faster than bare numpy
from scipy.spatial.distance import cdist
D = cdist(X, Y)
cdist can also deal with many, many distance measures as well as user-defined distance measures (although these are not optimized). Check the documentation linked above for details.
Y is X
For self-referring distances, scipy.spatial.distance.pdist works similar to cdist, but returns a 1-D condensed distance array, saving space on the symmetric distance matrix by only having each term once. You can convert this to a square matrix using squareform
from scipy.spatial.distance import pdist, squareform
D_cond = pdist(X)
D = squareform(D_cond)
#2. K Nearest Neighbors (KNN)
Only using numpy
We could use np.argpartition to get the k-nearest indices and use those to get the corresponding distance values. So, with D as the array holding the distance values obtained above, we would have -
if k == 1:
k_i = D.argmin(0)
else:
k_i = D.argpartition(k, axis = 0)[:k]
k_d = np.take_along_axis(D, k_i, axis = 0)
However we can speed this up a bit by not taking the square roots until we have reduced our dataset. np.sqrt is the slowest part of calculating the Euclidean norm, so we don't want to do that until the end.
D_sq = np.einsum('ij, ij ->i', X, X)[:, None] +\
np.einsum('ij, ij ->i', Y, Y) - 2 * X.dot(Y.T)
if k == 1:
k_i = D_sq.argmin(0)
else:
k_i = D_sq.argpartition(k, axis = 0)[:k]
k_d = np.sqrt(np.take_along_axis(D_sq, k_i, axis = 0))
Now, np.argpartition performs indirect partition and doesn't necessarily give us the elements in sorted order and only makes sure that the first k elements are the smallest ones. So, for a sorted output, we need to use argsort on the output from previous step -
sorted_idx = k_d.argsort(axis = 0)
k_i_sorted = np.take_along_axis(k_i, sorted_idx, axis = 0)
k_d_sorted = np.take_along_axis(k_d, sorted_idx, axis = 0)
If you only need, k_i, you never need the square root at all:
D_sq = np.einsum('ij, ij ->i', X, X)[:, None] +\
np.einsum('ij, ij ->i', Y, Y) - 2 * X.dot(Y.T)
if k == 1:
k_i = D_sq.argmin(0)
else:
k_i = D_sq.argpartition(k, axis = 0)[:k]
k_d_sq = np.take_along_axis(D_sq, k_i, axis = 0)
sorted_idx = k_d_sq.argsort(axis = 0)
k_i_sorted = np.take_along_axis(k_i, sorted_idx, axis = 0)
X is Y
In the above code, replace:
D_sq = np.einsum('ij, ij ->i', X, X)[:, None] +\
np.einsum('ij, ij ->i', Y, Y) - 2 * X.dot(Y.T)
with:
XX = np.einsum('ij, ij ->i', X, X)
D_sq = XX[:, None] + XX - 2 * X.dot(X.T))
Any Package
KD-Tree is a much faster method to find neighbors and constrained distances. Be aware the while KDTree is usually much faster than brute force solutions above for 3d (as long as oyu have more than 8 points), if you have n-dimensions, KDTree only scales well if you have more than 2**n points. For discussion and more advanced methods for high dimensions, see Here
The most recommended method for implementing KDTree is to use scipy's scipy.spatial.KDTree or scipy.spatial.cKDTree
from scipy.spatial import KDTree
X_tree = KDTree(X)
k_d, k_i = X_tree.query(Y, k = k)
Unfortunately scipy's KDTree implementation is slow and has a tendency to segfault for larger data sets. As pointed out by #HansMusgrave here, pykdtree increases the performance a lot, but is not as common an include as scipy and can only deal with Euclidean distance currently (while the KDTree in scipy can handle Minkowsi p-norms of any order)
X is Y
Use instead:
k_d, k_i = X_tree.query(X, k = k)
Arbitrary metrics
A BallTree has similar algorithmic properties to a KDTree. I'm not aware of a parallel/vectorized/fast BallTree in Python, but using scipy we can still have reasonable KNN queries for user-defined metrics. If available, builtin metrics will be much faster.
def d(a, b):
return max(np.abs(a-b))
tree = sklearn.neighbors.BallTree(X, metric=d)
k_d, k_i = tree.query(Y)
This answer will be wrong if d() is not a metric. The only reason a BallTree is faster than brute force is because the properties of a metric allow it to rule out some solutions. For truly arbitrary functions, brute force is actually necessary.
#3. Radius search
Only using numpy
The simplest method is just to use boolean indexing:
mask = D_sq < r**2
r_i, r_j = np.where(mask)
r_d = np.sqrt(D_sq[mask])
Any Package
Similar to above, you can use scipy.spatial.KDTree.query_ball_point
r_ij = X_tree.query_ball_point(Y, r = r)
or scipy.spatial.KDTree.query_ball_tree
Y_tree = KDTree(Y)
r_ij = X_tree.query_ball_tree(Y_tree, r = r)
Unfortunately r_ij ends up being a list of index arrays that are a bit difficult to untangle for later use.
Much easier is to use cKDTree's sparse_distance_matrix, which can output a coo_matrix
from scipy.spatial import cKDTree
X_cTree = cKDTree(X)
Y_cTree = cKDTree(Y)
D_coo = X_cTree.sparse_distance_matrix(Y_cTree, r = r, output_type = `coo_matrix`)
r_i = D_coo.row
r_j = D_coo.column
r_d = D_coo.data
This is an extraordinarily flexible format for the distance matrix, as it stays an actual matrix (if converted to csr) can also be used for many vectorized operations.

Doing many iterations of curve_fit in one go for piecewise function

I'm trying to perform what are many iterations of Scipy's curve_fit at once in order to avoid loops and therefore increase speed.
This is very similar to this problem, which was solved. However, the fact that the functions are piece-wise (discontinuous) makes so that that solution isn't applicable here.
Consider this example:
import numpy as np
from numpy import random as rng
from scipy.optimize import curve_fit
rng.seed(0)
N=20
X=np.logspace(-1,1,N)
Y = np.zeros((4, N))
for i in range(0,4):
b = i+1
a = b
print(a,b)
Y[i] = (X/b)**(-a) #+ 0.01 * rng.randn(6)
Y[i, X>b] = 1
This yields these arrays:
Which as you can see are discontinuous at X==b. I can retrieve the original values of a and b by using curve_fit iteratively:
def plaw(r, a, b):
""" Theoretical power law for the shape of the normalized conditional density """
import numpy as np
return np.piecewise(r, [r < b, r >= b], [lambda x: (x/b)**-a, lambda x: 1])
coeffs=[]
for ix in range(Y.shape[0]):
print(ix)
c0, pcov = curve_fit(plaw, X, Y[ix])
coeffs.append(c0)
But this process can be very slow depending of the size of X, Y and the loop, so I'm trying to speed things up by trying to get coeffs without the need for a loop. So far I haven't had any luck.
Things that might be important:
X and Y only contain positive values
a and b are always positive
Although the data to fit in this example is smooth (for the sake of simplicity), the real data has noise
EDIT
This is as far as I've gotten:
y=np.ma.masked_where(Y<1.01, Y)
lX = np.log(X)
lY = np.log(y)
A = np.vstack([lX, np.ones(len(lX))]).T
m,c=np.linalg.lstsq(A, lY.T)[0]
print('a=',-m)
print('b=',np.exp(-c/m))
But even without any noise the output is:
a= [0.18978965578339158 1.1353633705997466 2.220234483915197 3.3324502660995714]
b= [339.4090881838179 7.95073481873057 6.296592007396107 6.402567167503574]
Which is way worse than I was hoping to get.
Here are three approaches to speeding this up. You gave no desired speed up or accuracies, or even vector sizes, so buyer beware.
TL;DR
Timings:
len 1 2 3 4
1000 0.045 0.033 0.025 0.022
10000 0.290 0.097 0.029 0.023
100000 3.429 0.767 0.083 0.030
1000000 0.546 0.046
1) Original Method
2) Pre-estimate with Subset
3) M Newville [linear log-log estimate](https://stackoverflow.com/a/44975066/7311767)
4) Subset Estimate (Use Less Data)
Pre-estimate with Subset (Method 2):
A decent speedup can be achieved by simply running the curve_fit twice, where the first time uses a short subset of the data to get a quick estimate. That estimate is then used to seed a curve_fit with the entire dataset.
x, y = current_data
stride = int(max(1, len(x) / 200))
c0 = curve_fit(power_law, x[0:len(x):stride], y[0:len(y):stride])[0]
return curve_fit(power_law, x, y, p0=c0)[0]
M Newville linear log-log estimate (Method 3):
Using the log estimate proposed by M Newville, is also considerably faster. As the OP was concerned about the initial estimate method proposed by Newville, this method uses curve_fit with a subset to provide the estimate of the break point in the curve.
x, y = current_data
stride = int(max(1, len(x) / 200))
c0 = curve_fit(power_law, x[0:len(x):stride], y[0:len(y):stride])[0]
index_max = np.where(x > c0[1])[0][0]
log_x = np.log(x[:index_max])
log_y = np.log(y[:index_max])
result = linregress(log_x, log_y)
return -result[0], np.exp(-result[1] / result[0])
return (m, c), result
Use Less Data (Method 4):
Finally the seed mechanism used for the previous two methods provides pretty good estimates on the sample data. Of course it is sample data so your mileage may vary.
stride = int(max(1, len(x) / 200))
c0 = curve_fit(power_law, x[0:len(x):stride], y[0:len(y):stride])[0]
Test Code:
import numpy as np
from numpy import random as rng
from scipy.optimize import curve_fit
from scipy.stats import linregress
fit_data = {}
current_data = None
def data_for_fit(a, b, n):
key = a, b, n
if key not in fit_data:
rng.seed(0)
x = np.logspace(-1, 1, n)
y = np.clip((x / b) ** (-a) + 0.01 * rng.randn(n), 0.001, None)
y[x > b] = 1
fit_data[key] = x, y
return fit_data[key]
def power_law(r, a, b):
""" Power law for the shape of the normalized conditional density """
import numpy as np
return np.piecewise(
r, [r < b, r >= b], [lambda x: (x/b)**-a, lambda x: 1])
def method1():
x, y = current_data
return curve_fit(power_law, x, y)[0]
def method2():
x, y = current_data
return curve_fit(power_law, x, y, p0=method4()[0])
def method3():
x, y = current_data
c0, pcov = method4()
index_max = np.where(x > c0[1])[0][0]
log_x = np.log(x[:index_max])
log_y = np.log(y[:index_max])
result = linregress(log_x, log_y)
m, c = -result[0], np.exp(-result[1] / result[0])
return (m, c), result
def method4():
x, y = current_data
stride = int(max(1, len(x) / 200))
return curve_fit(power_law, x[0:len(x):stride], y[0:len(y):stride])
from timeit import timeit
def runit(stmt):
print("%s: %.3f %s" % (
stmt, timeit(stmt + '()', number=10,
setup='from __main__ import ' + stmt),
eval(stmt + '()')[0]
))
def runit_size(size):
print('Length: %d' % size)
if size <= 100000:
runit('method1')
runit('method2')
runit('method3')
runit('method4')
for i in (1000, 10000, 100000, 1000000):
current_data = data_for_fit(3, 3, i)
runit_size(i)
Two suggestions:
Use numpy.where (and possibly argmin) to find the X value at which the Y data becomes 1, or perhaps just slightly larger than 1, and truncate the data to that point -- effectively ignoring the data where Y=1.
That might be something like:
index_max = numpy.where(y < 1.2)[0][0]
x = y[:index_max]
y = y[:index_max]
Use the hint shown in your log-log plot that the power law is now linear in log-log. You don't need curve_fit, but can use scipy.stats.linregress on log(Y) vs log(Y). For your real work, that will at the very least give good starting values for a subsequent fit.
Following up on this and trying to follow your question, you might try something like:
import numpy as np
from scipy.stats import linregress
np.random.seed(0)
npts = 51
x = np.logspace(-2, 2, npts)
YTHRESH = 1.02
for i in range(5):
b = i + 1.0 + np.random.normal(scale=0.1)
a = b + np.random.random()
y = (x/b)**(-a) + np.random.normal(scale=0.0030, size=npts)
y[x>b] = 1.0
# to model exponential decay, first remove the values
# where y ~= 1 where the data is known to not decay...
imax = np.where(y < YTHRESH)[0][0]
# take log of this truncated x and y
_x = np.log(x[:imax])
_y = np.log(y[:imax])
# use linear regression on the log-log data:
out = linregress(_x, _y)
# map slope/intercept to scale, exponent
afit = -out.slope
bfit = np.exp(out.intercept/afit)
print(""" === Fit Example {i:3d}
a expected {a:4f}, got {afit:4f}
b expected {b:4f}, got {bfit:4f}
""".format(i=i+1, a=a, b=b, afit=afit, bfit=bfit))
Hopefully that's enough to get you going.

Using 3 'for' loops to integrate functions and return a 3d array in python is too slow, how can I optimize it?

I have 3 arrays of data to integrate over and need to return a 3d array. This is my code
import numpy as np
from scipy import integrate
def f(x, y, z):
return x + y + 2*z
a = np.arange(64)
b = np.arange(100)
c = np.arange(100)
result = []
for x in a:
for y in b:
for z in c:
result.append(integrate.quad(f, 0, x, (y, z))[0])
result_1 = np.array(result).reshape(len(a), len(b), len(c))
print(result_1)
It works, but this code is so slow and I need to handle problems bigger than this. Is there any other method, something like broadcasting? And the following is the function I need to handle:
import numpy as np
from scipy import integrate, special
def f(v,r):
alpha = 0.5
chi = 1/2*special.erf((r+1.2)/0.3)-1/2*special.erf((r-1.2)/0.3)
return 4/(np.sqrt(2*np.pi*alpha))*chi*np.exp(-v**2/(2*alpha))
def E_1(r):
def f_1(v,r):
return r*f(v,r)
a = 0
b = r
g = lambda x: -np.inf
h = lambda x: np.inf
return integrate.dblquad(f_1, a, b, g, h)
def E_f(tau, xi_1, xi_2):
return E_1(xi_1*np.cos(tau) + xi_2*np.sin(tau))[0]*(-np.sin(tau))
I need to input tau, xi_1, xi_2 as three arrays, and return to a 3D array. And its arrange like coordinate, every coordinate point corresponds to a result. Just like the first example.

How to perform cubic spline interpolation in python?

I have two lists to describe the function y(x):
x = [0,1,2,3,4,5]
y = [12,14,22,39,58,77]
I would like to perform cubic spline interpolation so that given some value u in the domain of x, e.g.
u = 1.25
I can find y(u).
I found this in SciPy but I am not sure how to use it.
Short answer:
from scipy import interpolate
def f(x):
x_points = [ 0, 1, 2, 3, 4, 5]
y_points = [12,14,22,39,58,77]
tck = interpolate.splrep(x_points, y_points)
return interpolate.splev(x, tck)
print(f(1.25))
Long answer:
scipy separates the steps involved in spline interpolation into two operations, most likely for computational efficiency.
The coefficients describing the spline curve are computed,
using splrep(). splrep returns an array of tuples containing the
coefficients.
These coefficients are passed into splev() to actually
evaluate the spline at the desired point x (in this example 1.25).
x can also be an array. Calling f([1.0, 1.25, 1.5]) returns the
interpolated points at 1, 1.25, and 1,5, respectively.
This approach is admittedly inconvenient for single evaluations, but since the most common use case is to start with a handful of function evaluation points, then to repeatedly use the spline to find interpolated values, it is usually quite useful in practice.
In case, scipy is not installed:
import numpy as np
from math import sqrt
def cubic_interp1d(x0, x, y):
"""
Interpolate a 1-D function using cubic splines.
x0 : a float or an 1d-array
x : (N,) array_like
A 1-D array of real/complex values.
y : (N,) array_like
A 1-D array of real values. The length of y along the
interpolation axis must be equal to the length of x.
Implement a trick to generate at first step the cholesky matrice L of
the tridiagonal matrice A (thus L is a bidiagonal matrice that
can be solved in two distinct loops).
additional ref: www.math.uh.edu/~jingqiu/math4364/spline.pdf
"""
x = np.asfarray(x)
y = np.asfarray(y)
# remove non finite values
# indexes = np.isfinite(x)
# x = x[indexes]
# y = y[indexes]
# check if sorted
if np.any(np.diff(x) < 0):
indexes = np.argsort(x)
x = x[indexes]
y = y[indexes]
size = len(x)
xdiff = np.diff(x)
ydiff = np.diff(y)
# allocate buffer matrices
Li = np.empty(size)
Li_1 = np.empty(size-1)
z = np.empty(size)
# fill diagonals Li and Li-1 and solve [L][y] = [B]
Li[0] = sqrt(2*xdiff[0])
Li_1[0] = 0.0
B0 = 0.0 # natural boundary
z[0] = B0 / Li[0]
for i in range(1, size-1, 1):
Li_1[i] = xdiff[i-1] / Li[i-1]
Li[i] = sqrt(2*(xdiff[i-1]+xdiff[i]) - Li_1[i-1] * Li_1[i-1])
Bi = 6*(ydiff[i]/xdiff[i] - ydiff[i-1]/xdiff[i-1])
z[i] = (Bi - Li_1[i-1]*z[i-1])/Li[i]
i = size - 1
Li_1[i-1] = xdiff[-1] / Li[i-1]
Li[i] = sqrt(2*xdiff[-1] - Li_1[i-1] * Li_1[i-1])
Bi = 0.0 # natural boundary
z[i] = (Bi - Li_1[i-1]*z[i-1])/Li[i]
# solve [L.T][x] = [y]
i = size-1
z[i] = z[i] / Li[i]
for i in range(size-2, -1, -1):
z[i] = (z[i] - Li_1[i-1]*z[i+1])/Li[i]
# find index
index = x.searchsorted(x0)
np.clip(index, 1, size-1, index)
xi1, xi0 = x[index], x[index-1]
yi1, yi0 = y[index], y[index-1]
zi1, zi0 = z[index], z[index-1]
hi1 = xi1 - xi0
# calculate cubic
f0 = zi0/(6*hi1)*(xi1-x0)**3 + \
zi1/(6*hi1)*(x0-xi0)**3 + \
(yi1/hi1 - zi1*hi1/6)*(x0-xi0) + \
(yi0/hi1 - zi0*hi1/6)*(xi1-x0)
return f0
if __name__ == '__main__':
import matplotlib.pyplot as plt
x = np.linspace(0, 10, 11)
y = np.sin(x)
plt.scatter(x, y)
x_new = np.linspace(0, 10, 201)
plt.plot(x_new, cubic_interp1d(x_new, x, y))
plt.show()
If you have scipy version >= 0.18.0 installed you can use CubicSpline function from scipy.interpolate for cubic spline interpolation.
You can check scipy version by running following commands in python:
#!/usr/bin/env python3
import scipy
scipy.version.version
If your scipy version is >= 0.18.0 you can run following example code for cubic spline interpolation:
#!/usr/bin/env python3
import numpy as np
from scipy.interpolate import CubicSpline
# calculate 5 natural cubic spline polynomials for 6 points
# (x,y) = (0,12) (1,14) (2,22) (3,39) (4,58) (5,77)
x = np.array([0, 1, 2, 3, 4, 5])
y = np.array([12,14,22,39,58,77])
# calculate natural cubic spline polynomials
cs = CubicSpline(x,y,bc_type='natural')
# show values of interpolation function at x=1.25
print('S(1.25) = ', cs(1.25))
## Aditional - find polynomial coefficients for different x regions
# if you want to print polynomial coefficients in form
# S0(0<=x<=1) = a0 + b0(x-x0) + c0(x-x0)^2 + d0(x-x0)^3
# S1(1< x<=2) = a1 + b1(x-x1) + c1(x-x1)^2 + d1(x-x1)^3
# ...
# S4(4< x<=5) = a4 + b4(x-x4) + c5(x-x4)^2 + d5(x-x4)^3
# x0 = 0; x1 = 1; x4 = 4; (start of x region interval)
# show values of a0, b0, c0, d0, a1, b1, c1, d1 ...
cs.c
# Polynomial coefficients for 0 <= x <= 1
a0 = cs.c.item(3,0)
b0 = cs.c.item(2,0)
c0 = cs.c.item(1,0)
d0 = cs.c.item(0,0)
# Polynomial coefficients for 1 < x <= 2
a1 = cs.c.item(3,1)
b1 = cs.c.item(2,1)
c1 = cs.c.item(1,1)
d1 = cs.c.item(0,1)
# ...
# Polynomial coefficients for 4 < x <= 5
a4 = cs.c.item(3,4)
b4 = cs.c.item(2,4)
c4 = cs.c.item(1,4)
d4 = cs.c.item(0,4)
# Print polynomial equations for different x regions
print('S0(0<=x<=1) = ', a0, ' + ', b0, '(x-0) + ', c0, '(x-0)^2 + ', d0, '(x-0)^3')
print('S1(1< x<=2) = ', a1, ' + ', b1, '(x-1) + ', c1, '(x-1)^2 + ', d1, '(x-1)^3')
print('...')
print('S5(4< x<=5) = ', a4, ' + ', b4, '(x-4) + ', c4, '(x-4)^2 + ', d4, '(x-4)^3')
# So we can calculate S(1.25) by using equation S1(1< x<=2)
print('S(1.25) = ', a1 + b1*0.25 + c1*(0.25**2) + d1*(0.25**3))
# Cubic spline interpolation calculus example
# https://www.youtube.com/watch?v=gT7F3TWihvk
Just putting this here if you want a dependency-free solution.
Code taken from an answer above: https://stackoverflow.com/a/48085583/36061
def my_cubic_interp1d(x0, x, y):
"""
Interpolate a 1-D function using cubic splines.
x0 : a 1d-array of floats to interpolate at
x : a 1-D array of floats sorted in increasing order
y : A 1-D array of floats. The length of y along the
interpolation axis must be equal to the length of x.
Implement a trick to generate at first step the cholesky matrice L of
the tridiagonal matrice A (thus L is a bidiagonal matrice that
can be solved in two distinct loops).
additional ref: www.math.uh.edu/~jingqiu/math4364/spline.pdf
# original function code at: https://stackoverflow.com/a/48085583/36061
This function is licenced under: Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0)
https://creativecommons.org/licenses/by-sa/3.0/
Original Author raphael valentin
Date 3 Jan 2018
Modifications made to remove numpy dependencies:
-all sub-functions by MR
This function, and all sub-functions, are licenced under: Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0)
Mod author: Matthew Rowles
Date 3 May 2021
"""
def diff(lst):
"""
numpy.diff with default settings
"""
size = len(lst)-1
r = [0]*size
for i in range(size):
r[i] = lst[i+1] - lst[i]
return r
def list_searchsorted(listToInsert, insertInto):
"""
numpy.searchsorted with default settings
"""
def float_searchsorted(floatToInsert, insertInto):
for i in range(len(insertInto)):
if floatToInsert <= insertInto[i]:
return i
return len(insertInto)
return [float_searchsorted(i, insertInto) for i in listToInsert]
def clip(lst, min_val, max_val, inPlace = False):
"""
numpy.clip
"""
if not inPlace:
lst = lst[:]
for i in range(len(lst)):
if lst[i] < min_val:
lst[i] = min_val
elif lst[i] > max_val:
lst[i] = max_val
return lst
def subtract(a,b):
"""
returns a - b
"""
return a - b
size = len(x)
xdiff = diff(x)
ydiff = diff(y)
# allocate buffer matrices
Li = [0]*size
Li_1 = [0]*(size-1)
z = [0]*(size)
# fill diagonals Li and Li-1 and solve [L][y] = [B]
Li[0] = sqrt(2*xdiff[0])
Li_1[0] = 0.0
B0 = 0.0 # natural boundary
z[0] = B0 / Li[0]
for i in range(1, size-1, 1):
Li_1[i] = xdiff[i-1] / Li[i-1]
Li[i] = sqrt(2*(xdiff[i-1]+xdiff[i]) - Li_1[i-1] * Li_1[i-1])
Bi = 6*(ydiff[i]/xdiff[i] - ydiff[i-1]/xdiff[i-1])
z[i] = (Bi - Li_1[i-1]*z[i-1])/Li[i]
i = size - 1
Li_1[i-1] = xdiff[-1] / Li[i-1]
Li[i] = sqrt(2*xdiff[-1] - Li_1[i-1] * Li_1[i-1])
Bi = 0.0 # natural boundary
z[i] = (Bi - Li_1[i-1]*z[i-1])/Li[i]
# solve [L.T][x] = [y]
i = size-1
z[i] = z[i] / Li[i]
for i in range(size-2, -1, -1):
z[i] = (z[i] - Li_1[i-1]*z[i+1])/Li[i]
# find index
index = list_searchsorted(x0,x)
index = clip(index, 1, size-1)
xi1 = [x[num] for num in index]
xi0 = [x[num-1] for num in index]
yi1 = [y[num] for num in index]
yi0 = [y[num-1] for num in index]
zi1 = [z[num] for num in index]
zi0 = [z[num-1] for num in index]
hi1 = list( map(subtract, xi1, xi0) )
# calculate cubic - all element-wise multiplication
f0 = [0]*len(hi1)
for j in range(len(f0)):
f0[j] = zi0[j]/(6*hi1[j])*(xi1[j]-x0[j])**3 + \
zi1[j]/(6*hi1[j])*(x0[j]-xi0[j])**3 + \
(yi1[j]/hi1[j] - zi1[j]*hi1[j]/6)*(x0[j]-xi0[j]) + \
(yi0[j]/hi1[j] - zi0[j]*hi1[j]/6)*(xi1[j]-x0[j])
return f0
Minimal python3 code:
from scipy import interpolate
if __name__ == '__main__':
x = [ 0, 1, 2, 3, 4, 5]
y = [12,14,22,39,58,77]
# tck : tuple (t,c,k) a tuple containing the vector of knots,
# the B-spline coefficients, and the degree of the spline.
tck = interpolate.splrep(x, y)
print(interpolate.splev(1.25, tck)) # Prints 15.203125000000002
print(interpolate.splev(...other_value_here..., tck))
Based on comment of cwhy and answer by youngmit
In my previous post, I wrote a code based on a Cholesky development to solve the matrix generated by the cubic algorithm. Unfortunately, due to the square root function, it may perform badly on some sets of points (typically a non-uniform set of points).
In the same spirit than previously, there is another idea using the Thomas algorithm (TDMA) (see https://en.wikipedia.org/wiki/Tridiagonal_matrix_algorithm) to solve partially the tridiagonal matrix during its definition loop. However, the condition to use TDMA is that it requires at least that the matrix shall be diagonally dominant. However, in our case, it shall be true since |bi| > |ai| + |ci| with ai = h[i], bi = 2*(h[i]+h[i+1]), ci = h[i+1], with h[i] unconditionally positive. (see https://www.cfd-online.com/Wiki/Tridiagonal_matrix_algorithm_-TDMA(Thomas_algorithm)
I refer again to the document from jingqiu (see my previous post, unfortunately the link is broken, but it is still possible to find it in the cache of the web).
An optimized version of the TDMA solver can be described as follows:
def TDMAsolver(a,b,c,d):
""" This function is licenced under: Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0)
https://creativecommons.org/licenses/by-sa/3.0/
Author raphael valentin
Date 25 Mar 2022
ref. https://www.cfd-online.com/Wiki/Tridiagonal_matrix_algorithm_-_TDMA_(Thomas_algorithm)
"""
n = len(d)
w = np.empty(n-1,float)
g = np.empty(n, float)
w[0] = c[0]/b[0]
g[0] = d[0]/b[0]
for i in range(1, n-1):
m = b[i] - a[i-1]*w[i-1]
w[i] = c[i] / m
g[i] = (d[i] - a[i-1]*g[i-1]) / m
g[n-1] = (d[n-1] - a[n-2]*g[n-2]) / (b[n-1] - a[n-2]*w[n-2])
for i in range(n-2, -1, -1):
g[i] = g[i] - w[i]*g[i+1]
return g
When it is possible to get each individual for ai, bi, ci, di, it becomes easy to combine the definitions of the natural cubic spline interpolator function within these 2 single loops.
def cubic_interpolate(x0, x, y):
""" Natural cubic spline interpolate function
This function is licenced under: Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0)
https://creativecommons.org/licenses/by-sa/3.0/
Author raphael valentin
Date 25 Mar 2022
"""
xdiff = np.diff(x)
dydx = np.diff(y)
dydx /= xdiff
n = size = len(x)
w = np.empty(n-1, float)
z = np.empty(n, float)
w[0] = 0.
z[0] = 0.
for i in range(1, n-1):
m = xdiff[i-1] * (2 - w[i-1]) + 2 * xdiff[i]
w[i] = xdiff[i] / m
z[i] = (6*(dydx[i] - dydx[i-1]) - xdiff[i-1]*z[i-1]) / m
z[-1] = 0.
for i in range(n-2, -1, -1):
z[i] = z[i] - w[i]*z[i+1]
# find index (it requires x0 is already sorted)
index = x.searchsorted(x0)
np.clip(index, 1, size-1, index)
xi1, xi0 = x[index], x[index-1]
yi1, yi0 = y[index], y[index-1]
zi1, zi0 = z[index], z[index-1]
hi1 = xi1 - xi0
# calculate cubic
f0 = zi0/(6*hi1)*(xi1-x0)**3 + \
zi1/(6*hi1)*(x0-xi0)**3 + \
(yi1/hi1 - zi1*hi1/6)*(x0-xi0) + \
(yi0/hi1 - zi0*hi1/6)*(xi1-x0)
return f0
This function gives the same results as the function/class CubicSpline from scipy.interpolate, as we can see in the next plot.
It is possible to implement as well the first and second analytical derivatives that can be described such way:
f1p = -zi0/(2*hi1)*(xi1-x0)**2 + zi1/(2*hi1)*(x0-xi0)**2 + (yi1/hi1 - zi1*hi1/6) + (yi0/hi1 - zi0*hi1/6)
f2p = zi0/hi1 * (xi1-x0) + zi1/hi1 * (x0-xi0)
Then, it is easy to verify that f2p[0] and f2p[-1] are equal to 0, then that the interpolator function yields natural splines.
An additional reference concerning natural spline:
https://faculty.ksu.edu.sa/sites/default/files/numerical_analysis_9th.pdf#page=167
An example of use:
import matplotlib.pyplot as plt
import numpy as np
x = [-8,-4.19,-3.54,-3.31,-2.56,-2.31,-1.66,-0.96,-0.22,0.62,1.21,3]
y = [-0.01,0.01,0.03,0.04,0.07,0.09,0.16,0.28,0.45,0.65,0.77,1]
x = np.asfarray(x)
y = np.asfarray(y)
plt.scatter(x, y)
x_new= np.linspace(min(x), max(x), 10000)
y_new = cubic_interpolate(x_new, x, y)
plt.plot(x_new, y_new)
from scipy.interpolate import CubicSpline
f = CubicSpline(x, y, bc_type='natural')
plt.plot(x_new, f(x_new), label='ref')
plt.legend()
plt.show()
In a conclusion, this updated algorithm shall perform interpolation with better stability and faster than the previous code (O(n)). Associated with numba or cython, it shall be even very fast. Finally, it is totally independent of Scipy.
Important, note that as most of algorithms, it is sometimes useful to normalize the data (e.g. against large or small number values) to get the best results. As well, in this code, I do not check nan values or ordered data.
Whatever, this update was a good lesson learning for me and I hope it can help someone. Let me know if you find something strange.
If you want to get the value
from scipy.interpolate import CubicSpline
import numpy as np
x = [-5,-4.19,-3.54,-3.31,-2.56,-2.31,-1.66,-0.96,-0.22,0.62,1.21,3]
y = [-0.01,0.01,0.03,0.04,0.07,0.09,0.16,0.28,0.45,0.65,0.77,1]
value = 2
#ascending order
if np.any(np.diff(x) < 0):
indexes = np.argsort(x).astype(int)
x = np.array(x)[indexes]
y = np.array(y)[indexes]
f = CubicSpline(x, y, bc_type='natural')
specificVal = f(value).item(0) #f(value) is numpy.ndarray!!
print(specificVal)
If you want to plot the interpolated function.
np.linspace third parameter increase the "accuracy".
from scipy.interpolate import CubicSpline
import numpy as np
import matplotlib.pyplot as plt
x = [-5,-4.19,-3.54,-3.31,-2.56,-2.31,-1.66,-0.96,-0.22,0.62,1.21,3]
y = [-0.01,0.01,0.03,0.04,0.07,0.09,0.16,0.28,0.45,0.65,0.77,1]
#ascending order
if np.any(np.diff(x) < 0):
indexes = np.argsort(x).astype(int)
x = np.array(x)[indexes]
y = np.array(y)[indexes]
f = CubicSpline(x, y, bc_type='natural')
x_new = np.linspace(min(x), max(x), 100)
y_new = f(x_new)
plt.plot(x_new, y_new)
plt.scatter(x, y)
plt.title('Cubic Spline Interpolation')
plt.show()
output:
Yes, as others have already noted, it should be as simple as
>>> from scipy.interpolate import CubicSpline
>>> CubicSpline(x,y)(u)
array(15.203125)
(you can, for example, convert it to float to get the value from a 0d NumPy array)
What has not been described yet is boundary conditions: the default ‘not-a-knot’ boundary conditions work best if you have zero knowledge about the data you’re going to interpolate.
If you see the following ‘features’ on the plot, you can fine-tune the boundary conditions to get a better result:
the first derivative vanishes at boundaries => bc_type=‘clamped’
the second derivative vanishes at boundaries => bc_type='natural'
the function is periodic => bc_type='periodic'
See my article for more details and an interactive demo.

Categories

Resources