def calculate_root(f: Polynomial, a, b, eps):
"""
Return root (assuming there's one) of f function
on the (a, b) interval using secant method
and also return number of iterations.
f's first two derivatives should be differentiable.
"""
assert f(a) * f(b) < 0
if f(a) > 0:
f = -f
if f(b) * f.deriv(2)(b) > 0:
x = a
c = b
elif f(a) * f.deriv(2)(a) > 0:
x = b
c = a
else:
raise Exception()
true_x = spo.brentq(f, a, b)
iter_count = 0
while abs(x - true_x) > eps and iter_count < MAX_ITERATION_COUNT:
x -= (c - x) / (f(c) - f(x)) * f(x)
iter_count += 1
return x, iter_count
def lagrange(nodes):
result = Polynomial([0])
w = Polynomial(np.poly(nodes)[::-1])
deriv = w.deriv(1)
for i in range(len(nodes)):
result = pl.polyadd(result,
make_l_k(i, nodes, w, deriv) * f(nodes[i]))[0]
return result
def lagrange(nodes):
result = Polynomial([0])
# np.poly строит полином по набору корней. Но порядок коэффициетов противоположный тому, который
# принимает конструктор класса Polynomial
w = Polynomial(np.poly(nodes)[::-1])
deriv = w.deriv(1)
for i in range(len(nodes)):
result = pl.polyadd(result,
make_l_k(i, nodes, w, deriv) * f(nodes[i]))[0]
# возвращается не только сумма
return result
def calculate_root(f: Polynomial, a, b, eps):
"""
Return root approximation calculated with Newton's method as well
as number of iterations made to calculate it.
Root of polynomial f is approximated in [a, b] interval until
error is smaller than epsilon eps.
"""
assert f(a) * f(b) < 0
df = f.deriv()
def newtons_lambda(x):
return -1 / df(x)
return sim.calculate_root(f, newtons_lambda, a, b, eps)
def make_a_k(nodes, p, a, b):
w = Polynomial(np.poly(nodes)[::-1])
result_coeffs = []
x = Symbol("x")
for k in range(len(nodes)):
poly_part = np.polydiv(w.coef, np.array([-nodes[k], 1]))[0] / w.deriv()(nodes[k])
# Вручную домножаем каждый член многоочлена на весовую функцию, чтобы потом взять интеграл
under_integral_list = []
for i, coef in enumerate(poly_part):
under_integral_list.append(coef * x**i * p(x))
under_integral = sum(under_integral_list)
# Берём интеграл на промежутке [a,b]
result_coeffs.append(integrate(under_integral, (x, a, b)))
return result_coeffs
def sturms_sequence(p: Polynomial):
"""
Returns Sturm's sequence of a given polynomial.
"""
# setting up the Sturm's sequence.
sturms_seq = []
sturms_seq.append(p)
sturms_seq.append(p.deriv())
# filling the Sturm's sequence list.
f = -Polynomial(poly.polydiv(sturms_seq[-2].coef, sturms_seq[-1].coef)[1])
while f.degree() != 0 or f.coef[0] != 0:
sturms_seq.append(f.copy())
f = -Polynomial(
poly.polydiv(sturms_seq[-2].coef, sturms_seq[-1].coef)[1])
return sturms_seq
class MPolynomial:
def __init__(self):
self.roots = [complex(-0.5, 0), complex(0.5, 0)]
self.update()
def update(self):
if len(self.roots)>0:
self.poly = Polynomial.fromroots(self.roots)
else:
self.poly = Polynomial([0])
def add_root(self, root):
self.roots.append(root)
self.update()
def remove_root(self, root_idx):
assert root_idx >= 0 and root_idx < len(self.roots), "remove_root invalid index"
del self.roots[root_idx]
self.update()
def get_roots(self):
return self.roots
def get_n_roots(self):
return len(self.roots)
def poly_as_str(self):
prev = "(float2)(0.0f,0.0f)"
for i,c in enumerate(self.poly.coef):
formula = "plus(mult((float2)({}f,{}f), complexPower(z,(float2)({}.0f,0.0f))),{})".format(c.real, c.imag, i,
prev)
prev = formula
return prev
def deriv_poly_as_str(self):
deriv = self.poly.deriv()
prev = "(float2)(0.0f,0.0f)"
for i, c in enumerate(deriv.coef):
formula = "plus(mult((float2)({}f,{}f), complexPower(z,(float2)({}.0f,0.0f))),{})".format(c.real, c.imag, i, prev)
prev = formula
return prev
]))
coef.reverse()
w = Polynomial(np.array(coef, dtype=np.float))
# have to explicitly specify dtype,
# otherwise root finder doesn't work
print([w(x) for x in range(1, 21)])
print(w.roots())
# This uses eigenvalues of the companion matrix for roots
from scipy.optimize import root
print(root(w, 21.0))
# This uses Optimization method root finding
for delta in (1e-8, 1e-6, 1e-4, 1e-2):
coef[20] = 1 + delta
w = Polynomial(np.array(coef, dtype=np.float))
print(root(w, 21.0))
coef[20] = 1
coef[19] = -210 - 2**(-23)
w = Polynomial(np.array(coef, dtype=np.float))
print(root(w, 16.1))
print(root(w, 17.1))
print(w.roots())
coef[19] = -210
w = Polynomial(np.array(coef, dtype=np.float))
wp = w.deriv()
[
print("%g" %
(sum([abs(coef[l]) * k**(l - 1)
for l in range(len(coef))]) / abs(wp(k))))
for k in (14, 16, 17, 20)
]
def measure_under_midline(
fl: xr.DataArray,
mid: Polynomial,
n_points: int = 100,
thickness: float = 0.0,
order=1,
norm_scale=1,
flatten=True,
) -> np.ndarray:
"""
Measure the intensity profile of the given image under the given midline at the given x-coordinates.
Parameters
----------
flatten
norm_scale
order
the interpolation order
fl
The fluorescence image to measure
mid
The midline under which to measure
n_points
The number of points to measure under
thickness
The thickness of the line to measure under.
Notes
-----
Using thickness is slower, depending on the amount of thickness
On my machine (2GHz Intel Core i5), as of 12/4/19:
0-thickness:
492 µs ± 16.6 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
2-thickness:
1.99 ms ± 65.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
10-thickness:
3.89 ms ± 92.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
Returns
-------
zs: np.ndarray
The intensity profile of the image measured under the midline at the given
x-coordinates.
"""
# Make sure the image orientation matches with the expected order of map_coordinates
try:
if thickness == 0:
xs, ys = mid.linspace(n=n_points)
fl = np.asarray(fl)
return ndi.map_coordinates(fl, np.stack([xs, ys]), order=order)
else:
# Gets a bit wonky, but makes sense
# We need to get the normal lines from each point in the midline
# then measure under those lines.
# First, get the coordinates of the midline
xs, ys = mid.linspace(n=n_points)
# Now, we get the angles of each normal vector
der = mid.deriv()
normal_slopes = -1 / der(xs)
normal_thetas = np.arctan(normal_slopes)
# We get the x and y components of the start/end of the normal vectors
mag = thickness / 2
x0 = np.cos(normal_thetas) * mag
y0 = np.sin(normal_thetas) * mag
x1 = np.cos(normal_thetas) * -mag
y1 = np.sin(normal_thetas) * -mag
# These are the actual coordinates of the starts/ends of the normal vectors as they move
# from (x,y) coordinates in the midline
xs0 = xs + x0
xs1 = xs + x1
ys0 = ys + y0
ys1 = ys + y1
# We need to measure in a consistent direction along the normal line
# if y0 < y1, we're going to be measuring in an opposite direction along the line... so we need flip the coordinates
for y0, y1, x0, x1, i in zip(ys0, ys1, xs0, xs1, range(len(xs0))):
if y0 < y1:
tx = xs0[i]
xs0[i] = xs1[i]
xs1[i] = tx
ty = ys0[i]
ys0[i] = ys1[i]
ys1[i] = ty
n_line_pts = thickness
all_xs = np.linspace(xs0, xs1, n_line_pts)
all_ys = np.linspace(ys0, ys1, n_line_pts)
straightened = ndi.map_coordinates(fl, [all_xs, all_ys], order=order)
if flatten:
# Create a normal distribution centered around 0 with the given scale (see scipy.norm.pdf)
# the distribution is then tiled to be the same shape as the straightened pharynx
# then, this resultant matrix is the weights for averaging
w = np.tile(
norm.pdf(np.linspace(-1, 1, n_line_pts), scale=norm_scale),
(n_points, 1),
).T
profile = np.average(straightened, axis=0, weights=w)
return profile
else:
return straightened
except AttributeError:
# This happens if the image is TL. Then it will have `None` instead of
# a midline object
pass
except Exception as e:
# Here, something actually went wrong
logging.warning(f"measuring under midline failed with error {e}")
return np.zeros((1, n_points))
