source: sasmodels/explore/realspace.py @ d5014e4

Last change on this file since d5014e4 was a1c32c2, checked in by Paul Kienzle <pkienzle@…>, 7 years ago

Use gauss-legendre integration for cross-checking against P(r)

  • Property mode set to 100644
File size: 15.8 KB
RevLine 
[cfa28d3]1from __future__ import division, print_function
2
3import time
4from copy import copy
5
6import numpy as np
7from numpy import pi, radians, sin, cos, sqrt
8from numpy.random import poisson, uniform
[a1c32c2]9from numpy.polynomial.legendre import leggauss
[cfa28d3]10from scipy.integrate import simps
11from scipy.special import j1 as J1
12
13# Definition of rotation matrices comes from wikipedia:
14#    https://en.wikipedia.org/wiki/Rotation_matrix#Basic_rotations
15def Rx(angle):
16    """Construct a matrix to rotate points about *x* by *angle* degrees."""
17    a = radians(angle)
18    R = [[1, 0, 0],
19         [0, +cos(a), -sin(a)],
20         [0, +sin(a), +cos(a)]]
21    return np.matrix(R)
22
23def Ry(angle):
24    """Construct a matrix to rotate points about *y* by *angle* degrees."""
25    a = radians(angle)
26    R = [[+cos(a), 0, +sin(a)],
27         [0, 1, 0],
28         [-sin(a), 0, +cos(a)]]
29    return np.matrix(R)
30
31def Rz(angle):
32    """Construct a matrix to rotate points about *z* by *angle* degrees."""
33    a = radians(angle)
34    R = [[+cos(a), -sin(a), 0],
35         [+sin(a), +cos(a), 0],
36         [0, 0, 1]]
37    return np.matrix(R)
38
39def rotation(theta, phi, psi):
40    """
41    Apply the jitter transform to a set of points.
42
43    Points are stored in a 3 x n numpy matrix, not a numpy array or tuple.
44    """
45    return Rx(phi)*Ry(theta)*Rz(psi)
46
47class Shape:
48    rotation = np.matrix([[1., 0, 0], [0, 1, 0], [0, 0, 1]])
49    center = np.array([0., 0., 0.])[:, None]
50    r_max = None
51
52    def volume(self):
53        # type: () -> float
54        raise NotImplementedError()
55
56    def sample(self, density):
57        # type: (float) -> np.ndarray[N], np.ndarray[N, 3]
58        raise NotImplementedError()
59
60    def rotate(self, theta, phi, psi):
61        self.rotation = rotation(theta, phi, psi) * self.rotation
62        return self
63
64    def shift(self, x, y, z):
65        self.center = self.center + np.array([x, y, z])[:, None]
66        return self
67
68    def _adjust(self, points):
69        points = np.asarray(self.rotation * np.matrix(points.T)) + self.center
70        return points.T
71
72    def r_bins(self, q, over_sampling=1, r_step=0.):
73        r_max = min(2 * pi / q[0], self.r_max)
74        if r_step == 0.:
75            r_step = 2 * pi / q[-1] / over_sampling
76        #r_step = 0.01
77        return np.arange(r_step, r_max, r_step)
78
79class Composite(Shape):
80    def __init__(self, shapes, center=(0, 0, 0), orientation=(0, 0, 0)):
81        self.shapes = shapes
82        self.rotate(*orientation)
83        self.shift(*center)
84
85        # Find the worst case distance between any two points amongst a set
86        # of shapes independent of orientation.  This could easily be a
87        # factor of two worse than necessary, e.g., a pair of thin rods
88        # end-to-end vs the same pair side-by-side.
89        distances = [((s1.r_max + s2.r_max)/2
90                      + sqrt(np.sum((s1.center - s2.center)**2)))
91                     for s1 in shapes
92                     for s2 in shapes]
93        self.r_max = max(distances + [s.r_max for s in shapes])
94
95    def volume(self):
96        return sum(shape.volume() for shape in self.shapes)
97
98    def sample(self, density):
99        values, points = zip(*(shape.sample(density) for shape in self.shapes))
100        return np.hstack(values), self._adjust(np.vstack(points))
101
102class Box(Shape):
103    def __init__(self, a, b, c,
104                 value, center=(0, 0, 0), orientation=(0, 0, 0)):
105        self.value = np.asarray(value)
106        self.rotate(*orientation)
107        self.shift(*center)
108        self.a, self.b, self.c = a, b, c
109        self._scale = np.array([a/2, b/2, c/2])[None, :]
110        self.r_max = sqrt(a**2 + b**2 + c**2)
111
112    def volume(self):
113        return self.a*self.b*self.c
114
115    def sample(self, density):
116        num_points = poisson(density*self.a*self.b*self.c)
117        points = self._scale*uniform(-1, 1, size=(num_points, 3))
118        values = self.value.repeat(points.shape[0])
119        return values, self._adjust(points)
120
121class EllipticalCylinder(Shape):
122    def __init__(self, ra, rb, length,
123                 value, center=(0, 0, 0), orientation=(0, 0, 0)):
124        self.value = np.asarray(value)
125        self.rotate(*orientation)
126        self.shift(*center)
127        self.ra, self.rb, self.length = ra, rb, length
128        self._scale = np.array([ra, rb, length/2])[None, :]
129        self.r_max = sqrt(4*max(ra, rb)**2 + length**2)
130
131    def volume(self):
132        return pi*self.ra*self.rb*self.length
133
134    def sample(self, density):
135        # density of the bounding box
136        num_points = poisson(density*4*self.ra*self.rb*self.length)
137        points = uniform(-1, 1, size=(num_points, 3))
138        radius = points[:, 0]**2 + points[:, 1]**2
139        points = self._scale*points[radius <= 1]
140        values = self.value.repeat(points.shape[0])
141        return values, self._adjust(points)
142
143class TriaxialEllipsoid(Shape):
144    def __init__(self, ra, rb, rc,
145                 value, center=(0, 0, 0), orientation=(0, 0, 0)):
146        self.value = np.asarray(value)
147        self.rotate(*orientation)
148        self.shift(*center)
149        self.ra, self.rb, self.rc = ra, rb, rc
150        self._scale = np.array([ra, rb, rc])[None, :]
151        self.r_max = 2*max(ra, rb, rc)
152
153    def volume(self):
154        return 4*pi/3 * self.ra * self.rb * self.rc
155
156    def sample(self, density):
157        # randomly sample from a box of side length 2*r, excluding anything
158        # not in the ellipsoid
159        num_points = poisson(density*8*self.ra*self.rb*self.rc)
160        points = uniform(-1, 1, size=(num_points, 3))
161        radius = np.sum(points**2, axis=1)
162        points = self._scale*points[radius <= 1]
163        values = self.value.repeat(points.shape[0])
164        return values, self._adjust(points)
165
166class Helix(Shape):
167    def __init__(self, helix_radius, helix_pitch, tube_radius, tube_length,
168                 value, center=(0, 0, 0), orientation=(0, 0, 0)):
169        self.value = np.asarray(value)
170        self.rotate(*orientation)
171        self.shift(*center)
172        self.helix_radius, self.helix_pitch = helix_radius, helix_pitch
173        self.tube_radius, self.tube_length = tube_radius, tube_length
174        helix_length = helix_pitch * tube_length/sqrt(helix_radius**2 + helix_pitch**2)
175        self.r_max = sqrt((2*helix_radius + 2*tube_radius)*2
176                          + (helix_length + 2*tube_radius)**2)
177
178    def volume(self):
179        # small tube radius approximation; for larger tubes need to account
180        # for the fact that the inner length is much shorter than the outer
181        # length
182        return pi*self.tube_radius**2*self.tube_length
183
184    def points(self, density):
185        num_points = poisson(density*4*self.tube_radius**2*self.tube_length)
186        points = uniform(-1, 1, size=(num_points, 3))
187        radius = points[:, 0]**2 + points[:, 1]**2
188        points = points[radius <= 1]
189
190        # Based on math stackexchange answer by Jyrki Lahtonen
191        #     https://math.stackexchange.com/a/461637
192        # with helix along z rather than x [so tuples in answer are (z, x, y)]
193        # and with random points in the cross section (p1, p2) rather than
194        # uniform points on the surface (cos u, sin u).
195        a, R = self.tube_radius, self.helix_radius
196        h = self.helix_pitch
197        scale = 1/sqrt(R**2 + h**2)
198        t = points[:, 3] * (self.tube_length * scale/2)
199        cos_t, sin_t = cos(t), sin(t)
200
201        # rx = R*cos_t
202        # ry = R*sin_t
203        # rz = h*t
204        # nx = -a * cos_t * points[:, 1]
205        # ny = -a * sin_t * points[:, 1]
206        # nz = 0
207        # bx = (a * h/scale) * sin_t * points[:, 2]
208        # by = (-a * h/scale) * cos_t * points[:, 2]
209        # bz = a*R/scale
210        # x = rx + nx + bx
211        # y = ry + ny + by
212        # z = rz + nz + bz
213        u, v = (R - a*points[:, 1]), (a * h/scale)*points[:, 2]
214        x = u * cos_t + v * sin_t
215        y = u * sin_t - v * cos_t
216        z = a*R/scale + h * t
217
218        points = np.hstack((x, y, z))
219        values = self.value.repeat(points.shape[0])
220        return values, self._adjust(points)
221
222def _calc_Pr_nonuniform(r, rho, points):
223    # Make Pr a little be bigger than necessary so that only distances
224    # min < d < max end up in Pr
225    n_max = len(r)+1
226    extended_Pr = np.zeros(n_max+1, 'd')
227    # r refers to bin centers; find corresponding bin edges
228    bins = bin_edges(r)
229    t_next = time.clock() + 3
230    for k, rho_k in enumerate(rho[:-1]):
231        distance = sqrt(np.sum((points[k] - points[k+1:])**2, axis=1))
232        weights = rho_k * rho[k+1:]
233        index = np.searchsorted(bins, distance)
234        # Note: indices may be duplicated, so "Pr[index] += w" will not work!!
235        extended_Pr += np.bincount(index, weights, n_max+1)
236        t = time.clock()
237        if t > t_next:
238            t_next = t + 3
239            print("processing %d of %d"%(k, len(rho)-1))
240    Pr = extended_Pr[1:-1]
241    return Pr / Pr.max()
242
243def calc_Pr(r, rho, points):
244    # P(r) with uniform steps in r is 3x faster; check if we are uniform
245    # before continuing
246    if np.max(np.abs(np.diff(r) - r[0])) > r[0]*0.01:
247        return _calc_Pr_nonuniform(r, rho, points)
248
249    # Make Pr a little be bigger than necessary so that only distances
250    # min < d < max end up in Pr
251    n_max = len(r)
252    extended_Pr = np.zeros(n_max+1, 'd')
253    t0 = time.clock()
254    t_next = t0 + 3
255    row_zero = np.zeros(len(rho), 'i')
256    for k, rho_k in enumerate(rho[:-1]):
257        distances = sqrt(np.sum((points[k] - points[k+1:])**2, axis=1))
258        weights = rho_k * rho[k+1:]
259        index = np.minimum(np.asarray(distances/r[0], 'i'), n_max)
260        # Note: indices may be duplicated, so "Pr[index] += w" will not work!!
261        extended_Pr += np.bincount(index, weights, n_max+1)
262        t = time.clock()
263        if t > t_next:
264            t_next = t + 3
265            print("processing %d of %d"%(k, len(rho)-1))
266    #print("time py:", time.clock() - t0)
267    Pr = extended_Pr[:-1]
268    # Make Pr independent of sampling density.  The factor of 2 comes because
269    # we are only accumulating the upper triangular distances.
270    #Pr = Pr * 2 / len(rho)**2
271    return Pr / Pr.max()
272
273    # Can get an additional 2x by going to C.  Cuda/OpenCL will allow even
274    # more speedup, though still bounded by the n^2 cose.
275    """
276void pdfcalc(int n, const double *pts, const double *rho,
277             int nPr, double *Pr, double rstep)
278{
279  int i,j;
280
281  for (i=0; i<n-2; i++) {
282    for (j=i+1; j<=n-1; j++) {
283      const double dxx=pts[3*i]-pts[3*j];
284      const double dyy=pts[3*i+1]-pts[3*j+1];
285      const double dzz=pts[3*i+2]-pts[3*j+2];
286      const double d=sqrt(dxx*dxx+dyy*dyy+dzz*dzz);
287      const int k=rint(d/rstep);
288      if (k < nPr) Pr[k]+=rho[i]*rho[j];
289    }
290  }
291}
292"""
293
294def j0(x):
295    return np.sinc(x/np.pi)
296
297def calc_Iq(q, r, Pr):
298    Iq = np.array([simps(Pr * j0(qk*r), r) for qk in q])
299    #Iq = np.array([np.trapz(Pr * j0(qk*r), r) for qk in q])
300    Iq /= Iq[0]
301    return Iq
302
303# NOTE: copied from sasmodels/resolution.py
304def bin_edges(x):
305    """
306    Determine bin edges from bin centers, assuming that edges are centered
307    between the bins.
308
309    Note: this uses the arithmetic mean, which may not be appropriate for
310    log-scaled data.
311    """
312    if len(x) < 2 or (np.diff(x) < 0).any():
313        raise ValueError("Expected bins to be an increasing set")
314    edges = np.hstack([
315        x[0]  - 0.5*(x[1]  - x[0]),  # first point minus half first interval
316        0.5*(x[1:] + x[:-1]),        # mid points of all central intervals
317        x[-1] + 0.5*(x[-1] - x[-2]), # last point plus half last interval
318        ])
319    return edges
320
321def plot_calc(r, Pr, q, Iq, theory=None):
322    import matplotlib.pyplot as plt
323    plt.subplot(211)
324    plt.plot(r, Pr, '-', label="Pr")
325    plt.xlabel('r (A)')
326    plt.ylabel('Pr (1/A^2)')
327    plt.subplot(212)
328    plt.loglog(q, Iq, '-', label='from Pr')
329    plt.xlabel('q (1/A')
330    plt.ylabel('Iq')
331    if theory is not None:
332        plt.loglog(theory[0], theory[1], '-', label='analytic')
333        plt.legend()
334
335def plot_points(rho, points):
336    import mpl_toolkits.mplot3d
337    import matplotlib.pyplot as plt
338
339    ax = plt.axes(projection='3d')
340    try:
341        ax.axis('square')
342    except Exception:
343        pass
344    n = len(points)
345    index = np.random.choice(n, size=1000) if n > 1000 else slice(None, None)
346    ax.scatter(points[index, 0], points[index, 1], points[index, 2], c=rho[index])
347    #low, high = points.min(axis=0), points.max(axis=0)
348    #ax.axis([low[0], high[0], low[1], high[1], low[2], high[2]])
349    ax.autoscale(True)
350
351def sas_2J1x_x(x):
352    with np.errstate(all='ignore'):
353        retvalue = 2*J1(x)/x
354    retvalue[x == 0] = 1.
355    return retvalue
356
357def sas_3j1x_x(x):
358    """return 3*j1(x)/x"""
359    with np.errstate(all='ignore'):
360        retvalue = 3*(sin(x) - x*cos(x))/x**3
361    retvalue[x == 0.] = 1.
362    return retvalue
363
364def cylinder_Iq(q, radius, length):
[a1c32c2]365    z, w = leggauss(76)
366    cos_alpha = (z+1)/2
367    sin_alpha = sqrt(1.0 - cos_alpha**2)
[cfa28d3]368    Iq = np.empty_like(q)
369    for k, qk in enumerate(q):
[a1c32c2]370        qab, qc = qk*sin_alpha, qk*cos_alpha
371        Fq = sas_2J1x_x(qab*radius) * j0(qc*length/2)
372        Iq[k] = np.sum(w*Fq**2)
[cfa28d3]373    Iq = Iq/Iq[0]
374    return Iq
375
376def sphere_Iq(q, radius):
377    Iq = sas_3j1x_x(q*radius)**2
378    return Iq/Iq[0]
379
380def csbox_Iq(q, a, b, c, da, db, dc, slda, sldb, sldc, sld_core):
[a1c32c2]381    z, w = leggauss(76)
382
[cfa28d3]383    sld_solvent = 0
384    overlapping = False
385    dr0 = sld_core - sld_solvent
386    drA, drB, drC = slda-sld_solvent, sldb-sld_solvent, sldc-sld_solvent
387    tA, tB, tC = a + 2*da, b + 2*db, c + 2*dc
388
[a1c32c2]389    outer_sum = np.zeros_like(q)
390    for cos_alpha, outer_w in zip((z+1)/2, w):
[cfa28d3]391        sin_alpha = sqrt(1.0-cos_alpha*cos_alpha)
392        qc = q*cos_alpha
393        siC = c*j0(c*qc/2)
394        siCt = tC*j0(tC*qc/2)
[a1c32c2]395        inner_sum = np.zeros_like(q)
396        for beta, inner_w in zip((z + 1)*pi/4, w):
[cfa28d3]397            qa, qb = q*sin_alpha*sin(beta), q*sin_alpha*cos(beta)
398            siA = a*j0(a*qa/2)
399            siB = b*j0(b*qb/2)
400            siAt = tA*j0(tA*qa/2)
401            siBt = tB*j0(tB*qb/2)
402            if overlapping:
[a1c32c2]403                Fq = (dr0*siA*siB*siC
404                      + drA*(siAt-siA)*siB*siC
405                      + drB*siAt*(siBt-siB)*siC
406                      + drC*siAt*siBt*(siCt-siC))
[cfa28d3]407            else:
[a1c32c2]408                Fq = (dr0*siA*siB*siC
409                      + drA*(siAt-siA)*siB*siC
410                      + drB*siA*(siBt-siB)*siC
411                      + drC*siA*siB*(siCt-siC))
412            inner_sum += inner_w * Fq**2
413        outer_sum += outer_w * inner_sum
414    Iq = outer_sum / 4  # = outer*um*zm*8.0/(4.0*M_PI)
[cfa28d3]415    return Iq/Iq[0]
416
417def check_shape(shape, fn=None):
418    rho_solvent = 0
419    q = np.logspace(-3, 0, 200)
420    r = shape.r_bins(q, r_step=0.01)
421    sampling_density = 15000 / shape.volume()
422    rho, points = shape.sample(sampling_density)
423    Pr = calc_Pr(r, rho-rho_solvent, points)
424    Iq = calc_Iq(q, r, Pr)
425    theory = (q, fn(q)) if fn is not None else None
426
427    import pylab
428    #plot_points(rho, points); pylab.figure()
429    plot_calc(r, Pr, q, Iq, theory=theory)
430    pylab.show()
431
432def check_cylinder(radius=25, length=125, rho=2.):
433    shape = EllipticalCylinder(radius, radius, length, rho)
434    fn = lambda q: cylinder_Iq(q, radius, length)
435    check_shape(shape, fn)
436
437def check_sphere(radius=125, rho=2):
438    shape = TriaxialEllipsoid(radius, radius, radius, rho)
439    fn = lambda q: sphere_Iq(q, radius)
440    check_shape(shape, fn)
441
442def check_csbox(a=10, b=20, c=30, da=1, db=2, dc=3, slda=1, sldb=2, sldc=3, sld_core=4):
443    core = Box(a, b, c, sld_core)
444    side_a = Box(da, b, c, slda, center=((a+da)/2, 0, 0))
445    side_b = Box(a, db, c, sldb, center=(0, (b+db)/2, 0))
446    side_c = Box(a, b, dc, sldc, center=(0, 0, (c+dc)/2))
447    side_a2 = copy(side_a).shift(-a-da, 0, 0)
448    side_b2 = copy(side_b).shift(0, -b-db, 0)
449    side_c2 = copy(side_c).shift(0, 0, -c-dc)
450    shape = Composite((core, side_a, side_b, side_c, side_a2, side_b2, side_c2))
451    fn = lambda q: csbox_Iq(q, a, b, c, da, db, dc, slda, sldb, sldc, sld_core)
452    check_shape(shape, fn)
453
454if __name__ == "__main__":
455    check_cylinder(radius=10, length=40)
456    #check_sphere()
457    #check_csbox()
458    #check_csbox(da=100, db=200, dc=300)
Note: See TracBrowser for help on using the repository browser.