forked from mrgt/EulerLagrangianOT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
EulerCommon.py
322 lines (270 loc) · 9.56 KB
/
EulerCommon.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
import sys
sys.path.append('..');
sys.path.append('../../MongeAmpere/PyMongeAmpere-build');
sys.path.append('../../MongeAmpere/PyMongeAmpere-build/lib');
import MongeAmpere as ma
import numpy as np
import scipy as sp
import scipy.optimize as opt
import multiprocessing
from scipy.spatial import ConvexHull
def distance_point_line(m, n, pt):
u = n - m
Mpt = pt - m
norm_u = np.linalg.norm(u)
dist = np.linalg.norm(Mpt - (np.inner(Mpt,u)/(norm_u*norm_u))*u)
return dist
# compute an admissible dual variable for the optimal transport
# problem between Y and dens, assuming that dens does not vanish on
# the convex hull of its vertices
def estimate_dual_variable(dens, Y):
X = dens.vertices
bX = np.mean(X, 0)
bY = np.mean(Y, 0)
# compute radius of Y wrt bY
rY = np.max(np.linalg.norm(Y - np.tile(bY,(Y.shape[0],1)), axis=1))
# search the largest inscribed circle in X centered on bX
chX = ConvexHull(X)
points = chX.points
simplices = chX.simplices
rX = float('inf')
for simplex in simplices:
d = distance_point_line(points[simplex[0]], points[simplex[1]], bX)
rX = min(d,rX)
ratio = min(rX / rY,1.0)
psi_tilde0 = (0.5 * ratio * (np.power(Y[:,0]-bY[0],2)+
np.power(Y[:,1]-bY[1],2)) +
bX[0]*Y[:,0] + bX[1]*Y[:,1])
psi0 = np.power(Y[:,0],2) + np.power(Y[:,1],2) - 2*psi_tilde0
return psi0
# project an ordered point cloud to the L^2-closest "incompressible"
# point cloud, i.e. compute the solution of the optimal transport
# problem between the density dens and the point cloud Z, and move
# each point to the centroid of the corresponding Laguerre cell
def project_on_incompressible(dens,Z,verbose=False):
N = Z.shape[0]
nu = np.ones(N) * dens.mass()/N
w0 = estimate_dual_variable(dens,Z)
w = ma.optimal_transport_2(dens, Z, nu, w0=w0, verbose=verbose)
return dens.lloyd(Z,w)[0],w
# compute projection on incompressible
# then compute the mass, centroids and second moment of Laguerre cells
def projection_on_incompressible_moments(dens, Z):
N = Z.shape[0]
nu = np.ones(N) * dens.mass()/N
w0= estimate_dual_variable(dens, Z)
w = ma.optimal_transport_2(dens, Z, nu, w0=w0, verbose=False)
return dens.moments(Z,w)
def squared_distance_to_incompressible(dens, s):
mass,cent,mom = projection_on_incompressible_moments(dens,s)
# energy = cxx + cyy + |s|^2 mass - 2 (cx sx + cy sy)
E = sum(mass * (np.power(s[:,0],2) + np.power(s[:,1],2)) +
mom[:,0] + mom[:,1] -
2 * (s[:,0] * cent[:,0] + s[:,1] * cent[:,1]));
g = 2 * (np.vstack((mass,mass)).T * s - cent);
return E,g
def sq_dist_to_incompressible(args):
shape,s = args
E,g = squared_distance_to_incompressible(ma.Density_2(shape),s)
return E,g
def euler_partial_energy(shape, S, s0, s1, lbda, parallel_map=None):
T = S.shape[0]
N = S.shape[1]
E = 0
g = np.zeros((T,N,2))
# we divide by N in the definition of alpha and gamma, because the
# L^2 distance should be weighted by a probability measure nb:
# this is not necessary for beta, because the rescaling is done
# in squared_distance_to_incompressible
alpha = float(T)/float(N)
beta = lbda
gamma = lbda/float(N)
# kinetic energy
for i in xrange(0,T-1):
dS = S[i+1,:,:] - S[i,:,:]
E = E + alpha * np.sum(np.sum(np.power(dS,2)))
g[i,:,:] = g[i,:,:] - 2 * alpha * dS
g[i+1,:,:] = g[i+1,:,:] + 2 * alpha * dS
# the computation of square distances to incompressibility
# constraints can be optionally parellized
if parallel_map is None:
parallel_map = map;
EG = parallel_map(sq_dist_to_incompressible, zip([shape for i in xrange(1,T-1)],
[S[i] for i in xrange(1,T-1)]))
# penalization of incompressibility constraint
for i in xrange(1,T-1):
Ed,gd = EG[i-1]
E = E + beta * Ed
g[i,:,:] = g[i,:,:] + beta * gd
# penalization of boundary conditions
bc0 = S[0,:,:] - s0
bc1 = S[T-1,:,:] - s1
E = E + gamma * (np.sum(np.sum(np.power(bc0,2))) +
np.sum(np.sum(np.power(bc1,2))))
g[0,:,:] = g[0,:,:] + 2 * gamma * bc0
g[T-1,:,:] = g[T-1,:,:] + 2 * gamma * bc1
# to save intermediary solutions, uncomment next line
# euler_save("/tmp/euler-save-temp.npz", shape=shape, S=S)
return E,g
def convert_shape(F,S,shape):
N = len(S)
E,g = F(np.reshape(S,shape))
return E, np.reshape(g, len(S))
from contextlib import closing
def euler_solve_lbfgs_step(shape, S0, s0, s1, lbda, pgtol=1e-10):
with closing(multiprocessing.Pool(processes=2)) as pool:
T = S0.shape[0]
N = S0.shape[1]
parallel_map = lambda f,v: pool.map(f,v)
F = lambda S: euler_partial_energy(shape, S, s0, s1, lbda,
parallel_map)
Fc = lambda S: convert_shape(F,S,(T,N,2))
S,f,d = opt.fmin_l_bfgs_b(Fc,np.reshape(S0,T*N*2),
iprint=1, pgtol=pgtol, factr=10, m=20);
pool.terminate()
return np.reshape(S,(T,N,2))
def euler_solve_lbfgs(shape, X, Y, p, k=2, nsmooth=4, sigma=0):
N = X.shape[0]
T = np.power(2,k)+1;
h = 1.0/np.sqrt(N); # 1/h^2 = N
S0 = np.zeros((2,N,2))
S0[0] = X;
S0[1] = Y;
for j in xrange(k):
T0 = S0.shape[0]
T = 2*T0-1
S = np.zeros((T,N,2))
print "\nADDING TIMESTEPS (T=%d)" % T
for i in xrange(T0-1):
S[2*i] = S0[i]
mean = (S0[i] + S0[i+1])/2
# when adding the first intermediate timestep, it might be
# necessary to perturb the mean of the two point clouds
# (this happens for the disk inversion, because in this
# case mean == 0)
if T0 == 2:
mean = mean + sigma*np.random.randn(N,2)
# initial guess for the inserted timesteps
S[2*i+1] = project_on_incompressible(ma.Density_2(shape), mean, verbose=True)
S[T-1] = S0[T0-1]
S0 = S
print "\nLBFGS OPTIMIZATION (T=%d)" % T
lbda = 1/np.power(h, p)
S0 = euler_solve_lbfgs_step(shape,S0,X,Y,lbda)
return S0
import os
def ensure_dir(d):
if not os.path.exists(d):
os.makedirs(d)
def euler_save(fname, **kwargs):
f = open(fname, "wb")
np.savez_compressed(f,**kwargs)
f.close()
def euler_load_experiment(fname):
f = open(fname, "rb")
v = np.load(f)
shape = v['shape']
X = v['X']
Y = v['Y']
f.close()
return shape, X, Y
def euler_load_result(fname):
f = open(fname, "rb")
v = np.load(f)
shape = v['shape']
S = v['S']
Sproj = v['Sproj']
f.close()
return shape, S, Sproj
# display code
def embed_solution_into_H1(S):
T = S.shape[0]
N = S.shape[1]
Sflat = np.zeros((N, T*2 + (T-1)*2))
for i in xrange(N):
Sflat[i,0:T*2] = np.reshape(S[:,i,:], (T*2))
for i in xrange(N):
Sflat[i,T*2:T*2+(T-1)*2] = T*np.reshape(S[1:T,i,:]-S[0:(T-1),i,:], ((T-1)*2))
return Sflat
def bounding_box(X):
xm = np.amin(X[:,0])
xM = np.amax(X[:,0])
ym = np.amin(X[:,1])
yM = np.amax(X[:,1])
return np.array([xm,xM, ym, yM])
def cut_vertically(X):
bb = bounding_box(X)
Y = X.copy()
Y[:,0] = (Y[:,0] - bb[0])/(bb[1]-bb[0])
Y[:,1] = (Y[:,1] - bb[2])/(bb[3]-bb[2])
ii = np.nonzero(Y[:,0] > .66)
jj = np.nonzero((Y[:,0] <= .66) & (Y[:,0] > .33))
kk = np.nonzero(Y[:,0] <= .33)
return ii,jj,kk
# display
def draw_voronoi_edges(E):
nan = float('nan')
N = E.shape[0]
x = np.zeros(3*N)
y = np.zeros(3*N)
for i in xrange(N):
x[3*i] = E[i,0]
x[3*i+1] = E[i,2]
x[3*i+2] = nan
y[3*i] = E[i,1]
y[3*i+1] = E[i,3]
y[3*i+2] = nan
return x,y
def draw_bbox(bbox):
nan = float('nan')
x = np.zeros(12)
y = np.zeros(12)
x0 = bbox[0]; x1 = bbox[2];
y0 = bbox[1]; y1 = bbox[3];
x[0] = x0; y[0] = y0;
x[1] = x1; y[1] = y0;
x[2] = nan; y[2] = nan;
x[3] = x1; y[3] = y0;
x[4] = x1; y[4] = y1;
x[5] = nan; y[5] = nan;
x[6] = x1; y[6] = y1;
x[7] = x0; y[7] = y1;
x[8] = nan; y[8] = nan;
x[9] = x0; y[9] = y1;
x[10] = x0; y[10] = y0;
x[11] = nan; y[11] = nan;
return x,y
def sqmom(V):
return np.sum(V[:,0] * V[:,0] + V[:,1] * V[:,1])
def gen_grid(bbox, N):
L = (bbox[2] - bbox[0])/2.
H = (bbox[3] - bbox[1])/2.
# nn = 2*K*L, mm = 2*K*H, mm*nn = N => 4K^2*L*H = N
K = np.sqrt(N/(4.*L*H))
nn = int(np.floor(2*K*L))
mm = int(np.floor(2*K*H))
x, y = np.meshgrid(np.linspace(bbox[0],bbox[2],nn),
np.linspace(bbox[1],bbox[3],mm))
N = int(np.floor(mm*nn))
X = np.vstack((np.reshape(x,N,1),np.reshape(y,N,1))).T
return X,N
def perform_euler_simulation(X, V, nt, dt, bname, force, energy, plot, integrator):
ensure_dir(bname)
X,A,P,w = force(X)
energies = np.zeros((nt,1))
for i in xrange(nt):
print(i)
plot(X, P, w, '%s/%03d.png' % (bname, i))
if integrator == "vv":
# Velocity-Verlet integrator
W = V + 0.5*A*dt # V(t+dt/2)
X = X + W*dt # X(t+dt)
X,A,P,w = force(X) # A(t+dt)
V = W + 0.5*A*dt # V(t+dt)
else:
V = V + A*dt # V(t+dt) = V(t) + dt A(t)
X = X + V*dt # X(t+dt) = X(t) + dt V(t+dt)
X,A,P,w = force(X)
energies[i,:] = energy(X,P,V)
print energies[i,:]
np.savetxt('%s/energies.txt' % (bname), energies, delimiter=",")