/
minres.py
428 lines (398 loc) · 15.1 KB
/
minres.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
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
"""
Note: this code is inspired from the following matlab source :
http://www.stanford.edu/group/SOL/software/minres.html
"""
import theano
import theano.tensor as TT
from theano import scan
import numpy
from pylearn2.utils import constantX
from pylearn2.expr.basic import multiple_switch, symGivens2, \
sqrt_inner_product, inner_product
# Messages that matches the flag value returned by the method
messages = [
' beta1 = 0. The exact solution is x = 0. ', # 0
' A solution to (poss. singular) Ax = b found, given rtol. ', # 1
' A least-squares solution was found, given rtol. ', # 2
' A solution to (poss. singular) Ax = b found, given eps. ', # 3
' A least-squares solution was found, given eps. ', # 4
' x has converged to an eigenvector. ', # 5
' xnorm has exceeded maxxnorm. ', # 6
' Acond has exceeded Acondlim. ', # 7
' The iteration limit was reached. ', # 8
' A least-squares solution for singular LS problem, given eps. ', # 9
' A least-squares solution for singular LS problem, given rtol.', # 10
' A null vector obtained, given rtol. ', # 11
' Numbers are too small to continue computation '] # 12
def minres(compute_Av,
bs,
rtol=constantX(1e-6),
maxit=20,
Ms=None,
shift=constantX(0.),
maxxnorm=constantX(1e15),
Acondlim=constantX(1e16),
profile=0):
"""
Attempts to find the minimum-length and minimum-residual-norm
solution :math:`x` to the system of linear equations :math:`A*x = b`
or least squares problem :math:`\\min||Ax-b||`.
The n-by-n coefficient matrix A must be symmetric (but need not be
positive definite or invertible). The right-hand-side column vector
b must have length n.
.. note:: This code is inspired from
http://www.stanford.edu/group/SOL/software/minres.html .
Parameters
----------
compute_Av : callable
Callable returing the symbolic expression for
`Av` (the product of matrix A with some vector v).
`v` should be a list of tensors, where the vector v means
the vector obtain by concatenating and flattening all tensors in v
bs : list
List of Theano expressions. We are looking to compute `A^-1\dot bs`.
rtol : float, optional
Specifies the tolerance of the method. Default is 1e-6.
maxit : int, positive, optional
Specifies the maximum number of iterations. Default is 20.
Ms : list
List of theano expression of same shape as `bs`. The method uses
these to precondition with diag(Ms)
shift : float, optional
Default is 0. Effectively solve the system (A - shift I) * x = b.
maxxnorm : float, positive, optional
Maximum bound on NORM(x). Default is 1e14.
Acondlim : float, positive, optional
Maximum bound on COND(A). Default is 1e15.
show : bool
If True, show iterations, otherwise suppress outputs. Default is
False.
Returns
-------
x : list
List of Theano tensor representing the solution
flag : tensor_like
Theano int scalar - convergence flag
0. beta1 = 0. The exact solution is x = 0.
1. A solution to (poss. singular) Ax = b found, given rtol.
2. Pseudoinverse solution for singular LS problem, given rtol.
3. A solution to (poss. singular) Ax = b found, given eps.
4. Pseudoinverse solution for singular LS problem, given eps.
5. x has converged to an eigenvector.
6. xnorm has exceeded maxxnorm.
7. Acond has exceeded Acondlim.
8. The iteration limit was reached.
9. 10. It is a least squares problem but no converged
solution yet.
iter : int
Iteration number at which x was computed: `0 <= iter <= maxit`.
relres : float
Real positive, the relative residual is defined as
NORM(b-A*x)/(NORM(A) * NORM(x) + NORM(b)),
computed recurrently here. If flag is 1 or 3, relres <= TOL.
relAres : float
Real positive, the relative-NORM(Ar) := NORM(Ar) / NORM(A)
computed recurrently here. If flag is 2 or 4, relAres <= TOL.
Anorm : float
Real positive, estimate of matrix 2-norm of A.
Acond : float
Real positive, estimate of condition number of A with respect to
2-norm.
xnorm : float
Non-negative positive, recurrently computed NORM(x)
Axnorm : float
Non-negative positive, recurrently computed NORM(A * x).
References
----------
.. [1] Choi, Sou-Cheng. Iterative Methods for Singular Linear
Equations and Least-Squares Problems, PhD Dissertation,
Stanford University, 2006.
"""
if not isinstance(bs, (tuple, list)):
bs = [bs]
return_as_list = False
else:
bs = list(bs)
return_as_list = True
eps = constantX(1e-23)
# Initialise
beta1 = sqrt_inner_product(bs)
#------------------------------------------------------------------
# Set up p and v for the first Lanczos vector v1.
# p = beta1 P' v1, where P = C**(-1).
# v is really P' v1.
#------------------------------------------------------------------
r3s = [b for b in bs]
r2s = [b for b in bs]
r1s = [b for b in bs]
if Ms is not None:
r3s = [b / m for b, m in zip(bs, Ms)]
beta1 = sqrt_inner_product(r3s, bs)
#------------------------------------------------------------------
## Initialize other quantities.
# Note that Anorm has been initialized by IsOpSym6.
# ------------------------------------------------------------------
bnorm = beta1
n_params = len(bs)
def loop(niter,
beta,
betan,
phi,
Acond,
cs,
dbarn,
eplnn,
rnorm,
sn,
Tnorm,
rnorml,
xnorm,
Dnorm,
gamma,
pnorm,
gammal,
Axnorm,
relrnorm,
relArnorml,
Anorm,
flag,
*args):
#-----------------------------------------------------------------
## Obtain quantities for the next Lanczos vector vk+1, k = 1, 2,...
# The general iteration is similar to the case k = 1 with v0 = 0:
#
# p1 = Operator * v1 - beta1 * v0,
# alpha1 = v1'p1,
# q2 = p2 - alpha1 * v1,
# beta2^2 = q2'q2,
# v2 = (1/beta2) q2.
#
# Again, p = betak P vk, where P = C**(-1).
# .... more description needed.
#-----------------------------------------------------------------
xs = args[0 * n_params: 1 * n_params]
r1s = args[1 * n_params: 2 * n_params]
r2s = args[2 * n_params: 3 * n_params]
r3s = args[3 * n_params: 4 * n_params]
dls = args[4 * n_params: 5 * n_params]
ds = args[5 * n_params: 6 * n_params]
betal = beta
beta = betan
vs = [r3 / beta for r3 in r3s]
r3s, upds = compute_Av(*vs)
r3s = [r3 - shift * v for r3, v in zip(r3s, vs)]
r3s = [TT.switch(TT.ge(niter, constantX(1.)),
r3 - (beta / betal) * r1,
r3) for r3, r1 in zip(r3s, r1s)]
alpha = inner_product(r3s, vs)
r3s = [r3 - (alpha / beta) * r2 for r3, r2 in zip(r3s, r2s)]
r1s = [r2 for r2 in r2s]
r2s = [r3 for r3 in r3s]
if Ms is not None:
r3s = [r3 / M for r3, M in zip(r3s, Ms)]
betan = sqrt_inner_product(r2s, r3s)
else:
betan = sqrt_inner_product(r3s)
pnorml = pnorm
pnorm = TT.switch(TT.eq(niter, constantX(0.)),
TT.sqrt(TT.sqr(alpha) + TT.sqr(betan)),
TT.sqrt(TT.sqr(alpha) + TT.sqr(betan) +
TT.sqr(beta)))
#-----------------------------------------------------------------
## Apply previous rotation Qk-1 to get
# [dlta_k epln_{k+1}] = [cs sn][dbar_k 0 ]
# [gbar_k dbar_{k+1} ] [sn -cs][alpha_k beta_{k+1}].
#-----------------------------------------------------------------
dbar = dbarn
epln = eplnn
dlta = cs * dbar + sn * alpha
gbar = sn * dbar - cs * alpha
eplnn = sn * betan
dbarn = -cs * betan
## Compute the current plane rotation Qk
gammal2 = gammal
gammal = gamma
cs, sn, gamma = symGivens2(gbar, betan)
tau = cs * phi
phi = sn * phi
Axnorm = TT.sqrt(TT.sqr(Axnorm) + TT.sqr(tau))
# Update d
dl2s = [dl for dl in dls]
dls = [d for d in ds]
ds = [TT.switch(TT.neq(gamma, constantX(0.)),
(v - epln * dl2 - dlta * dl) / gamma,
v)
for v, dl2, dl in zip(vs, dl2s, dls)]
d_norm = TT.switch(TT.neq(gamma, constantX(0.)),
sqrt_inner_product(ds),
constantX(numpy.inf))
# Update x except if it will become too big
xnorml = xnorm
dl2s = [x for x in xs]
xs = [x + tau * d for x, d in zip(xs, ds)]
xnorm = sqrt_inner_product(xs)
xs = [TT.switch(TT.ge(xnorm, maxxnorm),
dl2, x)
for dl2, x in zip(dl2s, xs)]
flag = TT.switch(TT.ge(xnorm, maxxnorm),
constantX(6.), flag)
# Estimate various norms
rnorml = rnorm # ||r_{k-1}||
Anorml = Anorm
Acondl = Acond
relrnorml = relrnorm
flag_no_6 = TT.neq(flag, constantX(6.))
Dnorm = TT.switch(flag_no_6,
TT.sqrt(TT.sqr(Dnorm) + TT.sqr(d_norm)),
Dnorm)
xnorm = TT.switch(flag_no_6, sqrt_inner_product(xs), xnorm)
rnorm = TT.switch(flag_no_6, phi, rnorm)
relrnorm = TT.switch(flag_no_6,
rnorm / (Anorm * xnorm + bnorm),
relrnorm)
Tnorm = TT.switch(flag_no_6,
TT.switch(TT.eq(niter, constantX(0.)),
TT.sqrt(TT.sqr(alpha) + TT.sqr(betan)),
TT.sqrt(TT.sqr(Tnorm) +
TT.sqr(beta) +
TT.sqr(alpha) +
TT.sqr(betan))),
Tnorm)
Anorm = TT.maximum(Anorm, pnorm)
Acond = Anorm * Dnorm
rootl = TT.sqrt(TT.sqr(gbar) + TT.sqr(dbarn))
Anorml = rnorml * rootl
relArnorml = rootl / Anorm
#---------------------------------------------------------------
# See if any of the stopping criteria are satisfied.
# In rare cases, flag is already -1 from above (Abar = const*I).
#---------------------------------------------------------------
epsx = Anorm * xnorm * eps
epsr = Anorm * xnorm * rtol
#Test for singular Hk (hence singular A)
# or x is already an LS solution (so again A must be singular).
t1 = constantX(1) + relrnorm
t2 = constantX(1) + relArnorml
flag = TT.switch(
TT.bitwise_or(TT.eq(flag, constantX(0)),
TT.eq(flag, constantX(6))),
multiple_switch(TT.le(t1, constantX(1)),
constantX(3),
TT.le(t2, constantX(1)),
constantX(4),
TT.le(relrnorm, rtol),
constantX(1),
TT.le(Anorm, constantX(1e-20)),
constantX(12),
TT.le(relArnorml, rtol),
constantX(10),
TT.ge(epsx, beta1),
constantX(5),
TT.ge(xnorm, maxxnorm),
constantX(6),
TT.ge(niter, TT.cast(maxit,
theano.config.floatX)),
constantX(8),
flag),
flag)
flag = TT.switch(TT.lt(Axnorm, rtol * Anorm * xnorm),
constantX(11.),
flag)
return [niter + constantX(1.),
beta,
betan,
phi,
Acond,
cs,
dbarn,
eplnn,
rnorm,
sn,
Tnorm,
rnorml,
xnorm,
Dnorm,
gamma,
pnorm,
gammal,
Axnorm,
relrnorm,
relArnorml,
Anorm,
flag] + xs + r1s + r2s + r3s + dls + ds, upds, \
theano.scan_module.scan_utils.until(TT.neq(flag, 0))
states = []
# 0 niter
states.append(constantX(0))
# 1 beta
states.append(constantX(0))
# 2 betan
states.append(beta1)
# 3 phi
states.append(beta1)
# 4 Acond
states.append(constantX(1))
# 5 cs
states.append(constantX(-1))
# 6 dbarn
states.append(constantX(0))
# 7 eplnn
states.append(constantX(0))
# 8 rnorm
states.append(beta1)
# 9 sn
states.append(constantX(0))
# 10 Tnorm
states.append(constantX(0))
# 11 rnorml
states.append(beta1)
# 12 xnorm
states.append(constantX(0))
# 13 Dnorm
states.append(constantX(0))
# 14 gamma
states.append(constantX(0))
# 15 pnorm
states.append(constantX(0))
# 16 gammal
states.append(constantX(0))
# 17 Axnorm
states.append(constantX(0))
# 18 relrnorm
states.append(constantX(1))
# 19 relArnorml
states.append(constantX(1))
# 20 Anorm
states.append(constantX(0))
# 21 flag
states.append(constantX(0))
xs = [TT.zeros_like(b) for b in bs]
ds = [TT.zeros_like(b) for b in bs]
dls = [TT.zeros_like(b) for b in bs]
rvals, loc_updates = scan(
loop,
outputs_info=(states + xs + r1s + r2s + r3s + dls + ds),
n_steps=maxit + numpy.int32(1),
name='minres',
profile=profile,
mode=theano.Mode(linker='cvm'))
assert isinstance(loc_updates, dict) and 'Ordered' in str(type(loc_updates))
niters = TT.cast(rvals[0][-1], 'int32')
flag = TT.cast(rvals[21][-1], 'int32')
relres = rvals[18][-1]
relAres = rvals[19][-1]
Anorm = rvals[20][-1]
Acond = rvals[4][-1]
xnorm = rvals[12][-1]
Axnorm = rvals[17][-1]
sol = [x[-1] for x in rvals[22: 22 + n_params]]
return (sol,
flag,
niters,
relres,
relAres,
Anorm,
Acond,
xnorm,
Axnorm,
loc_updates)