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
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
|
# -*- coding: utf-8 -*-
"""
Gromov-Wasserstein transport method
===================================
"""
# Author: Erwan Vautier <erwan.vautier@gmail.com>
# Nicolas Courty <ncourty@irisa.fr>
# Rémi Flamary <remi.flamary@unice.fr>
#
# License: MIT License
import numpy as np
from .bregman import sinkhorn
from .utils import dist
from .optim import cg
def init_matrix(C1,C2,T,p,q,loss_fun='square_loss'):
""" Return loss matrices and tensors for Gromov-Wasserstein fast computation
Returns the value of \mathcal{L}(C1,C2) \otimes T with the selected loss
function as the loss function of Gromow-Wasserstein discrepancy.
The matrices are computed as described in Proposition 1 in [12]
Where :
* C1 : Metric cost matrix in the source space
* C2 : Metric cost matrix in the target space
* T : A coupling between those two spaces
The square-loss function L(a,b)=(1/2)*|a-b|^2 is read as :
L(a,b) = f1(a)+f2(b)-h1(a)*h2(b) with :
* f1(a)=(a^2)/2
* f2(b)=(b^2)/2
* h1(a)=a
* h2(b)=b
The kl-loss function L(a,b)=(1/2)*|a-b|^2 is read as :
L(a,b) = f1(a)+f2(b)-h1(a)*h2(b) with :
* f1(a)=a*log(a)-a
* f2(b)=b
* h1(a)=a
* h2(b)=log(b)
Parameters
----------
C1 : ndarray, shape (ns, ns)
Metric cost matrix in the source space
C2 : ndarray, shape (nt, nt)
Metric costfr matrix in the target space
T : ndarray, shape (ns, nt)
Coupling between source and target spaces
p : ndarray, shape (ns,)
Returns
-------
constC : ndarray, shape (ns, nt)
Constant C matrix in Eq. (6)
hC1 : ndarray, shape (ns, ns)
h1(C1) matrix in Eq. (6)
hC2 : ndarray, shape (nt, nt)
h2(C) matrix in Eq. (6)
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
if loss_fun == 'square_loss':
def f1(a):
return (a**2)/2
def f2(b):
return (b**2)/2
def h1(a):
return a
def h2(b):
return b
elif loss_fun == 'kl_loss':
def f1(a):
return a * np.log(a + 1e-15) - a
def f2(b):
return b
def h1(a):
return a
def h2(b):
return np.log(b + 1e-15)
constC1=np.dot(np.dot(f1(C1),p.reshape(-1,1)),
np.ones(len(q)).reshape(1,-1))
constC2=np.dot(np.ones(len(p)).reshape(-1,1),
np.dot(q.reshape(1,-1),f2(C2).T))
constC=constC1+constC2
hC1 = h1(C1)
hC2 = h2(C2)
return constC,hC1,hC2
def tensor_product(constC,hC1,hC2,T):
""" Return the tensor for Gromov-Wasserstein fast computation
The tensor is computed as described in Proposition 1 Eq. (6) in [12].
Parameters
----------
constC : ndarray, shape (ns, nt)
Constant C matrix in Eq. (6)
hC1 : ndarray, shape (ns, ns)
h1(C1) matrix in Eq. (6)
hC2 : ndarray, shape (nt, nt)
h2(C) matrix in Eq. (6)
Returns
-------
tens : ndarray, shape (ns, nt)
\mathcal{L}(C1,C2) \otimes T tensor-matrix multiplication result
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
A=-np.dot(hC1, T).dot(hC2.T)
tens = constC+A
#tens -= tens.min()
return tens
def gwloss(constC,hC1,hC2,T):
""" Return the Loss for Gromov-Wasserstein
The loss is computed as described in Proposition 1 Eq. (6) in [12].
Parameters
----------
constC : ndarray, shape (ns, nt)
Constant C matrix in Eq. (6)
hC1 : ndarray, shape (ns, ns)
h1(C1) matrix in Eq. (6)
hC2 : ndarray, shape (nt, nt)
h2(C) matrix in Eq. (6)
T : ndarray, shape (ns, nt)
Current value of transport matrix T
Returns
-------
loss : float
Gromov Wasserstein loss
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
tens=tensor_product(constC,hC1,hC2,T)
return np.sum(tens*T)
def gwggrad(constC,hC1,hC2,T):
""" Return the gradient for Gromov-Wasserstein
The gradient is computed as described in Proposition 2 in [12].
Parameters
----------
constC : ndarray, shape (ns, nt)
Constant C matrix in Eq. (6)
hC1 : ndarray, shape (ns, ns)
h1(C1) matrix in Eq. (6)
hC2 : ndarray, shape (nt, nt)
h2(C) matrix in Eq. (6)
T : ndarray, shape (ns, nt)
Current value of transport matrix T
Returns
-------
grad : ndarray, shape (ns, nt)
Gromov Wasserstein gradient
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
return 2*tensor_product(constC,hC1,hC2,T) # [12] Prop. 2 misses a 2 factor
def gromov_wasserstein(C1,C2,p,q,loss_fun,log=False,**kwargs):
"""
Returns the gromov-wasserstein discrepancy between the two measured similarity matrices
(C1,p) and (C2,q)
The function solves the following optimization problem:
.. math::
\GW_Dist = \min_T \sum_{i,j,k,l} L(C1_{i,k},C2_{j,l})*T_{i,j}*T_{k,l}
Where :
C1 : Metric cost matrix in the source space
C2 : Metric cost matrix in the target space
p : distribution in the source space
q : distribution in the target space
L : loss function to account for the misfit between the similarity matrices
H : entropy
Parameters
----------
C1 : ndarray, shape (ns, ns)
Metric cost matrix in the source space
C2 : ndarray, shape (nt, nt)
Metric costfr matrix in the target space
p : ndarray, shape (ns,)
distribution in the source space
q : ndarray, shape (nt,)
distribution in the target space
loss_fun : string
loss function used for the solver either 'square_loss' or 'kl_loss'
max_iter : int, optional
Max number of iterations
tol : float, optional
Stop threshold on error (>0)
verbose : bool, optional
Print information along iterations
log : bool, optional
record log if True
Returns
-------
gw_dist : float
Gromov-Wasserstein distance
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
T = np.eye(len(p), len(q))
constC,hC1,hC2=init_matrix(C1,C2,T,p,q,loss_fun)
G0=p[:,None]*q[None,:]
def f(G):
return gwloss(constC,hC1,hC2,G)
def df(G):
return gwggrad(constC,hC1,hC2,G)
if log:
res,log=cg(p,q,0,alpha,f,df,G0,log=True,**kwargs)
log['gw_dist']=gwloss(constC,hC1,hC2,res)
return res,log
else:
return cg(p,q,0,alpha,f,df,G0,**kwargs)
def update_square_loss(p, lambdas, T, Cs):
"""
Updates C according to the L2 Loss kernel with the S Ts couplings
calculated at each iteration
Parameters
----------
p : ndarray, shape (N,)
masses in the targeted barycenter
lambdas : list of float
list of the S spaces' weights
T : list of S np.ndarray(ns,N)
the S Ts couplings calculated at each iteration
Cs : list of S ndarray, shape(ns,ns)
Metric cost matrices
Returns
----------
C : ndarray, shape (nt,nt)
updated C matrix
"""
tmpsum = sum([lambdas[s] * np.dot(T[s].T, Cs[s]).dot(T[s])
for s in range(len(T))])
ppt = np.outer(p, p)
return np.divide(tmpsum, ppt)
def update_kl_loss(p, lambdas, T, Cs):
"""
Updates C according to the KL Loss kernel with the S Ts couplings calculated at each iteration
Parameters
----------
p : ndarray, shape (N,)
weights in the targeted barycenter
lambdas : list of the S spaces' weights
T : list of S np.ndarray(ns,N)
the S Ts couplings calculated at each iteration
Cs : list of S ndarray, shape(ns,ns)
Metric cost matrices
Returns
----------
C : ndarray, shape (ns,ns)
updated C matrix
"""
tmpsum = sum([lambdas[s] * np.dot(T[s].T, Cs[s]).dot(T[s])
for s in range(len(T))])
ppt = np.outer(p, p)
return np.exp(np.divide(tmpsum, ppt))
def entropic_gromov_wasserstein(C1, C2, p, q, loss_fun, epsilon,
max_iter=1000, tol=1e-9, verbose=False, log=False):
"""
Returns the regularized gromov-wasserstein coupling between the two measured similarity matrices
(C1,p) and (C2,q)
The function solves the following optimization problem:
.. math::
\GW = arg\min_T \sum_{i,j,k,l} L(C1_{i,k},C2_{j,l})*T_{i,j}*T_{k,l}-\epsilon(H(T))
s.t. \GW 1 = p
\GW^T 1= q
\GW\geq 0
Where :
C1 : Metric cost matrix in the source space
C2 : Metric cost matrix in the target space
p : distribution in the source space
q : distribution in the target space
L : loss function to account for the misfit between the similarity matrices
H : entropy
Parameters
----------
C1 : ndarray, shape (ns, ns)
Metric cost matrix in the source space
C2 : ndarray, shape (nt, nt)
Metric costfr matrix in the target space
p : ndarray, shape (ns,)
distribution in the source space
q : ndarray, shape (nt,)
distribution in the target space
loss_fun : string
loss function used for the solver either 'square_loss' or 'kl_loss'
epsilon : float
Regularization term >0
max_iter : int, optional
Max number of iterations
tol : float, optional
Stop threshold on error (>0)
verbose : bool, optional
Print information along iterations
log : bool, optional
record log if True
Returns
-------
T : ndarray, shape (ns, nt)
coupling between the two spaces that minimizes :
\sum_{i,j,k,l} L(C1_{i,k},C2_{j,l})*T_{i,j}*T_{k,l}-\epsilon(H(T))
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
C1 = np.asarray(C1, dtype=np.float64)
C2 = np.asarray(C2, dtype=np.float64)
T = np.outer(p, q) # Initialization
constC,hC1,hC2=init_matrix(C1,C2,T,p,q,loss_fun)
cpt = 0
err = 1
if log:
log={'err':[]}
while (err > tol and cpt < max_iter):
Tprev = T
# compute the gradient
tens=gwggrad(constC,hC1,hC2,T)
T = sinkhorn(p, q, tens, epsilon)
if cpt % 10 == 0:
# we can speed up the process by checking for the error only all
# the 10th iterations
err = np.linalg.norm(T - Tprev)
if log:
log['err'].append(err)
if verbose:
if cpt % 200 == 0:
print('{:5s}|{:12s}'.format(
'It.', 'Err') + '\n' + '-' * 19)
print('{:5d}|{:8e}|'.format(cpt, err))
cpt += 1
if log:
log['gw_dist']=gwloss(constC,hC1,hC2,T)
return T, log
else:
return T
def entropic_gromov_wasserstein2(C1, C2, p, q, loss_fun, epsilon,
max_iter=1000, tol=1e-9, verbose=False, log=False):
"""
Returns the entropic gromov-wasserstein discrepancy between the two measured similarity matrices
(C1,p) and (C2,q)
The function solves the following optimization problem:
.. math::
\GW_Dist = \min_T \sum_{i,j,k,l} L(C1_{i,k},C2_{j,l})*T_{i,j}*T_{k,l}-\epsilon(H(T))
Where :
C1 : Metric cost matrix in the source space
C2 : Metric cost matrix in the target space
p : distribution in the source space
q : distribution in the target space
L : loss function to account for the misfit between the similarity matrices
H : entropy
Parameters
----------
C1 : ndarray, shape (ns, ns)
Metric cost matrix in the source space
C2 : ndarray, shape (nt, nt)
Metric costfr matrix in the target space
p : ndarray, shape (ns,)
distribution in the source space
q : ndarray, shape (nt,)
distribution in the target space
loss_fun : string
loss function used for the solver either 'square_loss' or 'kl_loss'
epsilon : float
Regularization term >0
max_iter : int, optional
Max number of iterations
tol : float, optional
Stop threshold on error (>0)
verbose : bool, optional
Print information along iterations
log : bool, optional
record log if True
Returns
-------
gw_dist : float
Gromov-Wasserstein distance
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
gw, logv = entropic_gromov_wasserstein(
C1, C2, p, q, loss_fun, epsilon, max_iter, tol, verbose, log=True)
log['T']=gw
if log:
return logv['gw_dist'], logv
else:
return logv['gw_dist']
def gromov_barycenters(N, Cs, ps, p, lambdas, loss_fun, epsilon,
max_iter=1000, tol=1e-9, verbose=False, log=False, init_C=None):
"""
Returns the gromov-wasserstein barycenters of S measured similarity matrices
(Cs)_{s=1}^{s=S}
The function solves the following optimization problem:
.. math::
C = argmin_C\in R^NxN \sum_s \lambda_s GW(C,Cs,p,ps)
Where :
Cs : metric cost matrix
ps : distribution
Parameters
----------
N : Integer
Size of the targeted barycenter
Cs : list of S np.ndarray(ns,ns)
Metric cost matrices
ps : list of S np.ndarray(ns,)
sample weights in the S spaces
p : ndarray, shape(N,)
weights in the targeted barycenter
lambdas : list of float
list of the S spaces' weights
loss_fun : tensor-matrix multiplication function based on specific loss function
update : function(p,lambdas,T,Cs) that updates C according to a specific Kernel
with the S Ts couplings calculated at each iteration
epsilon : float
Regularization term >0
max_iter : int, optional
Max number of iterations
tol : float, optional
Stop threshol on error (>0)
verbose : bool, optional
Print information along iterations
log : bool, optional
record log if True
init_C : bool, ndarray, shape(N,N)
random initial value for the C matrix provided by user
Returns
-------
C : ndarray, shape (N, N)
Similarity matrix in the barycenter space (permutated arbitrarily)
References
----------
.. [12] Peyré, Gabriel, Marco Cuturi, and Justin Solomon,
"Gromov-Wasserstein averaging of kernel and distance matrices."
International Conference on Machine Learning (ICML). 2016.
"""
S = len(Cs)
Cs = [np.asarray(Cs[s], dtype=np.float64) for s in range(S)]
lambdas = np.asarray(lambdas, dtype=np.float64)
# Initialization of C : random SPD matrix (if not provided by user)
if init_C is None:
xalea = np.random.randn(N, 2)
C = dist(xalea, xalea)
C /= C.max()
else:
C = init_C
cpt = 0
err = 1
error = []
while(err > tol and cpt < max_iter):
Cprev = C
T = [entropic_gromov_wasserstein(Cs[s], C, ps[s], p, loss_fun, epsilon,
max_iter, 1e-5, verbose, log) for s in range(S)]
if loss_fun == 'square_loss':
C = update_square_loss(p, lambdas, T, Cs)
elif loss_fun == 'kl_loss':
C = update_kl_loss(p, lambdas, T, Cs)
if cpt % 10 == 0:
# we can speed up the process by checking for the error only all
# the 10th iterations
err = np.linalg.norm(C - Cprev)
error.append(err)
if log:
log['err'].append(err)
if verbose:
if cpt % 200 == 0:
print('{:5s}|{:12s}'.format(
'It.', 'Err') + '\n' + '-' * 19)
print('{:5d}|{:8e}|'.format(cpt, err))
cpt += 1
return C
|