nested_blockmodel.py 33.8 KB
Newer Older
1
2
3
4
5
#! /usr/bin/env python
# -*- coding: utf-8 -*-
#
# graph_tool -- a general graph manipulation python module
#
Tiago Peixoto's avatar
Tiago Peixoto committed
6
# Copyright (C) 2006-2022 Tiago de Paula Peixoto <tiago@skewed.de>
7
#
8
9
10
11
# This program is free software; you can redistribute it and/or modify it under
# the terms of the GNU Lesser General Public License as published by the Free
# Software Foundation; either version 3 of the License, or (at your option) any
# later version.
12
#
13
14
15
16
# This program is distributed in the hope that it will be useful, but WITHOUT
# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
# FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
# details.
17
#
18
19
# You should have received a copy of the GNU Lesser General Public License
# along with this program. If not, see <http://www.gnu.org/licenses/>.
20

Alex Henrie's avatar
Alex Henrie committed
21
from .. import _prop, Graph, GraphView
22
23
24
25

from . base_states import _bm_test
from . base_states import *

26
27
28
29
30
31
32
33
34
35
36
37
38
39
from . blockmodel import *
from . overlap_blockmodel import *
from . layered_blockmodel import *

from numpy import *
import numpy
import copy

class NestedBlockState(object):
    r"""The nested stochastic block model state of a given graph.

    Parameters
    ----------
    g : :class:`~graph_tool.Graph`
Tiago Peixoto's avatar
Tiago Peixoto committed
40
        Graph to be modeled.
Tiago Peixoto's avatar
Tiago Peixoto committed
41
42
43
    bs : ``list`` of :class:`~graph_tool.VertexPropertyMap` or :class:`numpy.ndarray` (optional, default: ``None``)
        Hierarchical node partition. If not provided it will correspond to a
        single-group hierarchy of length :math:`\lceil\log_2(N)\rceil`.
44
    base_type : ``type`` (optional, default: :class:`~graph_tool.inference.BlockState`)
Tiago Peixoto's avatar
Tiago Peixoto committed
45
        State type for lowermost level
46
47
48
        (e.g. :class:`~graph_tool.inference.BlockState`,
        :class:`~graph_tool.inference.OverlapBlockState` or
        :class:`~graph_tool.inference.LayeredBlockState`)
49
50
51
52
53
54
    hstate_args : ``dict`` (optional, default: `{}`)
        Keyword arguments to be passed to the constructor of the higher-level
        states.
    hentropy_args : ``dict`` (optional, default: `{}`)
        Keyword arguments to be passed to the ``entropy()`` method of the
        higher-level states.
55
    state_args : ``dict`` (optional, default: ``{}``)
56
        Keyword arguments to be passed to base type constructor.
57
58
59
    **kwargs :  keyword arguments
        Keyword arguments to be passed to base type constructor. The
        ``state_args`` parameter overrides this.
Tiago Peixoto's avatar
Tiago Peixoto committed
60

61
    """
62

Tiago Peixoto's avatar
Tiago Peixoto committed
63
    def __init__(self, g, bs=None, base_type=BlockState, state_args={},
64
                 hstate_args={}, hentropy_args={}, **kwargs):
65
        self.g = g
Tiago Peixoto's avatar
Tiago Peixoto committed
66

67
68
69
70
71
        self.base_type = base_type
        if base_type is LayeredBlockState:
            self.Lrecdx = []
        else:
            self.Lrecdx = libcore.Vector_double()
72
        self.state_args = dict(kwargs, **state_args)
73
        self.state_args["Lrecdx"] = self.Lrecdx
74
75
76
77
        if "rec_params" not in self.state_args:
            recs = self.state_args.get("recs", None)
            if recs is not None:
                self.state_args["rec_params"] = ["microcanonical"] * len(recs)
78
        self.hstate_args = dict(dict(deg_corr=False, vweight="nonempty"),
79
                                **hstate_args)
80
        self.hstate_args["Lrecdx"] = self.Lrecdx
81
        self.hstate_args["copy_bg"] = False
82
83
84
85
86
87
88
89
        self.hentropy_args = dict(hentropy_args,
                                  adjacency=True,
                                  dense=True,
                                  multigraph=True,
                                  dl=True,
                                  partition_dl=True,
                                  degree_dl=True,
                                  degree_dl_kind="distributed",
90
                                  edges_dl=False,
91
                                  exact=True,
92
                                  recs=True,
93
                                  recs_dl=False,
94
                                  beta_dl=1.)
95

96
97
        self.levels = [base_type(g, b=bs[0] if bs is not None else None,
                                 **self.state_args)]
98

99
100
101
102
103
104
105
        if bs is None:
            if base_type is OverlapBlockState:
                N = 2 * self.levels[0].get_N()
            else:
                N = self.levels[0].get_N()
            L = int(numpy.ceil(numpy.log2(N)))
            bs = [None] * (L + 1)
106

107
        for i, b in enumerate(bs[1:]):
108
            state = self.levels[-1]
109
110
111
112
            args = self.hstate_args
            if i == len(bs[1:]) - 1:
                args = dict(args, clabel=None, pclabel=None)
            bstate = state.get_block_state(b=b, **args)
113
114
            self.levels.append(bstate)

115
116
        self._regen_Lrecdx()

117
        self._couple_levels(self.hentropy_args, None)
118

119
120
121
        if _bm_test():
            self._consistency_check()

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
    def _regen_Lrecdx(self, lstate=None):
        if lstate is None:
            levels = self.levels
            Lrecdx = self.Lrecdx
        else:
            levels = [s for s in self.levels]
            l, s = lstate
            levels[l] = s
            s = s.get_block_state(**dict(self.hstate_args,
                                         b=s.get_bclabel(),
                                         copy_bg=False))
            if l < len(levels) - 1:
                levels[l+1] = s
            else:
                levels.append(s)
            if self.base_type is LayeredBlockState:
                Lrecdx = [x.copy() for x in self.Lrecdx]
            else:
                Lrecdx = self.Lrecdx.copy()

        if self.base_type is not LayeredBlockState:
            Lrecdx.a = 0
            Lrecdx[0] = len([s for s in levels if s._state.get_B_E_D() > 0])
            for s in levels:
                Lrecdx.a[1:] += s.recdx.a * s._state.get_B_E_D()
                s.epsilon.a = levels[0].epsilon.a
            for s in levels:
                s.Lrecdx.a = Lrecdx.a
        else:
            Lrecdx[0].a = 0
            Lrecdx[0][0] = len([s for s in levels if s._state.get_B_E_D() > 0])
            for j in range(levels[0].C):
                Lrecdx[j+1].a = 0
                Lrecdx[j+1][0] = len([s for s in levels if s._state.get_layer(j).get_B_E_D() > 0])
            for s in levels:
                Lrecdx[0].a[1:] += s.recdx.a * s._state.get_B_E_D()
                s.epsilon.a = levels[0].epsilon.a
                for j in range(levels[0].C):
                    Lrecdx[j+1].a[1:] += s.layer_states[j].recdx.a * s._state.get_layer(j).get_B_E_D()
                    s.layer_states[j].epsilon.a = levels[0].epsilon.a

            for s in self.levels:
                for x, y in zip(s.Lrecdx, Lrecdx):
                    x.a = y.a

        if lstate is not None:
            return Lrecdx

170

171
172
173
    def _regen_levels(self):
        for l in range(1, len(self.levels)):
            state = self.levels[l]
174
175
            nstate = self.levels[l-1].get_block_state(b=state.b,
                                                      **self.hstate_args)
176
            self.levels[l] = nstate
177
        self._regen_Lrecdx()
178

179
180
181
    def __repr__(self):
        return "<NestedBlockState object, with base %s, and %d levels of sizes %s at 0x%x>" % \
            (repr(self.levels[0]), len(self.levels),
182
             str([(s.get_N(), s.get_nonempty_B()) for s in self.levels]), id(self))
183
184
185
186

    def __copy__(self):
        return self.copy()

187
    def copy(self, **kwargs):
188
189
        r"""Copies the block state. The parameters override the state properties,
        and have the same meaning as in the constructor."""
190
191
        state = dict(self.__getstate__(), **kwargs)
        return NestedBlockState(**state)
192
193

    def __getstate__(self):
194
        base_state = self.levels[0].__getstate__()
195
196
197
        base_state.pop("Lrecdx", None)
        base_state.pop("epsilon", None)
        base_state.pop("drec", None)
198
199
200
201
202
        state_args = dict(base_state, **self.state_args)
        state_args.pop("g", None)
        state_args.pop("b", None)
        state = dict(g=self.g, bs=self.get_bs(),
                     base_type=type(self.levels[0]),
203
                     hstate_args=self.hstate_args,
204
                     hentropy_args=self.hentropy_args,
205
                     state_args=state_args)
206
207
208
        return state

    def __setstate__(self, state):
209
        self.__init__(**state)
210

211
212
213
214
    def get_bs(self):
        """Get hierarchy levels as a list of :class:`numpy.ndarray` objects with the
        group memberships at each level.
        """
215
        return [s.b.fa.copy() for s in self.levels]
216

217
218
219
220
    def get_state(self):
        """Alias to :meth:`~NestedBlockState.get_bs`."""
        return self.get_bs()

221
222
223
224
225
    def set_state(self, bs):
        r"""Sets the internal nested partition of the state."""
        for i in range(len(bs)):
            self.levels[i].set_state(bs[i])

226
    def get_levels(self):
227
        """Get hierarchy levels as a list of :class:`~graph_tool.inference.BlockState`
228
229
230
        instances."""
        return self.levels

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
    def project_partition(self, j, l):
        """Project partition of level ``j`` onto level ``l``, and return it."""
        b = self.levels[l].b.copy()
        for i in range(l + 1, j + 1):
            clabel = self.levels[i].b.copy()
            pmap(b, clabel)
        return b

    def propagate_clabel(self, l):
        """Project base clabel to level ``l``."""
        clabel = self.levels[0].clabel.copy()
        for j in range(l):
            bg = self.levels[j].bg
            bclabel = bg.new_vertex_property("int")
            reverse_map(self.levels[j].b, bclabel)
            pmap(bclabel, clabel)
            clabel = bclabel
        return clabel

    def get_clabel(self, l):
        """Get clabel for level ``l``."""
        clabel = self.propagate_clabel(l)
        if l < len(self.levels) - 1:
            b = self.project_partition(l + 1, l)
            clabel.fa += (clabel.fa.max() + 1) * b.fa
        return clabel

    def _consistency_check(self):
        for l in range(1, len(self.levels)):
            b = self.levels[l].b.fa.copy()
            state = self.levels[l-1]
262
263
264
265
            args = self.hstate_args
            if l == len(self.levels) - 1:
                args = dict(args, clabel=None, pclabel=None)
            bstate = state.get_block_state(b=b, **args)
266
            b2 = bstate.b.fa.copy()
267
268
            b = contiguous_map(b)
            b2 = contiguous_map(b2)
269
            assert ((b == b2).all() and
270
271
272
                    math.isclose(bstate.entropy(dl=False),
                                 self.levels[l].entropy(dl=False),
                                 abs_tol=1e-8)), \
273
274
275
                "inconsistent level %d (%s %g,  %s %g): %s" % \
                (l, str(bstate), bstate.entropy(), str(self.levels[l]),
                 self.levels[l].entropy(), str(self))
276
277
            assert (bstate.get_N() >= bstate.get_nonempty_B()), \
                (l, bstate.get_N(), bstate.get_nonempty_B(), str(self))
278

279
    def level_entropy(self, l, bstate=None, **kwargs):
280
281
282
283
284
        """Compute the entropy of level ``l``."""

        if bstate is None:
            bstate = self.levels[l]

285
286
287
288
289
290
291
        kwargs = kwargs.copy()
        hentropy_args = dict(self.hentropy_args,
                             **kwargs.pop("hentropy_args", {}))
        hentropy_args_top = dict(dict(hentropy_args, edges_dl=True,
                                      recs_dl=True),
                                 **kwargs.pop("hentropy_args_top", {}))

292
        if l > 0:
293
294
295
296
            if l == (len(self.levels) - 1):
                eargs = hentropy_args_top
            else:
                eargs = hentropy_args
297
        else:
298
            eargs = dict(kwargs, edges_dl=False)
299

300
        S = bstate.entropy(**eargs)
301
302
303
304

        if l > 0:
            S *= kwargs.get("beta_dl", 1.)

305
306
        return S

307
    def _Lrecdx_entropy(self, Lrecdx=None):
308
309
        if self.base_type is not LayeredBlockState:
            S_D = 0
310

311
312
313
314
315
316
            if Lrecdx is None:
                Lrecdx = self.Lrecdx
                for s in self.levels:
                    B_E_D = s._state.get_B_E_D()
                    if B_E_D > 0:
                        S_D -= log(B_E_D)
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
            S = 0
            for i in range(len(self.levels[0].rec)):
                if self.levels[0].rec_types[i] != libinference.rec_type.real_normal:
                    continue
                assert not _bm_test() or Lrecdx[i+1] >= 0, (i, Lrecdx[i+1])
                S += -libinference.positive_w_log_P(Lrecdx[0], Lrecdx[i+1],
                                                    numpy.nan, numpy.nan,
                                                    self.levels[0].epsilon[i])
                S += S_D
            return S
        else:
            S_D = [0 for j in range(self.levels[0].C)]
            if Lrecdx is None:
                Lrecdx = self.Lrecdx
                for s in self.levels:
                    for j in range(self.levels[0].C):
                        B_E_D = s._state.get_layer(j).get_B_E_D()
                        if B_E_D > 0:
                            S_D[j] -= log(B_E_D)

            S = 0
            for i in range(len(self.levels[0].rec)):
                if self.levels[0].rec_types[i] != libinference.rec_type.real_normal:
                    continue
                for j in range(self.levels[0].C):
                    assert not _bm_test() or Lrecdx[j+1][i+1] >= 0, (i, j, Lrecdx[j+1][i+1])
                    S += -libinference.positive_w_log_P(Lrecdx[j+1][0],
                                                        Lrecdx[j+1][i+1],
                                                        numpy.nan, numpy.nan,
                                                        self.levels[0].epsilon[i])
                    S += S_D[j]
            return S
350

351
    @copy_state_wrap
352
    def entropy(self, **kwargs):
Tiago Peixoto's avatar
Tiago Peixoto committed
353
354
        """Compute the entropy of whole hierarchy.

355
356
        The keyword arguments are passed to the ``entropy()`` method of the
        underlying state objects
357
358
359
        (e.g. :class:`graph_tool.inference.BlockState.entropy`,
        :class:`graph_tool.inference.OverlapBlockState.entropy`, or
        :class:`graph_tool.inference.LayeredBlockState.entropy`).  """
360
361
        S = 0
        for l in range(len(self.levels)):
362
            S += self.level_entropy(l, **dict(kwargs, test=False))
363

364
        S += kwargs.get("beta_dl", 1.) * self._Lrecdx_entropy()
365

366
367
        return S

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
    def move_vertex(self, v, s):
        r"""Move vertex ``v`` to block ``s``."""
        self.levels[0].move_vertex(v, s)
        self._regen_levels()

    def remove_vertex(self, v):
        r"""Remove vertex ``v`` from its current group.

        This optionally accepts a list of vertices to remove.

        .. warning::

           This will leave the state in an inconsistent state before the vertex
           is returned to some other group, or if the same vertex is removed
           twice.
        """
        self.levels[0].remove_vertex(v)
        self._regen_levels()

    def add_vertex(self, v, r):
        r"""Add vertex ``v`` to block ``r``.

        This optionally accepts a list of vertices and blocks to add.

        .. warning::

           This can leave the state in an inconsistent state if a vertex is
           added twice to the same group.
        """
        self.levels[0].add_vertex(v, r)
        self._regen_levels()

400
    def get_edges_prob(self, missing, spurious=[], entropy_args={}):
401
        r"""Compute the joint log-probability of the missing and spurious edges given by
402
403
404
405
406
407
408
409
        ``missing`` and ``spurious`` (a list of ``(source, target)``
        tuples, or :meth:`~graph_tool.Edge` instances), together with the
        observed edges.

        More precisely, the log-likelihood returned is

        .. math::

410
            \ln \frac{P(\boldsymbol G + \delta \boldsymbol G | \boldsymbol b)}{P(\boldsymbol G| \boldsymbol b)}
411
412
413
414
415

        where :math:`\boldsymbol G + \delta \boldsymbol G` is the modified graph
        (with missing edges added and spurious edges deleted).

        The values in ``entropy_args`` are passed to
416
        :meth:`graph_tool.inference.BlockState.entropy()` to calculate the
417
418
        log-probability.
        """
419

420
421
422
423
424
425
426
        entropy_args = entropy_args.copy()
        hentropy_args = dict(self.hentropy_args,
                             **entropy_args.pop("hentropy_args", {}))
        hentropy_args_top = dict(dict(hentropy_args, edges_dl=True,
                                      recs_dl=True),
                                 **entropy_args.pop("hentropy_args_top", {}))

427
        L = 0
428
        for l, lstate in enumerate(self.levels):
429
            if l > 0:
430
431
432
433
                if l == (len(self.levels) - 1):
                    eargs = hentropy_args_top
                else:
                    eargs = hentropy_args
434
435
436
            else:
                eargs = entropy_args

437
438
439
440
            lstate._couple_state(None, None)
            if l > 0:
                lstate._state.sync_emat()
                lstate._state.clear_egroups()
441

442
            L += lstate.get_edges_prob(missing, spurious, entropy_args=eargs)
443
            if isinstance(self.levels[0], LayeredBlockState):
444
445
                missing = [(lstate.b[u], lstate.b[v], l_) for u, v, l_ in missing]
                spurious = [(lstate.b[u], lstate.b[v], l_) for u, v, l_ in spurious]
446
            else:
447
448
449
                missing = [(lstate.b[u], lstate.b[v]) for u, v in missing]
                spurious = [(lstate.b[u], lstate.b[v]) for u, v in spurious]

450
451
        return L

452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
    def get_bstack(self):
        """Return the nested levels as individual graphs.

        This returns a list of :class:`~graph_tool.Graph` instances
        representing the inferred hierarchy at each level. Each graph has two
        internal vertex and edge property maps named "count" which correspond to
        the vertex and edge counts at the lower level, respectively. Additionally,
        an internal vertex property map named "b" specifies the block partition.
        """

        bstack = []
        for l, bstate in enumerate(self.levels):
            cg = bstate.g
            if l == 0:
                cg = GraphView(cg, skip_properties=True)
            cg.vp["b"] = bstate.b.copy()
468
469
470
            if bstate.is_weighted:
                cg.ep["count"] = cg.own_property(bstate.eweight.copy())
                cg.vp["count"] = cg.own_property(bstate.vweight.copy())
471
472
473
474
            else:
                cg.ep["count"] = cg.new_ep("int", 1)

            bstack.append(cg)
475
            if bstate.get_N() == 1:
476
477
478
479
480
481
482
                break
        return bstack

    def project_level(self, l):
        """Project the partition at level ``l`` onto the lowest level, and return the
        corresponding state."""
        b = self.project_partition(l, 0)
483
        return self.levels[0].copy(b=b)
484
485
486
487

    def print_summary(self):
        """Print a hierarchy summary."""
        for l, state in enumerate(self.levels):
488
489
            print("l: %d, N: %d, B: %d" % (l, state.get_N(),
                                           state.get_nonempty_B()))
Tiago Peixoto's avatar
Tiago Peixoto committed
490
491
            if state.get_N() == 1:
                break
492

493
494
495
    def _couple_levels(self, hentropy_args, hentropy_args_top):
        if hentropy_args_top is None:
            hentropy_args_top = dict(hentropy_args, edges_dl=True, recs_dl=True)
496
        for l in range(len(self.levels) - 1):
497
498
499
500
            if l + 1 == len(self.levels) - 1:
                eargs = hentropy_args_top
            else:
                eargs = hentropy_args
501
502
            self.levels[l]._couple_state(self.levels[l + 1], eargs)

503
504
505
506
    def _clear_egroups(self):
        for lstate in self.levels:
            lstate._clear_egroups()

507
    def _h_sweep_gen(self, **kwargs):
508

509
        verbose = kwargs.get("verbose", False)
510
511
512
513
514
515
        entropy_args = dict(kwargs.get("entropy_args", {}), edges_dl=False)
        hentropy_args = dict(self.hentropy_args,
                             **entropy_args.pop("hentropy_args", {}))
        hentropy_args_top = dict(dict(hentropy_args, edges_dl=True,
                                      recs_dl=True),
                                 **entropy_args.pop("hentropy_args_top", {}))
516

517
        self._couple_levels(hentropy_args, hentropy_args_top)
518

519
520
        c = kwargs.get("c", None)

521
        lrange = list(kwargs.pop("ls", range(len(self.levels))))
522
523
        if kwargs.pop("ls_shuffle", True):
            numpy.random.shuffle(lrange)
524
        for l in lrange:
525
526
527
            if check_verbose(verbose):
                print(verbose_pad(verbose) + "level:", l)
            if l > 0:
528
529
530
531
                if l == len(self.levels) - 1:
                    eargs = hentropy_args_top
                else:
                    eargs = hentropy_args
532
533
534
            else:
                eargs = entropy_args

535
            if c is None:
536
                args = dict(kwargs, entropy_args=eargs)
537
            else:
538
                args = dict(kwargs, entropy_args=eargs, c=c[l])
539

540
541
            if l > 0 and "beta_dl" in entropy_args:
                args = dict(args, beta=args.get("beta", 1.) * entropy_args["beta_dl"])
542

543
544
545
546
547
548
549
550
551
            yield l, self.levels[l], args

    def _h_sweep(self, algo, **kwargs):
        entropy_args = kwargs.get("entropy_args", {})

        dS = 0
        nattempts = 0
        nmoves = 0

552
        for l, lstate, args in self._h_sweep_gen(**kwargs):
553

554
            ret = algo(self.levels[l], **dict(args, test=False))
555

556
557
558
559
560
561
            if l > 0 and "beta_dl" in entropy_args:
                dS += ret[0] * entropy_args["beta_dl"]
            else:
                dS += ret[0]
            nattempts += ret[1]
            nmoves += ret[2]
562

563
        return dS, nattempts, nmoves
564

565
566
567
    def _h_sweep_states(self, algo, **kwargs):
        entropy_args = kwargs.get("entropy_args", {})
        for l, lstate, args in self._h_sweep_gen(**kwargs):
568
569
            beta_dl = entropy_args.get("beta_dl", 1) if l > 0 else 1
            yield l, lstate, algo(self.levels[l], dispatch=False, **args), beta_dl
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586

    def _h_sweep_parallel_dispatch(states, sweeps, algo):
        ret = None
        for lsweep in zip(*sweeps):
            ls = [x[0] for x in lsweep]
            lstates = [x[1] for x in lsweep]
            lsweep_states = [x[2] for x in lsweep]
            beta_dl = [x[3] for x in lsweep]
            lret = algo(type(lstates[0]), lstates, lsweep_states)
            if ret is None:
                ret = lret
            else:
                ret = [(ret[i][0] + lret[i][0] * beta_dl[i],
                        ret[i][1] + lret[i][1],
                        ret[i][2] + lret[i][2]) for i in range(len(lret))]
        return ret

587
    @mcmc_sweep_wrap
588
589
    def mcmc_sweep(self, **kwargs):
        r"""Perform ``niter`` sweeps of a Metropolis-Hastings acceptance-rejection
Tiago Peixoto's avatar
Tiago Peixoto committed
590
        MCMC to sample hierarchical network partitions.
591
592

        The arguments accepted are the same as in
593
        :meth:`graph_tool.inference.BlockState.mcmc_sweep`.
594
595
596
597
598

        If the parameter ``c`` is a scalar, the values used at each level are
        ``c * 2 ** l`` for ``l`` in the range ``[0, L-1]``. Optionally, a list
        of values may be passed instead, which specifies the value of ``c[l]``
        to be used at each level.
Tiago Peixoto's avatar
Tiago Peixoto committed
599
600
601
602
603
604
605
606

        .. warning::

           This function performs ``niter`` sweeps at each hierarchical level
           once. This means that in order for the chain to equilibrate, we need
           to call this function several times, i.e. it is not enough to call
           it once with a large value of ``niter``.

607
        """
608

609
        c = kwargs.pop("c", 1)
610
        if not isinstance(c, collections.abc.Iterable):
611
            c = [c * 2 ** l for l in range(0, len(self.levels))]
Tiago Peixoto's avatar
Tiago Peixoto committed
612

613
        if kwargs.pop("dispatch", True):
614
615
            return self._h_sweep(lambda s, **a: s.mcmc_sweep(**a), c=c,
                                 **kwargs)
616
617
618
619
620
621
622
        else:
            return self._h_sweep_states(lambda s, **a: s.mcmc_sweep(**a),
                                        c=c, **kwargs)

    def _mcmc_sweep_parallel_dispatch(states, sweeps):
        algo = lambda s, lstates, lsweep_states: s._mcmc_sweep_parallel_dispatch(lstates, lsweep_states)
        return NestedBlockState._h_sweep_parallel_dispatch(states, sweeps, algo)
623

624
    @mcmc_sweep_wrap
625
626
627
628
629
    def multiflip_mcmc_sweep(self, **kwargs):
        r"""Perform ``niter`` sweeps of a Metropolis-Hastings acceptance-rejection MCMC
        with multiple moves to sample hierarchical network partitions.

        The arguments accepted are the same as in
630
        :meth:`graph_tool.inference.BlockState.multiflip_mcmc_sweep`.
631
632
633
634
635
636

        If the parameter ``c`` is a scalar, the values used at each level are
        ``c * 2 ** l`` for ``l`` in the range ``[0, L-1]``. Optionally, a list
        of values may be passed instead, which specifies the value of ``c[l]``
        to be used at each level.

Tiago Peixoto's avatar
Tiago Peixoto committed
637
638
639
640
641
642
643
        .. warning::

           This function performs ``niter`` sweeps at each hierarchical level
           once. This means that in order for the chain to equilibrate, we need
           to call this function several times, i.e. it is not enough to call
           it once with a large value of ``niter``.

644
645
        """

646
        kwargs["psingle"] = kwargs.get("psingle", self.levels[0].get_N())
647

648
        c = kwargs.pop("c", 1)
649
        if not isinstance(c, collections.abc.Iterable):
650
            c = [c * 2 ** l for l in range(0, len(self.levels))]
651

652
        if kwargs.pop("dispatch", True):
653
654
655
656
657
658
659
660
661
662
            def dispatch_level(s, **a):
                if s is not self.levels[0]:
                    a = dict(**a)
                    a.pop("B_min", None)
                    a.pop("B_max", None)
                    a.pop("b_min", None)
                    a.pop("b_max", None)
                return s.multiflip_mcmc_sweep(**a)

            return self._h_sweep(dispatch_level, c=c, **kwargs)
663
664
665
666
667
668
        else:
            return self._h_sweep_states(lambda s, **a: s.multiflip_mcmc_sweep(**a),
                                        c=c, **kwargs)

    def _multiflip_mcmc_sweep_parallel_dispatch(states, sweeps):
        algo = lambda s, lstates, lsweep_states: s._multiflip_mcmc_sweep_parallel_dispatch(lstates, lsweep_states)
669
670
        return NestedBlockState._h_sweep_parallel_dispatch(states, sweeps, algo)

671
    @mcmc_sweep_wrap
672
    def multilevel_mcmc_sweep(self, **kwargs):
Tiago Peixoto's avatar
Tiago Peixoto committed
673
674
675
676
        r"""Perform ``niter`` sweeps of a Metropolis-Hastings acceptance-rejection MCMC
        with multilevel moves to sample hierarchical network partitions.

        The arguments accepted are the same as in
677
        :meth:`graph_tool.inference.BlockState.multilevel_mcmc_sweep`.
Tiago Peixoto's avatar
Tiago Peixoto committed
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692

        If the parameter ``c`` is a scalar, the values used at each level are
        ``c * 2 ** l`` for ``l`` in the range ``[0, L-1]``. Optionally, a list
        of values may be passed instead, which specifies the value of ``c[l]``
        to be used at each level.

        .. warning::

           This function performs ``niter`` sweeps at each hierarchical level
           once. This means that in order for the chain to equilibrate, we need
           to call this function several times, i.e. it is not enough to call
           it once with a large value of ``niter``.

        """

693
694
695
696
697
698
699
        kwargs["psingle"] = kwargs.get("psingle", self.g.num_vertices())

        c = kwargs.pop("c", 1)
        if not isinstance(c, collections.abc.Iterable):
            c = [c * 2 ** l for l in range(0, len(self.levels))]

        if kwargs.pop("dispatch", True):
700
701
            return self._h_sweep(lambda s, **a: s.multilevel_mcmc_sweep(**a),
                                 c=c, **kwargs)
702
703
704
705
706
707
        else:
            return self._h_sweep_states(lambda s, **a: s.multilevel_mcmc_sweep(**a),
                                        c=c, **kwargs)

    def _multilevel_mcmc_sweep_parallel_dispatch(states, sweeps):
        algo = lambda s, lstates, lsweep_states: s._multilevel_mcmc_sweep_parallel_dispatch(lstates, lsweep_states)
708
        return NestedBlockState._h_sweep_parallel_dispatch(states, sweeps, algo)
709

710
    @mcmc_sweep_wrap
711
712
713
714
715
    def gibbs_sweep(self, **kwargs):
        r"""Perform ``niter`` sweeps of a rejection-free Gibbs sampling MCMC
        to sample network partitions.

        The arguments accepted are the same as in
716
        :meth:`graph_tool.inference.BlockState.gibbs_sweep`.
Tiago Peixoto's avatar
Tiago Peixoto committed
717
718
719
720
721
722
723
724

        .. warning::

           This function performs ``niter`` sweeps at each hierarchical level
           once. This means that in order for the chain to equilibrate, we need
           to call this function several times, i.e. it is not enough to call
           it once with a large value of ``niter``.

725
        """
726
727
        return self._h_sweep(lambda s, **a: s.gibbs_sweep(**a),
                             **kwargs)
728
729
730
731
732

    def _gibbs_sweep_parallel_dispatch(states, sweeps):
        algo = lambda s, lstates, lsweep_states: s._gibbs_sweep_parallel_dispatch(lstates, lsweep_states)
        return NestedBlockState._h_sweep_parallel_dispatch(states, sweeps, algo)

733
    @mcmc_sweep_wrap
734
    def multicanonical_sweep(self, m_state, **kwargs):
735
736
737
738
        r"""Perform ``niter`` sweeps of a non-Markovian multicanonical sampling using the
        Wang-Landau algorithm.

        The arguments accepted are the same as in
739
        :meth:`graph_tool.inference.BlockState.multicanonical_sweep`.
740
        """
741
742
743
744
745
746
747
748
749
750

        def sweep(s, **kwargs):
            S = 0
            for l, state in enumerate(self.levels):
                if s is state:
                    continue
                S += self.level_entropy(l)
            return s.multicanonical_sweep(m_state, entropy_offset=S, **kwargs)

        return self._h_sweep(sweep)
751

752
753
754
755
    def collect_partition_histogram(self, h=None, update=1):
        r"""Collect a histogram of partitions.

        This should be called multiple times, e.g. after repeated runs of the
756
        :meth:`graph_tool.inference.NestedBlockState.mcmc_sweep` function.
757
758
759

        Parameters
        ----------
760
        h : :class:`~graph_tool.inference.PartitionHist` (optional, default: ``None``)
761
762
763
764
765
766
767
            Partition histogram. If not provided, an empty histogram will be created.
        update : float (optional, default: ``1``)
            Each call increases the current count by the amount given by this
            parameter.

        Returns
        -------
768
        h : :class:`~graph_tool.inference.PartitionHist` (optional, default: ``None``)
769
770
771
772
773
774
775
776
777
778
            Updated Partition histogram.

        """

        if h is None:
            h = PartitionHist()
        bs = [_prop("v", state.g, state.b) for state in self.levels]
        libinference.collect_hierarchical_partitions(bs, h, update)
        return h

779
780
781
782
783
784
    def draw(self, **kwargs):
        r"""Convenience wrapper to :func:`~graph_tool.draw.draw_hierarchy` that
        draws the hierarchical state."""
        import graph_tool.draw
        return graph_tool.draw.draw_hierarchy(self, **kwargs)

785
def get_hierarchy_tree(state, empty_branches=False):
786
787
    r"""Obtain the nested hierarchical levels as a tree.

788
    This transforms a :class:`~graph_tool.inference.NestedBlockState` instance
789
790
791
792
793
    into a single :class:`~graph_tool.Graph` instance containing the hierarchy
    tree.

    Parameters
    ----------
794
    state : :class:`~graph_tool.inference.NestedBlockState`
795
       Nested block model state.
796
    empty_branches : ``bool`` (optional, default: ``False``)
797
798
799
800
801
802
803
804
805
       If ``empty_branches == False``, dangling branches at the upper layers
       will be pruned.

    Returns
    -------

    tree : :class:`~graph_tool.Graph`
       A directed graph, where vertices are blocks, and a directed edge points
       to an upper to a lower level in the hierarchy.
806
    label : :class:`~graph_tool.VertexPropertyMap`
807
       A vertex property map containing the block label for each node.
808
    order : :class:`~graph_tool.VertexPropertyMap`
809
810
811
812
813
814
815
816
817
818
819
820
821
822
       A vertex property map containing the relative ordering of each layer
       according to the total degree of the groups at the specific levels.
    """

    bstack = state.get_bstack()

    g = bstack[0]
    b = g.vp["b"]
    bstack = bstack[1:]

    if bstack[-1].num_vertices() > 1:
        bg = Graph(directed=g.is_directed())
        bg.add_vertex()
        e = bg.add_edge(0, 0)
823
824
825
        bg.vp.count = bg.new_vp("int", 1)
        bg.ep.count = bg.new_ep("int", g.ep.count.fa.sum())
        bg.vp.b = bg.new_vp("int", 0)
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
        bstack.append(bg)

    t = Graph()

    if g.get_vertex_filter()[0] is None:
        t.add_vertex(g.num_vertices())
    else:
        t.add_vertex(g.num_vertices(ignore_filter=True))
        filt = g.get_vertex_filter()
        t.set_vertex_filter(t.own_property(filt[0].copy()),
                            filt[1])
    label = t.vertex_index.copy("int")

    order = t.own_property(g.degree_property_map("total").copy())
    t_vertices = list(t.vertices())

    last_pos = 0
    for l, s in enumerate(bstack):
        pos = t.num_vertices()
        if s.num_vertices() > 1:
            t_vertices.extend(t.add_vertex(s.num_vertices()))
        else:
            t_vertices.append(t.add_vertex(s.num_vertices()))
        label.a[-s.num_vertices():] = arange(s.num_vertices())

        # relative ordering based on total degree
        count = s.ep["count"].copy("double")
        for e in s.edges():
            if e.source() == e.target():
                count[e] /= 2
        vs = []
        pvs = {}
        for vi in range(pos, t.num_vertices()):
            vs.append(t_vertices[vi])
            pvs[vs[-1]] = vi - pos
        vs = sorted(vs, key=lambda v: (s.vertex(pvs[v]).out_degree(count) +
                                       s.vertex(pvs[v]).in_degree(count)))
        for vi, v in enumerate(vs):
            order[v] = vi

        for vi, v in enumerate(g.vertices()):
            w = t_vertices[vi + last_pos]
868
869
870
871
            if s.num_vertices() == 1:
                u = t_vertices[pos]
            else:
                u = t_vertices[b[v] + pos]
872
873
874
875
            t.add_edge(u, w)

        last_pos = pos
        g = s
Tiago Peixoto's avatar
Tiago Peixoto committed
876
877
878
879
880
881
        if empty_branches:
            if g.num_vertices() == 1:
                break
        else:
            if g.vp.count.fa.sum() == 1:
                break
882
883
884
        b = g.vp["b"]

    if not empty_branches:
885
        vmask = t.new_vertex_property("bool", True)
886
887
888
889
890
891
        t = GraphView(t, vfilt=vmask)
        vmask = t.get_vertex_filter()[0]

        for vi in range(state.g.num_vertices(ignore_filter=True),
                        t.num_vertices()):
            v = t.vertex(t_vertices[vi])
892
893
            if v.out_degree() == 0:
                vmask[v] = False
894
895
896

        t.vp.label = label
        t.vp.order = order
897
        t = Graph(t, prune=True)
898
899
        label = t.vp.label
        order = t.vp.order
900
        del t.vp["label"]
901
        del t.vp["order"]
902
903
904
905

    return t, label, order

from . minimize import *